Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * kernel/sched.c | |
3 | * | |
4 | * Kernel scheduler and related syscalls | |
5 | * | |
6 | * Copyright (C) 1991-2002 Linus Torvalds | |
7 | * | |
8 | * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and | |
9 | * make semaphores SMP safe | |
10 | * 1998-11-19 Implemented schedule_timeout() and related stuff | |
11 | * by Andrea Arcangeli | |
12 | * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: | |
13 | * hybrid priority-list and round-robin design with | |
14 | * an array-switch method of distributing timeslices | |
15 | * and per-CPU runqueues. Cleanups and useful suggestions | |
16 | * by Davide Libenzi, preemptible kernel bits by Robert Love. | |
17 | * 2003-09-03 Interactivity tuning by Con Kolivas. | |
18 | * 2004-04-02 Scheduler domains code by Nick Piggin | |
c31f2e8a IM |
19 | * 2007-04-15 Work begun on replacing all interactivity tuning with a |
20 | * fair scheduling design by Con Kolivas. | |
21 | * 2007-05-05 Load balancing (smp-nice) and other improvements | |
22 | * by Peter Williams | |
23 | * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith | |
24 | * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri | |
b9131769 IM |
25 | * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, |
26 | * Thomas Gleixner, Mike Kravetz | |
1da177e4 LT |
27 | */ |
28 | ||
29 | #include <linux/mm.h> | |
30 | #include <linux/module.h> | |
31 | #include <linux/nmi.h> | |
32 | #include <linux/init.h> | |
dff06c15 | 33 | #include <linux/uaccess.h> |
1da177e4 LT |
34 | #include <linux/highmem.h> |
35 | #include <linux/smp_lock.h> | |
36 | #include <asm/mmu_context.h> | |
37 | #include <linux/interrupt.h> | |
c59ede7b | 38 | #include <linux/capability.h> |
1da177e4 LT |
39 | #include <linux/completion.h> |
40 | #include <linux/kernel_stat.h> | |
9a11b49a | 41 | #include <linux/debug_locks.h> |
1da177e4 LT |
42 | #include <linux/security.h> |
43 | #include <linux/notifier.h> | |
44 | #include <linux/profile.h> | |
7dfb7103 | 45 | #include <linux/freezer.h> |
198e2f18 | 46 | #include <linux/vmalloc.h> |
1da177e4 LT |
47 | #include <linux/blkdev.h> |
48 | #include <linux/delay.h> | |
b488893a | 49 | #include <linux/pid_namespace.h> |
1da177e4 LT |
50 | #include <linux/smp.h> |
51 | #include <linux/threads.h> | |
52 | #include <linux/timer.h> | |
53 | #include <linux/rcupdate.h> | |
54 | #include <linux/cpu.h> | |
55 | #include <linux/cpuset.h> | |
56 | #include <linux/percpu.h> | |
57 | #include <linux/kthread.h> | |
58 | #include <linux/seq_file.h> | |
e692ab53 | 59 | #include <linux/sysctl.h> |
1da177e4 LT |
60 | #include <linux/syscalls.h> |
61 | #include <linux/times.h> | |
8f0ab514 | 62 | #include <linux/tsacct_kern.h> |
c6fd91f0 | 63 | #include <linux/kprobes.h> |
0ff92245 | 64 | #include <linux/delayacct.h> |
5517d86b | 65 | #include <linux/reciprocal_div.h> |
dff06c15 | 66 | #include <linux/unistd.h> |
f5ff8422 | 67 | #include <linux/pagemap.h> |
8f4d37ec | 68 | #include <linux/hrtimer.h> |
30914a58 | 69 | #include <linux/tick.h> |
434d53b0 | 70 | #include <linux/bootmem.h> |
f00b45c1 PZ |
71 | #include <linux/debugfs.h> |
72 | #include <linux/ctype.h> | |
1da177e4 | 73 | |
5517d86b | 74 | #include <asm/tlb.h> |
838225b4 | 75 | #include <asm/irq_regs.h> |
1da177e4 LT |
76 | |
77 | /* | |
78 | * Convert user-nice values [ -20 ... 0 ... 19 ] | |
79 | * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ], | |
80 | * and back. | |
81 | */ | |
82 | #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20) | |
83 | #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20) | |
84 | #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio) | |
85 | ||
86 | /* | |
87 | * 'User priority' is the nice value converted to something we | |
88 | * can work with better when scaling various scheduler parameters, | |
89 | * it's a [ 0 ... 39 ] range. | |
90 | */ | |
91 | #define USER_PRIO(p) ((p)-MAX_RT_PRIO) | |
92 | #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio) | |
93 | #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO)) | |
94 | ||
95 | /* | |
d7876a08 | 96 | * Helpers for converting nanosecond timing to jiffy resolution |
1da177e4 | 97 | */ |
d6322faf | 98 | #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ)) |
1da177e4 | 99 | |
6aa645ea IM |
100 | #define NICE_0_LOAD SCHED_LOAD_SCALE |
101 | #define NICE_0_SHIFT SCHED_LOAD_SHIFT | |
102 | ||
1da177e4 LT |
103 | /* |
104 | * These are the 'tuning knobs' of the scheduler: | |
105 | * | |
a4ec24b4 | 106 | * default timeslice is 100 msecs (used only for SCHED_RR tasks). |
1da177e4 LT |
107 | * Timeslices get refilled after they expire. |
108 | */ | |
1da177e4 | 109 | #define DEF_TIMESLICE (100 * HZ / 1000) |
2dd73a4f | 110 | |
d0b27fa7 PZ |
111 | /* |
112 | * single value that denotes runtime == period, ie unlimited time. | |
113 | */ | |
114 | #define RUNTIME_INF ((u64)~0ULL) | |
115 | ||
5517d86b ED |
116 | #ifdef CONFIG_SMP |
117 | /* | |
118 | * Divide a load by a sched group cpu_power : (load / sg->__cpu_power) | |
119 | * Since cpu_power is a 'constant', we can use a reciprocal divide. | |
120 | */ | |
121 | static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load) | |
122 | { | |
123 | return reciprocal_divide(load, sg->reciprocal_cpu_power); | |
124 | } | |
125 | ||
126 | /* | |
127 | * Each time a sched group cpu_power is changed, | |
128 | * we must compute its reciprocal value | |
129 | */ | |
130 | static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val) | |
131 | { | |
132 | sg->__cpu_power += val; | |
133 | sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power); | |
134 | } | |
135 | #endif | |
136 | ||
e05606d3 IM |
137 | static inline int rt_policy(int policy) |
138 | { | |
139 | if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR)) | |
140 | return 1; | |
141 | return 0; | |
142 | } | |
143 | ||
144 | static inline int task_has_rt_policy(struct task_struct *p) | |
145 | { | |
146 | return rt_policy(p->policy); | |
147 | } | |
148 | ||
1da177e4 | 149 | /* |
6aa645ea | 150 | * This is the priority-queue data structure of the RT scheduling class: |
1da177e4 | 151 | */ |
6aa645ea IM |
152 | struct rt_prio_array { |
153 | DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */ | |
154 | struct list_head queue[MAX_RT_PRIO]; | |
155 | }; | |
156 | ||
d0b27fa7 | 157 | struct rt_bandwidth { |
ea736ed5 IM |
158 | /* nests inside the rq lock: */ |
159 | spinlock_t rt_runtime_lock; | |
160 | ktime_t rt_period; | |
161 | u64 rt_runtime; | |
162 | struct hrtimer rt_period_timer; | |
d0b27fa7 PZ |
163 | }; |
164 | ||
165 | static struct rt_bandwidth def_rt_bandwidth; | |
166 | ||
167 | static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun); | |
168 | ||
169 | static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer) | |
170 | { | |
171 | struct rt_bandwidth *rt_b = | |
172 | container_of(timer, struct rt_bandwidth, rt_period_timer); | |
173 | ktime_t now; | |
174 | int overrun; | |
175 | int idle = 0; | |
176 | ||
177 | for (;;) { | |
178 | now = hrtimer_cb_get_time(timer); | |
179 | overrun = hrtimer_forward(timer, now, rt_b->rt_period); | |
180 | ||
181 | if (!overrun) | |
182 | break; | |
183 | ||
184 | idle = do_sched_rt_period_timer(rt_b, overrun); | |
185 | } | |
186 | ||
187 | return idle ? HRTIMER_NORESTART : HRTIMER_RESTART; | |
188 | } | |
189 | ||
190 | static | |
191 | void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime) | |
192 | { | |
193 | rt_b->rt_period = ns_to_ktime(period); | |
194 | rt_b->rt_runtime = runtime; | |
195 | ||
ac086bc2 PZ |
196 | spin_lock_init(&rt_b->rt_runtime_lock); |
197 | ||
d0b27fa7 PZ |
198 | hrtimer_init(&rt_b->rt_period_timer, |
199 | CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
200 | rt_b->rt_period_timer.function = sched_rt_period_timer; | |
201 | rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
202 | } | |
203 | ||
204 | static void start_rt_bandwidth(struct rt_bandwidth *rt_b) | |
205 | { | |
206 | ktime_t now; | |
207 | ||
208 | if (rt_b->rt_runtime == RUNTIME_INF) | |
209 | return; | |
210 | ||
211 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
212 | return; | |
213 | ||
214 | spin_lock(&rt_b->rt_runtime_lock); | |
215 | for (;;) { | |
216 | if (hrtimer_active(&rt_b->rt_period_timer)) | |
217 | break; | |
218 | ||
219 | now = hrtimer_cb_get_time(&rt_b->rt_period_timer); | |
220 | hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period); | |
221 | hrtimer_start(&rt_b->rt_period_timer, | |
222 | rt_b->rt_period_timer.expires, | |
223 | HRTIMER_MODE_ABS); | |
224 | } | |
225 | spin_unlock(&rt_b->rt_runtime_lock); | |
226 | } | |
227 | ||
228 | #ifdef CONFIG_RT_GROUP_SCHED | |
229 | static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b) | |
230 | { | |
231 | hrtimer_cancel(&rt_b->rt_period_timer); | |
232 | } | |
233 | #endif | |
234 | ||
712555ee HC |
235 | /* |
236 | * sched_domains_mutex serializes calls to arch_init_sched_domains, | |
237 | * detach_destroy_domains and partition_sched_domains. | |
238 | */ | |
239 | static DEFINE_MUTEX(sched_domains_mutex); | |
240 | ||
052f1dc7 | 241 | #ifdef CONFIG_GROUP_SCHED |
29f59db3 | 242 | |
68318b8e SV |
243 | #include <linux/cgroup.h> |
244 | ||
29f59db3 SV |
245 | struct cfs_rq; |
246 | ||
6f505b16 PZ |
247 | static LIST_HEAD(task_groups); |
248 | ||
29f59db3 | 249 | /* task group related information */ |
4cf86d77 | 250 | struct task_group { |
052f1dc7 | 251 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
252 | struct cgroup_subsys_state css; |
253 | #endif | |
052f1dc7 PZ |
254 | |
255 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
29f59db3 SV |
256 | /* schedulable entities of this group on each cpu */ |
257 | struct sched_entity **se; | |
258 | /* runqueue "owned" by this group on each cpu */ | |
259 | struct cfs_rq **cfs_rq; | |
260 | unsigned long shares; | |
052f1dc7 PZ |
261 | #endif |
262 | ||
263 | #ifdef CONFIG_RT_GROUP_SCHED | |
264 | struct sched_rt_entity **rt_se; | |
265 | struct rt_rq **rt_rq; | |
266 | ||
d0b27fa7 | 267 | struct rt_bandwidth rt_bandwidth; |
052f1dc7 | 268 | #endif |
6b2d7700 | 269 | |
ae8393e5 | 270 | struct rcu_head rcu; |
6f505b16 | 271 | struct list_head list; |
f473aa5e PZ |
272 | |
273 | struct task_group *parent; | |
274 | struct list_head siblings; | |
275 | struct list_head children; | |
29f59db3 SV |
276 | }; |
277 | ||
354d60c2 | 278 | #ifdef CONFIG_USER_SCHED |
eff766a6 PZ |
279 | |
280 | /* | |
281 | * Root task group. | |
282 | * Every UID task group (including init_task_group aka UID-0) will | |
283 | * be a child to this group. | |
284 | */ | |
285 | struct task_group root_task_group; | |
286 | ||
052f1dc7 | 287 | #ifdef CONFIG_FAIR_GROUP_SCHED |
29f59db3 SV |
288 | /* Default task group's sched entity on each cpu */ |
289 | static DEFINE_PER_CPU(struct sched_entity, init_sched_entity); | |
290 | /* Default task group's cfs_rq on each cpu */ | |
291 | static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 PZ |
292 | #endif |
293 | ||
294 | #ifdef CONFIG_RT_GROUP_SCHED | |
295 | static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity); | |
296 | static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp; | |
052f1dc7 | 297 | #endif |
eff766a6 PZ |
298 | #else |
299 | #define root_task_group init_task_group | |
354d60c2 | 300 | #endif |
6f505b16 | 301 | |
8ed36996 | 302 | /* task_group_lock serializes add/remove of task groups and also changes to |
ec2c507f SV |
303 | * a task group's cpu shares. |
304 | */ | |
8ed36996 | 305 | static DEFINE_SPINLOCK(task_group_lock); |
ec2c507f | 306 | |
052f1dc7 | 307 | #ifdef CONFIG_FAIR_GROUP_SCHED |
052f1dc7 PZ |
308 | #ifdef CONFIG_USER_SCHED |
309 | # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD) | |
310 | #else | |
311 | # define INIT_TASK_GROUP_LOAD NICE_0_LOAD | |
312 | #endif | |
313 | ||
cb4ad1ff MX |
314 | /* |
315 | * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems. | |
316 | * (The default weight is 1024 - so there's no practical | |
317 | * limitation from this.) | |
318 | */ | |
18d95a28 | 319 | #define MIN_SHARES 2 |
cb4ad1ff | 320 | #define MAX_SHARES (ULONG_MAX - 1) |
18d95a28 | 321 | |
052f1dc7 PZ |
322 | static int init_task_group_load = INIT_TASK_GROUP_LOAD; |
323 | #endif | |
324 | ||
29f59db3 | 325 | /* Default task group. |
3a252015 | 326 | * Every task in system belong to this group at bootup. |
29f59db3 | 327 | */ |
434d53b0 | 328 | struct task_group init_task_group; |
29f59db3 SV |
329 | |
330 | /* return group to which a task belongs */ | |
4cf86d77 | 331 | static inline struct task_group *task_group(struct task_struct *p) |
29f59db3 | 332 | { |
4cf86d77 | 333 | struct task_group *tg; |
9b5b7751 | 334 | |
052f1dc7 | 335 | #ifdef CONFIG_USER_SCHED |
24e377a8 | 336 | tg = p->user->tg; |
052f1dc7 | 337 | #elif defined(CONFIG_CGROUP_SCHED) |
68318b8e SV |
338 | tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id), |
339 | struct task_group, css); | |
24e377a8 | 340 | #else |
41a2d6cf | 341 | tg = &init_task_group; |
24e377a8 | 342 | #endif |
9b5b7751 | 343 | return tg; |
29f59db3 SV |
344 | } |
345 | ||
346 | /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */ | |
6f505b16 | 347 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) |
29f59db3 | 348 | { |
052f1dc7 | 349 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ce96b5ac DA |
350 | p->se.cfs_rq = task_group(p)->cfs_rq[cpu]; |
351 | p->se.parent = task_group(p)->se[cpu]; | |
052f1dc7 | 352 | #endif |
6f505b16 | 353 | |
052f1dc7 | 354 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
355 | p->rt.rt_rq = task_group(p)->rt_rq[cpu]; |
356 | p->rt.parent = task_group(p)->rt_se[cpu]; | |
052f1dc7 | 357 | #endif |
29f59db3 SV |
358 | } |
359 | ||
360 | #else | |
361 | ||
6f505b16 | 362 | static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { } |
29f59db3 | 363 | |
052f1dc7 | 364 | #endif /* CONFIG_GROUP_SCHED */ |
29f59db3 | 365 | |
6aa645ea IM |
366 | /* CFS-related fields in a runqueue */ |
367 | struct cfs_rq { | |
368 | struct load_weight load; | |
369 | unsigned long nr_running; | |
370 | ||
6aa645ea | 371 | u64 exec_clock; |
e9acbff6 | 372 | u64 min_vruntime; |
6aa645ea IM |
373 | |
374 | struct rb_root tasks_timeline; | |
375 | struct rb_node *rb_leftmost; | |
4a55bd5e PZ |
376 | |
377 | struct list_head tasks; | |
378 | struct list_head *balance_iterator; | |
379 | ||
380 | /* | |
381 | * 'curr' points to currently running entity on this cfs_rq. | |
6aa645ea IM |
382 | * It is set to NULL otherwise (i.e when none are currently running). |
383 | */ | |
aa2ac252 | 384 | struct sched_entity *curr, *next; |
ddc97297 PZ |
385 | |
386 | unsigned long nr_spread_over; | |
387 | ||
62160e3f | 388 | #ifdef CONFIG_FAIR_GROUP_SCHED |
6aa645ea IM |
389 | struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */ |
390 | ||
41a2d6cf IM |
391 | /* |
392 | * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in | |
6aa645ea IM |
393 | * a hierarchy). Non-leaf lrqs hold other higher schedulable entities |
394 | * (like users, containers etc.) | |
395 | * | |
396 | * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This | |
397 | * list is used during load balance. | |
398 | */ | |
41a2d6cf IM |
399 | struct list_head leaf_cfs_rq_list; |
400 | struct task_group *tg; /* group that "owns" this runqueue */ | |
18d95a28 PZ |
401 | |
402 | #ifdef CONFIG_SMP | |
403 | unsigned long task_weight; | |
404 | unsigned long shares; | |
405 | /* | |
406 | * We need space to build a sched_domain wide view of the full task | |
407 | * group tree, in order to avoid depending on dynamic memory allocation | |
408 | * during the load balancing we place this in the per cpu task group | |
409 | * hierarchy. This limits the load balancing to one instance per cpu, | |
410 | * but more should not be needed anyway. | |
411 | */ | |
412 | struct aggregate_struct { | |
413 | /* | |
414 | * load = weight(cpus) * f(tg) | |
415 | * | |
416 | * Where f(tg) is the recursive weight fraction assigned to | |
417 | * this group. | |
418 | */ | |
419 | unsigned long load; | |
420 | ||
421 | /* | |
422 | * part of the group weight distributed to this span. | |
423 | */ | |
424 | unsigned long shares; | |
425 | ||
426 | /* | |
427 | * The sum of all runqueue weights within this span. | |
428 | */ | |
429 | unsigned long rq_weight; | |
430 | ||
431 | /* | |
432 | * Weight contributed by tasks; this is the part we can | |
433 | * influence by moving tasks around. | |
434 | */ | |
435 | unsigned long task_weight; | |
436 | } aggregate; | |
437 | #endif | |
6aa645ea IM |
438 | #endif |
439 | }; | |
1da177e4 | 440 | |
6aa645ea IM |
441 | /* Real-Time classes' related field in a runqueue: */ |
442 | struct rt_rq { | |
443 | struct rt_prio_array active; | |
63489e45 | 444 | unsigned long rt_nr_running; |
052f1dc7 | 445 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
6f505b16 PZ |
446 | int highest_prio; /* highest queued rt task prio */ |
447 | #endif | |
fa85ae24 | 448 | #ifdef CONFIG_SMP |
73fe6aae | 449 | unsigned long rt_nr_migratory; |
a22d7fc1 | 450 | int overloaded; |
fa85ae24 | 451 | #endif |
6f505b16 | 452 | int rt_throttled; |
fa85ae24 | 453 | u64 rt_time; |
ac086bc2 | 454 | u64 rt_runtime; |
ea736ed5 | 455 | /* Nests inside the rq lock: */ |
ac086bc2 | 456 | spinlock_t rt_runtime_lock; |
6f505b16 | 457 | |
052f1dc7 | 458 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc PZ |
459 | unsigned long rt_nr_boosted; |
460 | ||
6f505b16 PZ |
461 | struct rq *rq; |
462 | struct list_head leaf_rt_rq_list; | |
463 | struct task_group *tg; | |
464 | struct sched_rt_entity *rt_se; | |
465 | #endif | |
6aa645ea IM |
466 | }; |
467 | ||
57d885fe GH |
468 | #ifdef CONFIG_SMP |
469 | ||
470 | /* | |
471 | * We add the notion of a root-domain which will be used to define per-domain | |
0eab9146 IM |
472 | * variables. Each exclusive cpuset essentially defines an island domain by |
473 | * fully partitioning the member cpus from any other cpuset. Whenever a new | |
57d885fe GH |
474 | * exclusive cpuset is created, we also create and attach a new root-domain |
475 | * object. | |
476 | * | |
57d885fe GH |
477 | */ |
478 | struct root_domain { | |
479 | atomic_t refcount; | |
480 | cpumask_t span; | |
481 | cpumask_t online; | |
637f5085 | 482 | |
0eab9146 | 483 | /* |
637f5085 GH |
484 | * The "RT overload" flag: it gets set if a CPU has more than |
485 | * one runnable RT task. | |
486 | */ | |
487 | cpumask_t rto_mask; | |
0eab9146 | 488 | atomic_t rto_count; |
57d885fe GH |
489 | }; |
490 | ||
dc938520 GH |
491 | /* |
492 | * By default the system creates a single root-domain with all cpus as | |
493 | * members (mimicking the global state we have today). | |
494 | */ | |
57d885fe GH |
495 | static struct root_domain def_root_domain; |
496 | ||
497 | #endif | |
498 | ||
1da177e4 LT |
499 | /* |
500 | * This is the main, per-CPU runqueue data structure. | |
501 | * | |
502 | * Locking rule: those places that want to lock multiple runqueues | |
503 | * (such as the load balancing or the thread migration code), lock | |
504 | * acquire operations must be ordered by ascending &runqueue. | |
505 | */ | |
70b97a7f | 506 | struct rq { |
d8016491 IM |
507 | /* runqueue lock: */ |
508 | spinlock_t lock; | |
1da177e4 LT |
509 | |
510 | /* | |
511 | * nr_running and cpu_load should be in the same cacheline because | |
512 | * remote CPUs use both these fields when doing load calculation. | |
513 | */ | |
514 | unsigned long nr_running; | |
6aa645ea IM |
515 | #define CPU_LOAD_IDX_MAX 5 |
516 | unsigned long cpu_load[CPU_LOAD_IDX_MAX]; | |
bdecea3a | 517 | unsigned char idle_at_tick; |
46cb4b7c | 518 | #ifdef CONFIG_NO_HZ |
15934a37 | 519 | unsigned long last_tick_seen; |
46cb4b7c SS |
520 | unsigned char in_nohz_recently; |
521 | #endif | |
d8016491 IM |
522 | /* capture load from *all* tasks on this cpu: */ |
523 | struct load_weight load; | |
6aa645ea IM |
524 | unsigned long nr_load_updates; |
525 | u64 nr_switches; | |
526 | ||
527 | struct cfs_rq cfs; | |
6f505b16 | 528 | struct rt_rq rt; |
6f505b16 | 529 | |
6aa645ea | 530 | #ifdef CONFIG_FAIR_GROUP_SCHED |
d8016491 IM |
531 | /* list of leaf cfs_rq on this cpu: */ |
532 | struct list_head leaf_cfs_rq_list; | |
052f1dc7 PZ |
533 | #endif |
534 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 535 | struct list_head leaf_rt_rq_list; |
1da177e4 | 536 | #endif |
1da177e4 LT |
537 | |
538 | /* | |
539 | * This is part of a global counter where only the total sum | |
540 | * over all CPUs matters. A task can increase this counter on | |
541 | * one CPU and if it got migrated afterwards it may decrease | |
542 | * it on another CPU. Always updated under the runqueue lock: | |
543 | */ | |
544 | unsigned long nr_uninterruptible; | |
545 | ||
36c8b586 | 546 | struct task_struct *curr, *idle; |
c9819f45 | 547 | unsigned long next_balance; |
1da177e4 | 548 | struct mm_struct *prev_mm; |
6aa645ea | 549 | |
3e51f33f | 550 | u64 clock; |
6aa645ea | 551 | |
1da177e4 LT |
552 | atomic_t nr_iowait; |
553 | ||
554 | #ifdef CONFIG_SMP | |
0eab9146 | 555 | struct root_domain *rd; |
1da177e4 LT |
556 | struct sched_domain *sd; |
557 | ||
558 | /* For active balancing */ | |
559 | int active_balance; | |
560 | int push_cpu; | |
d8016491 IM |
561 | /* cpu of this runqueue: */ |
562 | int cpu; | |
1da177e4 | 563 | |
36c8b586 | 564 | struct task_struct *migration_thread; |
1da177e4 LT |
565 | struct list_head migration_queue; |
566 | #endif | |
567 | ||
8f4d37ec PZ |
568 | #ifdef CONFIG_SCHED_HRTICK |
569 | unsigned long hrtick_flags; | |
570 | ktime_t hrtick_expire; | |
571 | struct hrtimer hrtick_timer; | |
572 | #endif | |
573 | ||
1da177e4 LT |
574 | #ifdef CONFIG_SCHEDSTATS |
575 | /* latency stats */ | |
576 | struct sched_info rq_sched_info; | |
577 | ||
578 | /* sys_sched_yield() stats */ | |
480b9434 KC |
579 | unsigned int yld_exp_empty; |
580 | unsigned int yld_act_empty; | |
581 | unsigned int yld_both_empty; | |
582 | unsigned int yld_count; | |
1da177e4 LT |
583 | |
584 | /* schedule() stats */ | |
480b9434 KC |
585 | unsigned int sched_switch; |
586 | unsigned int sched_count; | |
587 | unsigned int sched_goidle; | |
1da177e4 LT |
588 | |
589 | /* try_to_wake_up() stats */ | |
480b9434 KC |
590 | unsigned int ttwu_count; |
591 | unsigned int ttwu_local; | |
b8efb561 IM |
592 | |
593 | /* BKL stats */ | |
480b9434 | 594 | unsigned int bkl_count; |
1da177e4 | 595 | #endif |
fcb99371 | 596 | struct lock_class_key rq_lock_key; |
1da177e4 LT |
597 | }; |
598 | ||
f34e3b61 | 599 | static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); |
1da177e4 | 600 | |
dd41f596 IM |
601 | static inline void check_preempt_curr(struct rq *rq, struct task_struct *p) |
602 | { | |
603 | rq->curr->sched_class->check_preempt_curr(rq, p); | |
604 | } | |
605 | ||
0a2966b4 CL |
606 | static inline int cpu_of(struct rq *rq) |
607 | { | |
608 | #ifdef CONFIG_SMP | |
609 | return rq->cpu; | |
610 | #else | |
611 | return 0; | |
612 | #endif | |
613 | } | |
614 | ||
674311d5 NP |
615 | /* |
616 | * The domain tree (rq->sd) is protected by RCU's quiescent state transition. | |
1a20ff27 | 617 | * See detach_destroy_domains: synchronize_sched for details. |
674311d5 NP |
618 | * |
619 | * The domain tree of any CPU may only be accessed from within | |
620 | * preempt-disabled sections. | |
621 | */ | |
48f24c4d IM |
622 | #define for_each_domain(cpu, __sd) \ |
623 | for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent) | |
1da177e4 LT |
624 | |
625 | #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu))) | |
626 | #define this_rq() (&__get_cpu_var(runqueues)) | |
627 | #define task_rq(p) cpu_rq(task_cpu(p)) | |
628 | #define cpu_curr(cpu) (cpu_rq(cpu)->curr) | |
629 | ||
3e51f33f PZ |
630 | static inline void update_rq_clock(struct rq *rq) |
631 | { | |
632 | rq->clock = sched_clock_cpu(cpu_of(rq)); | |
633 | } | |
634 | ||
bf5c91ba IM |
635 | /* |
636 | * Tunables that become constants when CONFIG_SCHED_DEBUG is off: | |
637 | */ | |
638 | #ifdef CONFIG_SCHED_DEBUG | |
639 | # define const_debug __read_mostly | |
640 | #else | |
641 | # define const_debug static const | |
642 | #endif | |
643 | ||
644 | /* | |
645 | * Debugging: various feature bits | |
646 | */ | |
f00b45c1 PZ |
647 | |
648 | #define SCHED_FEAT(name, enabled) \ | |
649 | __SCHED_FEAT_##name , | |
650 | ||
bf5c91ba | 651 | enum { |
f00b45c1 | 652 | #include "sched_features.h" |
bf5c91ba IM |
653 | }; |
654 | ||
f00b45c1 PZ |
655 | #undef SCHED_FEAT |
656 | ||
657 | #define SCHED_FEAT(name, enabled) \ | |
658 | (1UL << __SCHED_FEAT_##name) * enabled | | |
659 | ||
bf5c91ba | 660 | const_debug unsigned int sysctl_sched_features = |
f00b45c1 PZ |
661 | #include "sched_features.h" |
662 | 0; | |
663 | ||
664 | #undef SCHED_FEAT | |
665 | ||
666 | #ifdef CONFIG_SCHED_DEBUG | |
667 | #define SCHED_FEAT(name, enabled) \ | |
668 | #name , | |
669 | ||
983ed7a6 | 670 | static __read_mostly char *sched_feat_names[] = { |
f00b45c1 PZ |
671 | #include "sched_features.h" |
672 | NULL | |
673 | }; | |
674 | ||
675 | #undef SCHED_FEAT | |
676 | ||
983ed7a6 | 677 | static int sched_feat_open(struct inode *inode, struct file *filp) |
f00b45c1 PZ |
678 | { |
679 | filp->private_data = inode->i_private; | |
680 | return 0; | |
681 | } | |
682 | ||
683 | static ssize_t | |
684 | sched_feat_read(struct file *filp, char __user *ubuf, | |
685 | size_t cnt, loff_t *ppos) | |
686 | { | |
687 | char *buf; | |
688 | int r = 0; | |
689 | int len = 0; | |
690 | int i; | |
691 | ||
692 | for (i = 0; sched_feat_names[i]; i++) { | |
693 | len += strlen(sched_feat_names[i]); | |
694 | len += 4; | |
695 | } | |
696 | ||
697 | buf = kmalloc(len + 2, GFP_KERNEL); | |
698 | if (!buf) | |
699 | return -ENOMEM; | |
700 | ||
701 | for (i = 0; sched_feat_names[i]; i++) { | |
702 | if (sysctl_sched_features & (1UL << i)) | |
703 | r += sprintf(buf + r, "%s ", sched_feat_names[i]); | |
704 | else | |
c24b7c52 | 705 | r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]); |
f00b45c1 PZ |
706 | } |
707 | ||
708 | r += sprintf(buf + r, "\n"); | |
709 | WARN_ON(r >= len + 2); | |
710 | ||
711 | r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r); | |
712 | ||
713 | kfree(buf); | |
714 | ||
715 | return r; | |
716 | } | |
717 | ||
718 | static ssize_t | |
719 | sched_feat_write(struct file *filp, const char __user *ubuf, | |
720 | size_t cnt, loff_t *ppos) | |
721 | { | |
722 | char buf[64]; | |
723 | char *cmp = buf; | |
724 | int neg = 0; | |
725 | int i; | |
726 | ||
727 | if (cnt > 63) | |
728 | cnt = 63; | |
729 | ||
730 | if (copy_from_user(&buf, ubuf, cnt)) | |
731 | return -EFAULT; | |
732 | ||
733 | buf[cnt] = 0; | |
734 | ||
c24b7c52 | 735 | if (strncmp(buf, "NO_", 3) == 0) { |
f00b45c1 PZ |
736 | neg = 1; |
737 | cmp += 3; | |
738 | } | |
739 | ||
740 | for (i = 0; sched_feat_names[i]; i++) { | |
741 | int len = strlen(sched_feat_names[i]); | |
742 | ||
743 | if (strncmp(cmp, sched_feat_names[i], len) == 0) { | |
744 | if (neg) | |
745 | sysctl_sched_features &= ~(1UL << i); | |
746 | else | |
747 | sysctl_sched_features |= (1UL << i); | |
748 | break; | |
749 | } | |
750 | } | |
751 | ||
752 | if (!sched_feat_names[i]) | |
753 | return -EINVAL; | |
754 | ||
755 | filp->f_pos += cnt; | |
756 | ||
757 | return cnt; | |
758 | } | |
759 | ||
760 | static struct file_operations sched_feat_fops = { | |
761 | .open = sched_feat_open, | |
762 | .read = sched_feat_read, | |
763 | .write = sched_feat_write, | |
764 | }; | |
765 | ||
766 | static __init int sched_init_debug(void) | |
767 | { | |
f00b45c1 PZ |
768 | debugfs_create_file("sched_features", 0644, NULL, NULL, |
769 | &sched_feat_fops); | |
770 | ||
771 | return 0; | |
772 | } | |
773 | late_initcall(sched_init_debug); | |
774 | ||
775 | #endif | |
776 | ||
777 | #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x)) | |
bf5c91ba | 778 | |
b82d9fdd PZ |
779 | /* |
780 | * Number of tasks to iterate in a single balance run. | |
781 | * Limited because this is done with IRQs disabled. | |
782 | */ | |
783 | const_debug unsigned int sysctl_sched_nr_migrate = 32; | |
784 | ||
fa85ae24 | 785 | /* |
9f0c1e56 | 786 | * period over which we measure -rt task cpu usage in us. |
fa85ae24 PZ |
787 | * default: 1s |
788 | */ | |
9f0c1e56 | 789 | unsigned int sysctl_sched_rt_period = 1000000; |
fa85ae24 | 790 | |
6892b75e IM |
791 | static __read_mostly int scheduler_running; |
792 | ||
9f0c1e56 PZ |
793 | /* |
794 | * part of the period that we allow rt tasks to run in us. | |
795 | * default: 0.95s | |
796 | */ | |
797 | int sysctl_sched_rt_runtime = 950000; | |
fa85ae24 | 798 | |
d0b27fa7 PZ |
799 | static inline u64 global_rt_period(void) |
800 | { | |
801 | return (u64)sysctl_sched_rt_period * NSEC_PER_USEC; | |
802 | } | |
803 | ||
804 | static inline u64 global_rt_runtime(void) | |
805 | { | |
806 | if (sysctl_sched_rt_period < 0) | |
807 | return RUNTIME_INF; | |
808 | ||
809 | return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC; | |
810 | } | |
fa85ae24 | 811 | |
690229a0 | 812 | unsigned long long time_sync_thresh = 100000; |
27ec4407 IM |
813 | |
814 | static DEFINE_PER_CPU(unsigned long long, time_offset); | |
815 | static DEFINE_PER_CPU(unsigned long long, prev_cpu_time); | |
816 | ||
e436d800 | 817 | /* |
27ec4407 IM |
818 | * Global lock which we take every now and then to synchronize |
819 | * the CPUs time. This method is not warp-safe, but it's good | |
820 | * enough to synchronize slowly diverging time sources and thus | |
821 | * it's good enough for tracing: | |
e436d800 | 822 | */ |
27ec4407 IM |
823 | static DEFINE_SPINLOCK(time_sync_lock); |
824 | static unsigned long long prev_global_time; | |
825 | ||
dfbf4a1b | 826 | static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu) |
27ec4407 | 827 | { |
dfbf4a1b IM |
828 | /* |
829 | * We want this inlined, to not get tracer function calls | |
830 | * in this critical section: | |
831 | */ | |
832 | spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_); | |
833 | __raw_spin_lock(&time_sync_lock.raw_lock); | |
27ec4407 IM |
834 | |
835 | if (time < prev_global_time) { | |
836 | per_cpu(time_offset, cpu) += prev_global_time - time; | |
837 | time = prev_global_time; | |
838 | } else { | |
839 | prev_global_time = time; | |
840 | } | |
841 | ||
dfbf4a1b IM |
842 | __raw_spin_unlock(&time_sync_lock.raw_lock); |
843 | spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_); | |
27ec4407 IM |
844 | |
845 | return time; | |
846 | } | |
847 | ||
848 | static unsigned long long __cpu_clock(int cpu) | |
e436d800 | 849 | { |
e436d800 | 850 | unsigned long long now; |
e436d800 | 851 | |
8ced5f69 IM |
852 | /* |
853 | * Only call sched_clock() if the scheduler has already been | |
854 | * initialized (some code might call cpu_clock() very early): | |
855 | */ | |
6892b75e IM |
856 | if (unlikely(!scheduler_running)) |
857 | return 0; | |
858 | ||
3e51f33f | 859 | now = sched_clock_cpu(cpu); |
e436d800 IM |
860 | |
861 | return now; | |
862 | } | |
27ec4407 IM |
863 | |
864 | /* | |
865 | * For kernel-internal use: high-speed (but slightly incorrect) per-cpu | |
866 | * clock constructed from sched_clock(): | |
867 | */ | |
868 | unsigned long long cpu_clock(int cpu) | |
869 | { | |
870 | unsigned long long prev_cpu_time, time, delta_time; | |
dfbf4a1b | 871 | unsigned long flags; |
27ec4407 | 872 | |
dfbf4a1b | 873 | local_irq_save(flags); |
27ec4407 IM |
874 | prev_cpu_time = per_cpu(prev_cpu_time, cpu); |
875 | time = __cpu_clock(cpu) + per_cpu(time_offset, cpu); | |
876 | delta_time = time-prev_cpu_time; | |
877 | ||
dfbf4a1b | 878 | if (unlikely(delta_time > time_sync_thresh)) { |
27ec4407 | 879 | time = __sync_cpu_clock(time, cpu); |
dfbf4a1b IM |
880 | per_cpu(prev_cpu_time, cpu) = time; |
881 | } | |
882 | local_irq_restore(flags); | |
27ec4407 IM |
883 | |
884 | return time; | |
885 | } | |
a58f6f25 | 886 | EXPORT_SYMBOL_GPL(cpu_clock); |
e436d800 | 887 | |
1da177e4 | 888 | #ifndef prepare_arch_switch |
4866cde0 NP |
889 | # define prepare_arch_switch(next) do { } while (0) |
890 | #endif | |
891 | #ifndef finish_arch_switch | |
892 | # define finish_arch_switch(prev) do { } while (0) | |
893 | #endif | |
894 | ||
051a1d1a DA |
895 | static inline int task_current(struct rq *rq, struct task_struct *p) |
896 | { | |
897 | return rq->curr == p; | |
898 | } | |
899 | ||
4866cde0 | 900 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW |
70b97a7f | 901 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 | 902 | { |
051a1d1a | 903 | return task_current(rq, p); |
4866cde0 NP |
904 | } |
905 | ||
70b97a7f | 906 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
907 | { |
908 | } | |
909 | ||
70b97a7f | 910 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 | 911 | { |
da04c035 IM |
912 | #ifdef CONFIG_DEBUG_SPINLOCK |
913 | /* this is a valid case when another task releases the spinlock */ | |
914 | rq->lock.owner = current; | |
915 | #endif | |
8a25d5de IM |
916 | /* |
917 | * If we are tracking spinlock dependencies then we have to | |
918 | * fix up the runqueue lock - which gets 'carried over' from | |
919 | * prev into current: | |
920 | */ | |
921 | spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_); | |
922 | ||
4866cde0 NP |
923 | spin_unlock_irq(&rq->lock); |
924 | } | |
925 | ||
926 | #else /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
70b97a7f | 927 | static inline int task_running(struct rq *rq, struct task_struct *p) |
4866cde0 NP |
928 | { |
929 | #ifdef CONFIG_SMP | |
930 | return p->oncpu; | |
931 | #else | |
051a1d1a | 932 | return task_current(rq, p); |
4866cde0 NP |
933 | #endif |
934 | } | |
935 | ||
70b97a7f | 936 | static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next) |
4866cde0 NP |
937 | { |
938 | #ifdef CONFIG_SMP | |
939 | /* | |
940 | * We can optimise this out completely for !SMP, because the | |
941 | * SMP rebalancing from interrupt is the only thing that cares | |
942 | * here. | |
943 | */ | |
944 | next->oncpu = 1; | |
945 | #endif | |
946 | #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
947 | spin_unlock_irq(&rq->lock); | |
948 | #else | |
949 | spin_unlock(&rq->lock); | |
950 | #endif | |
951 | } | |
952 | ||
70b97a7f | 953 | static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev) |
4866cde0 NP |
954 | { |
955 | #ifdef CONFIG_SMP | |
956 | /* | |
957 | * After ->oncpu is cleared, the task can be moved to a different CPU. | |
958 | * We must ensure this doesn't happen until the switch is completely | |
959 | * finished. | |
960 | */ | |
961 | smp_wmb(); | |
962 | prev->oncpu = 0; | |
963 | #endif | |
964 | #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW | |
965 | local_irq_enable(); | |
1da177e4 | 966 | #endif |
4866cde0 NP |
967 | } |
968 | #endif /* __ARCH_WANT_UNLOCKED_CTXSW */ | |
1da177e4 | 969 | |
b29739f9 IM |
970 | /* |
971 | * __task_rq_lock - lock the runqueue a given task resides on. | |
972 | * Must be called interrupts disabled. | |
973 | */ | |
70b97a7f | 974 | static inline struct rq *__task_rq_lock(struct task_struct *p) |
b29739f9 IM |
975 | __acquires(rq->lock) |
976 | { | |
3a5c359a AK |
977 | for (;;) { |
978 | struct rq *rq = task_rq(p); | |
979 | spin_lock(&rq->lock); | |
980 | if (likely(rq == task_rq(p))) | |
981 | return rq; | |
b29739f9 | 982 | spin_unlock(&rq->lock); |
b29739f9 | 983 | } |
b29739f9 IM |
984 | } |
985 | ||
1da177e4 LT |
986 | /* |
987 | * task_rq_lock - lock the runqueue a given task resides on and disable | |
41a2d6cf | 988 | * interrupts. Note the ordering: we can safely lookup the task_rq without |
1da177e4 LT |
989 | * explicitly disabling preemption. |
990 | */ | |
70b97a7f | 991 | static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) |
1da177e4 LT |
992 | __acquires(rq->lock) |
993 | { | |
70b97a7f | 994 | struct rq *rq; |
1da177e4 | 995 | |
3a5c359a AK |
996 | for (;;) { |
997 | local_irq_save(*flags); | |
998 | rq = task_rq(p); | |
999 | spin_lock(&rq->lock); | |
1000 | if (likely(rq == task_rq(p))) | |
1001 | return rq; | |
1da177e4 | 1002 | spin_unlock_irqrestore(&rq->lock, *flags); |
1da177e4 | 1003 | } |
1da177e4 LT |
1004 | } |
1005 | ||
a9957449 | 1006 | static void __task_rq_unlock(struct rq *rq) |
b29739f9 IM |
1007 | __releases(rq->lock) |
1008 | { | |
1009 | spin_unlock(&rq->lock); | |
1010 | } | |
1011 | ||
70b97a7f | 1012 | static inline void task_rq_unlock(struct rq *rq, unsigned long *flags) |
1da177e4 LT |
1013 | __releases(rq->lock) |
1014 | { | |
1015 | spin_unlock_irqrestore(&rq->lock, *flags); | |
1016 | } | |
1017 | ||
1da177e4 | 1018 | /* |
cc2a73b5 | 1019 | * this_rq_lock - lock this runqueue and disable interrupts. |
1da177e4 | 1020 | */ |
a9957449 | 1021 | static struct rq *this_rq_lock(void) |
1da177e4 LT |
1022 | __acquires(rq->lock) |
1023 | { | |
70b97a7f | 1024 | struct rq *rq; |
1da177e4 LT |
1025 | |
1026 | local_irq_disable(); | |
1027 | rq = this_rq(); | |
1028 | spin_lock(&rq->lock); | |
1029 | ||
1030 | return rq; | |
1031 | } | |
1032 | ||
8f4d37ec PZ |
1033 | static void __resched_task(struct task_struct *p, int tif_bit); |
1034 | ||
1035 | static inline void resched_task(struct task_struct *p) | |
1036 | { | |
1037 | __resched_task(p, TIF_NEED_RESCHED); | |
1038 | } | |
1039 | ||
1040 | #ifdef CONFIG_SCHED_HRTICK | |
1041 | /* | |
1042 | * Use HR-timers to deliver accurate preemption points. | |
1043 | * | |
1044 | * Its all a bit involved since we cannot program an hrt while holding the | |
1045 | * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a | |
1046 | * reschedule event. | |
1047 | * | |
1048 | * When we get rescheduled we reprogram the hrtick_timer outside of the | |
1049 | * rq->lock. | |
1050 | */ | |
1051 | static inline void resched_hrt(struct task_struct *p) | |
1052 | { | |
1053 | __resched_task(p, TIF_HRTICK_RESCHED); | |
1054 | } | |
1055 | ||
1056 | static inline void resched_rq(struct rq *rq) | |
1057 | { | |
1058 | unsigned long flags; | |
1059 | ||
1060 | spin_lock_irqsave(&rq->lock, flags); | |
1061 | resched_task(rq->curr); | |
1062 | spin_unlock_irqrestore(&rq->lock, flags); | |
1063 | } | |
1064 | ||
1065 | enum { | |
1066 | HRTICK_SET, /* re-programm hrtick_timer */ | |
1067 | HRTICK_RESET, /* not a new slice */ | |
b328ca18 | 1068 | HRTICK_BLOCK, /* stop hrtick operations */ |
8f4d37ec PZ |
1069 | }; |
1070 | ||
1071 | /* | |
1072 | * Use hrtick when: | |
1073 | * - enabled by features | |
1074 | * - hrtimer is actually high res | |
1075 | */ | |
1076 | static inline int hrtick_enabled(struct rq *rq) | |
1077 | { | |
1078 | if (!sched_feat(HRTICK)) | |
1079 | return 0; | |
b328ca18 PZ |
1080 | if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags))) |
1081 | return 0; | |
8f4d37ec PZ |
1082 | return hrtimer_is_hres_active(&rq->hrtick_timer); |
1083 | } | |
1084 | ||
1085 | /* | |
1086 | * Called to set the hrtick timer state. | |
1087 | * | |
1088 | * called with rq->lock held and irqs disabled | |
1089 | */ | |
1090 | static void hrtick_start(struct rq *rq, u64 delay, int reset) | |
1091 | { | |
1092 | assert_spin_locked(&rq->lock); | |
1093 | ||
1094 | /* | |
1095 | * preempt at: now + delay | |
1096 | */ | |
1097 | rq->hrtick_expire = | |
1098 | ktime_add_ns(rq->hrtick_timer.base->get_time(), delay); | |
1099 | /* | |
1100 | * indicate we need to program the timer | |
1101 | */ | |
1102 | __set_bit(HRTICK_SET, &rq->hrtick_flags); | |
1103 | if (reset) | |
1104 | __set_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1105 | ||
1106 | /* | |
1107 | * New slices are called from the schedule path and don't need a | |
1108 | * forced reschedule. | |
1109 | */ | |
1110 | if (reset) | |
1111 | resched_hrt(rq->curr); | |
1112 | } | |
1113 | ||
1114 | static void hrtick_clear(struct rq *rq) | |
1115 | { | |
1116 | if (hrtimer_active(&rq->hrtick_timer)) | |
1117 | hrtimer_cancel(&rq->hrtick_timer); | |
1118 | } | |
1119 | ||
1120 | /* | |
1121 | * Update the timer from the possible pending state. | |
1122 | */ | |
1123 | static void hrtick_set(struct rq *rq) | |
1124 | { | |
1125 | ktime_t time; | |
1126 | int set, reset; | |
1127 | unsigned long flags; | |
1128 | ||
1129 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1130 | ||
1131 | spin_lock_irqsave(&rq->lock, flags); | |
1132 | set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags); | |
1133 | reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags); | |
1134 | time = rq->hrtick_expire; | |
1135 | clear_thread_flag(TIF_HRTICK_RESCHED); | |
1136 | spin_unlock_irqrestore(&rq->lock, flags); | |
1137 | ||
1138 | if (set) { | |
1139 | hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS); | |
1140 | if (reset && !hrtimer_active(&rq->hrtick_timer)) | |
1141 | resched_rq(rq); | |
1142 | } else | |
1143 | hrtick_clear(rq); | |
1144 | } | |
1145 | ||
1146 | /* | |
1147 | * High-resolution timer tick. | |
1148 | * Runs from hardirq context with interrupts disabled. | |
1149 | */ | |
1150 | static enum hrtimer_restart hrtick(struct hrtimer *timer) | |
1151 | { | |
1152 | struct rq *rq = container_of(timer, struct rq, hrtick_timer); | |
1153 | ||
1154 | WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); | |
1155 | ||
1156 | spin_lock(&rq->lock); | |
3e51f33f | 1157 | update_rq_clock(rq); |
8f4d37ec PZ |
1158 | rq->curr->sched_class->task_tick(rq, rq->curr, 1); |
1159 | spin_unlock(&rq->lock); | |
1160 | ||
1161 | return HRTIMER_NORESTART; | |
1162 | } | |
1163 | ||
b328ca18 PZ |
1164 | static void hotplug_hrtick_disable(int cpu) |
1165 | { | |
1166 | struct rq *rq = cpu_rq(cpu); | |
1167 | unsigned long flags; | |
1168 | ||
1169 | spin_lock_irqsave(&rq->lock, flags); | |
1170 | rq->hrtick_flags = 0; | |
1171 | __set_bit(HRTICK_BLOCK, &rq->hrtick_flags); | |
1172 | spin_unlock_irqrestore(&rq->lock, flags); | |
1173 | ||
1174 | hrtick_clear(rq); | |
1175 | } | |
1176 | ||
1177 | static void hotplug_hrtick_enable(int cpu) | |
1178 | { | |
1179 | struct rq *rq = cpu_rq(cpu); | |
1180 | unsigned long flags; | |
1181 | ||
1182 | spin_lock_irqsave(&rq->lock, flags); | |
1183 | __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags); | |
1184 | spin_unlock_irqrestore(&rq->lock, flags); | |
1185 | } | |
1186 | ||
1187 | static int | |
1188 | hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1189 | { | |
1190 | int cpu = (int)(long)hcpu; | |
1191 | ||
1192 | switch (action) { | |
1193 | case CPU_UP_CANCELED: | |
1194 | case CPU_UP_CANCELED_FROZEN: | |
1195 | case CPU_DOWN_PREPARE: | |
1196 | case CPU_DOWN_PREPARE_FROZEN: | |
1197 | case CPU_DEAD: | |
1198 | case CPU_DEAD_FROZEN: | |
1199 | hotplug_hrtick_disable(cpu); | |
1200 | return NOTIFY_OK; | |
1201 | ||
1202 | case CPU_UP_PREPARE: | |
1203 | case CPU_UP_PREPARE_FROZEN: | |
1204 | case CPU_DOWN_FAILED: | |
1205 | case CPU_DOWN_FAILED_FROZEN: | |
1206 | case CPU_ONLINE: | |
1207 | case CPU_ONLINE_FROZEN: | |
1208 | hotplug_hrtick_enable(cpu); | |
1209 | return NOTIFY_OK; | |
1210 | } | |
1211 | ||
1212 | return NOTIFY_DONE; | |
1213 | } | |
1214 | ||
1215 | static void init_hrtick(void) | |
1216 | { | |
1217 | hotcpu_notifier(hotplug_hrtick, 0); | |
1218 | } | |
1219 | ||
1220 | static void init_rq_hrtick(struct rq *rq) | |
8f4d37ec PZ |
1221 | { |
1222 | rq->hrtick_flags = 0; | |
1223 | hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
1224 | rq->hrtick_timer.function = hrtick; | |
1225 | rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ; | |
1226 | } | |
1227 | ||
1228 | void hrtick_resched(void) | |
1229 | { | |
1230 | struct rq *rq; | |
1231 | unsigned long flags; | |
1232 | ||
1233 | if (!test_thread_flag(TIF_HRTICK_RESCHED)) | |
1234 | return; | |
1235 | ||
1236 | local_irq_save(flags); | |
1237 | rq = cpu_rq(smp_processor_id()); | |
1238 | hrtick_set(rq); | |
1239 | local_irq_restore(flags); | |
1240 | } | |
1241 | #else | |
1242 | static inline void hrtick_clear(struct rq *rq) | |
1243 | { | |
1244 | } | |
1245 | ||
1246 | static inline void hrtick_set(struct rq *rq) | |
1247 | { | |
1248 | } | |
1249 | ||
1250 | static inline void init_rq_hrtick(struct rq *rq) | |
1251 | { | |
1252 | } | |
1253 | ||
1254 | void hrtick_resched(void) | |
1255 | { | |
1256 | } | |
b328ca18 PZ |
1257 | |
1258 | static inline void init_hrtick(void) | |
1259 | { | |
1260 | } | |
8f4d37ec PZ |
1261 | #endif |
1262 | ||
c24d20db IM |
1263 | /* |
1264 | * resched_task - mark a task 'to be rescheduled now'. | |
1265 | * | |
1266 | * On UP this means the setting of the need_resched flag, on SMP it | |
1267 | * might also involve a cross-CPU call to trigger the scheduler on | |
1268 | * the target CPU. | |
1269 | */ | |
1270 | #ifdef CONFIG_SMP | |
1271 | ||
1272 | #ifndef tsk_is_polling | |
1273 | #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) | |
1274 | #endif | |
1275 | ||
8f4d37ec | 1276 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1277 | { |
1278 | int cpu; | |
1279 | ||
1280 | assert_spin_locked(&task_rq(p)->lock); | |
1281 | ||
8f4d37ec | 1282 | if (unlikely(test_tsk_thread_flag(p, tif_bit))) |
c24d20db IM |
1283 | return; |
1284 | ||
8f4d37ec | 1285 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1286 | |
1287 | cpu = task_cpu(p); | |
1288 | if (cpu == smp_processor_id()) | |
1289 | return; | |
1290 | ||
1291 | /* NEED_RESCHED must be visible before we test polling */ | |
1292 | smp_mb(); | |
1293 | if (!tsk_is_polling(p)) | |
1294 | smp_send_reschedule(cpu); | |
1295 | } | |
1296 | ||
1297 | static void resched_cpu(int cpu) | |
1298 | { | |
1299 | struct rq *rq = cpu_rq(cpu); | |
1300 | unsigned long flags; | |
1301 | ||
1302 | if (!spin_trylock_irqsave(&rq->lock, flags)) | |
1303 | return; | |
1304 | resched_task(cpu_curr(cpu)); | |
1305 | spin_unlock_irqrestore(&rq->lock, flags); | |
1306 | } | |
06d8308c TG |
1307 | |
1308 | #ifdef CONFIG_NO_HZ | |
1309 | /* | |
1310 | * When add_timer_on() enqueues a timer into the timer wheel of an | |
1311 | * idle CPU then this timer might expire before the next timer event | |
1312 | * which is scheduled to wake up that CPU. In case of a completely | |
1313 | * idle system the next event might even be infinite time into the | |
1314 | * future. wake_up_idle_cpu() ensures that the CPU is woken up and | |
1315 | * leaves the inner idle loop so the newly added timer is taken into | |
1316 | * account when the CPU goes back to idle and evaluates the timer | |
1317 | * wheel for the next timer event. | |
1318 | */ | |
1319 | void wake_up_idle_cpu(int cpu) | |
1320 | { | |
1321 | struct rq *rq = cpu_rq(cpu); | |
1322 | ||
1323 | if (cpu == smp_processor_id()) | |
1324 | return; | |
1325 | ||
1326 | /* | |
1327 | * This is safe, as this function is called with the timer | |
1328 | * wheel base lock of (cpu) held. When the CPU is on the way | |
1329 | * to idle and has not yet set rq->curr to idle then it will | |
1330 | * be serialized on the timer wheel base lock and take the new | |
1331 | * timer into account automatically. | |
1332 | */ | |
1333 | if (rq->curr != rq->idle) | |
1334 | return; | |
1335 | ||
1336 | /* | |
1337 | * We can set TIF_RESCHED on the idle task of the other CPU | |
1338 | * lockless. The worst case is that the other CPU runs the | |
1339 | * idle task through an additional NOOP schedule() | |
1340 | */ | |
1341 | set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED); | |
1342 | ||
1343 | /* NEED_RESCHED must be visible before we test polling */ | |
1344 | smp_mb(); | |
1345 | if (!tsk_is_polling(rq->idle)) | |
1346 | smp_send_reschedule(cpu); | |
1347 | } | |
1348 | #endif | |
1349 | ||
c24d20db | 1350 | #else |
8f4d37ec | 1351 | static void __resched_task(struct task_struct *p, int tif_bit) |
c24d20db IM |
1352 | { |
1353 | assert_spin_locked(&task_rq(p)->lock); | |
8f4d37ec | 1354 | set_tsk_thread_flag(p, tif_bit); |
c24d20db IM |
1355 | } |
1356 | #endif | |
1357 | ||
45bf76df IM |
1358 | #if BITS_PER_LONG == 32 |
1359 | # define WMULT_CONST (~0UL) | |
1360 | #else | |
1361 | # define WMULT_CONST (1UL << 32) | |
1362 | #endif | |
1363 | ||
1364 | #define WMULT_SHIFT 32 | |
1365 | ||
194081eb IM |
1366 | /* |
1367 | * Shift right and round: | |
1368 | */ | |
cf2ab469 | 1369 | #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y)) |
194081eb | 1370 | |
8f1bc385 PZ |
1371 | /* |
1372 | * delta *= weight / lw | |
1373 | */ | |
cb1c4fc9 | 1374 | static unsigned long |
45bf76df IM |
1375 | calc_delta_mine(unsigned long delta_exec, unsigned long weight, |
1376 | struct load_weight *lw) | |
1377 | { | |
1378 | u64 tmp; | |
1379 | ||
e05510d0 PZ |
1380 | if (!lw->inv_weight) |
1381 | lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1); | |
45bf76df IM |
1382 | |
1383 | tmp = (u64)delta_exec * weight; | |
1384 | /* | |
1385 | * Check whether we'd overflow the 64-bit multiplication: | |
1386 | */ | |
194081eb | 1387 | if (unlikely(tmp > WMULT_CONST)) |
cf2ab469 | 1388 | tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight, |
194081eb IM |
1389 | WMULT_SHIFT/2); |
1390 | else | |
cf2ab469 | 1391 | tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT); |
45bf76df | 1392 | |
ecf691da | 1393 | return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX); |
45bf76df IM |
1394 | } |
1395 | ||
1091985b | 1396 | static inline void update_load_add(struct load_weight *lw, unsigned long inc) |
45bf76df IM |
1397 | { |
1398 | lw->weight += inc; | |
e89996ae | 1399 | lw->inv_weight = 0; |
45bf76df IM |
1400 | } |
1401 | ||
1091985b | 1402 | static inline void update_load_sub(struct load_weight *lw, unsigned long dec) |
45bf76df IM |
1403 | { |
1404 | lw->weight -= dec; | |
e89996ae | 1405 | lw->inv_weight = 0; |
45bf76df IM |
1406 | } |
1407 | ||
2dd73a4f PW |
1408 | /* |
1409 | * To aid in avoiding the subversion of "niceness" due to uneven distribution | |
1410 | * of tasks with abnormal "nice" values across CPUs the contribution that | |
1411 | * each task makes to its run queue's load is weighted according to its | |
41a2d6cf | 1412 | * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a |
2dd73a4f PW |
1413 | * scaled version of the new time slice allocation that they receive on time |
1414 | * slice expiry etc. | |
1415 | */ | |
1416 | ||
dd41f596 IM |
1417 | #define WEIGHT_IDLEPRIO 2 |
1418 | #define WMULT_IDLEPRIO (1 << 31) | |
1419 | ||
1420 | /* | |
1421 | * Nice levels are multiplicative, with a gentle 10% change for every | |
1422 | * nice level changed. I.e. when a CPU-bound task goes from nice 0 to | |
1423 | * nice 1, it will get ~10% less CPU time than another CPU-bound task | |
1424 | * that remained on nice 0. | |
1425 | * | |
1426 | * The "10% effect" is relative and cumulative: from _any_ nice level, | |
1427 | * if you go up 1 level, it's -10% CPU usage, if you go down 1 level | |
f9153ee6 IM |
1428 | * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25. |
1429 | * If a task goes up by ~10% and another task goes down by ~10% then | |
1430 | * the relative distance between them is ~25%.) | |
dd41f596 IM |
1431 | */ |
1432 | static const int prio_to_weight[40] = { | |
254753dc IM |
1433 | /* -20 */ 88761, 71755, 56483, 46273, 36291, |
1434 | /* -15 */ 29154, 23254, 18705, 14949, 11916, | |
1435 | /* -10 */ 9548, 7620, 6100, 4904, 3906, | |
1436 | /* -5 */ 3121, 2501, 1991, 1586, 1277, | |
1437 | /* 0 */ 1024, 820, 655, 526, 423, | |
1438 | /* 5 */ 335, 272, 215, 172, 137, | |
1439 | /* 10 */ 110, 87, 70, 56, 45, | |
1440 | /* 15 */ 36, 29, 23, 18, 15, | |
dd41f596 IM |
1441 | }; |
1442 | ||
5714d2de IM |
1443 | /* |
1444 | * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated. | |
1445 | * | |
1446 | * In cases where the weight does not change often, we can use the | |
1447 | * precalculated inverse to speed up arithmetics by turning divisions | |
1448 | * into multiplications: | |
1449 | */ | |
dd41f596 | 1450 | static const u32 prio_to_wmult[40] = { |
254753dc IM |
1451 | /* -20 */ 48388, 59856, 76040, 92818, 118348, |
1452 | /* -15 */ 147320, 184698, 229616, 287308, 360437, | |
1453 | /* -10 */ 449829, 563644, 704093, 875809, 1099582, | |
1454 | /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326, | |
1455 | /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587, | |
1456 | /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126, | |
1457 | /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717, | |
1458 | /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153, | |
dd41f596 | 1459 | }; |
2dd73a4f | 1460 | |
dd41f596 IM |
1461 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup); |
1462 | ||
1463 | /* | |
1464 | * runqueue iterator, to support SMP load-balancing between different | |
1465 | * scheduling classes, without having to expose their internal data | |
1466 | * structures to the load-balancing proper: | |
1467 | */ | |
1468 | struct rq_iterator { | |
1469 | void *arg; | |
1470 | struct task_struct *(*start)(void *); | |
1471 | struct task_struct *(*next)(void *); | |
1472 | }; | |
1473 | ||
e1d1484f PW |
1474 | #ifdef CONFIG_SMP |
1475 | static unsigned long | |
1476 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1477 | unsigned long max_load_move, struct sched_domain *sd, | |
1478 | enum cpu_idle_type idle, int *all_pinned, | |
1479 | int *this_best_prio, struct rq_iterator *iterator); | |
1480 | ||
1481 | static int | |
1482 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
1483 | struct sched_domain *sd, enum cpu_idle_type idle, | |
1484 | struct rq_iterator *iterator); | |
e1d1484f | 1485 | #endif |
dd41f596 | 1486 | |
d842de87 SV |
1487 | #ifdef CONFIG_CGROUP_CPUACCT |
1488 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime); | |
1489 | #else | |
1490 | static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {} | |
1491 | #endif | |
1492 | ||
18d95a28 PZ |
1493 | static inline void inc_cpu_load(struct rq *rq, unsigned long load) |
1494 | { | |
1495 | update_load_add(&rq->load, load); | |
1496 | } | |
1497 | ||
1498 | static inline void dec_cpu_load(struct rq *rq, unsigned long load) | |
1499 | { | |
1500 | update_load_sub(&rq->load, load); | |
1501 | } | |
1502 | ||
e7693a36 GH |
1503 | #ifdef CONFIG_SMP |
1504 | static unsigned long source_load(int cpu, int type); | |
1505 | static unsigned long target_load(int cpu, int type); | |
1506 | static unsigned long cpu_avg_load_per_task(int cpu); | |
1507 | static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd); | |
18d95a28 PZ |
1508 | |
1509 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1510 | ||
1511 | /* | |
1512 | * Group load balancing. | |
1513 | * | |
1514 | * We calculate a few balance domain wide aggregate numbers; load and weight. | |
1515 | * Given the pictures below, and assuming each item has equal weight: | |
1516 | * | |
1517 | * root 1 - thread | |
1518 | * / | \ A - group | |
1519 | * A 1 B | |
1520 | * /|\ / \ | |
1521 | * C 2 D 3 4 | |
1522 | * | | | |
1523 | * 5 6 | |
1524 | * | |
1525 | * load: | |
1526 | * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd, | |
1527 | * which equals 1/9-th of the total load. | |
1528 | * | |
1529 | * shares: | |
1530 | * The weight of this group on the selected cpus. | |
1531 | * | |
1532 | * rq_weight: | |
1533 | * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while | |
1534 | * B would get 2. | |
1535 | * | |
1536 | * task_weight: | |
1537 | * Part of the rq_weight contributed by tasks; all groups except B would | |
1538 | * get 1, B gets 2. | |
1539 | */ | |
1540 | ||
1541 | static inline struct aggregate_struct * | |
1542 | aggregate(struct task_group *tg, struct sched_domain *sd) | |
1543 | { | |
1544 | return &tg->cfs_rq[sd->first_cpu]->aggregate; | |
1545 | } | |
1546 | ||
1547 | typedef void (*aggregate_func)(struct task_group *, struct sched_domain *); | |
1548 | ||
1549 | /* | |
1550 | * Iterate the full tree, calling @down when first entering a node and @up when | |
1551 | * leaving it for the final time. | |
1552 | */ | |
1553 | static | |
1554 | void aggregate_walk_tree(aggregate_func down, aggregate_func up, | |
1555 | struct sched_domain *sd) | |
1556 | { | |
1557 | struct task_group *parent, *child; | |
1558 | ||
1559 | rcu_read_lock(); | |
1560 | parent = &root_task_group; | |
1561 | down: | |
1562 | (*down)(parent, sd); | |
1563 | list_for_each_entry_rcu(child, &parent->children, siblings) { | |
1564 | parent = child; | |
1565 | goto down; | |
1566 | ||
1567 | up: | |
1568 | continue; | |
1569 | } | |
1570 | (*up)(parent, sd); | |
1571 | ||
1572 | child = parent; | |
1573 | parent = parent->parent; | |
1574 | if (parent) | |
1575 | goto up; | |
1576 | rcu_read_unlock(); | |
1577 | } | |
1578 | ||
1579 | /* | |
1580 | * Calculate the aggregate runqueue weight. | |
1581 | */ | |
1582 | static | |
1583 | void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd) | |
1584 | { | |
1585 | unsigned long rq_weight = 0; | |
1586 | unsigned long task_weight = 0; | |
1587 | int i; | |
1588 | ||
1589 | for_each_cpu_mask(i, sd->span) { | |
1590 | rq_weight += tg->cfs_rq[i]->load.weight; | |
1591 | task_weight += tg->cfs_rq[i]->task_weight; | |
1592 | } | |
1593 | ||
1594 | aggregate(tg, sd)->rq_weight = rq_weight; | |
1595 | aggregate(tg, sd)->task_weight = task_weight; | |
1596 | } | |
1597 | ||
18d95a28 PZ |
1598 | /* |
1599 | * Compute the weight of this group on the given cpus. | |
1600 | */ | |
1601 | static | |
1602 | void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd) | |
1603 | { | |
1604 | unsigned long shares = 0; | |
1605 | int i; | |
1606 | ||
18d95a28 PZ |
1607 | for_each_cpu_mask(i, sd->span) |
1608 | shares += tg->cfs_rq[i]->shares; | |
1609 | ||
3f5087a2 PZ |
1610 | if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares) |
1611 | shares = tg->shares; | |
18d95a28 PZ |
1612 | |
1613 | aggregate(tg, sd)->shares = shares; | |
1614 | } | |
1615 | ||
1616 | /* | |
1617 | * Compute the load fraction assigned to this group, relies on the aggregate | |
1618 | * weight and this group's parent's load, i.e. top-down. | |
1619 | */ | |
1620 | static | |
1621 | void aggregate_group_load(struct task_group *tg, struct sched_domain *sd) | |
1622 | { | |
1623 | unsigned long load; | |
1624 | ||
1625 | if (!tg->parent) { | |
1626 | int i; | |
1627 | ||
1628 | load = 0; | |
1629 | for_each_cpu_mask(i, sd->span) | |
1630 | load += cpu_rq(i)->load.weight; | |
1631 | ||
1632 | } else { | |
1633 | load = aggregate(tg->parent, sd)->load; | |
1634 | ||
1635 | /* | |
1636 | * shares is our weight in the parent's rq so | |
1637 | * shares/parent->rq_weight gives our fraction of the load | |
1638 | */ | |
1639 | load *= aggregate(tg, sd)->shares; | |
1640 | load /= aggregate(tg->parent, sd)->rq_weight + 1; | |
1641 | } | |
1642 | ||
1643 | aggregate(tg, sd)->load = load; | |
1644 | } | |
1645 | ||
1646 | static void __set_se_shares(struct sched_entity *se, unsigned long shares); | |
1647 | ||
1648 | /* | |
1649 | * Calculate and set the cpu's group shares. | |
1650 | */ | |
1651 | static void | |
1652 | __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd, | |
1653 | int tcpu) | |
1654 | { | |
1655 | int boost = 0; | |
1656 | unsigned long shares; | |
1657 | unsigned long rq_weight; | |
1658 | ||
1659 | if (!tg->se[tcpu]) | |
1660 | return; | |
1661 | ||
1662 | rq_weight = tg->cfs_rq[tcpu]->load.weight; | |
1663 | ||
1664 | /* | |
1665 | * If there are currently no tasks on the cpu pretend there is one of | |
1666 | * average load so that when a new task gets to run here it will not | |
1667 | * get delayed by group starvation. | |
1668 | */ | |
1669 | if (!rq_weight) { | |
1670 | boost = 1; | |
1671 | rq_weight = NICE_0_LOAD; | |
1672 | } | |
1673 | ||
1674 | /* | |
1675 | * \Sum shares * rq_weight | |
1676 | * shares = ----------------------- | |
1677 | * \Sum rq_weight | |
1678 | * | |
1679 | */ | |
1680 | shares = aggregate(tg, sd)->shares * rq_weight; | |
1681 | shares /= aggregate(tg, sd)->rq_weight + 1; | |
1682 | ||
1683 | /* | |
1684 | * record the actual number of shares, not the boosted amount. | |
1685 | */ | |
1686 | tg->cfs_rq[tcpu]->shares = boost ? 0 : shares; | |
1687 | ||
1688 | if (shares < MIN_SHARES) | |
1689 | shares = MIN_SHARES; | |
cb4ad1ff MX |
1690 | else if (shares > MAX_SHARES) |
1691 | shares = MAX_SHARES; | |
18d95a28 PZ |
1692 | |
1693 | __set_se_shares(tg->se[tcpu], shares); | |
1694 | } | |
1695 | ||
1696 | /* | |
1697 | * Re-adjust the weights on the cpu the task came from and on the cpu the | |
1698 | * task went to. | |
1699 | */ | |
1700 | static void | |
1701 | __move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1702 | int scpu, int dcpu) | |
1703 | { | |
1704 | unsigned long shares; | |
1705 | ||
1706 | shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1707 | ||
1708 | __update_group_shares_cpu(tg, sd, scpu); | |
1709 | __update_group_shares_cpu(tg, sd, dcpu); | |
1710 | ||
1711 | /* | |
1712 | * ensure we never loose shares due to rounding errors in the | |
1713 | * above redistribution. | |
1714 | */ | |
1715 | shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares; | |
1716 | if (shares) | |
1717 | tg->cfs_rq[dcpu]->shares += shares; | |
1718 | } | |
1719 | ||
1720 | /* | |
1721 | * Because changing a group's shares changes the weight of the super-group | |
1722 | * we need to walk up the tree and change all shares until we hit the root. | |
1723 | */ | |
1724 | static void | |
1725 | move_group_shares(struct task_group *tg, struct sched_domain *sd, | |
1726 | int scpu, int dcpu) | |
1727 | { | |
1728 | while (tg) { | |
1729 | __move_group_shares(tg, sd, scpu, dcpu); | |
1730 | tg = tg->parent; | |
1731 | } | |
1732 | } | |
1733 | ||
1734 | static | |
1735 | void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd) | |
1736 | { | |
1737 | unsigned long shares = aggregate(tg, sd)->shares; | |
1738 | int i; | |
1739 | ||
1740 | for_each_cpu_mask(i, sd->span) { | |
1741 | struct rq *rq = cpu_rq(i); | |
1742 | unsigned long flags; | |
1743 | ||
1744 | spin_lock_irqsave(&rq->lock, flags); | |
1745 | __update_group_shares_cpu(tg, sd, i); | |
1746 | spin_unlock_irqrestore(&rq->lock, flags); | |
1747 | } | |
1748 | ||
1749 | aggregate_group_shares(tg, sd); | |
1750 | ||
1751 | /* | |
1752 | * ensure we never loose shares due to rounding errors in the | |
1753 | * above redistribution. | |
1754 | */ | |
1755 | shares -= aggregate(tg, sd)->shares; | |
1756 | if (shares) { | |
1757 | tg->cfs_rq[sd->first_cpu]->shares += shares; | |
1758 | aggregate(tg, sd)->shares += shares; | |
1759 | } | |
1760 | } | |
1761 | ||
1762 | /* | |
1763 | * Calculate the accumulative weight and recursive load of each task group | |
1764 | * while walking down the tree. | |
1765 | */ | |
1766 | static | |
1767 | void aggregate_get_down(struct task_group *tg, struct sched_domain *sd) | |
1768 | { | |
1769 | aggregate_group_weight(tg, sd); | |
1770 | aggregate_group_shares(tg, sd); | |
1771 | aggregate_group_load(tg, sd); | |
1772 | } | |
1773 | ||
1774 | /* | |
1775 | * Rebalance the cpu shares while walking back up the tree. | |
1776 | */ | |
1777 | static | |
1778 | void aggregate_get_up(struct task_group *tg, struct sched_domain *sd) | |
1779 | { | |
1780 | aggregate_group_set_shares(tg, sd); | |
1781 | } | |
1782 | ||
1783 | static DEFINE_PER_CPU(spinlock_t, aggregate_lock); | |
1784 | ||
1785 | static void __init init_aggregate(void) | |
1786 | { | |
1787 | int i; | |
1788 | ||
1789 | for_each_possible_cpu(i) | |
1790 | spin_lock_init(&per_cpu(aggregate_lock, i)); | |
1791 | } | |
1792 | ||
1793 | static int get_aggregate(struct sched_domain *sd) | |
1794 | { | |
1795 | if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu))) | |
1796 | return 0; | |
1797 | ||
1798 | aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd); | |
1799 | return 1; | |
1800 | } | |
1801 | ||
1802 | static void put_aggregate(struct sched_domain *sd) | |
1803 | { | |
1804 | spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu)); | |
1805 | } | |
1806 | ||
1807 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1808 | { | |
1809 | cfs_rq->shares = shares; | |
1810 | } | |
1811 | ||
1812 | #else | |
1813 | ||
1814 | static inline void init_aggregate(void) | |
1815 | { | |
1816 | } | |
1817 | ||
1818 | static inline int get_aggregate(struct sched_domain *sd) | |
1819 | { | |
1820 | return 0; | |
1821 | } | |
1822 | ||
1823 | static inline void put_aggregate(struct sched_domain *sd) | |
1824 | { | |
1825 | } | |
1826 | #endif | |
1827 | ||
1828 | #else /* CONFIG_SMP */ | |
1829 | ||
1830 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
1831 | static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares) | |
1832 | { | |
1833 | } | |
1834 | #endif | |
1835 | ||
e7693a36 GH |
1836 | #endif /* CONFIG_SMP */ |
1837 | ||
dd41f596 | 1838 | #include "sched_stats.h" |
dd41f596 | 1839 | #include "sched_idletask.c" |
5522d5d5 IM |
1840 | #include "sched_fair.c" |
1841 | #include "sched_rt.c" | |
dd41f596 IM |
1842 | #ifdef CONFIG_SCHED_DEBUG |
1843 | # include "sched_debug.c" | |
1844 | #endif | |
1845 | ||
1846 | #define sched_class_highest (&rt_sched_class) | |
1847 | ||
18d95a28 | 1848 | static void inc_nr_running(struct rq *rq) |
9c217245 IM |
1849 | { |
1850 | rq->nr_running++; | |
9c217245 IM |
1851 | } |
1852 | ||
18d95a28 | 1853 | static void dec_nr_running(struct rq *rq) |
9c217245 IM |
1854 | { |
1855 | rq->nr_running--; | |
9c217245 IM |
1856 | } |
1857 | ||
45bf76df IM |
1858 | static void set_load_weight(struct task_struct *p) |
1859 | { | |
1860 | if (task_has_rt_policy(p)) { | |
dd41f596 IM |
1861 | p->se.load.weight = prio_to_weight[0] * 2; |
1862 | p->se.load.inv_weight = prio_to_wmult[0] >> 1; | |
1863 | return; | |
1864 | } | |
45bf76df | 1865 | |
dd41f596 IM |
1866 | /* |
1867 | * SCHED_IDLE tasks get minimal weight: | |
1868 | */ | |
1869 | if (p->policy == SCHED_IDLE) { | |
1870 | p->se.load.weight = WEIGHT_IDLEPRIO; | |
1871 | p->se.load.inv_weight = WMULT_IDLEPRIO; | |
1872 | return; | |
1873 | } | |
71f8bd46 | 1874 | |
dd41f596 IM |
1875 | p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO]; |
1876 | p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO]; | |
71f8bd46 IM |
1877 | } |
1878 | ||
8159f87e | 1879 | static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup) |
71f8bd46 | 1880 | { |
dd41f596 | 1881 | sched_info_queued(p); |
fd390f6a | 1882 | p->sched_class->enqueue_task(rq, p, wakeup); |
dd41f596 | 1883 | p->se.on_rq = 1; |
71f8bd46 IM |
1884 | } |
1885 | ||
69be72c1 | 1886 | static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep) |
71f8bd46 | 1887 | { |
f02231e5 | 1888 | p->sched_class->dequeue_task(rq, p, sleep); |
dd41f596 | 1889 | p->se.on_rq = 0; |
71f8bd46 IM |
1890 | } |
1891 | ||
14531189 | 1892 | /* |
dd41f596 | 1893 | * __normal_prio - return the priority that is based on the static prio |
14531189 | 1894 | */ |
14531189 IM |
1895 | static inline int __normal_prio(struct task_struct *p) |
1896 | { | |
dd41f596 | 1897 | return p->static_prio; |
14531189 IM |
1898 | } |
1899 | ||
b29739f9 IM |
1900 | /* |
1901 | * Calculate the expected normal priority: i.e. priority | |
1902 | * without taking RT-inheritance into account. Might be | |
1903 | * boosted by interactivity modifiers. Changes upon fork, | |
1904 | * setprio syscalls, and whenever the interactivity | |
1905 | * estimator recalculates. | |
1906 | */ | |
36c8b586 | 1907 | static inline int normal_prio(struct task_struct *p) |
b29739f9 IM |
1908 | { |
1909 | int prio; | |
1910 | ||
e05606d3 | 1911 | if (task_has_rt_policy(p)) |
b29739f9 IM |
1912 | prio = MAX_RT_PRIO-1 - p->rt_priority; |
1913 | else | |
1914 | prio = __normal_prio(p); | |
1915 | return prio; | |
1916 | } | |
1917 | ||
1918 | /* | |
1919 | * Calculate the current priority, i.e. the priority | |
1920 | * taken into account by the scheduler. This value might | |
1921 | * be boosted by RT tasks, or might be boosted by | |
1922 | * interactivity modifiers. Will be RT if the task got | |
1923 | * RT-boosted. If not then it returns p->normal_prio. | |
1924 | */ | |
36c8b586 | 1925 | static int effective_prio(struct task_struct *p) |
b29739f9 IM |
1926 | { |
1927 | p->normal_prio = normal_prio(p); | |
1928 | /* | |
1929 | * If we are RT tasks or we were boosted to RT priority, | |
1930 | * keep the priority unchanged. Otherwise, update priority | |
1931 | * to the normal priority: | |
1932 | */ | |
1933 | if (!rt_prio(p->prio)) | |
1934 | return p->normal_prio; | |
1935 | return p->prio; | |
1936 | } | |
1937 | ||
1da177e4 | 1938 | /* |
dd41f596 | 1939 | * activate_task - move a task to the runqueue. |
1da177e4 | 1940 | */ |
dd41f596 | 1941 | static void activate_task(struct rq *rq, struct task_struct *p, int wakeup) |
1da177e4 | 1942 | { |
d9514f6c | 1943 | if (task_contributes_to_load(p)) |
dd41f596 | 1944 | rq->nr_uninterruptible--; |
1da177e4 | 1945 | |
8159f87e | 1946 | enqueue_task(rq, p, wakeup); |
18d95a28 | 1947 | inc_nr_running(rq); |
1da177e4 LT |
1948 | } |
1949 | ||
1da177e4 LT |
1950 | /* |
1951 | * deactivate_task - remove a task from the runqueue. | |
1952 | */ | |
2e1cb74a | 1953 | static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep) |
1da177e4 | 1954 | { |
d9514f6c | 1955 | if (task_contributes_to_load(p)) |
dd41f596 IM |
1956 | rq->nr_uninterruptible++; |
1957 | ||
69be72c1 | 1958 | dequeue_task(rq, p, sleep); |
18d95a28 | 1959 | dec_nr_running(rq); |
1da177e4 LT |
1960 | } |
1961 | ||
1da177e4 LT |
1962 | /** |
1963 | * task_curr - is this task currently executing on a CPU? | |
1964 | * @p: the task in question. | |
1965 | */ | |
36c8b586 | 1966 | inline int task_curr(const struct task_struct *p) |
1da177e4 LT |
1967 | { |
1968 | return cpu_curr(task_cpu(p)) == p; | |
1969 | } | |
1970 | ||
2dd73a4f PW |
1971 | /* Used instead of source_load when we know the type == 0 */ |
1972 | unsigned long weighted_cpuload(const int cpu) | |
1973 | { | |
495eca49 | 1974 | return cpu_rq(cpu)->load.weight; |
dd41f596 IM |
1975 | } |
1976 | ||
1977 | static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu) | |
1978 | { | |
6f505b16 | 1979 | set_task_rq(p, cpu); |
dd41f596 | 1980 | #ifdef CONFIG_SMP |
ce96b5ac DA |
1981 | /* |
1982 | * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be | |
1983 | * successfuly executed on another CPU. We must ensure that updates of | |
1984 | * per-task data have been completed by this moment. | |
1985 | */ | |
1986 | smp_wmb(); | |
dd41f596 | 1987 | task_thread_info(p)->cpu = cpu; |
dd41f596 | 1988 | #endif |
2dd73a4f PW |
1989 | } |
1990 | ||
cb469845 SR |
1991 | static inline void check_class_changed(struct rq *rq, struct task_struct *p, |
1992 | const struct sched_class *prev_class, | |
1993 | int oldprio, int running) | |
1994 | { | |
1995 | if (prev_class != p->sched_class) { | |
1996 | if (prev_class->switched_from) | |
1997 | prev_class->switched_from(rq, p, running); | |
1998 | p->sched_class->switched_to(rq, p, running); | |
1999 | } else | |
2000 | p->sched_class->prio_changed(rq, p, oldprio, running); | |
2001 | } | |
2002 | ||
1da177e4 | 2003 | #ifdef CONFIG_SMP |
c65cc870 | 2004 | |
cc367732 IM |
2005 | /* |
2006 | * Is this task likely cache-hot: | |
2007 | */ | |
e7693a36 | 2008 | static int |
cc367732 IM |
2009 | task_hot(struct task_struct *p, u64 now, struct sched_domain *sd) |
2010 | { | |
2011 | s64 delta; | |
2012 | ||
f540a608 IM |
2013 | /* |
2014 | * Buddy candidates are cache hot: | |
2015 | */ | |
d25ce4cd | 2016 | if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next)) |
f540a608 IM |
2017 | return 1; |
2018 | ||
cc367732 IM |
2019 | if (p->sched_class != &fair_sched_class) |
2020 | return 0; | |
2021 | ||
6bc1665b IM |
2022 | if (sysctl_sched_migration_cost == -1) |
2023 | return 1; | |
2024 | if (sysctl_sched_migration_cost == 0) | |
2025 | return 0; | |
2026 | ||
cc367732 IM |
2027 | delta = now - p->se.exec_start; |
2028 | ||
2029 | return delta < (s64)sysctl_sched_migration_cost; | |
2030 | } | |
2031 | ||
2032 | ||
dd41f596 | 2033 | void set_task_cpu(struct task_struct *p, unsigned int new_cpu) |
c65cc870 | 2034 | { |
dd41f596 IM |
2035 | int old_cpu = task_cpu(p); |
2036 | struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu); | |
2830cf8c SV |
2037 | struct cfs_rq *old_cfsrq = task_cfs_rq(p), |
2038 | *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu); | |
bbdba7c0 | 2039 | u64 clock_offset; |
dd41f596 IM |
2040 | |
2041 | clock_offset = old_rq->clock - new_rq->clock; | |
6cfb0d5d IM |
2042 | |
2043 | #ifdef CONFIG_SCHEDSTATS | |
2044 | if (p->se.wait_start) | |
2045 | p->se.wait_start -= clock_offset; | |
dd41f596 IM |
2046 | if (p->se.sleep_start) |
2047 | p->se.sleep_start -= clock_offset; | |
2048 | if (p->se.block_start) | |
2049 | p->se.block_start -= clock_offset; | |
cc367732 IM |
2050 | if (old_cpu != new_cpu) { |
2051 | schedstat_inc(p, se.nr_migrations); | |
2052 | if (task_hot(p, old_rq->clock, NULL)) | |
2053 | schedstat_inc(p, se.nr_forced2_migrations); | |
2054 | } | |
6cfb0d5d | 2055 | #endif |
2830cf8c SV |
2056 | p->se.vruntime -= old_cfsrq->min_vruntime - |
2057 | new_cfsrq->min_vruntime; | |
dd41f596 IM |
2058 | |
2059 | __set_task_cpu(p, new_cpu); | |
c65cc870 IM |
2060 | } |
2061 | ||
70b97a7f | 2062 | struct migration_req { |
1da177e4 | 2063 | struct list_head list; |
1da177e4 | 2064 | |
36c8b586 | 2065 | struct task_struct *task; |
1da177e4 LT |
2066 | int dest_cpu; |
2067 | ||
1da177e4 | 2068 | struct completion done; |
70b97a7f | 2069 | }; |
1da177e4 LT |
2070 | |
2071 | /* | |
2072 | * The task's runqueue lock must be held. | |
2073 | * Returns true if you have to wait for migration thread. | |
2074 | */ | |
36c8b586 | 2075 | static int |
70b97a7f | 2076 | migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req) |
1da177e4 | 2077 | { |
70b97a7f | 2078 | struct rq *rq = task_rq(p); |
1da177e4 LT |
2079 | |
2080 | /* | |
2081 | * If the task is not on a runqueue (and not running), then | |
2082 | * it is sufficient to simply update the task's cpu field. | |
2083 | */ | |
dd41f596 | 2084 | if (!p->se.on_rq && !task_running(rq, p)) { |
1da177e4 LT |
2085 | set_task_cpu(p, dest_cpu); |
2086 | return 0; | |
2087 | } | |
2088 | ||
2089 | init_completion(&req->done); | |
1da177e4 LT |
2090 | req->task = p; |
2091 | req->dest_cpu = dest_cpu; | |
2092 | list_add(&req->list, &rq->migration_queue); | |
48f24c4d | 2093 | |
1da177e4 LT |
2094 | return 1; |
2095 | } | |
2096 | ||
2097 | /* | |
2098 | * wait_task_inactive - wait for a thread to unschedule. | |
2099 | * | |
2100 | * The caller must ensure that the task *will* unschedule sometime soon, | |
2101 | * else this function might spin for a *long* time. This function can't | |
2102 | * be called with interrupts off, or it may introduce deadlock with | |
2103 | * smp_call_function() if an IPI is sent by the same process we are | |
2104 | * waiting to become inactive. | |
2105 | */ | |
36c8b586 | 2106 | void wait_task_inactive(struct task_struct *p) |
1da177e4 LT |
2107 | { |
2108 | unsigned long flags; | |
dd41f596 | 2109 | int running, on_rq; |
70b97a7f | 2110 | struct rq *rq; |
1da177e4 | 2111 | |
3a5c359a AK |
2112 | for (;;) { |
2113 | /* | |
2114 | * We do the initial early heuristics without holding | |
2115 | * any task-queue locks at all. We'll only try to get | |
2116 | * the runqueue lock when things look like they will | |
2117 | * work out! | |
2118 | */ | |
2119 | rq = task_rq(p); | |
fa490cfd | 2120 | |
3a5c359a AK |
2121 | /* |
2122 | * If the task is actively running on another CPU | |
2123 | * still, just relax and busy-wait without holding | |
2124 | * any locks. | |
2125 | * | |
2126 | * NOTE! Since we don't hold any locks, it's not | |
2127 | * even sure that "rq" stays as the right runqueue! | |
2128 | * But we don't care, since "task_running()" will | |
2129 | * return false if the runqueue has changed and p | |
2130 | * is actually now running somewhere else! | |
2131 | */ | |
2132 | while (task_running(rq, p)) | |
2133 | cpu_relax(); | |
fa490cfd | 2134 | |
3a5c359a AK |
2135 | /* |
2136 | * Ok, time to look more closely! We need the rq | |
2137 | * lock now, to be *sure*. If we're wrong, we'll | |
2138 | * just go back and repeat. | |
2139 | */ | |
2140 | rq = task_rq_lock(p, &flags); | |
2141 | running = task_running(rq, p); | |
2142 | on_rq = p->se.on_rq; | |
2143 | task_rq_unlock(rq, &flags); | |
fa490cfd | 2144 | |
3a5c359a AK |
2145 | /* |
2146 | * Was it really running after all now that we | |
2147 | * checked with the proper locks actually held? | |
2148 | * | |
2149 | * Oops. Go back and try again.. | |
2150 | */ | |
2151 | if (unlikely(running)) { | |
2152 | cpu_relax(); | |
2153 | continue; | |
2154 | } | |
fa490cfd | 2155 | |
3a5c359a AK |
2156 | /* |
2157 | * It's not enough that it's not actively running, | |
2158 | * it must be off the runqueue _entirely_, and not | |
2159 | * preempted! | |
2160 | * | |
2161 | * So if it wa still runnable (but just not actively | |
2162 | * running right now), it's preempted, and we should | |
2163 | * yield - it could be a while. | |
2164 | */ | |
2165 | if (unlikely(on_rq)) { | |
2166 | schedule_timeout_uninterruptible(1); | |
2167 | continue; | |
2168 | } | |
fa490cfd | 2169 | |
3a5c359a AK |
2170 | /* |
2171 | * Ahh, all good. It wasn't running, and it wasn't | |
2172 | * runnable, which means that it will never become | |
2173 | * running in the future either. We're all done! | |
2174 | */ | |
2175 | break; | |
2176 | } | |
1da177e4 LT |
2177 | } |
2178 | ||
2179 | /*** | |
2180 | * kick_process - kick a running thread to enter/exit the kernel | |
2181 | * @p: the to-be-kicked thread | |
2182 | * | |
2183 | * Cause a process which is running on another CPU to enter | |
2184 | * kernel-mode, without any delay. (to get signals handled.) | |
2185 | * | |
2186 | * NOTE: this function doesnt have to take the runqueue lock, | |
2187 | * because all it wants to ensure is that the remote task enters | |
2188 | * the kernel. If the IPI races and the task has been migrated | |
2189 | * to another CPU then no harm is done and the purpose has been | |
2190 | * achieved as well. | |
2191 | */ | |
36c8b586 | 2192 | void kick_process(struct task_struct *p) |
1da177e4 LT |
2193 | { |
2194 | int cpu; | |
2195 | ||
2196 | preempt_disable(); | |
2197 | cpu = task_cpu(p); | |
2198 | if ((cpu != smp_processor_id()) && task_curr(p)) | |
2199 | smp_send_reschedule(cpu); | |
2200 | preempt_enable(); | |
2201 | } | |
2202 | ||
2203 | /* | |
2dd73a4f PW |
2204 | * Return a low guess at the load of a migration-source cpu weighted |
2205 | * according to the scheduling class and "nice" value. | |
1da177e4 LT |
2206 | * |
2207 | * We want to under-estimate the load of migration sources, to | |
2208 | * balance conservatively. | |
2209 | */ | |
a9957449 | 2210 | static unsigned long source_load(int cpu, int type) |
1da177e4 | 2211 | { |
70b97a7f | 2212 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2213 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2214 | |
3b0bd9bc | 2215 | if (type == 0) |
dd41f596 | 2216 | return total; |
b910472d | 2217 | |
dd41f596 | 2218 | return min(rq->cpu_load[type-1], total); |
1da177e4 LT |
2219 | } |
2220 | ||
2221 | /* | |
2dd73a4f PW |
2222 | * Return a high guess at the load of a migration-target cpu weighted |
2223 | * according to the scheduling class and "nice" value. | |
1da177e4 | 2224 | */ |
a9957449 | 2225 | static unsigned long target_load(int cpu, int type) |
1da177e4 | 2226 | { |
70b97a7f | 2227 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2228 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f | 2229 | |
7897986b | 2230 | if (type == 0) |
dd41f596 | 2231 | return total; |
3b0bd9bc | 2232 | |
dd41f596 | 2233 | return max(rq->cpu_load[type-1], total); |
2dd73a4f PW |
2234 | } |
2235 | ||
2236 | /* | |
2237 | * Return the average load per task on the cpu's run queue | |
2238 | */ | |
e7693a36 | 2239 | static unsigned long cpu_avg_load_per_task(int cpu) |
2dd73a4f | 2240 | { |
70b97a7f | 2241 | struct rq *rq = cpu_rq(cpu); |
dd41f596 | 2242 | unsigned long total = weighted_cpuload(cpu); |
2dd73a4f PW |
2243 | unsigned long n = rq->nr_running; |
2244 | ||
dd41f596 | 2245 | return n ? total / n : SCHED_LOAD_SCALE; |
1da177e4 LT |
2246 | } |
2247 | ||
147cbb4b NP |
2248 | /* |
2249 | * find_idlest_group finds and returns the least busy CPU group within the | |
2250 | * domain. | |
2251 | */ | |
2252 | static struct sched_group * | |
2253 | find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu) | |
2254 | { | |
2255 | struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups; | |
2256 | unsigned long min_load = ULONG_MAX, this_load = 0; | |
2257 | int load_idx = sd->forkexec_idx; | |
2258 | int imbalance = 100 + (sd->imbalance_pct-100)/2; | |
2259 | ||
2260 | do { | |
2261 | unsigned long load, avg_load; | |
2262 | int local_group; | |
2263 | int i; | |
2264 | ||
da5a5522 BD |
2265 | /* Skip over this group if it has no CPUs allowed */ |
2266 | if (!cpus_intersects(group->cpumask, p->cpus_allowed)) | |
3a5c359a | 2267 | continue; |
da5a5522 | 2268 | |
147cbb4b | 2269 | local_group = cpu_isset(this_cpu, group->cpumask); |
147cbb4b NP |
2270 | |
2271 | /* Tally up the load of all CPUs in the group */ | |
2272 | avg_load = 0; | |
2273 | ||
2274 | for_each_cpu_mask(i, group->cpumask) { | |
2275 | /* Bias balancing toward cpus of our domain */ | |
2276 | if (local_group) | |
2277 | load = source_load(i, load_idx); | |
2278 | else | |
2279 | load = target_load(i, load_idx); | |
2280 | ||
2281 | avg_load += load; | |
2282 | } | |
2283 | ||
2284 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
2285 | avg_load = sg_div_cpu_power(group, |
2286 | avg_load * SCHED_LOAD_SCALE); | |
147cbb4b NP |
2287 | |
2288 | if (local_group) { | |
2289 | this_load = avg_load; | |
2290 | this = group; | |
2291 | } else if (avg_load < min_load) { | |
2292 | min_load = avg_load; | |
2293 | idlest = group; | |
2294 | } | |
3a5c359a | 2295 | } while (group = group->next, group != sd->groups); |
147cbb4b NP |
2296 | |
2297 | if (!idlest || 100*this_load < imbalance*min_load) | |
2298 | return NULL; | |
2299 | return idlest; | |
2300 | } | |
2301 | ||
2302 | /* | |
0feaece9 | 2303 | * find_idlest_cpu - find the idlest cpu among the cpus in group. |
147cbb4b | 2304 | */ |
95cdf3b7 | 2305 | static int |
7c16ec58 MT |
2306 | find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu, |
2307 | cpumask_t *tmp) | |
147cbb4b NP |
2308 | { |
2309 | unsigned long load, min_load = ULONG_MAX; | |
2310 | int idlest = -1; | |
2311 | int i; | |
2312 | ||
da5a5522 | 2313 | /* Traverse only the allowed CPUs */ |
7c16ec58 | 2314 | cpus_and(*tmp, group->cpumask, p->cpus_allowed); |
da5a5522 | 2315 | |
7c16ec58 | 2316 | for_each_cpu_mask(i, *tmp) { |
2dd73a4f | 2317 | load = weighted_cpuload(i); |
147cbb4b NP |
2318 | |
2319 | if (load < min_load || (load == min_load && i == this_cpu)) { | |
2320 | min_load = load; | |
2321 | idlest = i; | |
2322 | } | |
2323 | } | |
2324 | ||
2325 | return idlest; | |
2326 | } | |
2327 | ||
476d139c NP |
2328 | /* |
2329 | * sched_balance_self: balance the current task (running on cpu) in domains | |
2330 | * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and | |
2331 | * SD_BALANCE_EXEC. | |
2332 | * | |
2333 | * Balance, ie. select the least loaded group. | |
2334 | * | |
2335 | * Returns the target CPU number, or the same CPU if no balancing is needed. | |
2336 | * | |
2337 | * preempt must be disabled. | |
2338 | */ | |
2339 | static int sched_balance_self(int cpu, int flag) | |
2340 | { | |
2341 | struct task_struct *t = current; | |
2342 | struct sched_domain *tmp, *sd = NULL; | |
147cbb4b | 2343 | |
c96d145e | 2344 | for_each_domain(cpu, tmp) { |
9761eea8 IM |
2345 | /* |
2346 | * If power savings logic is enabled for a domain, stop there. | |
2347 | */ | |
5c45bf27 SS |
2348 | if (tmp->flags & SD_POWERSAVINGS_BALANCE) |
2349 | break; | |
476d139c NP |
2350 | if (tmp->flags & flag) |
2351 | sd = tmp; | |
c96d145e | 2352 | } |
476d139c NP |
2353 | |
2354 | while (sd) { | |
7c16ec58 | 2355 | cpumask_t span, tmpmask; |
476d139c | 2356 | struct sched_group *group; |
1a848870 SS |
2357 | int new_cpu, weight; |
2358 | ||
2359 | if (!(sd->flags & flag)) { | |
2360 | sd = sd->child; | |
2361 | continue; | |
2362 | } | |
476d139c NP |
2363 | |
2364 | span = sd->span; | |
2365 | group = find_idlest_group(sd, t, cpu); | |
1a848870 SS |
2366 | if (!group) { |
2367 | sd = sd->child; | |
2368 | continue; | |
2369 | } | |
476d139c | 2370 | |
7c16ec58 | 2371 | new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask); |
1a848870 SS |
2372 | if (new_cpu == -1 || new_cpu == cpu) { |
2373 | /* Now try balancing at a lower domain level of cpu */ | |
2374 | sd = sd->child; | |
2375 | continue; | |
2376 | } | |
476d139c | 2377 | |
1a848870 | 2378 | /* Now try balancing at a lower domain level of new_cpu */ |
476d139c | 2379 | cpu = new_cpu; |
476d139c NP |
2380 | sd = NULL; |
2381 | weight = cpus_weight(span); | |
2382 | for_each_domain(cpu, tmp) { | |
2383 | if (weight <= cpus_weight(tmp->span)) | |
2384 | break; | |
2385 | if (tmp->flags & flag) | |
2386 | sd = tmp; | |
2387 | } | |
2388 | /* while loop will break here if sd == NULL */ | |
2389 | } | |
2390 | ||
2391 | return cpu; | |
2392 | } | |
2393 | ||
2394 | #endif /* CONFIG_SMP */ | |
1da177e4 | 2395 | |
1da177e4 LT |
2396 | /*** |
2397 | * try_to_wake_up - wake up a thread | |
2398 | * @p: the to-be-woken-up thread | |
2399 | * @state: the mask of task states that can be woken | |
2400 | * @sync: do a synchronous wakeup? | |
2401 | * | |
2402 | * Put it on the run-queue if it's not already there. The "current" | |
2403 | * thread is always on the run-queue (except when the actual | |
2404 | * re-schedule is in progress), and as such you're allowed to do | |
2405 | * the simpler "current->state = TASK_RUNNING" to mark yourself | |
2406 | * runnable without the overhead of this. | |
2407 | * | |
2408 | * returns failure only if the task is already active. | |
2409 | */ | |
36c8b586 | 2410 | static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync) |
1da177e4 | 2411 | { |
cc367732 | 2412 | int cpu, orig_cpu, this_cpu, success = 0; |
1da177e4 LT |
2413 | unsigned long flags; |
2414 | long old_state; | |
70b97a7f | 2415 | struct rq *rq; |
1da177e4 | 2416 | |
b85d0667 IM |
2417 | if (!sched_feat(SYNC_WAKEUPS)) |
2418 | sync = 0; | |
2419 | ||
04e2f174 | 2420 | smp_wmb(); |
1da177e4 LT |
2421 | rq = task_rq_lock(p, &flags); |
2422 | old_state = p->state; | |
2423 | if (!(old_state & state)) | |
2424 | goto out; | |
2425 | ||
dd41f596 | 2426 | if (p->se.on_rq) |
1da177e4 LT |
2427 | goto out_running; |
2428 | ||
2429 | cpu = task_cpu(p); | |
cc367732 | 2430 | orig_cpu = cpu; |
1da177e4 LT |
2431 | this_cpu = smp_processor_id(); |
2432 | ||
2433 | #ifdef CONFIG_SMP | |
2434 | if (unlikely(task_running(rq, p))) | |
2435 | goto out_activate; | |
2436 | ||
5d2f5a61 DA |
2437 | cpu = p->sched_class->select_task_rq(p, sync); |
2438 | if (cpu != orig_cpu) { | |
2439 | set_task_cpu(p, cpu); | |
1da177e4 LT |
2440 | task_rq_unlock(rq, &flags); |
2441 | /* might preempt at this point */ | |
2442 | rq = task_rq_lock(p, &flags); | |
2443 | old_state = p->state; | |
2444 | if (!(old_state & state)) | |
2445 | goto out; | |
dd41f596 | 2446 | if (p->se.on_rq) |
1da177e4 LT |
2447 | goto out_running; |
2448 | ||
2449 | this_cpu = smp_processor_id(); | |
2450 | cpu = task_cpu(p); | |
2451 | } | |
2452 | ||
e7693a36 GH |
2453 | #ifdef CONFIG_SCHEDSTATS |
2454 | schedstat_inc(rq, ttwu_count); | |
2455 | if (cpu == this_cpu) | |
2456 | schedstat_inc(rq, ttwu_local); | |
2457 | else { | |
2458 | struct sched_domain *sd; | |
2459 | for_each_domain(this_cpu, sd) { | |
2460 | if (cpu_isset(cpu, sd->span)) { | |
2461 | schedstat_inc(sd, ttwu_wake_remote); | |
2462 | break; | |
2463 | } | |
2464 | } | |
2465 | } | |
e7693a36 GH |
2466 | #endif |
2467 | ||
1da177e4 LT |
2468 | out_activate: |
2469 | #endif /* CONFIG_SMP */ | |
cc367732 IM |
2470 | schedstat_inc(p, se.nr_wakeups); |
2471 | if (sync) | |
2472 | schedstat_inc(p, se.nr_wakeups_sync); | |
2473 | if (orig_cpu != cpu) | |
2474 | schedstat_inc(p, se.nr_wakeups_migrate); | |
2475 | if (cpu == this_cpu) | |
2476 | schedstat_inc(p, se.nr_wakeups_local); | |
2477 | else | |
2478 | schedstat_inc(p, se.nr_wakeups_remote); | |
2daa3577 | 2479 | update_rq_clock(rq); |
dd41f596 | 2480 | activate_task(rq, p, 1); |
1da177e4 LT |
2481 | success = 1; |
2482 | ||
2483 | out_running: | |
4ae7d5ce IM |
2484 | check_preempt_curr(rq, p); |
2485 | ||
1da177e4 | 2486 | p->state = TASK_RUNNING; |
9a897c5a SR |
2487 | #ifdef CONFIG_SMP |
2488 | if (p->sched_class->task_wake_up) | |
2489 | p->sched_class->task_wake_up(rq, p); | |
2490 | #endif | |
1da177e4 LT |
2491 | out: |
2492 | task_rq_unlock(rq, &flags); | |
2493 | ||
2494 | return success; | |
2495 | } | |
2496 | ||
7ad5b3a5 | 2497 | int wake_up_process(struct task_struct *p) |
1da177e4 | 2498 | { |
d9514f6c | 2499 | return try_to_wake_up(p, TASK_ALL, 0); |
1da177e4 | 2500 | } |
1da177e4 LT |
2501 | EXPORT_SYMBOL(wake_up_process); |
2502 | ||
7ad5b3a5 | 2503 | int wake_up_state(struct task_struct *p, unsigned int state) |
1da177e4 LT |
2504 | { |
2505 | return try_to_wake_up(p, state, 0); | |
2506 | } | |
2507 | ||
1da177e4 LT |
2508 | /* |
2509 | * Perform scheduler related setup for a newly forked process p. | |
2510 | * p is forked by current. | |
dd41f596 IM |
2511 | * |
2512 | * __sched_fork() is basic setup used by init_idle() too: | |
2513 | */ | |
2514 | static void __sched_fork(struct task_struct *p) | |
2515 | { | |
dd41f596 IM |
2516 | p->se.exec_start = 0; |
2517 | p->se.sum_exec_runtime = 0; | |
f6cf891c | 2518 | p->se.prev_sum_exec_runtime = 0; |
4ae7d5ce IM |
2519 | p->se.last_wakeup = 0; |
2520 | p->se.avg_overlap = 0; | |
6cfb0d5d IM |
2521 | |
2522 | #ifdef CONFIG_SCHEDSTATS | |
2523 | p->se.wait_start = 0; | |
dd41f596 IM |
2524 | p->se.sum_sleep_runtime = 0; |
2525 | p->se.sleep_start = 0; | |
dd41f596 IM |
2526 | p->se.block_start = 0; |
2527 | p->se.sleep_max = 0; | |
2528 | p->se.block_max = 0; | |
2529 | p->se.exec_max = 0; | |
eba1ed4b | 2530 | p->se.slice_max = 0; |
dd41f596 | 2531 | p->se.wait_max = 0; |
6cfb0d5d | 2532 | #endif |
476d139c | 2533 | |
fa717060 | 2534 | INIT_LIST_HEAD(&p->rt.run_list); |
dd41f596 | 2535 | p->se.on_rq = 0; |
4a55bd5e | 2536 | INIT_LIST_HEAD(&p->se.group_node); |
476d139c | 2537 | |
e107be36 AK |
2538 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2539 | INIT_HLIST_HEAD(&p->preempt_notifiers); | |
2540 | #endif | |
2541 | ||
1da177e4 LT |
2542 | /* |
2543 | * We mark the process as running here, but have not actually | |
2544 | * inserted it onto the runqueue yet. This guarantees that | |
2545 | * nobody will actually run it, and a signal or other external | |
2546 | * event cannot wake it up and insert it on the runqueue either. | |
2547 | */ | |
2548 | p->state = TASK_RUNNING; | |
dd41f596 IM |
2549 | } |
2550 | ||
2551 | /* | |
2552 | * fork()/clone()-time setup: | |
2553 | */ | |
2554 | void sched_fork(struct task_struct *p, int clone_flags) | |
2555 | { | |
2556 | int cpu = get_cpu(); | |
2557 | ||
2558 | __sched_fork(p); | |
2559 | ||
2560 | #ifdef CONFIG_SMP | |
2561 | cpu = sched_balance_self(cpu, SD_BALANCE_FORK); | |
2562 | #endif | |
02e4bac2 | 2563 | set_task_cpu(p, cpu); |
b29739f9 IM |
2564 | |
2565 | /* | |
2566 | * Make sure we do not leak PI boosting priority to the child: | |
2567 | */ | |
2568 | p->prio = current->normal_prio; | |
2ddbf952 HS |
2569 | if (!rt_prio(p->prio)) |
2570 | p->sched_class = &fair_sched_class; | |
b29739f9 | 2571 | |
52f17b6c | 2572 | #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) |
dd41f596 | 2573 | if (likely(sched_info_on())) |
52f17b6c | 2574 | memset(&p->sched_info, 0, sizeof(p->sched_info)); |
1da177e4 | 2575 | #endif |
d6077cb8 | 2576 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
4866cde0 NP |
2577 | p->oncpu = 0; |
2578 | #endif | |
1da177e4 | 2579 | #ifdef CONFIG_PREEMPT |
4866cde0 | 2580 | /* Want to start with kernel preemption disabled. */ |
a1261f54 | 2581 | task_thread_info(p)->preempt_count = 1; |
1da177e4 | 2582 | #endif |
476d139c | 2583 | put_cpu(); |
1da177e4 LT |
2584 | } |
2585 | ||
2586 | /* | |
2587 | * wake_up_new_task - wake up a newly created task for the first time. | |
2588 | * | |
2589 | * This function will do some initial scheduler statistics housekeeping | |
2590 | * that must be done for every newly created context, then puts the task | |
2591 | * on the runqueue and wakes it. | |
2592 | */ | |
7ad5b3a5 | 2593 | void wake_up_new_task(struct task_struct *p, unsigned long clone_flags) |
1da177e4 LT |
2594 | { |
2595 | unsigned long flags; | |
dd41f596 | 2596 | struct rq *rq; |
1da177e4 LT |
2597 | |
2598 | rq = task_rq_lock(p, &flags); | |
147cbb4b | 2599 | BUG_ON(p->state != TASK_RUNNING); |
a8e504d2 | 2600 | update_rq_clock(rq); |
1da177e4 LT |
2601 | |
2602 | p->prio = effective_prio(p); | |
2603 | ||
b9dca1e0 | 2604 | if (!p->sched_class->task_new || !current->se.on_rq) { |
dd41f596 | 2605 | activate_task(rq, p, 0); |
1da177e4 | 2606 | } else { |
1da177e4 | 2607 | /* |
dd41f596 IM |
2608 | * Let the scheduling class do new task startup |
2609 | * management (if any): | |
1da177e4 | 2610 | */ |
ee0827d8 | 2611 | p->sched_class->task_new(rq, p); |
18d95a28 | 2612 | inc_nr_running(rq); |
1da177e4 | 2613 | } |
dd41f596 | 2614 | check_preempt_curr(rq, p); |
9a897c5a SR |
2615 | #ifdef CONFIG_SMP |
2616 | if (p->sched_class->task_wake_up) | |
2617 | p->sched_class->task_wake_up(rq, p); | |
2618 | #endif | |
dd41f596 | 2619 | task_rq_unlock(rq, &flags); |
1da177e4 LT |
2620 | } |
2621 | ||
e107be36 AK |
2622 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
2623 | ||
2624 | /** | |
421cee29 RD |
2625 | * preempt_notifier_register - tell me when current is being being preempted & rescheduled |
2626 | * @notifier: notifier struct to register | |
e107be36 AK |
2627 | */ |
2628 | void preempt_notifier_register(struct preempt_notifier *notifier) | |
2629 | { | |
2630 | hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); | |
2631 | } | |
2632 | EXPORT_SYMBOL_GPL(preempt_notifier_register); | |
2633 | ||
2634 | /** | |
2635 | * preempt_notifier_unregister - no longer interested in preemption notifications | |
421cee29 | 2636 | * @notifier: notifier struct to unregister |
e107be36 AK |
2637 | * |
2638 | * This is safe to call from within a preemption notifier. | |
2639 | */ | |
2640 | void preempt_notifier_unregister(struct preempt_notifier *notifier) | |
2641 | { | |
2642 | hlist_del(¬ifier->link); | |
2643 | } | |
2644 | EXPORT_SYMBOL_GPL(preempt_notifier_unregister); | |
2645 | ||
2646 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2647 | { | |
2648 | struct preempt_notifier *notifier; | |
2649 | struct hlist_node *node; | |
2650 | ||
2651 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2652 | notifier->ops->sched_in(notifier, raw_smp_processor_id()); | |
2653 | } | |
2654 | ||
2655 | static void | |
2656 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2657 | struct task_struct *next) | |
2658 | { | |
2659 | struct preempt_notifier *notifier; | |
2660 | struct hlist_node *node; | |
2661 | ||
2662 | hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) | |
2663 | notifier->ops->sched_out(notifier, next); | |
2664 | } | |
2665 | ||
2666 | #else | |
2667 | ||
2668 | static void fire_sched_in_preempt_notifiers(struct task_struct *curr) | |
2669 | { | |
2670 | } | |
2671 | ||
2672 | static void | |
2673 | fire_sched_out_preempt_notifiers(struct task_struct *curr, | |
2674 | struct task_struct *next) | |
2675 | { | |
2676 | } | |
2677 | ||
2678 | #endif | |
2679 | ||
4866cde0 NP |
2680 | /** |
2681 | * prepare_task_switch - prepare to switch tasks | |
2682 | * @rq: the runqueue preparing to switch | |
421cee29 | 2683 | * @prev: the current task that is being switched out |
4866cde0 NP |
2684 | * @next: the task we are going to switch to. |
2685 | * | |
2686 | * This is called with the rq lock held and interrupts off. It must | |
2687 | * be paired with a subsequent finish_task_switch after the context | |
2688 | * switch. | |
2689 | * | |
2690 | * prepare_task_switch sets up locking and calls architecture specific | |
2691 | * hooks. | |
2692 | */ | |
e107be36 AK |
2693 | static inline void |
2694 | prepare_task_switch(struct rq *rq, struct task_struct *prev, | |
2695 | struct task_struct *next) | |
4866cde0 | 2696 | { |
e107be36 | 2697 | fire_sched_out_preempt_notifiers(prev, next); |
4866cde0 NP |
2698 | prepare_lock_switch(rq, next); |
2699 | prepare_arch_switch(next); | |
2700 | } | |
2701 | ||
1da177e4 LT |
2702 | /** |
2703 | * finish_task_switch - clean up after a task-switch | |
344babaa | 2704 | * @rq: runqueue associated with task-switch |
1da177e4 LT |
2705 | * @prev: the thread we just switched away from. |
2706 | * | |
4866cde0 NP |
2707 | * finish_task_switch must be called after the context switch, paired |
2708 | * with a prepare_task_switch call before the context switch. | |
2709 | * finish_task_switch will reconcile locking set up by prepare_task_switch, | |
2710 | * and do any other architecture-specific cleanup actions. | |
1da177e4 LT |
2711 | * |
2712 | * Note that we may have delayed dropping an mm in context_switch(). If | |
41a2d6cf | 2713 | * so, we finish that here outside of the runqueue lock. (Doing it |
1da177e4 LT |
2714 | * with the lock held can cause deadlocks; see schedule() for |
2715 | * details.) | |
2716 | */ | |
a9957449 | 2717 | static void finish_task_switch(struct rq *rq, struct task_struct *prev) |
1da177e4 LT |
2718 | __releases(rq->lock) |
2719 | { | |
1da177e4 | 2720 | struct mm_struct *mm = rq->prev_mm; |
55a101f8 | 2721 | long prev_state; |
1da177e4 LT |
2722 | |
2723 | rq->prev_mm = NULL; | |
2724 | ||
2725 | /* | |
2726 | * A task struct has one reference for the use as "current". | |
c394cc9f | 2727 | * If a task dies, then it sets TASK_DEAD in tsk->state and calls |
55a101f8 ON |
2728 | * schedule one last time. The schedule call will never return, and |
2729 | * the scheduled task must drop that reference. | |
c394cc9f | 2730 | * The test for TASK_DEAD must occur while the runqueue locks are |
1da177e4 LT |
2731 | * still held, otherwise prev could be scheduled on another cpu, die |
2732 | * there before we look at prev->state, and then the reference would | |
2733 | * be dropped twice. | |
2734 | * Manfred Spraul <manfred@colorfullife.com> | |
2735 | */ | |
55a101f8 | 2736 | prev_state = prev->state; |
4866cde0 NP |
2737 | finish_arch_switch(prev); |
2738 | finish_lock_switch(rq, prev); | |
9a897c5a SR |
2739 | #ifdef CONFIG_SMP |
2740 | if (current->sched_class->post_schedule) | |
2741 | current->sched_class->post_schedule(rq); | |
2742 | #endif | |
e8fa1362 | 2743 | |
e107be36 | 2744 | fire_sched_in_preempt_notifiers(current); |
1da177e4 LT |
2745 | if (mm) |
2746 | mmdrop(mm); | |
c394cc9f | 2747 | if (unlikely(prev_state == TASK_DEAD)) { |
c6fd91f0 | 2748 | /* |
2749 | * Remove function-return probe instances associated with this | |
2750 | * task and put them back on the free list. | |
9761eea8 | 2751 | */ |
c6fd91f0 | 2752 | kprobe_flush_task(prev); |
1da177e4 | 2753 | put_task_struct(prev); |
c6fd91f0 | 2754 | } |
1da177e4 LT |
2755 | } |
2756 | ||
2757 | /** | |
2758 | * schedule_tail - first thing a freshly forked thread must call. | |
2759 | * @prev: the thread we just switched away from. | |
2760 | */ | |
36c8b586 | 2761 | asmlinkage void schedule_tail(struct task_struct *prev) |
1da177e4 LT |
2762 | __releases(rq->lock) |
2763 | { | |
70b97a7f IM |
2764 | struct rq *rq = this_rq(); |
2765 | ||
4866cde0 NP |
2766 | finish_task_switch(rq, prev); |
2767 | #ifdef __ARCH_WANT_UNLOCKED_CTXSW | |
2768 | /* In this case, finish_task_switch does not reenable preemption */ | |
2769 | preempt_enable(); | |
2770 | #endif | |
1da177e4 | 2771 | if (current->set_child_tid) |
b488893a | 2772 | put_user(task_pid_vnr(current), current->set_child_tid); |
1da177e4 LT |
2773 | } |
2774 | ||
2775 | /* | |
2776 | * context_switch - switch to the new MM and the new | |
2777 | * thread's register state. | |
2778 | */ | |
dd41f596 | 2779 | static inline void |
70b97a7f | 2780 | context_switch(struct rq *rq, struct task_struct *prev, |
36c8b586 | 2781 | struct task_struct *next) |
1da177e4 | 2782 | { |
dd41f596 | 2783 | struct mm_struct *mm, *oldmm; |
1da177e4 | 2784 | |
e107be36 | 2785 | prepare_task_switch(rq, prev, next); |
dd41f596 IM |
2786 | mm = next->mm; |
2787 | oldmm = prev->active_mm; | |
9226d125 ZA |
2788 | /* |
2789 | * For paravirt, this is coupled with an exit in switch_to to | |
2790 | * combine the page table reload and the switch backend into | |
2791 | * one hypercall. | |
2792 | */ | |
2793 | arch_enter_lazy_cpu_mode(); | |
2794 | ||
dd41f596 | 2795 | if (unlikely(!mm)) { |
1da177e4 LT |
2796 | next->active_mm = oldmm; |
2797 | atomic_inc(&oldmm->mm_count); | |
2798 | enter_lazy_tlb(oldmm, next); | |
2799 | } else | |
2800 | switch_mm(oldmm, mm, next); | |
2801 | ||
dd41f596 | 2802 | if (unlikely(!prev->mm)) { |
1da177e4 | 2803 | prev->active_mm = NULL; |
1da177e4 LT |
2804 | rq->prev_mm = oldmm; |
2805 | } | |
3a5f5e48 IM |
2806 | /* |
2807 | * Since the runqueue lock will be released by the next | |
2808 | * task (which is an invalid locking op but in the case | |
2809 | * of the scheduler it's an obvious special-case), so we | |
2810 | * do an early lockdep release here: | |
2811 | */ | |
2812 | #ifndef __ARCH_WANT_UNLOCKED_CTXSW | |
8a25d5de | 2813 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
3a5f5e48 | 2814 | #endif |
1da177e4 LT |
2815 | |
2816 | /* Here we just switch the register state and the stack. */ | |
2817 | switch_to(prev, next, prev); | |
2818 | ||
dd41f596 IM |
2819 | barrier(); |
2820 | /* | |
2821 | * this_rq must be evaluated again because prev may have moved | |
2822 | * CPUs since it called schedule(), thus the 'rq' on its stack | |
2823 | * frame will be invalid. | |
2824 | */ | |
2825 | finish_task_switch(this_rq(), prev); | |
1da177e4 LT |
2826 | } |
2827 | ||
2828 | /* | |
2829 | * nr_running, nr_uninterruptible and nr_context_switches: | |
2830 | * | |
2831 | * externally visible scheduler statistics: current number of runnable | |
2832 | * threads, current number of uninterruptible-sleeping threads, total | |
2833 | * number of context switches performed since bootup. | |
2834 | */ | |
2835 | unsigned long nr_running(void) | |
2836 | { | |
2837 | unsigned long i, sum = 0; | |
2838 | ||
2839 | for_each_online_cpu(i) | |
2840 | sum += cpu_rq(i)->nr_running; | |
2841 | ||
2842 | return sum; | |
2843 | } | |
2844 | ||
2845 | unsigned long nr_uninterruptible(void) | |
2846 | { | |
2847 | unsigned long i, sum = 0; | |
2848 | ||
0a945022 | 2849 | for_each_possible_cpu(i) |
1da177e4 LT |
2850 | sum += cpu_rq(i)->nr_uninterruptible; |
2851 | ||
2852 | /* | |
2853 | * Since we read the counters lockless, it might be slightly | |
2854 | * inaccurate. Do not allow it to go below zero though: | |
2855 | */ | |
2856 | if (unlikely((long)sum < 0)) | |
2857 | sum = 0; | |
2858 | ||
2859 | return sum; | |
2860 | } | |
2861 | ||
2862 | unsigned long long nr_context_switches(void) | |
2863 | { | |
cc94abfc SR |
2864 | int i; |
2865 | unsigned long long sum = 0; | |
1da177e4 | 2866 | |
0a945022 | 2867 | for_each_possible_cpu(i) |
1da177e4 LT |
2868 | sum += cpu_rq(i)->nr_switches; |
2869 | ||
2870 | return sum; | |
2871 | } | |
2872 | ||
2873 | unsigned long nr_iowait(void) | |
2874 | { | |
2875 | unsigned long i, sum = 0; | |
2876 | ||
0a945022 | 2877 | for_each_possible_cpu(i) |
1da177e4 LT |
2878 | sum += atomic_read(&cpu_rq(i)->nr_iowait); |
2879 | ||
2880 | return sum; | |
2881 | } | |
2882 | ||
db1b1fef JS |
2883 | unsigned long nr_active(void) |
2884 | { | |
2885 | unsigned long i, running = 0, uninterruptible = 0; | |
2886 | ||
2887 | for_each_online_cpu(i) { | |
2888 | running += cpu_rq(i)->nr_running; | |
2889 | uninterruptible += cpu_rq(i)->nr_uninterruptible; | |
2890 | } | |
2891 | ||
2892 | if (unlikely((long)uninterruptible < 0)) | |
2893 | uninterruptible = 0; | |
2894 | ||
2895 | return running + uninterruptible; | |
2896 | } | |
2897 | ||
48f24c4d | 2898 | /* |
dd41f596 IM |
2899 | * Update rq->cpu_load[] statistics. This function is usually called every |
2900 | * scheduler tick (TICK_NSEC). | |
48f24c4d | 2901 | */ |
dd41f596 | 2902 | static void update_cpu_load(struct rq *this_rq) |
48f24c4d | 2903 | { |
495eca49 | 2904 | unsigned long this_load = this_rq->load.weight; |
dd41f596 IM |
2905 | int i, scale; |
2906 | ||
2907 | this_rq->nr_load_updates++; | |
dd41f596 IM |
2908 | |
2909 | /* Update our load: */ | |
2910 | for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { | |
2911 | unsigned long old_load, new_load; | |
2912 | ||
2913 | /* scale is effectively 1 << i now, and >> i divides by scale */ | |
2914 | ||
2915 | old_load = this_rq->cpu_load[i]; | |
2916 | new_load = this_load; | |
a25707f3 IM |
2917 | /* |
2918 | * Round up the averaging division if load is increasing. This | |
2919 | * prevents us from getting stuck on 9 if the load is 10, for | |
2920 | * example. | |
2921 | */ | |
2922 | if (new_load > old_load) | |
2923 | new_load += scale-1; | |
dd41f596 IM |
2924 | this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i; |
2925 | } | |
48f24c4d IM |
2926 | } |
2927 | ||
dd41f596 IM |
2928 | #ifdef CONFIG_SMP |
2929 | ||
1da177e4 LT |
2930 | /* |
2931 | * double_rq_lock - safely lock two runqueues | |
2932 | * | |
2933 | * Note this does not disable interrupts like task_rq_lock, | |
2934 | * you need to do so manually before calling. | |
2935 | */ | |
70b97a7f | 2936 | static void double_rq_lock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2937 | __acquires(rq1->lock) |
2938 | __acquires(rq2->lock) | |
2939 | { | |
054b9108 | 2940 | BUG_ON(!irqs_disabled()); |
1da177e4 LT |
2941 | if (rq1 == rq2) { |
2942 | spin_lock(&rq1->lock); | |
2943 | __acquire(rq2->lock); /* Fake it out ;) */ | |
2944 | } else { | |
c96d145e | 2945 | if (rq1 < rq2) { |
1da177e4 LT |
2946 | spin_lock(&rq1->lock); |
2947 | spin_lock(&rq2->lock); | |
2948 | } else { | |
2949 | spin_lock(&rq2->lock); | |
2950 | spin_lock(&rq1->lock); | |
2951 | } | |
2952 | } | |
6e82a3be IM |
2953 | update_rq_clock(rq1); |
2954 | update_rq_clock(rq2); | |
1da177e4 LT |
2955 | } |
2956 | ||
2957 | /* | |
2958 | * double_rq_unlock - safely unlock two runqueues | |
2959 | * | |
2960 | * Note this does not restore interrupts like task_rq_unlock, | |
2961 | * you need to do so manually after calling. | |
2962 | */ | |
70b97a7f | 2963 | static void double_rq_unlock(struct rq *rq1, struct rq *rq2) |
1da177e4 LT |
2964 | __releases(rq1->lock) |
2965 | __releases(rq2->lock) | |
2966 | { | |
2967 | spin_unlock(&rq1->lock); | |
2968 | if (rq1 != rq2) | |
2969 | spin_unlock(&rq2->lock); | |
2970 | else | |
2971 | __release(rq2->lock); | |
2972 | } | |
2973 | ||
2974 | /* | |
2975 | * double_lock_balance - lock the busiest runqueue, this_rq is locked already. | |
2976 | */ | |
e8fa1362 | 2977 | static int double_lock_balance(struct rq *this_rq, struct rq *busiest) |
1da177e4 LT |
2978 | __releases(this_rq->lock) |
2979 | __acquires(busiest->lock) | |
2980 | __acquires(this_rq->lock) | |
2981 | { | |
e8fa1362 SR |
2982 | int ret = 0; |
2983 | ||
054b9108 KK |
2984 | if (unlikely(!irqs_disabled())) { |
2985 | /* printk() doesn't work good under rq->lock */ | |
2986 | spin_unlock(&this_rq->lock); | |
2987 | BUG_ON(1); | |
2988 | } | |
1da177e4 | 2989 | if (unlikely(!spin_trylock(&busiest->lock))) { |
c96d145e | 2990 | if (busiest < this_rq) { |
1da177e4 LT |
2991 | spin_unlock(&this_rq->lock); |
2992 | spin_lock(&busiest->lock); | |
2993 | spin_lock(&this_rq->lock); | |
e8fa1362 | 2994 | ret = 1; |
1da177e4 LT |
2995 | } else |
2996 | spin_lock(&busiest->lock); | |
2997 | } | |
e8fa1362 | 2998 | return ret; |
1da177e4 LT |
2999 | } |
3000 | ||
1da177e4 LT |
3001 | /* |
3002 | * If dest_cpu is allowed for this process, migrate the task to it. | |
3003 | * This is accomplished by forcing the cpu_allowed mask to only | |
41a2d6cf | 3004 | * allow dest_cpu, which will force the cpu onto dest_cpu. Then |
1da177e4 LT |
3005 | * the cpu_allowed mask is restored. |
3006 | */ | |
36c8b586 | 3007 | static void sched_migrate_task(struct task_struct *p, int dest_cpu) |
1da177e4 | 3008 | { |
70b97a7f | 3009 | struct migration_req req; |
1da177e4 | 3010 | unsigned long flags; |
70b97a7f | 3011 | struct rq *rq; |
1da177e4 LT |
3012 | |
3013 | rq = task_rq_lock(p, &flags); | |
3014 | if (!cpu_isset(dest_cpu, p->cpus_allowed) | |
3015 | || unlikely(cpu_is_offline(dest_cpu))) | |
3016 | goto out; | |
3017 | ||
3018 | /* force the process onto the specified CPU */ | |
3019 | if (migrate_task(p, dest_cpu, &req)) { | |
3020 | /* Need to wait for migration thread (might exit: take ref). */ | |
3021 | struct task_struct *mt = rq->migration_thread; | |
36c8b586 | 3022 | |
1da177e4 LT |
3023 | get_task_struct(mt); |
3024 | task_rq_unlock(rq, &flags); | |
3025 | wake_up_process(mt); | |
3026 | put_task_struct(mt); | |
3027 | wait_for_completion(&req.done); | |
36c8b586 | 3028 | |
1da177e4 LT |
3029 | return; |
3030 | } | |
3031 | out: | |
3032 | task_rq_unlock(rq, &flags); | |
3033 | } | |
3034 | ||
3035 | /* | |
476d139c NP |
3036 | * sched_exec - execve() is a valuable balancing opportunity, because at |
3037 | * this point the task has the smallest effective memory and cache footprint. | |
1da177e4 LT |
3038 | */ |
3039 | void sched_exec(void) | |
3040 | { | |
1da177e4 | 3041 | int new_cpu, this_cpu = get_cpu(); |
476d139c | 3042 | new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC); |
1da177e4 | 3043 | put_cpu(); |
476d139c NP |
3044 | if (new_cpu != this_cpu) |
3045 | sched_migrate_task(current, new_cpu); | |
1da177e4 LT |
3046 | } |
3047 | ||
3048 | /* | |
3049 | * pull_task - move a task from a remote runqueue to the local runqueue. | |
3050 | * Both runqueues must be locked. | |
3051 | */ | |
dd41f596 IM |
3052 | static void pull_task(struct rq *src_rq, struct task_struct *p, |
3053 | struct rq *this_rq, int this_cpu) | |
1da177e4 | 3054 | { |
2e1cb74a | 3055 | deactivate_task(src_rq, p, 0); |
1da177e4 | 3056 | set_task_cpu(p, this_cpu); |
dd41f596 | 3057 | activate_task(this_rq, p, 0); |
1da177e4 LT |
3058 | /* |
3059 | * Note that idle threads have a prio of MAX_PRIO, for this test | |
3060 | * to be always true for them. | |
3061 | */ | |
dd41f596 | 3062 | check_preempt_curr(this_rq, p); |
1da177e4 LT |
3063 | } |
3064 | ||
3065 | /* | |
3066 | * can_migrate_task - may task p from runqueue rq be migrated to this_cpu? | |
3067 | */ | |
858119e1 | 3068 | static |
70b97a7f | 3069 | int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu, |
d15bcfdb | 3070 | struct sched_domain *sd, enum cpu_idle_type idle, |
95cdf3b7 | 3071 | int *all_pinned) |
1da177e4 LT |
3072 | { |
3073 | /* | |
3074 | * We do not migrate tasks that are: | |
3075 | * 1) running (obviously), or | |
3076 | * 2) cannot be migrated to this CPU due to cpus_allowed, or | |
3077 | * 3) are cache-hot on their current CPU. | |
3078 | */ | |
cc367732 IM |
3079 | if (!cpu_isset(this_cpu, p->cpus_allowed)) { |
3080 | schedstat_inc(p, se.nr_failed_migrations_affine); | |
1da177e4 | 3081 | return 0; |
cc367732 | 3082 | } |
81026794 NP |
3083 | *all_pinned = 0; |
3084 | ||
cc367732 IM |
3085 | if (task_running(rq, p)) { |
3086 | schedstat_inc(p, se.nr_failed_migrations_running); | |
81026794 | 3087 | return 0; |
cc367732 | 3088 | } |
1da177e4 | 3089 | |
da84d961 IM |
3090 | /* |
3091 | * Aggressive migration if: | |
3092 | * 1) task is cache cold, or | |
3093 | * 2) too many balance attempts have failed. | |
3094 | */ | |
3095 | ||
6bc1665b IM |
3096 | if (!task_hot(p, rq->clock, sd) || |
3097 | sd->nr_balance_failed > sd->cache_nice_tries) { | |
da84d961 | 3098 | #ifdef CONFIG_SCHEDSTATS |
cc367732 | 3099 | if (task_hot(p, rq->clock, sd)) { |
da84d961 | 3100 | schedstat_inc(sd, lb_hot_gained[idle]); |
cc367732 IM |
3101 | schedstat_inc(p, se.nr_forced_migrations); |
3102 | } | |
da84d961 IM |
3103 | #endif |
3104 | return 1; | |
3105 | } | |
3106 | ||
cc367732 IM |
3107 | if (task_hot(p, rq->clock, sd)) { |
3108 | schedstat_inc(p, se.nr_failed_migrations_hot); | |
da84d961 | 3109 | return 0; |
cc367732 | 3110 | } |
1da177e4 LT |
3111 | return 1; |
3112 | } | |
3113 | ||
e1d1484f PW |
3114 | static unsigned long |
3115 | balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3116 | unsigned long max_load_move, struct sched_domain *sd, | |
3117 | enum cpu_idle_type idle, int *all_pinned, | |
3118 | int *this_best_prio, struct rq_iterator *iterator) | |
1da177e4 | 3119 | { |
b82d9fdd | 3120 | int loops = 0, pulled = 0, pinned = 0, skip_for_load; |
dd41f596 IM |
3121 | struct task_struct *p; |
3122 | long rem_load_move = max_load_move; | |
1da177e4 | 3123 | |
e1d1484f | 3124 | if (max_load_move == 0) |
1da177e4 LT |
3125 | goto out; |
3126 | ||
81026794 NP |
3127 | pinned = 1; |
3128 | ||
1da177e4 | 3129 | /* |
dd41f596 | 3130 | * Start the load-balancing iterator: |
1da177e4 | 3131 | */ |
dd41f596 IM |
3132 | p = iterator->start(iterator->arg); |
3133 | next: | |
b82d9fdd | 3134 | if (!p || loops++ > sysctl_sched_nr_migrate) |
1da177e4 | 3135 | goto out; |
50ddd969 | 3136 | /* |
b82d9fdd | 3137 | * To help distribute high priority tasks across CPUs we don't |
50ddd969 PW |
3138 | * skip a task if it will be the highest priority task (i.e. smallest |
3139 | * prio value) on its new queue regardless of its load weight | |
3140 | */ | |
dd41f596 IM |
3141 | skip_for_load = (p->se.load.weight >> 1) > rem_load_move + |
3142 | SCHED_LOAD_SCALE_FUZZ; | |
a4ac01c3 | 3143 | if ((skip_for_load && p->prio >= *this_best_prio) || |
dd41f596 | 3144 | !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { |
dd41f596 IM |
3145 | p = iterator->next(iterator->arg); |
3146 | goto next; | |
1da177e4 LT |
3147 | } |
3148 | ||
dd41f596 | 3149 | pull_task(busiest, p, this_rq, this_cpu); |
1da177e4 | 3150 | pulled++; |
dd41f596 | 3151 | rem_load_move -= p->se.load.weight; |
1da177e4 | 3152 | |
2dd73a4f | 3153 | /* |
b82d9fdd | 3154 | * We only want to steal up to the prescribed amount of weighted load. |
2dd73a4f | 3155 | */ |
e1d1484f | 3156 | if (rem_load_move > 0) { |
a4ac01c3 PW |
3157 | if (p->prio < *this_best_prio) |
3158 | *this_best_prio = p->prio; | |
dd41f596 IM |
3159 | p = iterator->next(iterator->arg); |
3160 | goto next; | |
1da177e4 LT |
3161 | } |
3162 | out: | |
3163 | /* | |
e1d1484f | 3164 | * Right now, this is one of only two places pull_task() is called, |
1da177e4 LT |
3165 | * so we can safely collect pull_task() stats here rather than |
3166 | * inside pull_task(). | |
3167 | */ | |
3168 | schedstat_add(sd, lb_gained[idle], pulled); | |
81026794 NP |
3169 | |
3170 | if (all_pinned) | |
3171 | *all_pinned = pinned; | |
e1d1484f PW |
3172 | |
3173 | return max_load_move - rem_load_move; | |
1da177e4 LT |
3174 | } |
3175 | ||
dd41f596 | 3176 | /* |
43010659 PW |
3177 | * move_tasks tries to move up to max_load_move weighted load from busiest to |
3178 | * this_rq, as part of a balancing operation within domain "sd". | |
3179 | * Returns 1 if successful and 0 otherwise. | |
dd41f596 IM |
3180 | * |
3181 | * Called with both runqueues locked. | |
3182 | */ | |
3183 | static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
43010659 | 3184 | unsigned long max_load_move, |
dd41f596 IM |
3185 | struct sched_domain *sd, enum cpu_idle_type idle, |
3186 | int *all_pinned) | |
3187 | { | |
5522d5d5 | 3188 | const struct sched_class *class = sched_class_highest; |
43010659 | 3189 | unsigned long total_load_moved = 0; |
a4ac01c3 | 3190 | int this_best_prio = this_rq->curr->prio; |
dd41f596 IM |
3191 | |
3192 | do { | |
43010659 PW |
3193 | total_load_moved += |
3194 | class->load_balance(this_rq, this_cpu, busiest, | |
e1d1484f | 3195 | max_load_move - total_load_moved, |
a4ac01c3 | 3196 | sd, idle, all_pinned, &this_best_prio); |
dd41f596 | 3197 | class = class->next; |
43010659 | 3198 | } while (class && max_load_move > total_load_moved); |
dd41f596 | 3199 | |
43010659 PW |
3200 | return total_load_moved > 0; |
3201 | } | |
3202 | ||
e1d1484f PW |
3203 | static int |
3204 | iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3205 | struct sched_domain *sd, enum cpu_idle_type idle, | |
3206 | struct rq_iterator *iterator) | |
3207 | { | |
3208 | struct task_struct *p = iterator->start(iterator->arg); | |
3209 | int pinned = 0; | |
3210 | ||
3211 | while (p) { | |
3212 | if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) { | |
3213 | pull_task(busiest, p, this_rq, this_cpu); | |
3214 | /* | |
3215 | * Right now, this is only the second place pull_task() | |
3216 | * is called, so we can safely collect pull_task() | |
3217 | * stats here rather than inside pull_task(). | |
3218 | */ | |
3219 | schedstat_inc(sd, lb_gained[idle]); | |
3220 | ||
3221 | return 1; | |
3222 | } | |
3223 | p = iterator->next(iterator->arg); | |
3224 | } | |
3225 | ||
3226 | return 0; | |
3227 | } | |
3228 | ||
43010659 PW |
3229 | /* |
3230 | * move_one_task tries to move exactly one task from busiest to this_rq, as | |
3231 | * part of active balancing operations within "domain". | |
3232 | * Returns 1 if successful and 0 otherwise. | |
3233 | * | |
3234 | * Called with both runqueues locked. | |
3235 | */ | |
3236 | static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest, | |
3237 | struct sched_domain *sd, enum cpu_idle_type idle) | |
3238 | { | |
5522d5d5 | 3239 | const struct sched_class *class; |
43010659 PW |
3240 | |
3241 | for (class = sched_class_highest; class; class = class->next) | |
e1d1484f | 3242 | if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle)) |
43010659 PW |
3243 | return 1; |
3244 | ||
3245 | return 0; | |
dd41f596 IM |
3246 | } |
3247 | ||
1da177e4 LT |
3248 | /* |
3249 | * find_busiest_group finds and returns the busiest CPU group within the | |
48f24c4d IM |
3250 | * domain. It calculates and returns the amount of weighted load which |
3251 | * should be moved to restore balance via the imbalance parameter. | |
1da177e4 LT |
3252 | */ |
3253 | static struct sched_group * | |
3254 | find_busiest_group(struct sched_domain *sd, int this_cpu, | |
dd41f596 | 3255 | unsigned long *imbalance, enum cpu_idle_type idle, |
7c16ec58 | 3256 | int *sd_idle, const cpumask_t *cpus, int *balance) |
1da177e4 LT |
3257 | { |
3258 | struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups; | |
3259 | unsigned long max_load, avg_load, total_load, this_load, total_pwr; | |
0c117f1b | 3260 | unsigned long max_pull; |
2dd73a4f PW |
3261 | unsigned long busiest_load_per_task, busiest_nr_running; |
3262 | unsigned long this_load_per_task, this_nr_running; | |
908a7c1b | 3263 | int load_idx, group_imb = 0; |
5c45bf27 SS |
3264 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
3265 | int power_savings_balance = 1; | |
3266 | unsigned long leader_nr_running = 0, min_load_per_task = 0; | |
3267 | unsigned long min_nr_running = ULONG_MAX; | |
3268 | struct sched_group *group_min = NULL, *group_leader = NULL; | |
3269 | #endif | |
1da177e4 LT |
3270 | |
3271 | max_load = this_load = total_load = total_pwr = 0; | |
2dd73a4f PW |
3272 | busiest_load_per_task = busiest_nr_running = 0; |
3273 | this_load_per_task = this_nr_running = 0; | |
d15bcfdb | 3274 | if (idle == CPU_NOT_IDLE) |
7897986b | 3275 | load_idx = sd->busy_idx; |
d15bcfdb | 3276 | else if (idle == CPU_NEWLY_IDLE) |
7897986b NP |
3277 | load_idx = sd->newidle_idx; |
3278 | else | |
3279 | load_idx = sd->idle_idx; | |
1da177e4 LT |
3280 | |
3281 | do { | |
908a7c1b | 3282 | unsigned long load, group_capacity, max_cpu_load, min_cpu_load; |
1da177e4 LT |
3283 | int local_group; |
3284 | int i; | |
908a7c1b | 3285 | int __group_imb = 0; |
783609c6 | 3286 | unsigned int balance_cpu = -1, first_idle_cpu = 0; |
2dd73a4f | 3287 | unsigned long sum_nr_running, sum_weighted_load; |
1da177e4 LT |
3288 | |
3289 | local_group = cpu_isset(this_cpu, group->cpumask); | |
3290 | ||
783609c6 SS |
3291 | if (local_group) |
3292 | balance_cpu = first_cpu(group->cpumask); | |
3293 | ||
1da177e4 | 3294 | /* Tally up the load of all CPUs in the group */ |
2dd73a4f | 3295 | sum_weighted_load = sum_nr_running = avg_load = 0; |
908a7c1b KC |
3296 | max_cpu_load = 0; |
3297 | min_cpu_load = ~0UL; | |
1da177e4 LT |
3298 | |
3299 | for_each_cpu_mask(i, group->cpumask) { | |
0a2966b4 CL |
3300 | struct rq *rq; |
3301 | ||
3302 | if (!cpu_isset(i, *cpus)) | |
3303 | continue; | |
3304 | ||
3305 | rq = cpu_rq(i); | |
2dd73a4f | 3306 | |
9439aab8 | 3307 | if (*sd_idle && rq->nr_running) |
5969fe06 NP |
3308 | *sd_idle = 0; |
3309 | ||
1da177e4 | 3310 | /* Bias balancing toward cpus of our domain */ |
783609c6 SS |
3311 | if (local_group) { |
3312 | if (idle_cpu(i) && !first_idle_cpu) { | |
3313 | first_idle_cpu = 1; | |
3314 | balance_cpu = i; | |
3315 | } | |
3316 | ||
a2000572 | 3317 | load = target_load(i, load_idx); |
908a7c1b | 3318 | } else { |
a2000572 | 3319 | load = source_load(i, load_idx); |
908a7c1b KC |
3320 | if (load > max_cpu_load) |
3321 | max_cpu_load = load; | |
3322 | if (min_cpu_load > load) | |
3323 | min_cpu_load = load; | |
3324 | } | |
1da177e4 LT |
3325 | |
3326 | avg_load += load; | |
2dd73a4f | 3327 | sum_nr_running += rq->nr_running; |
dd41f596 | 3328 | sum_weighted_load += weighted_cpuload(i); |
1da177e4 LT |
3329 | } |
3330 | ||
783609c6 SS |
3331 | /* |
3332 | * First idle cpu or the first cpu(busiest) in this sched group | |
3333 | * is eligible for doing load balancing at this and above | |
9439aab8 SS |
3334 | * domains. In the newly idle case, we will allow all the cpu's |
3335 | * to do the newly idle load balance. | |
783609c6 | 3336 | */ |
9439aab8 SS |
3337 | if (idle != CPU_NEWLY_IDLE && local_group && |
3338 | balance_cpu != this_cpu && balance) { | |
783609c6 SS |
3339 | *balance = 0; |
3340 | goto ret; | |
3341 | } | |
3342 | ||
1da177e4 | 3343 | total_load += avg_load; |
5517d86b | 3344 | total_pwr += group->__cpu_power; |
1da177e4 LT |
3345 | |
3346 | /* Adjust by relative CPU power of the group */ | |
5517d86b ED |
3347 | avg_load = sg_div_cpu_power(group, |
3348 | avg_load * SCHED_LOAD_SCALE); | |
1da177e4 | 3349 | |
908a7c1b KC |
3350 | if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE) |
3351 | __group_imb = 1; | |
3352 | ||
5517d86b | 3353 | group_capacity = group->__cpu_power / SCHED_LOAD_SCALE; |
5c45bf27 | 3354 | |
1da177e4 LT |
3355 | if (local_group) { |
3356 | this_load = avg_load; | |
3357 | this = group; | |
2dd73a4f PW |
3358 | this_nr_running = sum_nr_running; |
3359 | this_load_per_task = sum_weighted_load; | |
3360 | } else if (avg_load > max_load && | |
908a7c1b | 3361 | (sum_nr_running > group_capacity || __group_imb)) { |
1da177e4 LT |
3362 | max_load = avg_load; |
3363 | busiest = group; | |
2dd73a4f PW |
3364 | busiest_nr_running = sum_nr_running; |
3365 | busiest_load_per_task = sum_weighted_load; | |
908a7c1b | 3366 | group_imb = __group_imb; |
1da177e4 | 3367 | } |
5c45bf27 SS |
3368 | |
3369 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) | |
3370 | /* | |
3371 | * Busy processors will not participate in power savings | |
3372 | * balance. | |
3373 | */ | |
dd41f596 IM |
3374 | if (idle == CPU_NOT_IDLE || |
3375 | !(sd->flags & SD_POWERSAVINGS_BALANCE)) | |
3376 | goto group_next; | |
5c45bf27 SS |
3377 | |
3378 | /* | |
3379 | * If the local group is idle or completely loaded | |
3380 | * no need to do power savings balance at this domain | |
3381 | */ | |
3382 | if (local_group && (this_nr_running >= group_capacity || | |
3383 | !this_nr_running)) | |
3384 | power_savings_balance = 0; | |
3385 | ||
dd41f596 | 3386 | /* |
5c45bf27 SS |
3387 | * If a group is already running at full capacity or idle, |
3388 | * don't include that group in power savings calculations | |
dd41f596 IM |
3389 | */ |
3390 | if (!power_savings_balance || sum_nr_running >= group_capacity | |
5c45bf27 | 3391 | || !sum_nr_running) |
dd41f596 | 3392 | goto group_next; |
5c45bf27 | 3393 | |
dd41f596 | 3394 | /* |
5c45bf27 | 3395 | * Calculate the group which has the least non-idle load. |
dd41f596 IM |
3396 | * This is the group from where we need to pick up the load |
3397 | * for saving power | |
3398 | */ | |
3399 | if ((sum_nr_running < min_nr_running) || | |
3400 | (sum_nr_running == min_nr_running && | |
5c45bf27 SS |
3401 | first_cpu(group->cpumask) < |
3402 | first_cpu(group_min->cpumask))) { | |
dd41f596 IM |
3403 | group_min = group; |
3404 | min_nr_running = sum_nr_running; | |
5c45bf27 SS |
3405 | min_load_per_task = sum_weighted_load / |
3406 | sum_nr_running; | |
dd41f596 | 3407 | } |
5c45bf27 | 3408 | |
dd41f596 | 3409 | /* |
5c45bf27 | 3410 | * Calculate the group which is almost near its |
dd41f596 IM |
3411 | * capacity but still has some space to pick up some load |
3412 | * from other group and save more power | |
3413 | */ | |
3414 | if (sum_nr_running <= group_capacity - 1) { | |
3415 | if (sum_nr_running > leader_nr_running || | |
3416 | (sum_nr_running == leader_nr_running && | |
3417 | first_cpu(group->cpumask) > | |
3418 | first_cpu(group_leader->cpumask))) { | |
3419 | group_leader = group; | |
3420 | leader_nr_running = sum_nr_running; | |
3421 | } | |
48f24c4d | 3422 | } |
5c45bf27 SS |
3423 | group_next: |
3424 | #endif | |
1da177e4 LT |
3425 | group = group->next; |
3426 | } while (group != sd->groups); | |
3427 | ||
2dd73a4f | 3428 | if (!busiest || this_load >= max_load || busiest_nr_running == 0) |
1da177e4 LT |
3429 | goto out_balanced; |
3430 | ||
3431 | avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr; | |
3432 | ||
3433 | if (this_load >= avg_load || | |
3434 | 100*max_load <= sd->imbalance_pct*this_load) | |
3435 | goto out_balanced; | |
3436 | ||
2dd73a4f | 3437 | busiest_load_per_task /= busiest_nr_running; |
908a7c1b KC |
3438 | if (group_imb) |
3439 | busiest_load_per_task = min(busiest_load_per_task, avg_load); | |
3440 | ||
1da177e4 LT |
3441 | /* |
3442 | * We're trying to get all the cpus to the average_load, so we don't | |
3443 | * want to push ourselves above the average load, nor do we wish to | |
3444 | * reduce the max loaded cpu below the average load, as either of these | |
3445 | * actions would just result in more rebalancing later, and ping-pong | |
3446 | * tasks around. Thus we look for the minimum possible imbalance. | |
3447 | * Negative imbalances (*we* are more loaded than anyone else) will | |
3448 | * be counted as no imbalance for these purposes -- we can't fix that | |
41a2d6cf | 3449 | * by pulling tasks to us. Be careful of negative numbers as they'll |
1da177e4 LT |
3450 | * appear as very large values with unsigned longs. |
3451 | */ | |
2dd73a4f PW |
3452 | if (max_load <= busiest_load_per_task) |
3453 | goto out_balanced; | |
3454 | ||
3455 | /* | |
3456 | * In the presence of smp nice balancing, certain scenarios can have | |
3457 | * max load less than avg load(as we skip the groups at or below | |
3458 | * its cpu_power, while calculating max_load..) | |
3459 | */ | |
3460 | if (max_load < avg_load) { | |
3461 | *imbalance = 0; | |
3462 | goto small_imbalance; | |
3463 | } | |
0c117f1b SS |
3464 | |
3465 | /* Don't want to pull so many tasks that a group would go idle */ | |
2dd73a4f | 3466 | max_pull = min(max_load - avg_load, max_load - busiest_load_per_task); |
0c117f1b | 3467 | |
1da177e4 | 3468 | /* How much load to actually move to equalise the imbalance */ |
5517d86b ED |
3469 | *imbalance = min(max_pull * busiest->__cpu_power, |
3470 | (avg_load - this_load) * this->__cpu_power) | |
1da177e4 LT |
3471 | / SCHED_LOAD_SCALE; |
3472 | ||
2dd73a4f PW |
3473 | /* |
3474 | * if *imbalance is less than the average load per runnable task | |
3475 | * there is no gaurantee that any tasks will be moved so we'll have | |
3476 | * a think about bumping its value to force at least one task to be | |
3477 | * moved | |
3478 | */ | |
7fd0d2dd | 3479 | if (*imbalance < busiest_load_per_task) { |
48f24c4d | 3480 | unsigned long tmp, pwr_now, pwr_move; |
2dd73a4f PW |
3481 | unsigned int imbn; |
3482 | ||
3483 | small_imbalance: | |
3484 | pwr_move = pwr_now = 0; | |
3485 | imbn = 2; | |
3486 | if (this_nr_running) { | |
3487 | this_load_per_task /= this_nr_running; | |
3488 | if (busiest_load_per_task > this_load_per_task) | |
3489 | imbn = 1; | |
3490 | } else | |
3491 | this_load_per_task = SCHED_LOAD_SCALE; | |
1da177e4 | 3492 | |
dd41f596 IM |
3493 | if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >= |
3494 | busiest_load_per_task * imbn) { | |
2dd73a4f | 3495 | *imbalance = busiest_load_per_task; |
1da177e4 LT |
3496 | return busiest; |
3497 | } | |
3498 | ||
3499 | /* | |
3500 | * OK, we don't have enough imbalance to justify moving tasks, | |
3501 | * however we may be able to increase total CPU power used by | |
3502 | * moving them. | |
3503 | */ | |
3504 | ||
5517d86b ED |
3505 | pwr_now += busiest->__cpu_power * |
3506 | min(busiest_load_per_task, max_load); | |
3507 | pwr_now += this->__cpu_power * | |
3508 | min(this_load_per_task, this_load); | |
1da177e4 LT |
3509 | pwr_now /= SCHED_LOAD_SCALE; |
3510 | ||
3511 | /* Amount of load we'd subtract */ | |
5517d86b ED |
3512 | tmp = sg_div_cpu_power(busiest, |
3513 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
1da177e4 | 3514 | if (max_load > tmp) |
5517d86b | 3515 | pwr_move += busiest->__cpu_power * |
2dd73a4f | 3516 | min(busiest_load_per_task, max_load - tmp); |
1da177e4 LT |
3517 | |
3518 | /* Amount of load we'd add */ | |
5517d86b | 3519 | if (max_load * busiest->__cpu_power < |
33859f7f | 3520 | busiest_load_per_task * SCHED_LOAD_SCALE) |
5517d86b ED |
3521 | tmp = sg_div_cpu_power(this, |
3522 | max_load * busiest->__cpu_power); | |
1da177e4 | 3523 | else |
5517d86b ED |
3524 | tmp = sg_div_cpu_power(this, |
3525 | busiest_load_per_task * SCHED_LOAD_SCALE); | |
3526 | pwr_move += this->__cpu_power * | |
3527 | min(this_load_per_task, this_load + tmp); | |
1da177e4 LT |
3528 | pwr_move /= SCHED_LOAD_SCALE; |
3529 | ||
3530 | /* Move if we gain throughput */ | |
7fd0d2dd SS |
3531 | if (pwr_move > pwr_now) |
3532 | *imbalance = busiest_load_per_task; | |
1da177e4 LT |
3533 | } |
3534 | ||
1da177e4 LT |
3535 | return busiest; |
3536 | ||
3537 | out_balanced: | |
5c45bf27 | 3538 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
d15bcfdb | 3539 | if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE)) |
5c45bf27 | 3540 | goto ret; |
1da177e4 | 3541 | |
5c45bf27 SS |
3542 | if (this == group_leader && group_leader != group_min) { |
3543 | *imbalance = min_load_per_task; | |
3544 | return group_min; | |
3545 | } | |
5c45bf27 | 3546 | #endif |
783609c6 | 3547 | ret: |
1da177e4 LT |
3548 | *imbalance = 0; |
3549 | return NULL; | |
3550 | } | |
3551 | ||
3552 | /* | |
3553 | * find_busiest_queue - find the busiest runqueue among the cpus in group. | |
3554 | */ | |
70b97a7f | 3555 | static struct rq * |
d15bcfdb | 3556 | find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle, |
7c16ec58 | 3557 | unsigned long imbalance, const cpumask_t *cpus) |
1da177e4 | 3558 | { |
70b97a7f | 3559 | struct rq *busiest = NULL, *rq; |
2dd73a4f | 3560 | unsigned long max_load = 0; |
1da177e4 LT |
3561 | int i; |
3562 | ||
3563 | for_each_cpu_mask(i, group->cpumask) { | |
dd41f596 | 3564 | unsigned long wl; |
0a2966b4 CL |
3565 | |
3566 | if (!cpu_isset(i, *cpus)) | |
3567 | continue; | |
3568 | ||
48f24c4d | 3569 | rq = cpu_rq(i); |
dd41f596 | 3570 | wl = weighted_cpuload(i); |
2dd73a4f | 3571 | |
dd41f596 | 3572 | if (rq->nr_running == 1 && wl > imbalance) |
2dd73a4f | 3573 | continue; |
1da177e4 | 3574 | |
dd41f596 IM |
3575 | if (wl > max_load) { |
3576 | max_load = wl; | |
48f24c4d | 3577 | busiest = rq; |
1da177e4 LT |
3578 | } |
3579 | } | |
3580 | ||
3581 | return busiest; | |
3582 | } | |
3583 | ||
77391d71 NP |
3584 | /* |
3585 | * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but | |
3586 | * so long as it is large enough. | |
3587 | */ | |
3588 | #define MAX_PINNED_INTERVAL 512 | |
3589 | ||
1da177e4 LT |
3590 | /* |
3591 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3592 | * tasks if there is an imbalance. | |
1da177e4 | 3593 | */ |
70b97a7f | 3594 | static int load_balance(int this_cpu, struct rq *this_rq, |
d15bcfdb | 3595 | struct sched_domain *sd, enum cpu_idle_type idle, |
7c16ec58 | 3596 | int *balance, cpumask_t *cpus) |
1da177e4 | 3597 | { |
43010659 | 3598 | int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0; |
1da177e4 | 3599 | struct sched_group *group; |
1da177e4 | 3600 | unsigned long imbalance; |
70b97a7f | 3601 | struct rq *busiest; |
fe2eea3f | 3602 | unsigned long flags; |
18d95a28 | 3603 | int unlock_aggregate; |
5969fe06 | 3604 | |
7c16ec58 MT |
3605 | cpus_setall(*cpus); |
3606 | ||
18d95a28 PZ |
3607 | unlock_aggregate = get_aggregate(sd); |
3608 | ||
89c4710e SS |
3609 | /* |
3610 | * When power savings policy is enabled for the parent domain, idle | |
3611 | * sibling can pick up load irrespective of busy siblings. In this case, | |
dd41f596 | 3612 | * let the state of idle sibling percolate up as CPU_IDLE, instead of |
d15bcfdb | 3613 | * portraying it as CPU_NOT_IDLE. |
89c4710e | 3614 | */ |
d15bcfdb | 3615 | if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3616 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3617 | sd_idle = 1; |
1da177e4 | 3618 | |
2d72376b | 3619 | schedstat_inc(sd, lb_count[idle]); |
1da177e4 | 3620 | |
0a2966b4 CL |
3621 | redo: |
3622 | group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle, | |
7c16ec58 | 3623 | cpus, balance); |
783609c6 | 3624 | |
06066714 | 3625 | if (*balance == 0) |
783609c6 | 3626 | goto out_balanced; |
783609c6 | 3627 | |
1da177e4 LT |
3628 | if (!group) { |
3629 | schedstat_inc(sd, lb_nobusyg[idle]); | |
3630 | goto out_balanced; | |
3631 | } | |
3632 | ||
7c16ec58 | 3633 | busiest = find_busiest_queue(group, idle, imbalance, cpus); |
1da177e4 LT |
3634 | if (!busiest) { |
3635 | schedstat_inc(sd, lb_nobusyq[idle]); | |
3636 | goto out_balanced; | |
3637 | } | |
3638 | ||
db935dbd | 3639 | BUG_ON(busiest == this_rq); |
1da177e4 LT |
3640 | |
3641 | schedstat_add(sd, lb_imbalance[idle], imbalance); | |
3642 | ||
43010659 | 3643 | ld_moved = 0; |
1da177e4 LT |
3644 | if (busiest->nr_running > 1) { |
3645 | /* | |
3646 | * Attempt to move tasks. If find_busiest_group has found | |
3647 | * an imbalance but busiest->nr_running <= 1, the group is | |
43010659 | 3648 | * still unbalanced. ld_moved simply stays zero, so it is |
1da177e4 LT |
3649 | * correctly treated as an imbalance. |
3650 | */ | |
fe2eea3f | 3651 | local_irq_save(flags); |
e17224bf | 3652 | double_rq_lock(this_rq, busiest); |
43010659 | 3653 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
48f24c4d | 3654 | imbalance, sd, idle, &all_pinned); |
e17224bf | 3655 | double_rq_unlock(this_rq, busiest); |
fe2eea3f | 3656 | local_irq_restore(flags); |
81026794 | 3657 | |
46cb4b7c SS |
3658 | /* |
3659 | * some other cpu did the load balance for us. | |
3660 | */ | |
43010659 | 3661 | if (ld_moved && this_cpu != smp_processor_id()) |
46cb4b7c SS |
3662 | resched_cpu(this_cpu); |
3663 | ||
81026794 | 3664 | /* All tasks on this runqueue were pinned by CPU affinity */ |
0a2966b4 | 3665 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3666 | cpu_clear(cpu_of(busiest), *cpus); |
3667 | if (!cpus_empty(*cpus)) | |
0a2966b4 | 3668 | goto redo; |
81026794 | 3669 | goto out_balanced; |
0a2966b4 | 3670 | } |
1da177e4 | 3671 | } |
81026794 | 3672 | |
43010659 | 3673 | if (!ld_moved) { |
1da177e4 LT |
3674 | schedstat_inc(sd, lb_failed[idle]); |
3675 | sd->nr_balance_failed++; | |
3676 | ||
3677 | if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) { | |
1da177e4 | 3678 | |
fe2eea3f | 3679 | spin_lock_irqsave(&busiest->lock, flags); |
fa3b6ddc SS |
3680 | |
3681 | /* don't kick the migration_thread, if the curr | |
3682 | * task on busiest cpu can't be moved to this_cpu | |
3683 | */ | |
3684 | if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) { | |
fe2eea3f | 3685 | spin_unlock_irqrestore(&busiest->lock, flags); |
fa3b6ddc SS |
3686 | all_pinned = 1; |
3687 | goto out_one_pinned; | |
3688 | } | |
3689 | ||
1da177e4 LT |
3690 | if (!busiest->active_balance) { |
3691 | busiest->active_balance = 1; | |
3692 | busiest->push_cpu = this_cpu; | |
81026794 | 3693 | active_balance = 1; |
1da177e4 | 3694 | } |
fe2eea3f | 3695 | spin_unlock_irqrestore(&busiest->lock, flags); |
81026794 | 3696 | if (active_balance) |
1da177e4 LT |
3697 | wake_up_process(busiest->migration_thread); |
3698 | ||
3699 | /* | |
3700 | * We've kicked active balancing, reset the failure | |
3701 | * counter. | |
3702 | */ | |
39507451 | 3703 | sd->nr_balance_failed = sd->cache_nice_tries+1; |
1da177e4 | 3704 | } |
81026794 | 3705 | } else |
1da177e4 LT |
3706 | sd->nr_balance_failed = 0; |
3707 | ||
81026794 | 3708 | if (likely(!active_balance)) { |
1da177e4 LT |
3709 | /* We were unbalanced, so reset the balancing interval */ |
3710 | sd->balance_interval = sd->min_interval; | |
81026794 NP |
3711 | } else { |
3712 | /* | |
3713 | * If we've begun active balancing, start to back off. This | |
3714 | * case may not be covered by the all_pinned logic if there | |
3715 | * is only 1 task on the busy runqueue (because we don't call | |
3716 | * move_tasks). | |
3717 | */ | |
3718 | if (sd->balance_interval < sd->max_interval) | |
3719 | sd->balance_interval *= 2; | |
1da177e4 LT |
3720 | } |
3721 | ||
43010659 | 3722 | if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3723 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3724 | ld_moved = -1; |
3725 | ||
3726 | goto out; | |
1da177e4 LT |
3727 | |
3728 | out_balanced: | |
1da177e4 LT |
3729 | schedstat_inc(sd, lb_balanced[idle]); |
3730 | ||
16cfb1c0 | 3731 | sd->nr_balance_failed = 0; |
fa3b6ddc SS |
3732 | |
3733 | out_one_pinned: | |
1da177e4 | 3734 | /* tune up the balancing interval */ |
77391d71 NP |
3735 | if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) || |
3736 | (sd->balance_interval < sd->max_interval)) | |
1da177e4 LT |
3737 | sd->balance_interval *= 2; |
3738 | ||
48f24c4d | 3739 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3740 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
18d95a28 PZ |
3741 | ld_moved = -1; |
3742 | else | |
3743 | ld_moved = 0; | |
3744 | out: | |
3745 | if (unlock_aggregate) | |
3746 | put_aggregate(sd); | |
3747 | return ld_moved; | |
1da177e4 LT |
3748 | } |
3749 | ||
3750 | /* | |
3751 | * Check this_cpu to ensure it is balanced within domain. Attempt to move | |
3752 | * tasks if there is an imbalance. | |
3753 | * | |
d15bcfdb | 3754 | * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE). |
1da177e4 LT |
3755 | * this_rq is locked. |
3756 | */ | |
48f24c4d | 3757 | static int |
7c16ec58 MT |
3758 | load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd, |
3759 | cpumask_t *cpus) | |
1da177e4 LT |
3760 | { |
3761 | struct sched_group *group; | |
70b97a7f | 3762 | struct rq *busiest = NULL; |
1da177e4 | 3763 | unsigned long imbalance; |
43010659 | 3764 | int ld_moved = 0; |
5969fe06 | 3765 | int sd_idle = 0; |
969bb4e4 | 3766 | int all_pinned = 0; |
7c16ec58 MT |
3767 | |
3768 | cpus_setall(*cpus); | |
5969fe06 | 3769 | |
89c4710e SS |
3770 | /* |
3771 | * When power savings policy is enabled for the parent domain, idle | |
3772 | * sibling can pick up load irrespective of busy siblings. In this case, | |
3773 | * let the state of idle sibling percolate up as IDLE, instead of | |
d15bcfdb | 3774 | * portraying it as CPU_NOT_IDLE. |
89c4710e SS |
3775 | */ |
3776 | if (sd->flags & SD_SHARE_CPUPOWER && | |
3777 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 | 3778 | sd_idle = 1; |
1da177e4 | 3779 | |
2d72376b | 3780 | schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]); |
0a2966b4 | 3781 | redo: |
d15bcfdb | 3782 | group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE, |
7c16ec58 | 3783 | &sd_idle, cpus, NULL); |
1da177e4 | 3784 | if (!group) { |
d15bcfdb | 3785 | schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3786 | goto out_balanced; |
1da177e4 LT |
3787 | } |
3788 | ||
7c16ec58 | 3789 | busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus); |
db935dbd | 3790 | if (!busiest) { |
d15bcfdb | 3791 | schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]); |
16cfb1c0 | 3792 | goto out_balanced; |
1da177e4 LT |
3793 | } |
3794 | ||
db935dbd NP |
3795 | BUG_ON(busiest == this_rq); |
3796 | ||
d15bcfdb | 3797 | schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance); |
d6d5cfaf | 3798 | |
43010659 | 3799 | ld_moved = 0; |
d6d5cfaf NP |
3800 | if (busiest->nr_running > 1) { |
3801 | /* Attempt to move tasks */ | |
3802 | double_lock_balance(this_rq, busiest); | |
6e82a3be IM |
3803 | /* this_rq->clock is already updated */ |
3804 | update_rq_clock(busiest); | |
43010659 | 3805 | ld_moved = move_tasks(this_rq, this_cpu, busiest, |
969bb4e4 SS |
3806 | imbalance, sd, CPU_NEWLY_IDLE, |
3807 | &all_pinned); | |
d6d5cfaf | 3808 | spin_unlock(&busiest->lock); |
0a2966b4 | 3809 | |
969bb4e4 | 3810 | if (unlikely(all_pinned)) { |
7c16ec58 MT |
3811 | cpu_clear(cpu_of(busiest), *cpus); |
3812 | if (!cpus_empty(*cpus)) | |
0a2966b4 CL |
3813 | goto redo; |
3814 | } | |
d6d5cfaf NP |
3815 | } |
3816 | ||
43010659 | 3817 | if (!ld_moved) { |
d15bcfdb | 3818 | schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]); |
89c4710e SS |
3819 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
3820 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) | |
5969fe06 NP |
3821 | return -1; |
3822 | } else | |
16cfb1c0 | 3823 | sd->nr_balance_failed = 0; |
1da177e4 | 3824 | |
43010659 | 3825 | return ld_moved; |
16cfb1c0 NP |
3826 | |
3827 | out_balanced: | |
d15bcfdb | 3828 | schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]); |
48f24c4d | 3829 | if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER && |
89c4710e | 3830 | !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE)) |
5969fe06 | 3831 | return -1; |
16cfb1c0 | 3832 | sd->nr_balance_failed = 0; |
48f24c4d | 3833 | |
16cfb1c0 | 3834 | return 0; |
1da177e4 LT |
3835 | } |
3836 | ||
3837 | /* | |
3838 | * idle_balance is called by schedule() if this_cpu is about to become | |
3839 | * idle. Attempts to pull tasks from other CPUs. | |
3840 | */ | |
70b97a7f | 3841 | static void idle_balance(int this_cpu, struct rq *this_rq) |
1da177e4 LT |
3842 | { |
3843 | struct sched_domain *sd; | |
dd41f596 IM |
3844 | int pulled_task = -1; |
3845 | unsigned long next_balance = jiffies + HZ; | |
7c16ec58 | 3846 | cpumask_t tmpmask; |
1da177e4 LT |
3847 | |
3848 | for_each_domain(this_cpu, sd) { | |
92c4ca5c CL |
3849 | unsigned long interval; |
3850 | ||
3851 | if (!(sd->flags & SD_LOAD_BALANCE)) | |
3852 | continue; | |
3853 | ||
3854 | if (sd->flags & SD_BALANCE_NEWIDLE) | |
48f24c4d | 3855 | /* If we've pulled tasks over stop searching: */ |
7c16ec58 MT |
3856 | pulled_task = load_balance_newidle(this_cpu, this_rq, |
3857 | sd, &tmpmask); | |
92c4ca5c CL |
3858 | |
3859 | interval = msecs_to_jiffies(sd->balance_interval); | |
3860 | if (time_after(next_balance, sd->last_balance + interval)) | |
3861 | next_balance = sd->last_balance + interval; | |
3862 | if (pulled_task) | |
3863 | break; | |
1da177e4 | 3864 | } |
dd41f596 | 3865 | if (pulled_task || time_after(jiffies, this_rq->next_balance)) { |
1bd77f2d CL |
3866 | /* |
3867 | * We are going idle. next_balance may be set based on | |
3868 | * a busy processor. So reset next_balance. | |
3869 | */ | |
3870 | this_rq->next_balance = next_balance; | |
dd41f596 | 3871 | } |
1da177e4 LT |
3872 | } |
3873 | ||
3874 | /* | |
3875 | * active_load_balance is run by migration threads. It pushes running tasks | |
3876 | * off the busiest CPU onto idle CPUs. It requires at least 1 task to be | |
3877 | * running on each physical CPU where possible, and avoids physical / | |
3878 | * logical imbalances. | |
3879 | * | |
3880 | * Called with busiest_rq locked. | |
3881 | */ | |
70b97a7f | 3882 | static void active_load_balance(struct rq *busiest_rq, int busiest_cpu) |
1da177e4 | 3883 | { |
39507451 | 3884 | int target_cpu = busiest_rq->push_cpu; |
70b97a7f IM |
3885 | struct sched_domain *sd; |
3886 | struct rq *target_rq; | |
39507451 | 3887 | |
48f24c4d | 3888 | /* Is there any task to move? */ |
39507451 | 3889 | if (busiest_rq->nr_running <= 1) |
39507451 NP |
3890 | return; |
3891 | ||
3892 | target_rq = cpu_rq(target_cpu); | |
1da177e4 LT |
3893 | |
3894 | /* | |
39507451 | 3895 | * This condition is "impossible", if it occurs |
41a2d6cf | 3896 | * we need to fix it. Originally reported by |
39507451 | 3897 | * Bjorn Helgaas on a 128-cpu setup. |
1da177e4 | 3898 | */ |
39507451 | 3899 | BUG_ON(busiest_rq == target_rq); |
1da177e4 | 3900 | |
39507451 NP |
3901 | /* move a task from busiest_rq to target_rq */ |
3902 | double_lock_balance(busiest_rq, target_rq); | |
6e82a3be IM |
3903 | update_rq_clock(busiest_rq); |
3904 | update_rq_clock(target_rq); | |
39507451 NP |
3905 | |
3906 | /* Search for an sd spanning us and the target CPU. */ | |
c96d145e | 3907 | for_each_domain(target_cpu, sd) { |
39507451 | 3908 | if ((sd->flags & SD_LOAD_BALANCE) && |
48f24c4d | 3909 | cpu_isset(busiest_cpu, sd->span)) |
39507451 | 3910 | break; |
c96d145e | 3911 | } |
39507451 | 3912 | |
48f24c4d | 3913 | if (likely(sd)) { |
2d72376b | 3914 | schedstat_inc(sd, alb_count); |
39507451 | 3915 | |
43010659 PW |
3916 | if (move_one_task(target_rq, target_cpu, busiest_rq, |
3917 | sd, CPU_IDLE)) | |
48f24c4d IM |
3918 | schedstat_inc(sd, alb_pushed); |
3919 | else | |
3920 | schedstat_inc(sd, alb_failed); | |
3921 | } | |
39507451 | 3922 | spin_unlock(&target_rq->lock); |
1da177e4 LT |
3923 | } |
3924 | ||
46cb4b7c SS |
3925 | #ifdef CONFIG_NO_HZ |
3926 | static struct { | |
3927 | atomic_t load_balancer; | |
41a2d6cf | 3928 | cpumask_t cpu_mask; |
46cb4b7c SS |
3929 | } nohz ____cacheline_aligned = { |
3930 | .load_balancer = ATOMIC_INIT(-1), | |
3931 | .cpu_mask = CPU_MASK_NONE, | |
3932 | }; | |
3933 | ||
7835b98b | 3934 | /* |
46cb4b7c SS |
3935 | * This routine will try to nominate the ilb (idle load balancing) |
3936 | * owner among the cpus whose ticks are stopped. ilb owner will do the idle | |
3937 | * load balancing on behalf of all those cpus. If all the cpus in the system | |
3938 | * go into this tickless mode, then there will be no ilb owner (as there is | |
3939 | * no need for one) and all the cpus will sleep till the next wakeup event | |
3940 | * arrives... | |
3941 | * | |
3942 | * For the ilb owner, tick is not stopped. And this tick will be used | |
3943 | * for idle load balancing. ilb owner will still be part of | |
3944 | * nohz.cpu_mask.. | |
7835b98b | 3945 | * |
46cb4b7c SS |
3946 | * While stopping the tick, this cpu will become the ilb owner if there |
3947 | * is no other owner. And will be the owner till that cpu becomes busy | |
3948 | * or if all cpus in the system stop their ticks at which point | |
3949 | * there is no need for ilb owner. | |
3950 | * | |
3951 | * When the ilb owner becomes busy, it nominates another owner, during the | |
3952 | * next busy scheduler_tick() | |
3953 | */ | |
3954 | int select_nohz_load_balancer(int stop_tick) | |
3955 | { | |
3956 | int cpu = smp_processor_id(); | |
3957 | ||
3958 | if (stop_tick) { | |
3959 | cpu_set(cpu, nohz.cpu_mask); | |
3960 | cpu_rq(cpu)->in_nohz_recently = 1; | |
3961 | ||
3962 | /* | |
3963 | * If we are going offline and still the leader, give up! | |
3964 | */ | |
3965 | if (cpu_is_offline(cpu) && | |
3966 | atomic_read(&nohz.load_balancer) == cpu) { | |
3967 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3968 | BUG(); | |
3969 | return 0; | |
3970 | } | |
3971 | ||
3972 | /* time for ilb owner also to sleep */ | |
3973 | if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
3974 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3975 | atomic_set(&nohz.load_balancer, -1); | |
3976 | return 0; | |
3977 | } | |
3978 | ||
3979 | if (atomic_read(&nohz.load_balancer) == -1) { | |
3980 | /* make me the ilb owner */ | |
3981 | if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1) | |
3982 | return 1; | |
3983 | } else if (atomic_read(&nohz.load_balancer) == cpu) | |
3984 | return 1; | |
3985 | } else { | |
3986 | if (!cpu_isset(cpu, nohz.cpu_mask)) | |
3987 | return 0; | |
3988 | ||
3989 | cpu_clear(cpu, nohz.cpu_mask); | |
3990 | ||
3991 | if (atomic_read(&nohz.load_balancer) == cpu) | |
3992 | if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu) | |
3993 | BUG(); | |
3994 | } | |
3995 | return 0; | |
3996 | } | |
3997 | #endif | |
3998 | ||
3999 | static DEFINE_SPINLOCK(balancing); | |
4000 | ||
4001 | /* | |
7835b98b CL |
4002 | * It checks each scheduling domain to see if it is due to be balanced, |
4003 | * and initiates a balancing operation if so. | |
4004 | * | |
4005 | * Balancing parameters are set up in arch_init_sched_domains. | |
4006 | */ | |
a9957449 | 4007 | static void rebalance_domains(int cpu, enum cpu_idle_type idle) |
7835b98b | 4008 | { |
46cb4b7c SS |
4009 | int balance = 1; |
4010 | struct rq *rq = cpu_rq(cpu); | |
7835b98b CL |
4011 | unsigned long interval; |
4012 | struct sched_domain *sd; | |
46cb4b7c | 4013 | /* Earliest time when we have to do rebalance again */ |
c9819f45 | 4014 | unsigned long next_balance = jiffies + 60*HZ; |
f549da84 | 4015 | int update_next_balance = 0; |
7c16ec58 | 4016 | cpumask_t tmp; |
1da177e4 | 4017 | |
46cb4b7c | 4018 | for_each_domain(cpu, sd) { |
1da177e4 LT |
4019 | if (!(sd->flags & SD_LOAD_BALANCE)) |
4020 | continue; | |
4021 | ||
4022 | interval = sd->balance_interval; | |
d15bcfdb | 4023 | if (idle != CPU_IDLE) |
1da177e4 LT |
4024 | interval *= sd->busy_factor; |
4025 | ||
4026 | /* scale ms to jiffies */ | |
4027 | interval = msecs_to_jiffies(interval); | |
4028 | if (unlikely(!interval)) | |
4029 | interval = 1; | |
dd41f596 IM |
4030 | if (interval > HZ*NR_CPUS/10) |
4031 | interval = HZ*NR_CPUS/10; | |
4032 | ||
1da177e4 | 4033 | |
08c183f3 CL |
4034 | if (sd->flags & SD_SERIALIZE) { |
4035 | if (!spin_trylock(&balancing)) | |
4036 | goto out; | |
4037 | } | |
4038 | ||
c9819f45 | 4039 | if (time_after_eq(jiffies, sd->last_balance + interval)) { |
7c16ec58 | 4040 | if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) { |
fa3b6ddc SS |
4041 | /* |
4042 | * We've pulled tasks over so either we're no | |
5969fe06 NP |
4043 | * longer idle, or one of our SMT siblings is |
4044 | * not idle. | |
4045 | */ | |
d15bcfdb | 4046 | idle = CPU_NOT_IDLE; |
1da177e4 | 4047 | } |
1bd77f2d | 4048 | sd->last_balance = jiffies; |
1da177e4 | 4049 | } |
08c183f3 CL |
4050 | if (sd->flags & SD_SERIALIZE) |
4051 | spin_unlock(&balancing); | |
4052 | out: | |
f549da84 | 4053 | if (time_after(next_balance, sd->last_balance + interval)) { |
c9819f45 | 4054 | next_balance = sd->last_balance + interval; |
f549da84 SS |
4055 | update_next_balance = 1; |
4056 | } | |
783609c6 SS |
4057 | |
4058 | /* | |
4059 | * Stop the load balance at this level. There is another | |
4060 | * CPU in our sched group which is doing load balancing more | |
4061 | * actively. | |
4062 | */ | |
4063 | if (!balance) | |
4064 | break; | |
1da177e4 | 4065 | } |
f549da84 SS |
4066 | |
4067 | /* | |
4068 | * next_balance will be updated only when there is a need. | |
4069 | * When the cpu is attached to null domain for ex, it will not be | |
4070 | * updated. | |
4071 | */ | |
4072 | if (likely(update_next_balance)) | |
4073 | rq->next_balance = next_balance; | |
46cb4b7c SS |
4074 | } |
4075 | ||
4076 | /* | |
4077 | * run_rebalance_domains is triggered when needed from the scheduler tick. | |
4078 | * In CONFIG_NO_HZ case, the idle load balance owner will do the | |
4079 | * rebalancing for all the cpus for whom scheduler ticks are stopped. | |
4080 | */ | |
4081 | static void run_rebalance_domains(struct softirq_action *h) | |
4082 | { | |
dd41f596 IM |
4083 | int this_cpu = smp_processor_id(); |
4084 | struct rq *this_rq = cpu_rq(this_cpu); | |
4085 | enum cpu_idle_type idle = this_rq->idle_at_tick ? | |
4086 | CPU_IDLE : CPU_NOT_IDLE; | |
46cb4b7c | 4087 | |
dd41f596 | 4088 | rebalance_domains(this_cpu, idle); |
46cb4b7c SS |
4089 | |
4090 | #ifdef CONFIG_NO_HZ | |
4091 | /* | |
4092 | * If this cpu is the owner for idle load balancing, then do the | |
4093 | * balancing on behalf of the other idle cpus whose ticks are | |
4094 | * stopped. | |
4095 | */ | |
dd41f596 IM |
4096 | if (this_rq->idle_at_tick && |
4097 | atomic_read(&nohz.load_balancer) == this_cpu) { | |
46cb4b7c SS |
4098 | cpumask_t cpus = nohz.cpu_mask; |
4099 | struct rq *rq; | |
4100 | int balance_cpu; | |
4101 | ||
dd41f596 | 4102 | cpu_clear(this_cpu, cpus); |
46cb4b7c SS |
4103 | for_each_cpu_mask(balance_cpu, cpus) { |
4104 | /* | |
4105 | * If this cpu gets work to do, stop the load balancing | |
4106 | * work being done for other cpus. Next load | |
4107 | * balancing owner will pick it up. | |
4108 | */ | |
4109 | if (need_resched()) | |
4110 | break; | |
4111 | ||
de0cf899 | 4112 | rebalance_domains(balance_cpu, CPU_IDLE); |
46cb4b7c SS |
4113 | |
4114 | rq = cpu_rq(balance_cpu); | |
dd41f596 IM |
4115 | if (time_after(this_rq->next_balance, rq->next_balance)) |
4116 | this_rq->next_balance = rq->next_balance; | |
46cb4b7c SS |
4117 | } |
4118 | } | |
4119 | #endif | |
4120 | } | |
4121 | ||
4122 | /* | |
4123 | * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing. | |
4124 | * | |
4125 | * In case of CONFIG_NO_HZ, this is the place where we nominate a new | |
4126 | * idle load balancing owner or decide to stop the periodic load balancing, | |
4127 | * if the whole system is idle. | |
4128 | */ | |
dd41f596 | 4129 | static inline void trigger_load_balance(struct rq *rq, int cpu) |
46cb4b7c | 4130 | { |
46cb4b7c SS |
4131 | #ifdef CONFIG_NO_HZ |
4132 | /* | |
4133 | * If we were in the nohz mode recently and busy at the current | |
4134 | * scheduler tick, then check if we need to nominate new idle | |
4135 | * load balancer. | |
4136 | */ | |
4137 | if (rq->in_nohz_recently && !rq->idle_at_tick) { | |
4138 | rq->in_nohz_recently = 0; | |
4139 | ||
4140 | if (atomic_read(&nohz.load_balancer) == cpu) { | |
4141 | cpu_clear(cpu, nohz.cpu_mask); | |
4142 | atomic_set(&nohz.load_balancer, -1); | |
4143 | } | |
4144 | ||
4145 | if (atomic_read(&nohz.load_balancer) == -1) { | |
4146 | /* | |
4147 | * simple selection for now: Nominate the | |
4148 | * first cpu in the nohz list to be the next | |
4149 | * ilb owner. | |
4150 | * | |
4151 | * TBD: Traverse the sched domains and nominate | |
4152 | * the nearest cpu in the nohz.cpu_mask. | |
4153 | */ | |
4154 | int ilb = first_cpu(nohz.cpu_mask); | |
4155 | ||
434d53b0 | 4156 | if (ilb < nr_cpu_ids) |
46cb4b7c SS |
4157 | resched_cpu(ilb); |
4158 | } | |
4159 | } | |
4160 | ||
4161 | /* | |
4162 | * If this cpu is idle and doing idle load balancing for all the | |
4163 | * cpus with ticks stopped, is it time for that to stop? | |
4164 | */ | |
4165 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu && | |
4166 | cpus_weight(nohz.cpu_mask) == num_online_cpus()) { | |
4167 | resched_cpu(cpu); | |
4168 | return; | |
4169 | } | |
4170 | ||
4171 | /* | |
4172 | * If this cpu is idle and the idle load balancing is done by | |
4173 | * someone else, then no need raise the SCHED_SOFTIRQ | |
4174 | */ | |
4175 | if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu && | |
4176 | cpu_isset(cpu, nohz.cpu_mask)) | |
4177 | return; | |
4178 | #endif | |
4179 | if (time_after_eq(jiffies, rq->next_balance)) | |
4180 | raise_softirq(SCHED_SOFTIRQ); | |
1da177e4 | 4181 | } |
dd41f596 IM |
4182 | |
4183 | #else /* CONFIG_SMP */ | |
4184 | ||
1da177e4 LT |
4185 | /* |
4186 | * on UP we do not need to balance between CPUs: | |
4187 | */ | |
70b97a7f | 4188 | static inline void idle_balance(int cpu, struct rq *rq) |
1da177e4 LT |
4189 | { |
4190 | } | |
dd41f596 | 4191 | |
1da177e4 LT |
4192 | #endif |
4193 | ||
1da177e4 LT |
4194 | DEFINE_PER_CPU(struct kernel_stat, kstat); |
4195 | ||
4196 | EXPORT_PER_CPU_SYMBOL(kstat); | |
4197 | ||
4198 | /* | |
41b86e9c IM |
4199 | * Return p->sum_exec_runtime plus any more ns on the sched_clock |
4200 | * that have not yet been banked in case the task is currently running. | |
1da177e4 | 4201 | */ |
41b86e9c | 4202 | unsigned long long task_sched_runtime(struct task_struct *p) |
1da177e4 | 4203 | { |
1da177e4 | 4204 | unsigned long flags; |
41b86e9c IM |
4205 | u64 ns, delta_exec; |
4206 | struct rq *rq; | |
48f24c4d | 4207 | |
41b86e9c IM |
4208 | rq = task_rq_lock(p, &flags); |
4209 | ns = p->se.sum_exec_runtime; | |
051a1d1a | 4210 | if (task_current(rq, p)) { |
a8e504d2 IM |
4211 | update_rq_clock(rq); |
4212 | delta_exec = rq->clock - p->se.exec_start; | |
41b86e9c IM |
4213 | if ((s64)delta_exec > 0) |
4214 | ns += delta_exec; | |
4215 | } | |
4216 | task_rq_unlock(rq, &flags); | |
48f24c4d | 4217 | |
1da177e4 LT |
4218 | return ns; |
4219 | } | |
4220 | ||
1da177e4 LT |
4221 | /* |
4222 | * Account user cpu time to a process. | |
4223 | * @p: the process that the cpu time gets accounted to | |
1da177e4 LT |
4224 | * @cputime: the cpu time spent in user space since the last update |
4225 | */ | |
4226 | void account_user_time(struct task_struct *p, cputime_t cputime) | |
4227 | { | |
4228 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4229 | cputime64_t tmp; | |
4230 | ||
4231 | p->utime = cputime_add(p->utime, cputime); | |
4232 | ||
4233 | /* Add user time to cpustat. */ | |
4234 | tmp = cputime_to_cputime64(cputime); | |
4235 | if (TASK_NICE(p) > 0) | |
4236 | cpustat->nice = cputime64_add(cpustat->nice, tmp); | |
4237 | else | |
4238 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4239 | } | |
4240 | ||
94886b84 LV |
4241 | /* |
4242 | * Account guest cpu time to a process. | |
4243 | * @p: the process that the cpu time gets accounted to | |
4244 | * @cputime: the cpu time spent in virtual machine since the last update | |
4245 | */ | |
f7402e03 | 4246 | static void account_guest_time(struct task_struct *p, cputime_t cputime) |
94886b84 LV |
4247 | { |
4248 | cputime64_t tmp; | |
4249 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4250 | ||
4251 | tmp = cputime_to_cputime64(cputime); | |
4252 | ||
4253 | p->utime = cputime_add(p->utime, cputime); | |
4254 | p->gtime = cputime_add(p->gtime, cputime); | |
4255 | ||
4256 | cpustat->user = cputime64_add(cpustat->user, tmp); | |
4257 | cpustat->guest = cputime64_add(cpustat->guest, tmp); | |
4258 | } | |
4259 | ||
c66f08be MN |
4260 | /* |
4261 | * Account scaled user cpu time to a process. | |
4262 | * @p: the process that the cpu time gets accounted to | |
4263 | * @cputime: the cpu time spent in user space since the last update | |
4264 | */ | |
4265 | void account_user_time_scaled(struct task_struct *p, cputime_t cputime) | |
4266 | { | |
4267 | p->utimescaled = cputime_add(p->utimescaled, cputime); | |
4268 | } | |
4269 | ||
1da177e4 LT |
4270 | /* |
4271 | * Account system cpu time to a process. | |
4272 | * @p: the process that the cpu time gets accounted to | |
4273 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4274 | * @cputime: the cpu time spent in kernel space since the last update | |
4275 | */ | |
4276 | void account_system_time(struct task_struct *p, int hardirq_offset, | |
4277 | cputime_t cputime) | |
4278 | { | |
4279 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
70b97a7f | 4280 | struct rq *rq = this_rq(); |
1da177e4 LT |
4281 | cputime64_t tmp; |
4282 | ||
983ed7a6 HH |
4283 | if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { |
4284 | account_guest_time(p, cputime); | |
4285 | return; | |
4286 | } | |
94886b84 | 4287 | |
1da177e4 LT |
4288 | p->stime = cputime_add(p->stime, cputime); |
4289 | ||
4290 | /* Add system time to cpustat. */ | |
4291 | tmp = cputime_to_cputime64(cputime); | |
4292 | if (hardirq_count() - hardirq_offset) | |
4293 | cpustat->irq = cputime64_add(cpustat->irq, tmp); | |
4294 | else if (softirq_count()) | |
4295 | cpustat->softirq = cputime64_add(cpustat->softirq, tmp); | |
cfb52856 | 4296 | else if (p != rq->idle) |
1da177e4 | 4297 | cpustat->system = cputime64_add(cpustat->system, tmp); |
cfb52856 | 4298 | else if (atomic_read(&rq->nr_iowait) > 0) |
1da177e4 LT |
4299 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); |
4300 | else | |
4301 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
4302 | /* Account for system time used */ | |
4303 | acct_update_integrals(p); | |
1da177e4 LT |
4304 | } |
4305 | ||
c66f08be MN |
4306 | /* |
4307 | * Account scaled system cpu time to a process. | |
4308 | * @p: the process that the cpu time gets accounted to | |
4309 | * @hardirq_offset: the offset to subtract from hardirq_count() | |
4310 | * @cputime: the cpu time spent in kernel space since the last update | |
4311 | */ | |
4312 | void account_system_time_scaled(struct task_struct *p, cputime_t cputime) | |
4313 | { | |
4314 | p->stimescaled = cputime_add(p->stimescaled, cputime); | |
4315 | } | |
4316 | ||
1da177e4 LT |
4317 | /* |
4318 | * Account for involuntary wait time. | |
4319 | * @p: the process from which the cpu time has been stolen | |
4320 | * @steal: the cpu time spent in involuntary wait | |
4321 | */ | |
4322 | void account_steal_time(struct task_struct *p, cputime_t steal) | |
4323 | { | |
4324 | struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat; | |
4325 | cputime64_t tmp = cputime_to_cputime64(steal); | |
70b97a7f | 4326 | struct rq *rq = this_rq(); |
1da177e4 LT |
4327 | |
4328 | if (p == rq->idle) { | |
4329 | p->stime = cputime_add(p->stime, steal); | |
4330 | if (atomic_read(&rq->nr_iowait) > 0) | |
4331 | cpustat->iowait = cputime64_add(cpustat->iowait, tmp); | |
4332 | else | |
4333 | cpustat->idle = cputime64_add(cpustat->idle, tmp); | |
cfb52856 | 4334 | } else |
1da177e4 LT |
4335 | cpustat->steal = cputime64_add(cpustat->steal, tmp); |
4336 | } | |
4337 | ||
7835b98b CL |
4338 | /* |
4339 | * This function gets called by the timer code, with HZ frequency. | |
4340 | * We call it with interrupts disabled. | |
4341 | * | |
4342 | * It also gets called by the fork code, when changing the parent's | |
4343 | * timeslices. | |
4344 | */ | |
4345 | void scheduler_tick(void) | |
4346 | { | |
7835b98b CL |
4347 | int cpu = smp_processor_id(); |
4348 | struct rq *rq = cpu_rq(cpu); | |
dd41f596 | 4349 | struct task_struct *curr = rq->curr; |
3e51f33f PZ |
4350 | |
4351 | sched_clock_tick(); | |
dd41f596 IM |
4352 | |
4353 | spin_lock(&rq->lock); | |
3e51f33f | 4354 | update_rq_clock(rq); |
f1a438d8 | 4355 | update_cpu_load(rq); |
fa85ae24 | 4356 | curr->sched_class->task_tick(rq, curr, 0); |
dd41f596 | 4357 | spin_unlock(&rq->lock); |
7835b98b | 4358 | |
e418e1c2 | 4359 | #ifdef CONFIG_SMP |
dd41f596 IM |
4360 | rq->idle_at_tick = idle_cpu(cpu); |
4361 | trigger_load_balance(rq, cpu); | |
e418e1c2 | 4362 | #endif |
1da177e4 LT |
4363 | } |
4364 | ||
1da177e4 LT |
4365 | #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT) |
4366 | ||
43627582 | 4367 | void __kprobes add_preempt_count(int val) |
1da177e4 LT |
4368 | { |
4369 | /* | |
4370 | * Underflow? | |
4371 | */ | |
9a11b49a IM |
4372 | if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) |
4373 | return; | |
1da177e4 LT |
4374 | preempt_count() += val; |
4375 | /* | |
4376 | * Spinlock count overflowing soon? | |
4377 | */ | |
33859f7f MOS |
4378 | DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= |
4379 | PREEMPT_MASK - 10); | |
1da177e4 LT |
4380 | } |
4381 | EXPORT_SYMBOL(add_preempt_count); | |
4382 | ||
43627582 | 4383 | void __kprobes sub_preempt_count(int val) |
1da177e4 LT |
4384 | { |
4385 | /* | |
4386 | * Underflow? | |
4387 | */ | |
9a11b49a IM |
4388 | if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) |
4389 | return; | |
1da177e4 LT |
4390 | /* |
4391 | * Is the spinlock portion underflowing? | |
4392 | */ | |
9a11b49a IM |
4393 | if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && |
4394 | !(preempt_count() & PREEMPT_MASK))) | |
4395 | return; | |
4396 | ||
1da177e4 LT |
4397 | preempt_count() -= val; |
4398 | } | |
4399 | EXPORT_SYMBOL(sub_preempt_count); | |
4400 | ||
4401 | #endif | |
4402 | ||
4403 | /* | |
dd41f596 | 4404 | * Print scheduling while atomic bug: |
1da177e4 | 4405 | */ |
dd41f596 | 4406 | static noinline void __schedule_bug(struct task_struct *prev) |
1da177e4 | 4407 | { |
838225b4 SS |
4408 | struct pt_regs *regs = get_irq_regs(); |
4409 | ||
4410 | printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", | |
4411 | prev->comm, prev->pid, preempt_count()); | |
4412 | ||
dd41f596 IM |
4413 | debug_show_held_locks(prev); |
4414 | if (irqs_disabled()) | |
4415 | print_irqtrace_events(prev); | |
838225b4 SS |
4416 | |
4417 | if (regs) | |
4418 | show_regs(regs); | |
4419 | else | |
4420 | dump_stack(); | |
dd41f596 | 4421 | } |
1da177e4 | 4422 | |
dd41f596 IM |
4423 | /* |
4424 | * Various schedule()-time debugging checks and statistics: | |
4425 | */ | |
4426 | static inline void schedule_debug(struct task_struct *prev) | |
4427 | { | |
1da177e4 | 4428 | /* |
41a2d6cf | 4429 | * Test if we are atomic. Since do_exit() needs to call into |
1da177e4 LT |
4430 | * schedule() atomically, we ignore that path for now. |
4431 | * Otherwise, whine if we are scheduling when we should not be. | |
4432 | */ | |
dd41f596 IM |
4433 | if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state)) |
4434 | __schedule_bug(prev); | |
4435 | ||
1da177e4 LT |
4436 | profile_hit(SCHED_PROFILING, __builtin_return_address(0)); |
4437 | ||
2d72376b | 4438 | schedstat_inc(this_rq(), sched_count); |
b8efb561 IM |
4439 | #ifdef CONFIG_SCHEDSTATS |
4440 | if (unlikely(prev->lock_depth >= 0)) { | |
2d72376b IM |
4441 | schedstat_inc(this_rq(), bkl_count); |
4442 | schedstat_inc(prev, sched_info.bkl_count); | |
b8efb561 IM |
4443 | } |
4444 | #endif | |
dd41f596 IM |
4445 | } |
4446 | ||
4447 | /* | |
4448 | * Pick up the highest-prio task: | |
4449 | */ | |
4450 | static inline struct task_struct * | |
ff95f3df | 4451 | pick_next_task(struct rq *rq, struct task_struct *prev) |
dd41f596 | 4452 | { |
5522d5d5 | 4453 | const struct sched_class *class; |
dd41f596 | 4454 | struct task_struct *p; |
1da177e4 LT |
4455 | |
4456 | /* | |
dd41f596 IM |
4457 | * Optimization: we know that if all tasks are in |
4458 | * the fair class we can call that function directly: | |
1da177e4 | 4459 | */ |
dd41f596 | 4460 | if (likely(rq->nr_running == rq->cfs.nr_running)) { |
fb8d4724 | 4461 | p = fair_sched_class.pick_next_task(rq); |
dd41f596 IM |
4462 | if (likely(p)) |
4463 | return p; | |
1da177e4 LT |
4464 | } |
4465 | ||
dd41f596 IM |
4466 | class = sched_class_highest; |
4467 | for ( ; ; ) { | |
fb8d4724 | 4468 | p = class->pick_next_task(rq); |
dd41f596 IM |
4469 | if (p) |
4470 | return p; | |
4471 | /* | |
4472 | * Will never be NULL as the idle class always | |
4473 | * returns a non-NULL p: | |
4474 | */ | |
4475 | class = class->next; | |
4476 | } | |
4477 | } | |
1da177e4 | 4478 | |
dd41f596 IM |
4479 | /* |
4480 | * schedule() is the main scheduler function. | |
4481 | */ | |
4482 | asmlinkage void __sched schedule(void) | |
4483 | { | |
4484 | struct task_struct *prev, *next; | |
67ca7bde | 4485 | unsigned long *switch_count; |
dd41f596 | 4486 | struct rq *rq; |
dd41f596 IM |
4487 | int cpu; |
4488 | ||
4489 | need_resched: | |
4490 | preempt_disable(); | |
4491 | cpu = smp_processor_id(); | |
4492 | rq = cpu_rq(cpu); | |
4493 | rcu_qsctr_inc(cpu); | |
4494 | prev = rq->curr; | |
4495 | switch_count = &prev->nivcsw; | |
4496 | ||
4497 | release_kernel_lock(prev); | |
4498 | need_resched_nonpreemptible: | |
4499 | ||
4500 | schedule_debug(prev); | |
1da177e4 | 4501 | |
8f4d37ec PZ |
4502 | hrtick_clear(rq); |
4503 | ||
1e819950 IM |
4504 | /* |
4505 | * Do the rq-clock update outside the rq lock: | |
4506 | */ | |
4507 | local_irq_disable(); | |
3e51f33f | 4508 | update_rq_clock(rq); |
1e819950 IM |
4509 | spin_lock(&rq->lock); |
4510 | clear_tsk_need_resched(prev); | |
1da177e4 | 4511 | |
1da177e4 | 4512 | if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { |
1da177e4 | 4513 | if (unlikely((prev->state & TASK_INTERRUPTIBLE) && |
23e3c3cd | 4514 | signal_pending(prev))) { |
1da177e4 | 4515 | prev->state = TASK_RUNNING; |
dd41f596 | 4516 | } else { |
2e1cb74a | 4517 | deactivate_task(rq, prev, 1); |
1da177e4 | 4518 | } |
dd41f596 | 4519 | switch_count = &prev->nvcsw; |
1da177e4 LT |
4520 | } |
4521 | ||
9a897c5a SR |
4522 | #ifdef CONFIG_SMP |
4523 | if (prev->sched_class->pre_schedule) | |
4524 | prev->sched_class->pre_schedule(rq, prev); | |
4525 | #endif | |
f65eda4f | 4526 | |
dd41f596 | 4527 | if (unlikely(!rq->nr_running)) |
1da177e4 | 4528 | idle_balance(cpu, rq); |
1da177e4 | 4529 | |
31ee529c | 4530 | prev->sched_class->put_prev_task(rq, prev); |
ff95f3df | 4531 | next = pick_next_task(rq, prev); |
1da177e4 | 4532 | |
1da177e4 | 4533 | if (likely(prev != next)) { |
673a90a1 DS |
4534 | sched_info_switch(prev, next); |
4535 | ||
1da177e4 LT |
4536 | rq->nr_switches++; |
4537 | rq->curr = next; | |
4538 | ++*switch_count; | |
4539 | ||
dd41f596 | 4540 | context_switch(rq, prev, next); /* unlocks the rq */ |
8f4d37ec PZ |
4541 | /* |
4542 | * the context switch might have flipped the stack from under | |
4543 | * us, hence refresh the local variables. | |
4544 | */ | |
4545 | cpu = smp_processor_id(); | |
4546 | rq = cpu_rq(cpu); | |
1da177e4 LT |
4547 | } else |
4548 | spin_unlock_irq(&rq->lock); | |
4549 | ||
8f4d37ec PZ |
4550 | hrtick_set(rq); |
4551 | ||
4552 | if (unlikely(reacquire_kernel_lock(current) < 0)) | |
1da177e4 | 4553 | goto need_resched_nonpreemptible; |
8f4d37ec | 4554 | |
1da177e4 LT |
4555 | preempt_enable_no_resched(); |
4556 | if (unlikely(test_thread_flag(TIF_NEED_RESCHED))) | |
4557 | goto need_resched; | |
4558 | } | |
1da177e4 LT |
4559 | EXPORT_SYMBOL(schedule); |
4560 | ||
4561 | #ifdef CONFIG_PREEMPT | |
4562 | /* | |
2ed6e34f | 4563 | * this is the entry point to schedule() from in-kernel preemption |
41a2d6cf | 4564 | * off of preempt_enable. Kernel preemptions off return from interrupt |
1da177e4 LT |
4565 | * occur there and call schedule directly. |
4566 | */ | |
4567 | asmlinkage void __sched preempt_schedule(void) | |
4568 | { | |
4569 | struct thread_info *ti = current_thread_info(); | |
6478d880 | 4570 | |
1da177e4 LT |
4571 | /* |
4572 | * If there is a non-zero preempt_count or interrupts are disabled, | |
41a2d6cf | 4573 | * we do not want to preempt the current task. Just return.. |
1da177e4 | 4574 | */ |
beed33a8 | 4575 | if (likely(ti->preempt_count || irqs_disabled())) |
1da177e4 LT |
4576 | return; |
4577 | ||
3a5c359a AK |
4578 | do { |
4579 | add_preempt_count(PREEMPT_ACTIVE); | |
3a5c359a | 4580 | schedule(); |
3a5c359a | 4581 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4582 | |
3a5c359a AK |
4583 | /* |
4584 | * Check again in case we missed a preemption opportunity | |
4585 | * between schedule and now. | |
4586 | */ | |
4587 | barrier(); | |
4588 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 | 4589 | } |
1da177e4 LT |
4590 | EXPORT_SYMBOL(preempt_schedule); |
4591 | ||
4592 | /* | |
2ed6e34f | 4593 | * this is the entry point to schedule() from kernel preemption |
1da177e4 LT |
4594 | * off of irq context. |
4595 | * Note, that this is called and return with irqs disabled. This will | |
4596 | * protect us against recursive calling from irq. | |
4597 | */ | |
4598 | asmlinkage void __sched preempt_schedule_irq(void) | |
4599 | { | |
4600 | struct thread_info *ti = current_thread_info(); | |
6478d880 | 4601 | |
2ed6e34f | 4602 | /* Catch callers which need to be fixed */ |
1da177e4 LT |
4603 | BUG_ON(ti->preempt_count || !irqs_disabled()); |
4604 | ||
3a5c359a AK |
4605 | do { |
4606 | add_preempt_count(PREEMPT_ACTIVE); | |
3a5c359a AK |
4607 | local_irq_enable(); |
4608 | schedule(); | |
4609 | local_irq_disable(); | |
3a5c359a | 4610 | sub_preempt_count(PREEMPT_ACTIVE); |
1da177e4 | 4611 | |
3a5c359a AK |
4612 | /* |
4613 | * Check again in case we missed a preemption opportunity | |
4614 | * between schedule and now. | |
4615 | */ | |
4616 | barrier(); | |
4617 | } while (unlikely(test_thread_flag(TIF_NEED_RESCHED))); | |
1da177e4 LT |
4618 | } |
4619 | ||
4620 | #endif /* CONFIG_PREEMPT */ | |
4621 | ||
95cdf3b7 IM |
4622 | int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, |
4623 | void *key) | |
1da177e4 | 4624 | { |
48f24c4d | 4625 | return try_to_wake_up(curr->private, mode, sync); |
1da177e4 | 4626 | } |
1da177e4 LT |
4627 | EXPORT_SYMBOL(default_wake_function); |
4628 | ||
4629 | /* | |
41a2d6cf IM |
4630 | * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just |
4631 | * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve | |
1da177e4 LT |
4632 | * number) then we wake all the non-exclusive tasks and one exclusive task. |
4633 | * | |
4634 | * There are circumstances in which we can try to wake a task which has already | |
41a2d6cf | 4635 | * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns |
1da177e4 LT |
4636 | * zero in this (rare) case, and we handle it by continuing to scan the queue. |
4637 | */ | |
4638 | static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, | |
4639 | int nr_exclusive, int sync, void *key) | |
4640 | { | |
2e45874c | 4641 | wait_queue_t *curr, *next; |
1da177e4 | 4642 | |
2e45874c | 4643 | list_for_each_entry_safe(curr, next, &q->task_list, task_list) { |
48f24c4d IM |
4644 | unsigned flags = curr->flags; |
4645 | ||
1da177e4 | 4646 | if (curr->func(curr, mode, sync, key) && |
48f24c4d | 4647 | (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) |
1da177e4 LT |
4648 | break; |
4649 | } | |
4650 | } | |
4651 | ||
4652 | /** | |
4653 | * __wake_up - wake up threads blocked on a waitqueue. | |
4654 | * @q: the waitqueue | |
4655 | * @mode: which threads | |
4656 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
67be2dd1 | 4657 | * @key: is directly passed to the wakeup function |
1da177e4 | 4658 | */ |
7ad5b3a5 | 4659 | void __wake_up(wait_queue_head_t *q, unsigned int mode, |
95cdf3b7 | 4660 | int nr_exclusive, void *key) |
1da177e4 LT |
4661 | { |
4662 | unsigned long flags; | |
4663 | ||
4664 | spin_lock_irqsave(&q->lock, flags); | |
4665 | __wake_up_common(q, mode, nr_exclusive, 0, key); | |
4666 | spin_unlock_irqrestore(&q->lock, flags); | |
4667 | } | |
1da177e4 LT |
4668 | EXPORT_SYMBOL(__wake_up); |
4669 | ||
4670 | /* | |
4671 | * Same as __wake_up but called with the spinlock in wait_queue_head_t held. | |
4672 | */ | |
7ad5b3a5 | 4673 | void __wake_up_locked(wait_queue_head_t *q, unsigned int mode) |
1da177e4 LT |
4674 | { |
4675 | __wake_up_common(q, mode, 1, 0, NULL); | |
4676 | } | |
4677 | ||
4678 | /** | |
67be2dd1 | 4679 | * __wake_up_sync - wake up threads blocked on a waitqueue. |
1da177e4 LT |
4680 | * @q: the waitqueue |
4681 | * @mode: which threads | |
4682 | * @nr_exclusive: how many wake-one or wake-many threads to wake up | |
4683 | * | |
4684 | * The sync wakeup differs that the waker knows that it will schedule | |
4685 | * away soon, so while the target thread will be woken up, it will not | |
4686 | * be migrated to another CPU - ie. the two threads are 'synchronized' | |
4687 | * with each other. This can prevent needless bouncing between CPUs. | |
4688 | * | |
4689 | * On UP it can prevent extra preemption. | |
4690 | */ | |
7ad5b3a5 | 4691 | void |
95cdf3b7 | 4692 | __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) |
1da177e4 LT |
4693 | { |
4694 | unsigned long flags; | |
4695 | int sync = 1; | |
4696 | ||
4697 | if (unlikely(!q)) | |
4698 | return; | |
4699 | ||
4700 | if (unlikely(!nr_exclusive)) | |
4701 | sync = 0; | |
4702 | ||
4703 | spin_lock_irqsave(&q->lock, flags); | |
4704 | __wake_up_common(q, mode, nr_exclusive, sync, NULL); | |
4705 | spin_unlock_irqrestore(&q->lock, flags); | |
4706 | } | |
4707 | EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ | |
4708 | ||
b15136e9 | 4709 | void complete(struct completion *x) |
1da177e4 LT |
4710 | { |
4711 | unsigned long flags; | |
4712 | ||
4713 | spin_lock_irqsave(&x->wait.lock, flags); | |
4714 | x->done++; | |
d9514f6c | 4715 | __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); |
1da177e4 LT |
4716 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4717 | } | |
4718 | EXPORT_SYMBOL(complete); | |
4719 | ||
b15136e9 | 4720 | void complete_all(struct completion *x) |
1da177e4 LT |
4721 | { |
4722 | unsigned long flags; | |
4723 | ||
4724 | spin_lock_irqsave(&x->wait.lock, flags); | |
4725 | x->done += UINT_MAX/2; | |
d9514f6c | 4726 | __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); |
1da177e4 LT |
4727 | spin_unlock_irqrestore(&x->wait.lock, flags); |
4728 | } | |
4729 | EXPORT_SYMBOL(complete_all); | |
4730 | ||
8cbbe86d AK |
4731 | static inline long __sched |
4732 | do_wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4733 | { |
1da177e4 LT |
4734 | if (!x->done) { |
4735 | DECLARE_WAITQUEUE(wait, current); | |
4736 | ||
4737 | wait.flags |= WQ_FLAG_EXCLUSIVE; | |
4738 | __add_wait_queue_tail(&x->wait, &wait); | |
4739 | do { | |
009e577e MW |
4740 | if ((state == TASK_INTERRUPTIBLE && |
4741 | signal_pending(current)) || | |
4742 | (state == TASK_KILLABLE && | |
4743 | fatal_signal_pending(current))) { | |
8cbbe86d AK |
4744 | __remove_wait_queue(&x->wait, &wait); |
4745 | return -ERESTARTSYS; | |
4746 | } | |
4747 | __set_current_state(state); | |
1da177e4 LT |
4748 | spin_unlock_irq(&x->wait.lock); |
4749 | timeout = schedule_timeout(timeout); | |
4750 | spin_lock_irq(&x->wait.lock); | |
4751 | if (!timeout) { | |
4752 | __remove_wait_queue(&x->wait, &wait); | |
8cbbe86d | 4753 | return timeout; |
1da177e4 LT |
4754 | } |
4755 | } while (!x->done); | |
4756 | __remove_wait_queue(&x->wait, &wait); | |
4757 | } | |
4758 | x->done--; | |
1da177e4 LT |
4759 | return timeout; |
4760 | } | |
1da177e4 | 4761 | |
8cbbe86d AK |
4762 | static long __sched |
4763 | wait_for_common(struct completion *x, long timeout, int state) | |
1da177e4 | 4764 | { |
1da177e4 LT |
4765 | might_sleep(); |
4766 | ||
4767 | spin_lock_irq(&x->wait.lock); | |
8cbbe86d | 4768 | timeout = do_wait_for_common(x, timeout, state); |
1da177e4 | 4769 | spin_unlock_irq(&x->wait.lock); |
8cbbe86d AK |
4770 | return timeout; |
4771 | } | |
1da177e4 | 4772 | |
b15136e9 | 4773 | void __sched wait_for_completion(struct completion *x) |
8cbbe86d AK |
4774 | { |
4775 | wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); | |
1da177e4 | 4776 | } |
8cbbe86d | 4777 | EXPORT_SYMBOL(wait_for_completion); |
1da177e4 | 4778 | |
b15136e9 | 4779 | unsigned long __sched |
8cbbe86d | 4780 | wait_for_completion_timeout(struct completion *x, unsigned long timeout) |
1da177e4 | 4781 | { |
8cbbe86d | 4782 | return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); |
1da177e4 | 4783 | } |
8cbbe86d | 4784 | EXPORT_SYMBOL(wait_for_completion_timeout); |
1da177e4 | 4785 | |
8cbbe86d | 4786 | int __sched wait_for_completion_interruptible(struct completion *x) |
0fec171c | 4787 | { |
51e97990 AK |
4788 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); |
4789 | if (t == -ERESTARTSYS) | |
4790 | return t; | |
4791 | return 0; | |
0fec171c | 4792 | } |
8cbbe86d | 4793 | EXPORT_SYMBOL(wait_for_completion_interruptible); |
1da177e4 | 4794 | |
b15136e9 | 4795 | unsigned long __sched |
8cbbe86d AK |
4796 | wait_for_completion_interruptible_timeout(struct completion *x, |
4797 | unsigned long timeout) | |
0fec171c | 4798 | { |
8cbbe86d | 4799 | return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); |
0fec171c | 4800 | } |
8cbbe86d | 4801 | EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); |
1da177e4 | 4802 | |
009e577e MW |
4803 | int __sched wait_for_completion_killable(struct completion *x) |
4804 | { | |
4805 | long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); | |
4806 | if (t == -ERESTARTSYS) | |
4807 | return t; | |
4808 | return 0; | |
4809 | } | |
4810 | EXPORT_SYMBOL(wait_for_completion_killable); | |
4811 | ||
8cbbe86d AK |
4812 | static long __sched |
4813 | sleep_on_common(wait_queue_head_t *q, int state, long timeout) | |
1da177e4 | 4814 | { |
0fec171c IM |
4815 | unsigned long flags; |
4816 | wait_queue_t wait; | |
4817 | ||
4818 | init_waitqueue_entry(&wait, current); | |
1da177e4 | 4819 | |
8cbbe86d | 4820 | __set_current_state(state); |
1da177e4 | 4821 | |
8cbbe86d AK |
4822 | spin_lock_irqsave(&q->lock, flags); |
4823 | __add_wait_queue(q, &wait); | |
4824 | spin_unlock(&q->lock); | |
4825 | timeout = schedule_timeout(timeout); | |
4826 | spin_lock_irq(&q->lock); | |
4827 | __remove_wait_queue(q, &wait); | |
4828 | spin_unlock_irqrestore(&q->lock, flags); | |
4829 | ||
4830 | return timeout; | |
4831 | } | |
4832 | ||
4833 | void __sched interruptible_sleep_on(wait_queue_head_t *q) | |
4834 | { | |
4835 | sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); | |
1da177e4 | 4836 | } |
1da177e4 LT |
4837 | EXPORT_SYMBOL(interruptible_sleep_on); |
4838 | ||
0fec171c | 4839 | long __sched |
95cdf3b7 | 4840 | interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4841 | { |
8cbbe86d | 4842 | return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); |
1da177e4 | 4843 | } |
1da177e4 LT |
4844 | EXPORT_SYMBOL(interruptible_sleep_on_timeout); |
4845 | ||
0fec171c | 4846 | void __sched sleep_on(wait_queue_head_t *q) |
1da177e4 | 4847 | { |
8cbbe86d | 4848 | sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); |
1da177e4 | 4849 | } |
1da177e4 LT |
4850 | EXPORT_SYMBOL(sleep_on); |
4851 | ||
0fec171c | 4852 | long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) |
1da177e4 | 4853 | { |
8cbbe86d | 4854 | return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); |
1da177e4 | 4855 | } |
1da177e4 LT |
4856 | EXPORT_SYMBOL(sleep_on_timeout); |
4857 | ||
b29739f9 IM |
4858 | #ifdef CONFIG_RT_MUTEXES |
4859 | ||
4860 | /* | |
4861 | * rt_mutex_setprio - set the current priority of a task | |
4862 | * @p: task | |
4863 | * @prio: prio value (kernel-internal form) | |
4864 | * | |
4865 | * This function changes the 'effective' priority of a task. It does | |
4866 | * not touch ->normal_prio like __setscheduler(). | |
4867 | * | |
4868 | * Used by the rt_mutex code to implement priority inheritance logic. | |
4869 | */ | |
36c8b586 | 4870 | void rt_mutex_setprio(struct task_struct *p, int prio) |
b29739f9 IM |
4871 | { |
4872 | unsigned long flags; | |
83b699ed | 4873 | int oldprio, on_rq, running; |
70b97a7f | 4874 | struct rq *rq; |
cb469845 | 4875 | const struct sched_class *prev_class = p->sched_class; |
b29739f9 IM |
4876 | |
4877 | BUG_ON(prio < 0 || prio > MAX_PRIO); | |
4878 | ||
4879 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 4880 | update_rq_clock(rq); |
b29739f9 | 4881 | |
d5f9f942 | 4882 | oldprio = p->prio; |
dd41f596 | 4883 | on_rq = p->se.on_rq; |
051a1d1a | 4884 | running = task_current(rq, p); |
0e1f3483 | 4885 | if (on_rq) |
69be72c1 | 4886 | dequeue_task(rq, p, 0); |
0e1f3483 HS |
4887 | if (running) |
4888 | p->sched_class->put_prev_task(rq, p); | |
dd41f596 IM |
4889 | |
4890 | if (rt_prio(prio)) | |
4891 | p->sched_class = &rt_sched_class; | |
4892 | else | |
4893 | p->sched_class = &fair_sched_class; | |
4894 | ||
b29739f9 IM |
4895 | p->prio = prio; |
4896 | ||
0e1f3483 HS |
4897 | if (running) |
4898 | p->sched_class->set_curr_task(rq); | |
dd41f596 | 4899 | if (on_rq) { |
8159f87e | 4900 | enqueue_task(rq, p, 0); |
cb469845 SR |
4901 | |
4902 | check_class_changed(rq, p, prev_class, oldprio, running); | |
b29739f9 IM |
4903 | } |
4904 | task_rq_unlock(rq, &flags); | |
4905 | } | |
4906 | ||
4907 | #endif | |
4908 | ||
36c8b586 | 4909 | void set_user_nice(struct task_struct *p, long nice) |
1da177e4 | 4910 | { |
dd41f596 | 4911 | int old_prio, delta, on_rq; |
1da177e4 | 4912 | unsigned long flags; |
70b97a7f | 4913 | struct rq *rq; |
1da177e4 LT |
4914 | |
4915 | if (TASK_NICE(p) == nice || nice < -20 || nice > 19) | |
4916 | return; | |
4917 | /* | |
4918 | * We have to be careful, if called from sys_setpriority(), | |
4919 | * the task might be in the middle of scheduling on another CPU. | |
4920 | */ | |
4921 | rq = task_rq_lock(p, &flags); | |
a8e504d2 | 4922 | update_rq_clock(rq); |
1da177e4 LT |
4923 | /* |
4924 | * The RT priorities are set via sched_setscheduler(), but we still | |
4925 | * allow the 'normal' nice value to be set - but as expected | |
4926 | * it wont have any effect on scheduling until the task is | |
dd41f596 | 4927 | * SCHED_FIFO/SCHED_RR: |
1da177e4 | 4928 | */ |
e05606d3 | 4929 | if (task_has_rt_policy(p)) { |
1da177e4 LT |
4930 | p->static_prio = NICE_TO_PRIO(nice); |
4931 | goto out_unlock; | |
4932 | } | |
dd41f596 | 4933 | on_rq = p->se.on_rq; |
18d95a28 | 4934 | if (on_rq) |
69be72c1 | 4935 | dequeue_task(rq, p, 0); |
1da177e4 | 4936 | |
1da177e4 | 4937 | p->static_prio = NICE_TO_PRIO(nice); |
2dd73a4f | 4938 | set_load_weight(p); |
b29739f9 IM |
4939 | old_prio = p->prio; |
4940 | p->prio = effective_prio(p); | |
4941 | delta = p->prio - old_prio; | |
1da177e4 | 4942 | |
dd41f596 | 4943 | if (on_rq) { |
8159f87e | 4944 | enqueue_task(rq, p, 0); |
1da177e4 | 4945 | /* |
d5f9f942 AM |
4946 | * If the task increased its priority or is running and |
4947 | * lowered its priority, then reschedule its CPU: | |
1da177e4 | 4948 | */ |
d5f9f942 | 4949 | if (delta < 0 || (delta > 0 && task_running(rq, p))) |
1da177e4 LT |
4950 | resched_task(rq->curr); |
4951 | } | |
4952 | out_unlock: | |
4953 | task_rq_unlock(rq, &flags); | |
4954 | } | |
1da177e4 LT |
4955 | EXPORT_SYMBOL(set_user_nice); |
4956 | ||
e43379f1 MM |
4957 | /* |
4958 | * can_nice - check if a task can reduce its nice value | |
4959 | * @p: task | |
4960 | * @nice: nice value | |
4961 | */ | |
36c8b586 | 4962 | int can_nice(const struct task_struct *p, const int nice) |
e43379f1 | 4963 | { |
024f4747 MM |
4964 | /* convert nice value [19,-20] to rlimit style value [1,40] */ |
4965 | int nice_rlim = 20 - nice; | |
48f24c4d | 4966 | |
e43379f1 MM |
4967 | return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur || |
4968 | capable(CAP_SYS_NICE)); | |
4969 | } | |
4970 | ||
1da177e4 LT |
4971 | #ifdef __ARCH_WANT_SYS_NICE |
4972 | ||
4973 | /* | |
4974 | * sys_nice - change the priority of the current process. | |
4975 | * @increment: priority increment | |
4976 | * | |
4977 | * sys_setpriority is a more generic, but much slower function that | |
4978 | * does similar things. | |
4979 | */ | |
4980 | asmlinkage long sys_nice(int increment) | |
4981 | { | |
48f24c4d | 4982 | long nice, retval; |
1da177e4 LT |
4983 | |
4984 | /* | |
4985 | * Setpriority might change our priority at the same moment. | |
4986 | * We don't have to worry. Conceptually one call occurs first | |
4987 | * and we have a single winner. | |
4988 | */ | |
e43379f1 MM |
4989 | if (increment < -40) |
4990 | increment = -40; | |
1da177e4 LT |
4991 | if (increment > 40) |
4992 | increment = 40; | |
4993 | ||
4994 | nice = PRIO_TO_NICE(current->static_prio) + increment; | |
4995 | if (nice < -20) | |
4996 | nice = -20; | |
4997 | if (nice > 19) | |
4998 | nice = 19; | |
4999 | ||
e43379f1 MM |
5000 | if (increment < 0 && !can_nice(current, nice)) |
5001 | return -EPERM; | |
5002 | ||
1da177e4 LT |
5003 | retval = security_task_setnice(current, nice); |
5004 | if (retval) | |
5005 | return retval; | |
5006 | ||
5007 | set_user_nice(current, nice); | |
5008 | return 0; | |
5009 | } | |
5010 | ||
5011 | #endif | |
5012 | ||
5013 | /** | |
5014 | * task_prio - return the priority value of a given task. | |
5015 | * @p: the task in question. | |
5016 | * | |
5017 | * This is the priority value as seen by users in /proc. | |
5018 | * RT tasks are offset by -200. Normal tasks are centered | |
5019 | * around 0, value goes from -16 to +15. | |
5020 | */ | |
36c8b586 | 5021 | int task_prio(const struct task_struct *p) |
1da177e4 LT |
5022 | { |
5023 | return p->prio - MAX_RT_PRIO; | |
5024 | } | |
5025 | ||
5026 | /** | |
5027 | * task_nice - return the nice value of a given task. | |
5028 | * @p: the task in question. | |
5029 | */ | |
36c8b586 | 5030 | int task_nice(const struct task_struct *p) |
1da177e4 LT |
5031 | { |
5032 | return TASK_NICE(p); | |
5033 | } | |
150d8bed | 5034 | EXPORT_SYMBOL(task_nice); |
1da177e4 LT |
5035 | |
5036 | /** | |
5037 | * idle_cpu - is a given cpu idle currently? | |
5038 | * @cpu: the processor in question. | |
5039 | */ | |
5040 | int idle_cpu(int cpu) | |
5041 | { | |
5042 | return cpu_curr(cpu) == cpu_rq(cpu)->idle; | |
5043 | } | |
5044 | ||
1da177e4 LT |
5045 | /** |
5046 | * idle_task - return the idle task for a given cpu. | |
5047 | * @cpu: the processor in question. | |
5048 | */ | |
36c8b586 | 5049 | struct task_struct *idle_task(int cpu) |
1da177e4 LT |
5050 | { |
5051 | return cpu_rq(cpu)->idle; | |
5052 | } | |
5053 | ||
5054 | /** | |
5055 | * find_process_by_pid - find a process with a matching PID value. | |
5056 | * @pid: the pid in question. | |
5057 | */ | |
a9957449 | 5058 | static struct task_struct *find_process_by_pid(pid_t pid) |
1da177e4 | 5059 | { |
228ebcbe | 5060 | return pid ? find_task_by_vpid(pid) : current; |
1da177e4 LT |
5061 | } |
5062 | ||
5063 | /* Actually do priority change: must hold rq lock. */ | |
dd41f596 IM |
5064 | static void |
5065 | __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) | |
1da177e4 | 5066 | { |
dd41f596 | 5067 | BUG_ON(p->se.on_rq); |
48f24c4d | 5068 | |
1da177e4 | 5069 | p->policy = policy; |
dd41f596 IM |
5070 | switch (p->policy) { |
5071 | case SCHED_NORMAL: | |
5072 | case SCHED_BATCH: | |
5073 | case SCHED_IDLE: | |
5074 | p->sched_class = &fair_sched_class; | |
5075 | break; | |
5076 | case SCHED_FIFO: | |
5077 | case SCHED_RR: | |
5078 | p->sched_class = &rt_sched_class; | |
5079 | break; | |
5080 | } | |
5081 | ||
1da177e4 | 5082 | p->rt_priority = prio; |
b29739f9 IM |
5083 | p->normal_prio = normal_prio(p); |
5084 | /* we are holding p->pi_lock already */ | |
5085 | p->prio = rt_mutex_getprio(p); | |
2dd73a4f | 5086 | set_load_weight(p); |
1da177e4 LT |
5087 | } |
5088 | ||
5089 | /** | |
72fd4a35 | 5090 | * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. |
1da177e4 LT |
5091 | * @p: the task in question. |
5092 | * @policy: new policy. | |
5093 | * @param: structure containing the new RT priority. | |
5fe1d75f | 5094 | * |
72fd4a35 | 5095 | * NOTE that the task may be already dead. |
1da177e4 | 5096 | */ |
95cdf3b7 IM |
5097 | int sched_setscheduler(struct task_struct *p, int policy, |
5098 | struct sched_param *param) | |
1da177e4 | 5099 | { |
83b699ed | 5100 | int retval, oldprio, oldpolicy = -1, on_rq, running; |
1da177e4 | 5101 | unsigned long flags; |
cb469845 | 5102 | const struct sched_class *prev_class = p->sched_class; |
70b97a7f | 5103 | struct rq *rq; |
1da177e4 | 5104 | |
66e5393a SR |
5105 | /* may grab non-irq protected spin_locks */ |
5106 | BUG_ON(in_interrupt()); | |
1da177e4 LT |
5107 | recheck: |
5108 | /* double check policy once rq lock held */ | |
5109 | if (policy < 0) | |
5110 | policy = oldpolicy = p->policy; | |
5111 | else if (policy != SCHED_FIFO && policy != SCHED_RR && | |
dd41f596 IM |
5112 | policy != SCHED_NORMAL && policy != SCHED_BATCH && |
5113 | policy != SCHED_IDLE) | |
b0a9499c | 5114 | return -EINVAL; |
1da177e4 LT |
5115 | /* |
5116 | * Valid priorities for SCHED_FIFO and SCHED_RR are | |
dd41f596 IM |
5117 | * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, |
5118 | * SCHED_BATCH and SCHED_IDLE is 0. | |
1da177e4 LT |
5119 | */ |
5120 | if (param->sched_priority < 0 || | |
95cdf3b7 | 5121 | (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || |
d46523ea | 5122 | (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) |
1da177e4 | 5123 | return -EINVAL; |
e05606d3 | 5124 | if (rt_policy(policy) != (param->sched_priority != 0)) |
1da177e4 LT |
5125 | return -EINVAL; |
5126 | ||
37e4ab3f OC |
5127 | /* |
5128 | * Allow unprivileged RT tasks to decrease priority: | |
5129 | */ | |
5130 | if (!capable(CAP_SYS_NICE)) { | |
e05606d3 | 5131 | if (rt_policy(policy)) { |
8dc3e909 | 5132 | unsigned long rlim_rtprio; |
8dc3e909 ON |
5133 | |
5134 | if (!lock_task_sighand(p, &flags)) | |
5135 | return -ESRCH; | |
5136 | rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur; | |
5137 | unlock_task_sighand(p, &flags); | |
5138 | ||
5139 | /* can't set/change the rt policy */ | |
5140 | if (policy != p->policy && !rlim_rtprio) | |
5141 | return -EPERM; | |
5142 | ||
5143 | /* can't increase priority */ | |
5144 | if (param->sched_priority > p->rt_priority && | |
5145 | param->sched_priority > rlim_rtprio) | |
5146 | return -EPERM; | |
5147 | } | |
dd41f596 IM |
5148 | /* |
5149 | * Like positive nice levels, dont allow tasks to | |
5150 | * move out of SCHED_IDLE either: | |
5151 | */ | |
5152 | if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) | |
5153 | return -EPERM; | |
5fe1d75f | 5154 | |
37e4ab3f OC |
5155 | /* can't change other user's priorities */ |
5156 | if ((current->euid != p->euid) && | |
5157 | (current->euid != p->uid)) | |
5158 | return -EPERM; | |
5159 | } | |
1da177e4 | 5160 | |
b68aa230 PZ |
5161 | #ifdef CONFIG_RT_GROUP_SCHED |
5162 | /* | |
5163 | * Do not allow realtime tasks into groups that have no runtime | |
5164 | * assigned. | |
5165 | */ | |
d0b27fa7 | 5166 | if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
5167 | return -EPERM; |
5168 | #endif | |
5169 | ||
1da177e4 LT |
5170 | retval = security_task_setscheduler(p, policy, param); |
5171 | if (retval) | |
5172 | return retval; | |
b29739f9 IM |
5173 | /* |
5174 | * make sure no PI-waiters arrive (or leave) while we are | |
5175 | * changing the priority of the task: | |
5176 | */ | |
5177 | spin_lock_irqsave(&p->pi_lock, flags); | |
1da177e4 LT |
5178 | /* |
5179 | * To be able to change p->policy safely, the apropriate | |
5180 | * runqueue lock must be held. | |
5181 | */ | |
b29739f9 | 5182 | rq = __task_rq_lock(p); |
1da177e4 LT |
5183 | /* recheck policy now with rq lock held */ |
5184 | if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { | |
5185 | policy = oldpolicy = -1; | |
b29739f9 IM |
5186 | __task_rq_unlock(rq); |
5187 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
1da177e4 LT |
5188 | goto recheck; |
5189 | } | |
2daa3577 | 5190 | update_rq_clock(rq); |
dd41f596 | 5191 | on_rq = p->se.on_rq; |
051a1d1a | 5192 | running = task_current(rq, p); |
0e1f3483 | 5193 | if (on_rq) |
2e1cb74a | 5194 | deactivate_task(rq, p, 0); |
0e1f3483 HS |
5195 | if (running) |
5196 | p->sched_class->put_prev_task(rq, p); | |
f6b53205 | 5197 | |
1da177e4 | 5198 | oldprio = p->prio; |
dd41f596 | 5199 | __setscheduler(rq, p, policy, param->sched_priority); |
f6b53205 | 5200 | |
0e1f3483 HS |
5201 | if (running) |
5202 | p->sched_class->set_curr_task(rq); | |
dd41f596 IM |
5203 | if (on_rq) { |
5204 | activate_task(rq, p, 0); | |
cb469845 SR |
5205 | |
5206 | check_class_changed(rq, p, prev_class, oldprio, running); | |
1da177e4 | 5207 | } |
b29739f9 IM |
5208 | __task_rq_unlock(rq); |
5209 | spin_unlock_irqrestore(&p->pi_lock, flags); | |
5210 | ||
95e02ca9 TG |
5211 | rt_mutex_adjust_pi(p); |
5212 | ||
1da177e4 LT |
5213 | return 0; |
5214 | } | |
5215 | EXPORT_SYMBOL_GPL(sched_setscheduler); | |
5216 | ||
95cdf3b7 IM |
5217 | static int |
5218 | do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5219 | { |
1da177e4 LT |
5220 | struct sched_param lparam; |
5221 | struct task_struct *p; | |
36c8b586 | 5222 | int retval; |
1da177e4 LT |
5223 | |
5224 | if (!param || pid < 0) | |
5225 | return -EINVAL; | |
5226 | if (copy_from_user(&lparam, param, sizeof(struct sched_param))) | |
5227 | return -EFAULT; | |
5fe1d75f ON |
5228 | |
5229 | rcu_read_lock(); | |
5230 | retval = -ESRCH; | |
1da177e4 | 5231 | p = find_process_by_pid(pid); |
5fe1d75f ON |
5232 | if (p != NULL) |
5233 | retval = sched_setscheduler(p, policy, &lparam); | |
5234 | rcu_read_unlock(); | |
36c8b586 | 5235 | |
1da177e4 LT |
5236 | return retval; |
5237 | } | |
5238 | ||
5239 | /** | |
5240 | * sys_sched_setscheduler - set/change the scheduler policy and RT priority | |
5241 | * @pid: the pid in question. | |
5242 | * @policy: new policy. | |
5243 | * @param: structure containing the new RT priority. | |
5244 | */ | |
41a2d6cf IM |
5245 | asmlinkage long |
5246 | sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) | |
1da177e4 | 5247 | { |
c21761f1 JB |
5248 | /* negative values for policy are not valid */ |
5249 | if (policy < 0) | |
5250 | return -EINVAL; | |
5251 | ||
1da177e4 LT |
5252 | return do_sched_setscheduler(pid, policy, param); |
5253 | } | |
5254 | ||
5255 | /** | |
5256 | * sys_sched_setparam - set/change the RT priority of a thread | |
5257 | * @pid: the pid in question. | |
5258 | * @param: structure containing the new RT priority. | |
5259 | */ | |
5260 | asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param) | |
5261 | { | |
5262 | return do_sched_setscheduler(pid, -1, param); | |
5263 | } | |
5264 | ||
5265 | /** | |
5266 | * sys_sched_getscheduler - get the policy (scheduling class) of a thread | |
5267 | * @pid: the pid in question. | |
5268 | */ | |
5269 | asmlinkage long sys_sched_getscheduler(pid_t pid) | |
5270 | { | |
36c8b586 | 5271 | struct task_struct *p; |
3a5c359a | 5272 | int retval; |
1da177e4 LT |
5273 | |
5274 | if (pid < 0) | |
3a5c359a | 5275 | return -EINVAL; |
1da177e4 LT |
5276 | |
5277 | retval = -ESRCH; | |
5278 | read_lock(&tasklist_lock); | |
5279 | p = find_process_by_pid(pid); | |
5280 | if (p) { | |
5281 | retval = security_task_getscheduler(p); | |
5282 | if (!retval) | |
5283 | retval = p->policy; | |
5284 | } | |
5285 | read_unlock(&tasklist_lock); | |
1da177e4 LT |
5286 | return retval; |
5287 | } | |
5288 | ||
5289 | /** | |
5290 | * sys_sched_getscheduler - get the RT priority of a thread | |
5291 | * @pid: the pid in question. | |
5292 | * @param: structure containing the RT priority. | |
5293 | */ | |
5294 | asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param) | |
5295 | { | |
5296 | struct sched_param lp; | |
36c8b586 | 5297 | struct task_struct *p; |
3a5c359a | 5298 | int retval; |
1da177e4 LT |
5299 | |
5300 | if (!param || pid < 0) | |
3a5c359a | 5301 | return -EINVAL; |
1da177e4 LT |
5302 | |
5303 | read_lock(&tasklist_lock); | |
5304 | p = find_process_by_pid(pid); | |
5305 | retval = -ESRCH; | |
5306 | if (!p) | |
5307 | goto out_unlock; | |
5308 | ||
5309 | retval = security_task_getscheduler(p); | |
5310 | if (retval) | |
5311 | goto out_unlock; | |
5312 | ||
5313 | lp.sched_priority = p->rt_priority; | |
5314 | read_unlock(&tasklist_lock); | |
5315 | ||
5316 | /* | |
5317 | * This one might sleep, we cannot do it with a spinlock held ... | |
5318 | */ | |
5319 | retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; | |
5320 | ||
1da177e4 LT |
5321 | return retval; |
5322 | ||
5323 | out_unlock: | |
5324 | read_unlock(&tasklist_lock); | |
5325 | return retval; | |
5326 | } | |
5327 | ||
b53e921b | 5328 | long sched_setaffinity(pid_t pid, const cpumask_t *in_mask) |
1da177e4 | 5329 | { |
1da177e4 | 5330 | cpumask_t cpus_allowed; |
b53e921b | 5331 | cpumask_t new_mask = *in_mask; |
36c8b586 IM |
5332 | struct task_struct *p; |
5333 | int retval; | |
1da177e4 | 5334 | |
95402b38 | 5335 | get_online_cpus(); |
1da177e4 LT |
5336 | read_lock(&tasklist_lock); |
5337 | ||
5338 | p = find_process_by_pid(pid); | |
5339 | if (!p) { | |
5340 | read_unlock(&tasklist_lock); | |
95402b38 | 5341 | put_online_cpus(); |
1da177e4 LT |
5342 | return -ESRCH; |
5343 | } | |
5344 | ||
5345 | /* | |
5346 | * It is not safe to call set_cpus_allowed with the | |
41a2d6cf | 5347 | * tasklist_lock held. We will bump the task_struct's |
1da177e4 LT |
5348 | * usage count and then drop tasklist_lock. |
5349 | */ | |
5350 | get_task_struct(p); | |
5351 | read_unlock(&tasklist_lock); | |
5352 | ||
5353 | retval = -EPERM; | |
5354 | if ((current->euid != p->euid) && (current->euid != p->uid) && | |
5355 | !capable(CAP_SYS_NICE)) | |
5356 | goto out_unlock; | |
5357 | ||
e7834f8f DQ |
5358 | retval = security_task_setscheduler(p, 0, NULL); |
5359 | if (retval) | |
5360 | goto out_unlock; | |
5361 | ||
f9a86fcb | 5362 | cpuset_cpus_allowed(p, &cpus_allowed); |
1da177e4 | 5363 | cpus_and(new_mask, new_mask, cpus_allowed); |
8707d8b8 | 5364 | again: |
7c16ec58 | 5365 | retval = set_cpus_allowed_ptr(p, &new_mask); |
1da177e4 | 5366 | |
8707d8b8 | 5367 | if (!retval) { |
f9a86fcb | 5368 | cpuset_cpus_allowed(p, &cpus_allowed); |
8707d8b8 PM |
5369 | if (!cpus_subset(new_mask, cpus_allowed)) { |
5370 | /* | |
5371 | * We must have raced with a concurrent cpuset | |
5372 | * update. Just reset the cpus_allowed to the | |
5373 | * cpuset's cpus_allowed | |
5374 | */ | |
5375 | new_mask = cpus_allowed; | |
5376 | goto again; | |
5377 | } | |
5378 | } | |
1da177e4 LT |
5379 | out_unlock: |
5380 | put_task_struct(p); | |
95402b38 | 5381 | put_online_cpus(); |
1da177e4 LT |
5382 | return retval; |
5383 | } | |
5384 | ||
5385 | static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, | |
5386 | cpumask_t *new_mask) | |
5387 | { | |
5388 | if (len < sizeof(cpumask_t)) { | |
5389 | memset(new_mask, 0, sizeof(cpumask_t)); | |
5390 | } else if (len > sizeof(cpumask_t)) { | |
5391 | len = sizeof(cpumask_t); | |
5392 | } | |
5393 | return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; | |
5394 | } | |
5395 | ||
5396 | /** | |
5397 | * sys_sched_setaffinity - set the cpu affinity of a process | |
5398 | * @pid: pid of the process | |
5399 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5400 | * @user_mask_ptr: user-space pointer to the new cpu mask | |
5401 | */ | |
5402 | asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len, | |
5403 | unsigned long __user *user_mask_ptr) | |
5404 | { | |
5405 | cpumask_t new_mask; | |
5406 | int retval; | |
5407 | ||
5408 | retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask); | |
5409 | if (retval) | |
5410 | return retval; | |
5411 | ||
b53e921b | 5412 | return sched_setaffinity(pid, &new_mask); |
1da177e4 LT |
5413 | } |
5414 | ||
5415 | /* | |
5416 | * Represents all cpu's present in the system | |
5417 | * In systems capable of hotplug, this map could dynamically grow | |
5418 | * as new cpu's are detected in the system via any platform specific | |
5419 | * method, such as ACPI for e.g. | |
5420 | */ | |
5421 | ||
4cef0c61 | 5422 | cpumask_t cpu_present_map __read_mostly; |
1da177e4 LT |
5423 | EXPORT_SYMBOL(cpu_present_map); |
5424 | ||
5425 | #ifndef CONFIG_SMP | |
4cef0c61 | 5426 | cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 GB |
5427 | EXPORT_SYMBOL(cpu_online_map); |
5428 | ||
4cef0c61 | 5429 | cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL; |
e16b38f7 | 5430 | EXPORT_SYMBOL(cpu_possible_map); |
1da177e4 LT |
5431 | #endif |
5432 | ||
5433 | long sched_getaffinity(pid_t pid, cpumask_t *mask) | |
5434 | { | |
36c8b586 | 5435 | struct task_struct *p; |
1da177e4 | 5436 | int retval; |
1da177e4 | 5437 | |
95402b38 | 5438 | get_online_cpus(); |
1da177e4 LT |
5439 | read_lock(&tasklist_lock); |
5440 | ||
5441 | retval = -ESRCH; | |
5442 | p = find_process_by_pid(pid); | |
5443 | if (!p) | |
5444 | goto out_unlock; | |
5445 | ||
e7834f8f DQ |
5446 | retval = security_task_getscheduler(p); |
5447 | if (retval) | |
5448 | goto out_unlock; | |
5449 | ||
2f7016d9 | 5450 | cpus_and(*mask, p->cpus_allowed, cpu_online_map); |
1da177e4 LT |
5451 | |
5452 | out_unlock: | |
5453 | read_unlock(&tasklist_lock); | |
95402b38 | 5454 | put_online_cpus(); |
1da177e4 | 5455 | |
9531b62f | 5456 | return retval; |
1da177e4 LT |
5457 | } |
5458 | ||
5459 | /** | |
5460 | * sys_sched_getaffinity - get the cpu affinity of a process | |
5461 | * @pid: pid of the process | |
5462 | * @len: length in bytes of the bitmask pointed to by user_mask_ptr | |
5463 | * @user_mask_ptr: user-space pointer to hold the current cpu mask | |
5464 | */ | |
5465 | asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len, | |
5466 | unsigned long __user *user_mask_ptr) | |
5467 | { | |
5468 | int ret; | |
5469 | cpumask_t mask; | |
5470 | ||
5471 | if (len < sizeof(cpumask_t)) | |
5472 | return -EINVAL; | |
5473 | ||
5474 | ret = sched_getaffinity(pid, &mask); | |
5475 | if (ret < 0) | |
5476 | return ret; | |
5477 | ||
5478 | if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t))) | |
5479 | return -EFAULT; | |
5480 | ||
5481 | return sizeof(cpumask_t); | |
5482 | } | |
5483 | ||
5484 | /** | |
5485 | * sys_sched_yield - yield the current processor to other threads. | |
5486 | * | |
dd41f596 IM |
5487 | * This function yields the current CPU to other tasks. If there are no |
5488 | * other threads running on this CPU then this function will return. | |
1da177e4 LT |
5489 | */ |
5490 | asmlinkage long sys_sched_yield(void) | |
5491 | { | |
70b97a7f | 5492 | struct rq *rq = this_rq_lock(); |
1da177e4 | 5493 | |
2d72376b | 5494 | schedstat_inc(rq, yld_count); |
4530d7ab | 5495 | current->sched_class->yield_task(rq); |
1da177e4 LT |
5496 | |
5497 | /* | |
5498 | * Since we are going to call schedule() anyway, there's | |
5499 | * no need to preempt or enable interrupts: | |
5500 | */ | |
5501 | __release(rq->lock); | |
8a25d5de | 5502 | spin_release(&rq->lock.dep_map, 1, _THIS_IP_); |
1da177e4 LT |
5503 | _raw_spin_unlock(&rq->lock); |
5504 | preempt_enable_no_resched(); | |
5505 | ||
5506 | schedule(); | |
5507 | ||
5508 | return 0; | |
5509 | } | |
5510 | ||
e7b38404 | 5511 | static void __cond_resched(void) |
1da177e4 | 5512 | { |
8e0a43d8 IM |
5513 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP |
5514 | __might_sleep(__FILE__, __LINE__); | |
5515 | #endif | |
5bbcfd90 IM |
5516 | /* |
5517 | * The BKS might be reacquired before we have dropped | |
5518 | * PREEMPT_ACTIVE, which could trigger a second | |
5519 | * cond_resched() call. | |
5520 | */ | |
1da177e4 LT |
5521 | do { |
5522 | add_preempt_count(PREEMPT_ACTIVE); | |
5523 | schedule(); | |
5524 | sub_preempt_count(PREEMPT_ACTIVE); | |
5525 | } while (need_resched()); | |
5526 | } | |
5527 | ||
02b67cc3 | 5528 | int __sched _cond_resched(void) |
1da177e4 | 5529 | { |
9414232f IM |
5530 | if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) && |
5531 | system_state == SYSTEM_RUNNING) { | |
1da177e4 LT |
5532 | __cond_resched(); |
5533 | return 1; | |
5534 | } | |
5535 | return 0; | |
5536 | } | |
02b67cc3 | 5537 | EXPORT_SYMBOL(_cond_resched); |
1da177e4 LT |
5538 | |
5539 | /* | |
5540 | * cond_resched_lock() - if a reschedule is pending, drop the given lock, | |
5541 | * call schedule, and on return reacquire the lock. | |
5542 | * | |
41a2d6cf | 5543 | * This works OK both with and without CONFIG_PREEMPT. We do strange low-level |
1da177e4 LT |
5544 | * operations here to prevent schedule() from being called twice (once via |
5545 | * spin_unlock(), once by hand). | |
5546 | */ | |
95cdf3b7 | 5547 | int cond_resched_lock(spinlock_t *lock) |
1da177e4 | 5548 | { |
95c354fe | 5549 | int resched = need_resched() && system_state == SYSTEM_RUNNING; |
6df3cecb JK |
5550 | int ret = 0; |
5551 | ||
95c354fe | 5552 | if (spin_needbreak(lock) || resched) { |
1da177e4 | 5553 | spin_unlock(lock); |
95c354fe NP |
5554 | if (resched && need_resched()) |
5555 | __cond_resched(); | |
5556 | else | |
5557 | cpu_relax(); | |
6df3cecb | 5558 | ret = 1; |
1da177e4 | 5559 | spin_lock(lock); |
1da177e4 | 5560 | } |
6df3cecb | 5561 | return ret; |
1da177e4 | 5562 | } |
1da177e4 LT |
5563 | EXPORT_SYMBOL(cond_resched_lock); |
5564 | ||
5565 | int __sched cond_resched_softirq(void) | |
5566 | { | |
5567 | BUG_ON(!in_softirq()); | |
5568 | ||
9414232f | 5569 | if (need_resched() && system_state == SYSTEM_RUNNING) { |
98d82567 | 5570 | local_bh_enable(); |
1da177e4 LT |
5571 | __cond_resched(); |
5572 | local_bh_disable(); | |
5573 | return 1; | |
5574 | } | |
5575 | return 0; | |
5576 | } | |
1da177e4 LT |
5577 | EXPORT_SYMBOL(cond_resched_softirq); |
5578 | ||
1da177e4 LT |
5579 | /** |
5580 | * yield - yield the current processor to other threads. | |
5581 | * | |
72fd4a35 | 5582 | * This is a shortcut for kernel-space yielding - it marks the |
1da177e4 LT |
5583 | * thread runnable and calls sys_sched_yield(). |
5584 | */ | |
5585 | void __sched yield(void) | |
5586 | { | |
5587 | set_current_state(TASK_RUNNING); | |
5588 | sys_sched_yield(); | |
5589 | } | |
1da177e4 LT |
5590 | EXPORT_SYMBOL(yield); |
5591 | ||
5592 | /* | |
41a2d6cf | 5593 | * This task is about to go to sleep on IO. Increment rq->nr_iowait so |
1da177e4 LT |
5594 | * that process accounting knows that this is a task in IO wait state. |
5595 | * | |
5596 | * But don't do that if it is a deliberate, throttling IO wait (this task | |
5597 | * has set its backing_dev_info: the queue against which it should throttle) | |
5598 | */ | |
5599 | void __sched io_schedule(void) | |
5600 | { | |
70b97a7f | 5601 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 | 5602 | |
0ff92245 | 5603 | delayacct_blkio_start(); |
1da177e4 LT |
5604 | atomic_inc(&rq->nr_iowait); |
5605 | schedule(); | |
5606 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5607 | delayacct_blkio_end(); |
1da177e4 | 5608 | } |
1da177e4 LT |
5609 | EXPORT_SYMBOL(io_schedule); |
5610 | ||
5611 | long __sched io_schedule_timeout(long timeout) | |
5612 | { | |
70b97a7f | 5613 | struct rq *rq = &__raw_get_cpu_var(runqueues); |
1da177e4 LT |
5614 | long ret; |
5615 | ||
0ff92245 | 5616 | delayacct_blkio_start(); |
1da177e4 LT |
5617 | atomic_inc(&rq->nr_iowait); |
5618 | ret = schedule_timeout(timeout); | |
5619 | atomic_dec(&rq->nr_iowait); | |
0ff92245 | 5620 | delayacct_blkio_end(); |
1da177e4 LT |
5621 | return ret; |
5622 | } | |
5623 | ||
5624 | /** | |
5625 | * sys_sched_get_priority_max - return maximum RT priority. | |
5626 | * @policy: scheduling class. | |
5627 | * | |
5628 | * this syscall returns the maximum rt_priority that can be used | |
5629 | * by a given scheduling class. | |
5630 | */ | |
5631 | asmlinkage long sys_sched_get_priority_max(int policy) | |
5632 | { | |
5633 | int ret = -EINVAL; | |
5634 | ||
5635 | switch (policy) { | |
5636 | case SCHED_FIFO: | |
5637 | case SCHED_RR: | |
5638 | ret = MAX_USER_RT_PRIO-1; | |
5639 | break; | |
5640 | case SCHED_NORMAL: | |
b0a9499c | 5641 | case SCHED_BATCH: |
dd41f596 | 5642 | case SCHED_IDLE: |
1da177e4 LT |
5643 | ret = 0; |
5644 | break; | |
5645 | } | |
5646 | return ret; | |
5647 | } | |
5648 | ||
5649 | /** | |
5650 | * sys_sched_get_priority_min - return minimum RT priority. | |
5651 | * @policy: scheduling class. | |
5652 | * | |
5653 | * this syscall returns the minimum rt_priority that can be used | |
5654 | * by a given scheduling class. | |
5655 | */ | |
5656 | asmlinkage long sys_sched_get_priority_min(int policy) | |
5657 | { | |
5658 | int ret = -EINVAL; | |
5659 | ||
5660 | switch (policy) { | |
5661 | case SCHED_FIFO: | |
5662 | case SCHED_RR: | |
5663 | ret = 1; | |
5664 | break; | |
5665 | case SCHED_NORMAL: | |
b0a9499c | 5666 | case SCHED_BATCH: |
dd41f596 | 5667 | case SCHED_IDLE: |
1da177e4 LT |
5668 | ret = 0; |
5669 | } | |
5670 | return ret; | |
5671 | } | |
5672 | ||
5673 | /** | |
5674 | * sys_sched_rr_get_interval - return the default timeslice of a process. | |
5675 | * @pid: pid of the process. | |
5676 | * @interval: userspace pointer to the timeslice value. | |
5677 | * | |
5678 | * this syscall writes the default timeslice value of a given process | |
5679 | * into the user-space timespec buffer. A value of '0' means infinity. | |
5680 | */ | |
5681 | asmlinkage | |
5682 | long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval) | |
5683 | { | |
36c8b586 | 5684 | struct task_struct *p; |
a4ec24b4 | 5685 | unsigned int time_slice; |
3a5c359a | 5686 | int retval; |
1da177e4 | 5687 | struct timespec t; |
1da177e4 LT |
5688 | |
5689 | if (pid < 0) | |
3a5c359a | 5690 | return -EINVAL; |
1da177e4 LT |
5691 | |
5692 | retval = -ESRCH; | |
5693 | read_lock(&tasklist_lock); | |
5694 | p = find_process_by_pid(pid); | |
5695 | if (!p) | |
5696 | goto out_unlock; | |
5697 | ||
5698 | retval = security_task_getscheduler(p); | |
5699 | if (retval) | |
5700 | goto out_unlock; | |
5701 | ||
77034937 IM |
5702 | /* |
5703 | * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER | |
5704 | * tasks that are on an otherwise idle runqueue: | |
5705 | */ | |
5706 | time_slice = 0; | |
5707 | if (p->policy == SCHED_RR) { | |
a4ec24b4 | 5708 | time_slice = DEF_TIMESLICE; |
1868f958 | 5709 | } else if (p->policy != SCHED_FIFO) { |
a4ec24b4 DA |
5710 | struct sched_entity *se = &p->se; |
5711 | unsigned long flags; | |
5712 | struct rq *rq; | |
5713 | ||
5714 | rq = task_rq_lock(p, &flags); | |
77034937 IM |
5715 | if (rq->cfs.load.weight) |
5716 | time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se)); | |
a4ec24b4 DA |
5717 | task_rq_unlock(rq, &flags); |
5718 | } | |
1da177e4 | 5719 | read_unlock(&tasklist_lock); |
a4ec24b4 | 5720 | jiffies_to_timespec(time_slice, &t); |
1da177e4 | 5721 | retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; |
1da177e4 | 5722 | return retval; |
3a5c359a | 5723 | |
1da177e4 LT |
5724 | out_unlock: |
5725 | read_unlock(&tasklist_lock); | |
5726 | return retval; | |
5727 | } | |
5728 | ||
2ed6e34f | 5729 | static const char stat_nam[] = "RSDTtZX"; |
36c8b586 | 5730 | |
82a1fcb9 | 5731 | void sched_show_task(struct task_struct *p) |
1da177e4 | 5732 | { |
1da177e4 | 5733 | unsigned long free = 0; |
36c8b586 | 5734 | unsigned state; |
1da177e4 | 5735 | |
1da177e4 | 5736 | state = p->state ? __ffs(p->state) + 1 : 0; |
cc4ea795 | 5737 | printk(KERN_INFO "%-13.13s %c", p->comm, |
2ed6e34f | 5738 | state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); |
4bd77321 | 5739 | #if BITS_PER_LONG == 32 |
1da177e4 | 5740 | if (state == TASK_RUNNING) |
cc4ea795 | 5741 | printk(KERN_CONT " running "); |
1da177e4 | 5742 | else |
cc4ea795 | 5743 | printk(KERN_CONT " %08lx ", thread_saved_pc(p)); |
1da177e4 LT |
5744 | #else |
5745 | if (state == TASK_RUNNING) | |
cc4ea795 | 5746 | printk(KERN_CONT " running task "); |
1da177e4 | 5747 | else |
cc4ea795 | 5748 | printk(KERN_CONT " %016lx ", thread_saved_pc(p)); |
1da177e4 LT |
5749 | #endif |
5750 | #ifdef CONFIG_DEBUG_STACK_USAGE | |
5751 | { | |
10ebffde | 5752 | unsigned long *n = end_of_stack(p); |
1da177e4 LT |
5753 | while (!*n) |
5754 | n++; | |
10ebffde | 5755 | free = (unsigned long)n - (unsigned long)end_of_stack(p); |
1da177e4 LT |
5756 | } |
5757 | #endif | |
ba25f9dc | 5758 | printk(KERN_CONT "%5lu %5d %6d\n", free, |
fcfd50af | 5759 | task_pid_nr(p), task_pid_nr(p->real_parent)); |
1da177e4 | 5760 | |
5fb5e6de | 5761 | show_stack(p, NULL); |
1da177e4 LT |
5762 | } |
5763 | ||
e59e2ae2 | 5764 | void show_state_filter(unsigned long state_filter) |
1da177e4 | 5765 | { |
36c8b586 | 5766 | struct task_struct *g, *p; |
1da177e4 | 5767 | |
4bd77321 IM |
5768 | #if BITS_PER_LONG == 32 |
5769 | printk(KERN_INFO | |
5770 | " task PC stack pid father\n"); | |
1da177e4 | 5771 | #else |
4bd77321 IM |
5772 | printk(KERN_INFO |
5773 | " task PC stack pid father\n"); | |
1da177e4 LT |
5774 | #endif |
5775 | read_lock(&tasklist_lock); | |
5776 | do_each_thread(g, p) { | |
5777 | /* | |
5778 | * reset the NMI-timeout, listing all files on a slow | |
5779 | * console might take alot of time: | |
5780 | */ | |
5781 | touch_nmi_watchdog(); | |
39bc89fd | 5782 | if (!state_filter || (p->state & state_filter)) |
82a1fcb9 | 5783 | sched_show_task(p); |
1da177e4 LT |
5784 | } while_each_thread(g, p); |
5785 | ||
04c9167f JF |
5786 | touch_all_softlockup_watchdogs(); |
5787 | ||
dd41f596 IM |
5788 | #ifdef CONFIG_SCHED_DEBUG |
5789 | sysrq_sched_debug_show(); | |
5790 | #endif | |
1da177e4 | 5791 | read_unlock(&tasklist_lock); |
e59e2ae2 IM |
5792 | /* |
5793 | * Only show locks if all tasks are dumped: | |
5794 | */ | |
5795 | if (state_filter == -1) | |
5796 | debug_show_all_locks(); | |
1da177e4 LT |
5797 | } |
5798 | ||
1df21055 IM |
5799 | void __cpuinit init_idle_bootup_task(struct task_struct *idle) |
5800 | { | |
dd41f596 | 5801 | idle->sched_class = &idle_sched_class; |
1df21055 IM |
5802 | } |
5803 | ||
f340c0d1 IM |
5804 | /** |
5805 | * init_idle - set up an idle thread for a given CPU | |
5806 | * @idle: task in question | |
5807 | * @cpu: cpu the idle task belongs to | |
5808 | * | |
5809 | * NOTE: this function does not set the idle thread's NEED_RESCHED | |
5810 | * flag, to make booting more robust. | |
5811 | */ | |
5c1e1767 | 5812 | void __cpuinit init_idle(struct task_struct *idle, int cpu) |
1da177e4 | 5813 | { |
70b97a7f | 5814 | struct rq *rq = cpu_rq(cpu); |
1da177e4 LT |
5815 | unsigned long flags; |
5816 | ||
dd41f596 IM |
5817 | __sched_fork(idle); |
5818 | idle->se.exec_start = sched_clock(); | |
5819 | ||
b29739f9 | 5820 | idle->prio = idle->normal_prio = MAX_PRIO; |
1da177e4 | 5821 | idle->cpus_allowed = cpumask_of_cpu(cpu); |
dd41f596 | 5822 | __set_task_cpu(idle, cpu); |
1da177e4 LT |
5823 | |
5824 | spin_lock_irqsave(&rq->lock, flags); | |
5825 | rq->curr = rq->idle = idle; | |
4866cde0 NP |
5826 | #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW) |
5827 | idle->oncpu = 1; | |
5828 | #endif | |
1da177e4 LT |
5829 | spin_unlock_irqrestore(&rq->lock, flags); |
5830 | ||
5831 | /* Set the preempt count _outside_ the spinlocks! */ | |
8e3e076c LT |
5832 | #if defined(CONFIG_PREEMPT) |
5833 | task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0); | |
5834 | #else | |
a1261f54 | 5835 | task_thread_info(idle)->preempt_count = 0; |
8e3e076c | 5836 | #endif |
dd41f596 IM |
5837 | /* |
5838 | * The idle tasks have their own, simple scheduling class: | |
5839 | */ | |
5840 | idle->sched_class = &idle_sched_class; | |
1da177e4 LT |
5841 | } |
5842 | ||
5843 | /* | |
5844 | * In a system that switches off the HZ timer nohz_cpu_mask | |
5845 | * indicates which cpus entered this state. This is used | |
5846 | * in the rcu update to wait only for active cpus. For system | |
5847 | * which do not switch off the HZ timer nohz_cpu_mask should | |
5848 | * always be CPU_MASK_NONE. | |
5849 | */ | |
5850 | cpumask_t nohz_cpu_mask = CPU_MASK_NONE; | |
5851 | ||
19978ca6 IM |
5852 | /* |
5853 | * Increase the granularity value when there are more CPUs, | |
5854 | * because with more CPUs the 'effective latency' as visible | |
5855 | * to users decreases. But the relationship is not linear, | |
5856 | * so pick a second-best guess by going with the log2 of the | |
5857 | * number of CPUs. | |
5858 | * | |
5859 | * This idea comes from the SD scheduler of Con Kolivas: | |
5860 | */ | |
5861 | static inline void sched_init_granularity(void) | |
5862 | { | |
5863 | unsigned int factor = 1 + ilog2(num_online_cpus()); | |
5864 | const unsigned long limit = 200000000; | |
5865 | ||
5866 | sysctl_sched_min_granularity *= factor; | |
5867 | if (sysctl_sched_min_granularity > limit) | |
5868 | sysctl_sched_min_granularity = limit; | |
5869 | ||
5870 | sysctl_sched_latency *= factor; | |
5871 | if (sysctl_sched_latency > limit) | |
5872 | sysctl_sched_latency = limit; | |
5873 | ||
5874 | sysctl_sched_wakeup_granularity *= factor; | |
19978ca6 IM |
5875 | } |
5876 | ||
1da177e4 LT |
5877 | #ifdef CONFIG_SMP |
5878 | /* | |
5879 | * This is how migration works: | |
5880 | * | |
70b97a7f | 5881 | * 1) we queue a struct migration_req structure in the source CPU's |
1da177e4 LT |
5882 | * runqueue and wake up that CPU's migration thread. |
5883 | * 2) we down() the locked semaphore => thread blocks. | |
5884 | * 3) migration thread wakes up (implicitly it forces the migrated | |
5885 | * thread off the CPU) | |
5886 | * 4) it gets the migration request and checks whether the migrated | |
5887 | * task is still in the wrong runqueue. | |
5888 | * 5) if it's in the wrong runqueue then the migration thread removes | |
5889 | * it and puts it into the right queue. | |
5890 | * 6) migration thread up()s the semaphore. | |
5891 | * 7) we wake up and the migration is done. | |
5892 | */ | |
5893 | ||
5894 | /* | |
5895 | * Change a given task's CPU affinity. Migrate the thread to a | |
5896 | * proper CPU and schedule it away if the CPU it's executing on | |
5897 | * is removed from the allowed bitmask. | |
5898 | * | |
5899 | * NOTE: the caller must have a valid reference to the task, the | |
41a2d6cf | 5900 | * task must not exit() & deallocate itself prematurely. The |
1da177e4 LT |
5901 | * call is not atomic; no spinlocks may be held. |
5902 | */ | |
cd8ba7cd | 5903 | int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask) |
1da177e4 | 5904 | { |
70b97a7f | 5905 | struct migration_req req; |
1da177e4 | 5906 | unsigned long flags; |
70b97a7f | 5907 | struct rq *rq; |
48f24c4d | 5908 | int ret = 0; |
1da177e4 LT |
5909 | |
5910 | rq = task_rq_lock(p, &flags); | |
cd8ba7cd | 5911 | if (!cpus_intersects(*new_mask, cpu_online_map)) { |
1da177e4 LT |
5912 | ret = -EINVAL; |
5913 | goto out; | |
5914 | } | |
5915 | ||
73fe6aae | 5916 | if (p->sched_class->set_cpus_allowed) |
cd8ba7cd | 5917 | p->sched_class->set_cpus_allowed(p, new_mask); |
73fe6aae | 5918 | else { |
cd8ba7cd MT |
5919 | p->cpus_allowed = *new_mask; |
5920 | p->rt.nr_cpus_allowed = cpus_weight(*new_mask); | |
73fe6aae GH |
5921 | } |
5922 | ||
1da177e4 | 5923 | /* Can the task run on the task's current CPU? If so, we're done */ |
cd8ba7cd | 5924 | if (cpu_isset(task_cpu(p), *new_mask)) |
1da177e4 LT |
5925 | goto out; |
5926 | ||
cd8ba7cd | 5927 | if (migrate_task(p, any_online_cpu(*new_mask), &req)) { |
1da177e4 LT |
5928 | /* Need help from migration thread: drop lock and wait. */ |
5929 | task_rq_unlock(rq, &flags); | |
5930 | wake_up_process(rq->migration_thread); | |
5931 | wait_for_completion(&req.done); | |
5932 | tlb_migrate_finish(p->mm); | |
5933 | return 0; | |
5934 | } | |
5935 | out: | |
5936 | task_rq_unlock(rq, &flags); | |
48f24c4d | 5937 | |
1da177e4 LT |
5938 | return ret; |
5939 | } | |
cd8ba7cd | 5940 | EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); |
1da177e4 LT |
5941 | |
5942 | /* | |
41a2d6cf | 5943 | * Move (not current) task off this cpu, onto dest cpu. We're doing |
1da177e4 LT |
5944 | * this because either it can't run here any more (set_cpus_allowed() |
5945 | * away from this CPU, or CPU going down), or because we're | |
5946 | * attempting to rebalance this task on exec (sched_exec). | |
5947 | * | |
5948 | * So we race with normal scheduler movements, but that's OK, as long | |
5949 | * as the task is no longer on this CPU. | |
efc30814 KK |
5950 | * |
5951 | * Returns non-zero if task was successfully migrated. | |
1da177e4 | 5952 | */ |
efc30814 | 5953 | static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) |
1da177e4 | 5954 | { |
70b97a7f | 5955 | struct rq *rq_dest, *rq_src; |
dd41f596 | 5956 | int ret = 0, on_rq; |
1da177e4 LT |
5957 | |
5958 | if (unlikely(cpu_is_offline(dest_cpu))) | |
efc30814 | 5959 | return ret; |
1da177e4 LT |
5960 | |
5961 | rq_src = cpu_rq(src_cpu); | |
5962 | rq_dest = cpu_rq(dest_cpu); | |
5963 | ||
5964 | double_rq_lock(rq_src, rq_dest); | |
5965 | /* Already moved. */ | |
5966 | if (task_cpu(p) != src_cpu) | |
5967 | goto out; | |
5968 | /* Affinity changed (again). */ | |
5969 | if (!cpu_isset(dest_cpu, p->cpus_allowed)) | |
5970 | goto out; | |
5971 | ||
dd41f596 | 5972 | on_rq = p->se.on_rq; |
6e82a3be | 5973 | if (on_rq) |
2e1cb74a | 5974 | deactivate_task(rq_src, p, 0); |
6e82a3be | 5975 | |
1da177e4 | 5976 | set_task_cpu(p, dest_cpu); |
dd41f596 IM |
5977 | if (on_rq) { |
5978 | activate_task(rq_dest, p, 0); | |
5979 | check_preempt_curr(rq_dest, p); | |
1da177e4 | 5980 | } |
efc30814 | 5981 | ret = 1; |
1da177e4 LT |
5982 | out: |
5983 | double_rq_unlock(rq_src, rq_dest); | |
efc30814 | 5984 | return ret; |
1da177e4 LT |
5985 | } |
5986 | ||
5987 | /* | |
5988 | * migration_thread - this is a highprio system thread that performs | |
5989 | * thread migration by bumping thread off CPU then 'pushing' onto | |
5990 | * another runqueue. | |
5991 | */ | |
95cdf3b7 | 5992 | static int migration_thread(void *data) |
1da177e4 | 5993 | { |
1da177e4 | 5994 | int cpu = (long)data; |
70b97a7f | 5995 | struct rq *rq; |
1da177e4 LT |
5996 | |
5997 | rq = cpu_rq(cpu); | |
5998 | BUG_ON(rq->migration_thread != current); | |
5999 | ||
6000 | set_current_state(TASK_INTERRUPTIBLE); | |
6001 | while (!kthread_should_stop()) { | |
70b97a7f | 6002 | struct migration_req *req; |
1da177e4 | 6003 | struct list_head *head; |
1da177e4 | 6004 | |
1da177e4 LT |
6005 | spin_lock_irq(&rq->lock); |
6006 | ||
6007 | if (cpu_is_offline(cpu)) { | |
6008 | spin_unlock_irq(&rq->lock); | |
6009 | goto wait_to_die; | |
6010 | } | |
6011 | ||
6012 | if (rq->active_balance) { | |
6013 | active_load_balance(rq, cpu); | |
6014 | rq->active_balance = 0; | |
6015 | } | |
6016 | ||
6017 | head = &rq->migration_queue; | |
6018 | ||
6019 | if (list_empty(head)) { | |
6020 | spin_unlock_irq(&rq->lock); | |
6021 | schedule(); | |
6022 | set_current_state(TASK_INTERRUPTIBLE); | |
6023 | continue; | |
6024 | } | |
70b97a7f | 6025 | req = list_entry(head->next, struct migration_req, list); |
1da177e4 LT |
6026 | list_del_init(head->next); |
6027 | ||
674311d5 NP |
6028 | spin_unlock(&rq->lock); |
6029 | __migrate_task(req->task, cpu, req->dest_cpu); | |
6030 | local_irq_enable(); | |
1da177e4 LT |
6031 | |
6032 | complete(&req->done); | |
6033 | } | |
6034 | __set_current_state(TASK_RUNNING); | |
6035 | return 0; | |
6036 | ||
6037 | wait_to_die: | |
6038 | /* Wait for kthread_stop */ | |
6039 | set_current_state(TASK_INTERRUPTIBLE); | |
6040 | while (!kthread_should_stop()) { | |
6041 | schedule(); | |
6042 | set_current_state(TASK_INTERRUPTIBLE); | |
6043 | } | |
6044 | __set_current_state(TASK_RUNNING); | |
6045 | return 0; | |
6046 | } | |
6047 | ||
6048 | #ifdef CONFIG_HOTPLUG_CPU | |
f7b4cddc ON |
6049 | |
6050 | static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu) | |
6051 | { | |
6052 | int ret; | |
6053 | ||
6054 | local_irq_disable(); | |
6055 | ret = __migrate_task(p, src_cpu, dest_cpu); | |
6056 | local_irq_enable(); | |
6057 | return ret; | |
6058 | } | |
6059 | ||
054b9108 | 6060 | /* |
3a4fa0a2 | 6061 | * Figure out where task on dead CPU should go, use force if necessary. |
054b9108 KK |
6062 | * NOTE: interrupts should be disabled by the caller |
6063 | */ | |
48f24c4d | 6064 | static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p) |
1da177e4 | 6065 | { |
efc30814 | 6066 | unsigned long flags; |
1da177e4 | 6067 | cpumask_t mask; |
70b97a7f IM |
6068 | struct rq *rq; |
6069 | int dest_cpu; | |
1da177e4 | 6070 | |
3a5c359a AK |
6071 | do { |
6072 | /* On same node? */ | |
6073 | mask = node_to_cpumask(cpu_to_node(dead_cpu)); | |
6074 | cpus_and(mask, mask, p->cpus_allowed); | |
6075 | dest_cpu = any_online_cpu(mask); | |
6076 | ||
6077 | /* On any allowed CPU? */ | |
434d53b0 | 6078 | if (dest_cpu >= nr_cpu_ids) |
3a5c359a AK |
6079 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6080 | ||
6081 | /* No more Mr. Nice Guy. */ | |
434d53b0 | 6082 | if (dest_cpu >= nr_cpu_ids) { |
f9a86fcb MT |
6083 | cpumask_t cpus_allowed; |
6084 | ||
6085 | cpuset_cpus_allowed_locked(p, &cpus_allowed); | |
470fd646 CW |
6086 | /* |
6087 | * Try to stay on the same cpuset, where the | |
6088 | * current cpuset may be a subset of all cpus. | |
6089 | * The cpuset_cpus_allowed_locked() variant of | |
41a2d6cf | 6090 | * cpuset_cpus_allowed() will not block. It must be |
470fd646 CW |
6091 | * called within calls to cpuset_lock/cpuset_unlock. |
6092 | */ | |
3a5c359a | 6093 | rq = task_rq_lock(p, &flags); |
470fd646 | 6094 | p->cpus_allowed = cpus_allowed; |
3a5c359a AK |
6095 | dest_cpu = any_online_cpu(p->cpus_allowed); |
6096 | task_rq_unlock(rq, &flags); | |
1da177e4 | 6097 | |
3a5c359a AK |
6098 | /* |
6099 | * Don't tell them about moving exiting tasks or | |
6100 | * kernel threads (both mm NULL), since they never | |
6101 | * leave kernel. | |
6102 | */ | |
41a2d6cf | 6103 | if (p->mm && printk_ratelimit()) { |
3a5c359a AK |
6104 | printk(KERN_INFO "process %d (%s) no " |
6105 | "longer affine to cpu%d\n", | |
41a2d6cf IM |
6106 | task_pid_nr(p), p->comm, dead_cpu); |
6107 | } | |
3a5c359a | 6108 | } |
f7b4cddc | 6109 | } while (!__migrate_task_irq(p, dead_cpu, dest_cpu)); |
1da177e4 LT |
6110 | } |
6111 | ||
6112 | /* | |
6113 | * While a dead CPU has no uninterruptible tasks queued at this point, | |
6114 | * it might still have a nonzero ->nr_uninterruptible counter, because | |
6115 | * for performance reasons the counter is not stricly tracking tasks to | |
6116 | * their home CPUs. So we just add the counter to another CPU's counter, | |
6117 | * to keep the global sum constant after CPU-down: | |
6118 | */ | |
70b97a7f | 6119 | static void migrate_nr_uninterruptible(struct rq *rq_src) |
1da177e4 | 6120 | { |
7c16ec58 | 6121 | struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR)); |
1da177e4 LT |
6122 | unsigned long flags; |
6123 | ||
6124 | local_irq_save(flags); | |
6125 | double_rq_lock(rq_src, rq_dest); | |
6126 | rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; | |
6127 | rq_src->nr_uninterruptible = 0; | |
6128 | double_rq_unlock(rq_src, rq_dest); | |
6129 | local_irq_restore(flags); | |
6130 | } | |
6131 | ||
6132 | /* Run through task list and migrate tasks from the dead cpu. */ | |
6133 | static void migrate_live_tasks(int src_cpu) | |
6134 | { | |
48f24c4d | 6135 | struct task_struct *p, *t; |
1da177e4 | 6136 | |
f7b4cddc | 6137 | read_lock(&tasklist_lock); |
1da177e4 | 6138 | |
48f24c4d IM |
6139 | do_each_thread(t, p) { |
6140 | if (p == current) | |
1da177e4 LT |
6141 | continue; |
6142 | ||
48f24c4d IM |
6143 | if (task_cpu(p) == src_cpu) |
6144 | move_task_off_dead_cpu(src_cpu, p); | |
6145 | } while_each_thread(t, p); | |
1da177e4 | 6146 | |
f7b4cddc | 6147 | read_unlock(&tasklist_lock); |
1da177e4 LT |
6148 | } |
6149 | ||
dd41f596 IM |
6150 | /* |
6151 | * Schedules idle task to be the next runnable task on current CPU. | |
94bc9a7b DA |
6152 | * It does so by boosting its priority to highest possible. |
6153 | * Used by CPU offline code. | |
1da177e4 LT |
6154 | */ |
6155 | void sched_idle_next(void) | |
6156 | { | |
48f24c4d | 6157 | int this_cpu = smp_processor_id(); |
70b97a7f | 6158 | struct rq *rq = cpu_rq(this_cpu); |
1da177e4 LT |
6159 | struct task_struct *p = rq->idle; |
6160 | unsigned long flags; | |
6161 | ||
6162 | /* cpu has to be offline */ | |
48f24c4d | 6163 | BUG_ON(cpu_online(this_cpu)); |
1da177e4 | 6164 | |
48f24c4d IM |
6165 | /* |
6166 | * Strictly not necessary since rest of the CPUs are stopped by now | |
6167 | * and interrupts disabled on the current cpu. | |
1da177e4 LT |
6168 | */ |
6169 | spin_lock_irqsave(&rq->lock, flags); | |
6170 | ||
dd41f596 | 6171 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
48f24c4d | 6172 | |
94bc9a7b DA |
6173 | update_rq_clock(rq); |
6174 | activate_task(rq, p, 0); | |
1da177e4 LT |
6175 | |
6176 | spin_unlock_irqrestore(&rq->lock, flags); | |
6177 | } | |
6178 | ||
48f24c4d IM |
6179 | /* |
6180 | * Ensures that the idle task is using init_mm right before its cpu goes | |
1da177e4 LT |
6181 | * offline. |
6182 | */ | |
6183 | void idle_task_exit(void) | |
6184 | { | |
6185 | struct mm_struct *mm = current->active_mm; | |
6186 | ||
6187 | BUG_ON(cpu_online(smp_processor_id())); | |
6188 | ||
6189 | if (mm != &init_mm) | |
6190 | switch_mm(mm, &init_mm, current); | |
6191 | mmdrop(mm); | |
6192 | } | |
6193 | ||
054b9108 | 6194 | /* called under rq->lock with disabled interrupts */ |
36c8b586 | 6195 | static void migrate_dead(unsigned int dead_cpu, struct task_struct *p) |
1da177e4 | 6196 | { |
70b97a7f | 6197 | struct rq *rq = cpu_rq(dead_cpu); |
1da177e4 LT |
6198 | |
6199 | /* Must be exiting, otherwise would be on tasklist. */ | |
270f722d | 6200 | BUG_ON(!p->exit_state); |
1da177e4 LT |
6201 | |
6202 | /* Cannot have done final schedule yet: would have vanished. */ | |
c394cc9f | 6203 | BUG_ON(p->state == TASK_DEAD); |
1da177e4 | 6204 | |
48f24c4d | 6205 | get_task_struct(p); |
1da177e4 LT |
6206 | |
6207 | /* | |
6208 | * Drop lock around migration; if someone else moves it, | |
41a2d6cf | 6209 | * that's OK. No task can be added to this CPU, so iteration is |
1da177e4 LT |
6210 | * fine. |
6211 | */ | |
f7b4cddc | 6212 | spin_unlock_irq(&rq->lock); |
48f24c4d | 6213 | move_task_off_dead_cpu(dead_cpu, p); |
f7b4cddc | 6214 | spin_lock_irq(&rq->lock); |
1da177e4 | 6215 | |
48f24c4d | 6216 | put_task_struct(p); |
1da177e4 LT |
6217 | } |
6218 | ||
6219 | /* release_task() removes task from tasklist, so we won't find dead tasks. */ | |
6220 | static void migrate_dead_tasks(unsigned int dead_cpu) | |
6221 | { | |
70b97a7f | 6222 | struct rq *rq = cpu_rq(dead_cpu); |
dd41f596 | 6223 | struct task_struct *next; |
48f24c4d | 6224 | |
dd41f596 IM |
6225 | for ( ; ; ) { |
6226 | if (!rq->nr_running) | |
6227 | break; | |
a8e504d2 | 6228 | update_rq_clock(rq); |
ff95f3df | 6229 | next = pick_next_task(rq, rq->curr); |
dd41f596 IM |
6230 | if (!next) |
6231 | break; | |
6232 | migrate_dead(dead_cpu, next); | |
e692ab53 | 6233 | |
1da177e4 LT |
6234 | } |
6235 | } | |
6236 | #endif /* CONFIG_HOTPLUG_CPU */ | |
6237 | ||
e692ab53 NP |
6238 | #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) |
6239 | ||
6240 | static struct ctl_table sd_ctl_dir[] = { | |
e0361851 AD |
6241 | { |
6242 | .procname = "sched_domain", | |
c57baf1e | 6243 | .mode = 0555, |
e0361851 | 6244 | }, |
38605cae | 6245 | {0, }, |
e692ab53 NP |
6246 | }; |
6247 | ||
6248 | static struct ctl_table sd_ctl_root[] = { | |
e0361851 | 6249 | { |
c57baf1e | 6250 | .ctl_name = CTL_KERN, |
e0361851 | 6251 | .procname = "kernel", |
c57baf1e | 6252 | .mode = 0555, |
e0361851 AD |
6253 | .child = sd_ctl_dir, |
6254 | }, | |
38605cae | 6255 | {0, }, |
e692ab53 NP |
6256 | }; |
6257 | ||
6258 | static struct ctl_table *sd_alloc_ctl_entry(int n) | |
6259 | { | |
6260 | struct ctl_table *entry = | |
5cf9f062 | 6261 | kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); |
e692ab53 | 6262 | |
e692ab53 NP |
6263 | return entry; |
6264 | } | |
6265 | ||
6382bc90 MM |
6266 | static void sd_free_ctl_entry(struct ctl_table **tablep) |
6267 | { | |
cd790076 | 6268 | struct ctl_table *entry; |
6382bc90 | 6269 | |
cd790076 MM |
6270 | /* |
6271 | * In the intermediate directories, both the child directory and | |
6272 | * procname are dynamically allocated and could fail but the mode | |
41a2d6cf | 6273 | * will always be set. In the lowest directory the names are |
cd790076 MM |
6274 | * static strings and all have proc handlers. |
6275 | */ | |
6276 | for (entry = *tablep; entry->mode; entry++) { | |
6382bc90 MM |
6277 | if (entry->child) |
6278 | sd_free_ctl_entry(&entry->child); | |
cd790076 MM |
6279 | if (entry->proc_handler == NULL) |
6280 | kfree(entry->procname); | |
6281 | } | |
6382bc90 MM |
6282 | |
6283 | kfree(*tablep); | |
6284 | *tablep = NULL; | |
6285 | } | |
6286 | ||
e692ab53 | 6287 | static void |
e0361851 | 6288 | set_table_entry(struct ctl_table *entry, |
e692ab53 NP |
6289 | const char *procname, void *data, int maxlen, |
6290 | mode_t mode, proc_handler *proc_handler) | |
6291 | { | |
e692ab53 NP |
6292 | entry->procname = procname; |
6293 | entry->data = data; | |
6294 | entry->maxlen = maxlen; | |
6295 | entry->mode = mode; | |
6296 | entry->proc_handler = proc_handler; | |
6297 | } | |
6298 | ||
6299 | static struct ctl_table * | |
6300 | sd_alloc_ctl_domain_table(struct sched_domain *sd) | |
6301 | { | |
ace8b3d6 | 6302 | struct ctl_table *table = sd_alloc_ctl_entry(12); |
e692ab53 | 6303 | |
ad1cdc1d MM |
6304 | if (table == NULL) |
6305 | return NULL; | |
6306 | ||
e0361851 | 6307 | set_table_entry(&table[0], "min_interval", &sd->min_interval, |
e692ab53 | 6308 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6309 | set_table_entry(&table[1], "max_interval", &sd->max_interval, |
e692ab53 | 6310 | sizeof(long), 0644, proc_doulongvec_minmax); |
e0361851 | 6311 | set_table_entry(&table[2], "busy_idx", &sd->busy_idx, |
e692ab53 | 6312 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6313 | set_table_entry(&table[3], "idle_idx", &sd->idle_idx, |
e692ab53 | 6314 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6315 | set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, |
e692ab53 | 6316 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6317 | set_table_entry(&table[5], "wake_idx", &sd->wake_idx, |
e692ab53 | 6318 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6319 | set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, |
e692ab53 | 6320 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6321 | set_table_entry(&table[7], "busy_factor", &sd->busy_factor, |
e692ab53 | 6322 | sizeof(int), 0644, proc_dointvec_minmax); |
e0361851 | 6323 | set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, |
e692ab53 | 6324 | sizeof(int), 0644, proc_dointvec_minmax); |
ace8b3d6 | 6325 | set_table_entry(&table[9], "cache_nice_tries", |
e692ab53 NP |
6326 | &sd->cache_nice_tries, |
6327 | sizeof(int), 0644, proc_dointvec_minmax); | |
ace8b3d6 | 6328 | set_table_entry(&table[10], "flags", &sd->flags, |
e692ab53 | 6329 | sizeof(int), 0644, proc_dointvec_minmax); |
6323469f | 6330 | /* &table[11] is terminator */ |
e692ab53 NP |
6331 | |
6332 | return table; | |
6333 | } | |
6334 | ||
9a4e7159 | 6335 | static ctl_table *sd_alloc_ctl_cpu_table(int cpu) |
e692ab53 NP |
6336 | { |
6337 | struct ctl_table *entry, *table; | |
6338 | struct sched_domain *sd; | |
6339 | int domain_num = 0, i; | |
6340 | char buf[32]; | |
6341 | ||
6342 | for_each_domain(cpu, sd) | |
6343 | domain_num++; | |
6344 | entry = table = sd_alloc_ctl_entry(domain_num + 1); | |
ad1cdc1d MM |
6345 | if (table == NULL) |
6346 | return NULL; | |
e692ab53 NP |
6347 | |
6348 | i = 0; | |
6349 | for_each_domain(cpu, sd) { | |
6350 | snprintf(buf, 32, "domain%d", i); | |
e692ab53 | 6351 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6352 | entry->mode = 0555; |
e692ab53 NP |
6353 | entry->child = sd_alloc_ctl_domain_table(sd); |
6354 | entry++; | |
6355 | i++; | |
6356 | } | |
6357 | return table; | |
6358 | } | |
6359 | ||
6360 | static struct ctl_table_header *sd_sysctl_header; | |
6382bc90 | 6361 | static void register_sched_domain_sysctl(void) |
e692ab53 NP |
6362 | { |
6363 | int i, cpu_num = num_online_cpus(); | |
6364 | struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); | |
6365 | char buf[32]; | |
6366 | ||
7378547f MM |
6367 | WARN_ON(sd_ctl_dir[0].child); |
6368 | sd_ctl_dir[0].child = entry; | |
6369 | ||
ad1cdc1d MM |
6370 | if (entry == NULL) |
6371 | return; | |
6372 | ||
97b6ea7b | 6373 | for_each_online_cpu(i) { |
e692ab53 | 6374 | snprintf(buf, 32, "cpu%d", i); |
e692ab53 | 6375 | entry->procname = kstrdup(buf, GFP_KERNEL); |
c57baf1e | 6376 | entry->mode = 0555; |
e692ab53 | 6377 | entry->child = sd_alloc_ctl_cpu_table(i); |
97b6ea7b | 6378 | entry++; |
e692ab53 | 6379 | } |
7378547f MM |
6380 | |
6381 | WARN_ON(sd_sysctl_header); | |
e692ab53 NP |
6382 | sd_sysctl_header = register_sysctl_table(sd_ctl_root); |
6383 | } | |
6382bc90 | 6384 | |
7378547f | 6385 | /* may be called multiple times per register */ |
6382bc90 MM |
6386 | static void unregister_sched_domain_sysctl(void) |
6387 | { | |
7378547f MM |
6388 | if (sd_sysctl_header) |
6389 | unregister_sysctl_table(sd_sysctl_header); | |
6382bc90 | 6390 | sd_sysctl_header = NULL; |
7378547f MM |
6391 | if (sd_ctl_dir[0].child) |
6392 | sd_free_ctl_entry(&sd_ctl_dir[0].child); | |
6382bc90 | 6393 | } |
e692ab53 | 6394 | #else |
6382bc90 MM |
6395 | static void register_sched_domain_sysctl(void) |
6396 | { | |
6397 | } | |
6398 | static void unregister_sched_domain_sysctl(void) | |
e692ab53 NP |
6399 | { |
6400 | } | |
6401 | #endif | |
6402 | ||
1da177e4 LT |
6403 | /* |
6404 | * migration_call - callback that gets triggered when a CPU is added. | |
6405 | * Here we can start up the necessary migration thread for the new CPU. | |
6406 | */ | |
48f24c4d IM |
6407 | static int __cpuinit |
6408 | migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
1da177e4 | 6409 | { |
1da177e4 | 6410 | struct task_struct *p; |
48f24c4d | 6411 | int cpu = (long)hcpu; |
1da177e4 | 6412 | unsigned long flags; |
70b97a7f | 6413 | struct rq *rq; |
1da177e4 LT |
6414 | |
6415 | switch (action) { | |
5be9361c | 6416 | |
1da177e4 | 6417 | case CPU_UP_PREPARE: |
8bb78442 | 6418 | case CPU_UP_PREPARE_FROZEN: |
dd41f596 | 6419 | p = kthread_create(migration_thread, hcpu, "migration/%d", cpu); |
1da177e4 LT |
6420 | if (IS_ERR(p)) |
6421 | return NOTIFY_BAD; | |
1da177e4 LT |
6422 | kthread_bind(p, cpu); |
6423 | /* Must be high prio: stop_machine expects to yield to it. */ | |
6424 | rq = task_rq_lock(p, &flags); | |
dd41f596 | 6425 | __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1); |
1da177e4 LT |
6426 | task_rq_unlock(rq, &flags); |
6427 | cpu_rq(cpu)->migration_thread = p; | |
6428 | break; | |
48f24c4d | 6429 | |
1da177e4 | 6430 | case CPU_ONLINE: |
8bb78442 | 6431 | case CPU_ONLINE_FROZEN: |
3a4fa0a2 | 6432 | /* Strictly unnecessary, as first user will wake it. */ |
1da177e4 | 6433 | wake_up_process(cpu_rq(cpu)->migration_thread); |
1f94ef59 GH |
6434 | |
6435 | /* Update our root-domain */ | |
6436 | rq = cpu_rq(cpu); | |
6437 | spin_lock_irqsave(&rq->lock, flags); | |
6438 | if (rq->rd) { | |
6439 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6440 | cpu_set(cpu, rq->rd->online); | |
6441 | } | |
6442 | spin_unlock_irqrestore(&rq->lock, flags); | |
1da177e4 | 6443 | break; |
48f24c4d | 6444 | |
1da177e4 LT |
6445 | #ifdef CONFIG_HOTPLUG_CPU |
6446 | case CPU_UP_CANCELED: | |
8bb78442 | 6447 | case CPU_UP_CANCELED_FROZEN: |
fc75cdfa HC |
6448 | if (!cpu_rq(cpu)->migration_thread) |
6449 | break; | |
41a2d6cf | 6450 | /* Unbind it from offline cpu so it can run. Fall thru. */ |
a4c4af7c HC |
6451 | kthread_bind(cpu_rq(cpu)->migration_thread, |
6452 | any_online_cpu(cpu_online_map)); | |
1da177e4 LT |
6453 | kthread_stop(cpu_rq(cpu)->migration_thread); |
6454 | cpu_rq(cpu)->migration_thread = NULL; | |
6455 | break; | |
48f24c4d | 6456 | |
1da177e4 | 6457 | case CPU_DEAD: |
8bb78442 | 6458 | case CPU_DEAD_FROZEN: |
470fd646 | 6459 | cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */ |
1da177e4 LT |
6460 | migrate_live_tasks(cpu); |
6461 | rq = cpu_rq(cpu); | |
6462 | kthread_stop(rq->migration_thread); | |
6463 | rq->migration_thread = NULL; | |
6464 | /* Idle task back to normal (off runqueue, low prio) */ | |
d2da272a | 6465 | spin_lock_irq(&rq->lock); |
a8e504d2 | 6466 | update_rq_clock(rq); |
2e1cb74a | 6467 | deactivate_task(rq, rq->idle, 0); |
1da177e4 | 6468 | rq->idle->static_prio = MAX_PRIO; |
dd41f596 IM |
6469 | __setscheduler(rq, rq->idle, SCHED_NORMAL, 0); |
6470 | rq->idle->sched_class = &idle_sched_class; | |
1da177e4 | 6471 | migrate_dead_tasks(cpu); |
d2da272a | 6472 | spin_unlock_irq(&rq->lock); |
470fd646 | 6473 | cpuset_unlock(); |
1da177e4 LT |
6474 | migrate_nr_uninterruptible(rq); |
6475 | BUG_ON(rq->nr_running != 0); | |
6476 | ||
41a2d6cf IM |
6477 | /* |
6478 | * No need to migrate the tasks: it was best-effort if | |
6479 | * they didn't take sched_hotcpu_mutex. Just wake up | |
6480 | * the requestors. | |
6481 | */ | |
1da177e4 LT |
6482 | spin_lock_irq(&rq->lock); |
6483 | while (!list_empty(&rq->migration_queue)) { | |
70b97a7f IM |
6484 | struct migration_req *req; |
6485 | ||
1da177e4 | 6486 | req = list_entry(rq->migration_queue.next, |
70b97a7f | 6487 | struct migration_req, list); |
1da177e4 LT |
6488 | list_del_init(&req->list); |
6489 | complete(&req->done); | |
6490 | } | |
6491 | spin_unlock_irq(&rq->lock); | |
6492 | break; | |
57d885fe | 6493 | |
08f503b0 GH |
6494 | case CPU_DYING: |
6495 | case CPU_DYING_FROZEN: | |
57d885fe GH |
6496 | /* Update our root-domain */ |
6497 | rq = cpu_rq(cpu); | |
6498 | spin_lock_irqsave(&rq->lock, flags); | |
6499 | if (rq->rd) { | |
6500 | BUG_ON(!cpu_isset(cpu, rq->rd->span)); | |
6501 | cpu_clear(cpu, rq->rd->online); | |
6502 | } | |
6503 | spin_unlock_irqrestore(&rq->lock, flags); | |
6504 | break; | |
1da177e4 LT |
6505 | #endif |
6506 | } | |
6507 | return NOTIFY_OK; | |
6508 | } | |
6509 | ||
6510 | /* Register at highest priority so that task migration (migrate_all_tasks) | |
6511 | * happens before everything else. | |
6512 | */ | |
26c2143b | 6513 | static struct notifier_block __cpuinitdata migration_notifier = { |
1da177e4 LT |
6514 | .notifier_call = migration_call, |
6515 | .priority = 10 | |
6516 | }; | |
6517 | ||
e6fe6649 | 6518 | void __init migration_init(void) |
1da177e4 LT |
6519 | { |
6520 | void *cpu = (void *)(long)smp_processor_id(); | |
07dccf33 | 6521 | int err; |
48f24c4d IM |
6522 | |
6523 | /* Start one for the boot CPU: */ | |
07dccf33 AM |
6524 | err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); |
6525 | BUG_ON(err == NOTIFY_BAD); | |
1da177e4 LT |
6526 | migration_call(&migration_notifier, CPU_ONLINE, cpu); |
6527 | register_cpu_notifier(&migration_notifier); | |
1da177e4 LT |
6528 | } |
6529 | #endif | |
6530 | ||
6531 | #ifdef CONFIG_SMP | |
476f3534 | 6532 | |
3e9830dc | 6533 | #ifdef CONFIG_SCHED_DEBUG |
4dcf6aff | 6534 | |
7c16ec58 MT |
6535 | static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, |
6536 | cpumask_t *groupmask) | |
1da177e4 | 6537 | { |
4dcf6aff | 6538 | struct sched_group *group = sd->groups; |
434d53b0 | 6539 | char str[256]; |
1da177e4 | 6540 | |
434d53b0 | 6541 | cpulist_scnprintf(str, sizeof(str), sd->span); |
7c16ec58 | 6542 | cpus_clear(*groupmask); |
4dcf6aff IM |
6543 | |
6544 | printk(KERN_DEBUG "%*s domain %d: ", level, "", level); | |
6545 | ||
6546 | if (!(sd->flags & SD_LOAD_BALANCE)) { | |
6547 | printk("does not load-balance\n"); | |
6548 | if (sd->parent) | |
6549 | printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" | |
6550 | " has parent"); | |
6551 | return -1; | |
41c7ce9a NP |
6552 | } |
6553 | ||
4dcf6aff IM |
6554 | printk(KERN_CONT "span %s\n", str); |
6555 | ||
6556 | if (!cpu_isset(cpu, sd->span)) { | |
6557 | printk(KERN_ERR "ERROR: domain->span does not contain " | |
6558 | "CPU%d\n", cpu); | |
6559 | } | |
6560 | if (!cpu_isset(cpu, group->cpumask)) { | |
6561 | printk(KERN_ERR "ERROR: domain->groups does not contain" | |
6562 | " CPU%d\n", cpu); | |
6563 | } | |
1da177e4 | 6564 | |
4dcf6aff | 6565 | printk(KERN_DEBUG "%*s groups:", level + 1, ""); |
1da177e4 | 6566 | do { |
4dcf6aff IM |
6567 | if (!group) { |
6568 | printk("\n"); | |
6569 | printk(KERN_ERR "ERROR: group is NULL\n"); | |
1da177e4 LT |
6570 | break; |
6571 | } | |
6572 | ||
4dcf6aff IM |
6573 | if (!group->__cpu_power) { |
6574 | printk(KERN_CONT "\n"); | |
6575 | printk(KERN_ERR "ERROR: domain->cpu_power not " | |
6576 | "set\n"); | |
6577 | break; | |
6578 | } | |
1da177e4 | 6579 | |
4dcf6aff IM |
6580 | if (!cpus_weight(group->cpumask)) { |
6581 | printk(KERN_CONT "\n"); | |
6582 | printk(KERN_ERR "ERROR: empty group\n"); | |
6583 | break; | |
6584 | } | |
1da177e4 | 6585 | |
7c16ec58 | 6586 | if (cpus_intersects(*groupmask, group->cpumask)) { |
4dcf6aff IM |
6587 | printk(KERN_CONT "\n"); |
6588 | printk(KERN_ERR "ERROR: repeated CPUs\n"); | |
6589 | break; | |
6590 | } | |
1da177e4 | 6591 | |
7c16ec58 | 6592 | cpus_or(*groupmask, *groupmask, group->cpumask); |
1da177e4 | 6593 | |
434d53b0 | 6594 | cpulist_scnprintf(str, sizeof(str), group->cpumask); |
4dcf6aff | 6595 | printk(KERN_CONT " %s", str); |
1da177e4 | 6596 | |
4dcf6aff IM |
6597 | group = group->next; |
6598 | } while (group != sd->groups); | |
6599 | printk(KERN_CONT "\n"); | |
1da177e4 | 6600 | |
7c16ec58 | 6601 | if (!cpus_equal(sd->span, *groupmask)) |
4dcf6aff | 6602 | printk(KERN_ERR "ERROR: groups don't span domain->span\n"); |
1da177e4 | 6603 | |
7c16ec58 | 6604 | if (sd->parent && !cpus_subset(*groupmask, sd->parent->span)) |
4dcf6aff IM |
6605 | printk(KERN_ERR "ERROR: parent span is not a superset " |
6606 | "of domain->span\n"); | |
6607 | return 0; | |
6608 | } | |
1da177e4 | 6609 | |
4dcf6aff IM |
6610 | static void sched_domain_debug(struct sched_domain *sd, int cpu) |
6611 | { | |
7c16ec58 | 6612 | cpumask_t *groupmask; |
4dcf6aff | 6613 | int level = 0; |
1da177e4 | 6614 | |
4dcf6aff IM |
6615 | if (!sd) { |
6616 | printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); | |
6617 | return; | |
6618 | } | |
1da177e4 | 6619 | |
4dcf6aff IM |
6620 | printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); |
6621 | ||
7c16ec58 MT |
6622 | groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL); |
6623 | if (!groupmask) { | |
6624 | printk(KERN_DEBUG "Cannot load-balance (out of memory)\n"); | |
6625 | return; | |
6626 | } | |
6627 | ||
4dcf6aff | 6628 | for (;;) { |
7c16ec58 | 6629 | if (sched_domain_debug_one(sd, cpu, level, groupmask)) |
4dcf6aff | 6630 | break; |
1da177e4 LT |
6631 | level++; |
6632 | sd = sd->parent; | |
33859f7f | 6633 | if (!sd) |
4dcf6aff IM |
6634 | break; |
6635 | } | |
7c16ec58 | 6636 | kfree(groupmask); |
1da177e4 LT |
6637 | } |
6638 | #else | |
48f24c4d | 6639 | # define sched_domain_debug(sd, cpu) do { } while (0) |
1da177e4 LT |
6640 | #endif |
6641 | ||
1a20ff27 | 6642 | static int sd_degenerate(struct sched_domain *sd) |
245af2c7 SS |
6643 | { |
6644 | if (cpus_weight(sd->span) == 1) | |
6645 | return 1; | |
6646 | ||
6647 | /* Following flags need at least 2 groups */ | |
6648 | if (sd->flags & (SD_LOAD_BALANCE | | |
6649 | SD_BALANCE_NEWIDLE | | |
6650 | SD_BALANCE_FORK | | |
89c4710e SS |
6651 | SD_BALANCE_EXEC | |
6652 | SD_SHARE_CPUPOWER | | |
6653 | SD_SHARE_PKG_RESOURCES)) { | |
245af2c7 SS |
6654 | if (sd->groups != sd->groups->next) |
6655 | return 0; | |
6656 | } | |
6657 | ||
6658 | /* Following flags don't use groups */ | |
6659 | if (sd->flags & (SD_WAKE_IDLE | | |
6660 | SD_WAKE_AFFINE | | |
6661 | SD_WAKE_BALANCE)) | |
6662 | return 0; | |
6663 | ||
6664 | return 1; | |
6665 | } | |
6666 | ||
48f24c4d IM |
6667 | static int |
6668 | sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) | |
245af2c7 SS |
6669 | { |
6670 | unsigned long cflags = sd->flags, pflags = parent->flags; | |
6671 | ||
6672 | if (sd_degenerate(parent)) | |
6673 | return 1; | |
6674 | ||
6675 | if (!cpus_equal(sd->span, parent->span)) | |
6676 | return 0; | |
6677 | ||
6678 | /* Does parent contain flags not in child? */ | |
6679 | /* WAKE_BALANCE is a subset of WAKE_AFFINE */ | |
6680 | if (cflags & SD_WAKE_AFFINE) | |
6681 | pflags &= ~SD_WAKE_BALANCE; | |
6682 | /* Flags needing groups don't count if only 1 group in parent */ | |
6683 | if (parent->groups == parent->groups->next) { | |
6684 | pflags &= ~(SD_LOAD_BALANCE | | |
6685 | SD_BALANCE_NEWIDLE | | |
6686 | SD_BALANCE_FORK | | |
89c4710e SS |
6687 | SD_BALANCE_EXEC | |
6688 | SD_SHARE_CPUPOWER | | |
6689 | SD_SHARE_PKG_RESOURCES); | |
245af2c7 SS |
6690 | } |
6691 | if (~cflags & pflags) | |
6692 | return 0; | |
6693 | ||
6694 | return 1; | |
6695 | } | |
6696 | ||
57d885fe GH |
6697 | static void rq_attach_root(struct rq *rq, struct root_domain *rd) |
6698 | { | |
6699 | unsigned long flags; | |
6700 | const struct sched_class *class; | |
6701 | ||
6702 | spin_lock_irqsave(&rq->lock, flags); | |
6703 | ||
6704 | if (rq->rd) { | |
6705 | struct root_domain *old_rd = rq->rd; | |
6706 | ||
0eab9146 | 6707 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6708 | if (class->leave_domain) |
6709 | class->leave_domain(rq); | |
0eab9146 | 6710 | } |
57d885fe | 6711 | |
dc938520 GH |
6712 | cpu_clear(rq->cpu, old_rd->span); |
6713 | cpu_clear(rq->cpu, old_rd->online); | |
6714 | ||
57d885fe GH |
6715 | if (atomic_dec_and_test(&old_rd->refcount)) |
6716 | kfree(old_rd); | |
6717 | } | |
6718 | ||
6719 | atomic_inc(&rd->refcount); | |
6720 | rq->rd = rd; | |
6721 | ||
dc938520 | 6722 | cpu_set(rq->cpu, rd->span); |
1f94ef59 GH |
6723 | if (cpu_isset(rq->cpu, cpu_online_map)) |
6724 | cpu_set(rq->cpu, rd->online); | |
dc938520 | 6725 | |
0eab9146 | 6726 | for (class = sched_class_highest; class; class = class->next) { |
57d885fe GH |
6727 | if (class->join_domain) |
6728 | class->join_domain(rq); | |
0eab9146 | 6729 | } |
57d885fe GH |
6730 | |
6731 | spin_unlock_irqrestore(&rq->lock, flags); | |
6732 | } | |
6733 | ||
dc938520 | 6734 | static void init_rootdomain(struct root_domain *rd) |
57d885fe GH |
6735 | { |
6736 | memset(rd, 0, sizeof(*rd)); | |
6737 | ||
dc938520 GH |
6738 | cpus_clear(rd->span); |
6739 | cpus_clear(rd->online); | |
57d885fe GH |
6740 | } |
6741 | ||
6742 | static void init_defrootdomain(void) | |
6743 | { | |
dc938520 | 6744 | init_rootdomain(&def_root_domain); |
57d885fe GH |
6745 | atomic_set(&def_root_domain.refcount, 1); |
6746 | } | |
6747 | ||
dc938520 | 6748 | static struct root_domain *alloc_rootdomain(void) |
57d885fe GH |
6749 | { |
6750 | struct root_domain *rd; | |
6751 | ||
6752 | rd = kmalloc(sizeof(*rd), GFP_KERNEL); | |
6753 | if (!rd) | |
6754 | return NULL; | |
6755 | ||
dc938520 | 6756 | init_rootdomain(rd); |
57d885fe GH |
6757 | |
6758 | return rd; | |
6759 | } | |
6760 | ||
1da177e4 | 6761 | /* |
0eab9146 | 6762 | * Attach the domain 'sd' to 'cpu' as its base domain. Callers must |
1da177e4 LT |
6763 | * hold the hotplug lock. |
6764 | */ | |
0eab9146 IM |
6765 | static void |
6766 | cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) | |
1da177e4 | 6767 | { |
70b97a7f | 6768 | struct rq *rq = cpu_rq(cpu); |
245af2c7 SS |
6769 | struct sched_domain *tmp; |
6770 | ||
6771 | /* Remove the sched domains which do not contribute to scheduling. */ | |
6772 | for (tmp = sd; tmp; tmp = tmp->parent) { | |
6773 | struct sched_domain *parent = tmp->parent; | |
6774 | if (!parent) | |
6775 | break; | |
1a848870 | 6776 | if (sd_parent_degenerate(tmp, parent)) { |
245af2c7 | 6777 | tmp->parent = parent->parent; |
1a848870 SS |
6778 | if (parent->parent) |
6779 | parent->parent->child = tmp; | |
6780 | } | |
245af2c7 SS |
6781 | } |
6782 | ||
1a848870 | 6783 | if (sd && sd_degenerate(sd)) { |
245af2c7 | 6784 | sd = sd->parent; |
1a848870 SS |
6785 | if (sd) |
6786 | sd->child = NULL; | |
6787 | } | |
1da177e4 LT |
6788 | |
6789 | sched_domain_debug(sd, cpu); | |
6790 | ||
57d885fe | 6791 | rq_attach_root(rq, rd); |
674311d5 | 6792 | rcu_assign_pointer(rq->sd, sd); |
1da177e4 LT |
6793 | } |
6794 | ||
6795 | /* cpus with isolated domains */ | |
67af63a6 | 6796 | static cpumask_t cpu_isolated_map = CPU_MASK_NONE; |
1da177e4 LT |
6797 | |
6798 | /* Setup the mask of cpus configured for isolated domains */ | |
6799 | static int __init isolated_cpu_setup(char *str) | |
6800 | { | |
6801 | int ints[NR_CPUS], i; | |
6802 | ||
6803 | str = get_options(str, ARRAY_SIZE(ints), ints); | |
6804 | cpus_clear(cpu_isolated_map); | |
6805 | for (i = 1; i <= ints[0]; i++) | |
6806 | if (ints[i] < NR_CPUS) | |
6807 | cpu_set(ints[i], cpu_isolated_map); | |
6808 | return 1; | |
6809 | } | |
6810 | ||
8927f494 | 6811 | __setup("isolcpus=", isolated_cpu_setup); |
1da177e4 LT |
6812 | |
6813 | /* | |
6711cab4 SS |
6814 | * init_sched_build_groups takes the cpumask we wish to span, and a pointer |
6815 | * to a function which identifies what group(along with sched group) a CPU | |
6816 | * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS | |
6817 | * (due to the fact that we keep track of groups covered with a cpumask_t). | |
1da177e4 LT |
6818 | * |
6819 | * init_sched_build_groups will build a circular linked list of the groups | |
6820 | * covered by the given span, and will set each group's ->cpumask correctly, | |
6821 | * and ->cpu_power to 0. | |
6822 | */ | |
a616058b | 6823 | static void |
7c16ec58 | 6824 | init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map, |
6711cab4 | 6825 | int (*group_fn)(int cpu, const cpumask_t *cpu_map, |
7c16ec58 MT |
6826 | struct sched_group **sg, |
6827 | cpumask_t *tmpmask), | |
6828 | cpumask_t *covered, cpumask_t *tmpmask) | |
1da177e4 LT |
6829 | { |
6830 | struct sched_group *first = NULL, *last = NULL; | |
1da177e4 LT |
6831 | int i; |
6832 | ||
7c16ec58 MT |
6833 | cpus_clear(*covered); |
6834 | ||
6835 | for_each_cpu_mask(i, *span) { | |
6711cab4 | 6836 | struct sched_group *sg; |
7c16ec58 | 6837 | int group = group_fn(i, cpu_map, &sg, tmpmask); |
1da177e4 LT |
6838 | int j; |
6839 | ||
7c16ec58 | 6840 | if (cpu_isset(i, *covered)) |
1da177e4 LT |
6841 | continue; |
6842 | ||
7c16ec58 | 6843 | cpus_clear(sg->cpumask); |
5517d86b | 6844 | sg->__cpu_power = 0; |
1da177e4 | 6845 | |
7c16ec58 MT |
6846 | for_each_cpu_mask(j, *span) { |
6847 | if (group_fn(j, cpu_map, NULL, tmpmask) != group) | |
1da177e4 LT |
6848 | continue; |
6849 | ||
7c16ec58 | 6850 | cpu_set(j, *covered); |
1da177e4 LT |
6851 | cpu_set(j, sg->cpumask); |
6852 | } | |
6853 | if (!first) | |
6854 | first = sg; | |
6855 | if (last) | |
6856 | last->next = sg; | |
6857 | last = sg; | |
6858 | } | |
6859 | last->next = first; | |
6860 | } | |
6861 | ||
9c1cfda2 | 6862 | #define SD_NODES_PER_DOMAIN 16 |
1da177e4 | 6863 | |
9c1cfda2 | 6864 | #ifdef CONFIG_NUMA |
198e2f18 | 6865 | |
9c1cfda2 JH |
6866 | /** |
6867 | * find_next_best_node - find the next node to include in a sched_domain | |
6868 | * @node: node whose sched_domain we're building | |
6869 | * @used_nodes: nodes already in the sched_domain | |
6870 | * | |
41a2d6cf | 6871 | * Find the next node to include in a given scheduling domain. Simply |
9c1cfda2 JH |
6872 | * finds the closest node not already in the @used_nodes map. |
6873 | * | |
6874 | * Should use nodemask_t. | |
6875 | */ | |
c5f59f08 | 6876 | static int find_next_best_node(int node, nodemask_t *used_nodes) |
9c1cfda2 JH |
6877 | { |
6878 | int i, n, val, min_val, best_node = 0; | |
6879 | ||
6880 | min_val = INT_MAX; | |
6881 | ||
6882 | for (i = 0; i < MAX_NUMNODES; i++) { | |
6883 | /* Start at @node */ | |
6884 | n = (node + i) % MAX_NUMNODES; | |
6885 | ||
6886 | if (!nr_cpus_node(n)) | |
6887 | continue; | |
6888 | ||
6889 | /* Skip already used nodes */ | |
c5f59f08 | 6890 | if (node_isset(n, *used_nodes)) |
9c1cfda2 JH |
6891 | continue; |
6892 | ||
6893 | /* Simple min distance search */ | |
6894 | val = node_distance(node, n); | |
6895 | ||
6896 | if (val < min_val) { | |
6897 | min_val = val; | |
6898 | best_node = n; | |
6899 | } | |
6900 | } | |
6901 | ||
c5f59f08 | 6902 | node_set(best_node, *used_nodes); |
9c1cfda2 JH |
6903 | return best_node; |
6904 | } | |
6905 | ||
6906 | /** | |
6907 | * sched_domain_node_span - get a cpumask for a node's sched_domain | |
6908 | * @node: node whose cpumask we're constructing | |
73486722 | 6909 | * @span: resulting cpumask |
9c1cfda2 | 6910 | * |
41a2d6cf | 6911 | * Given a node, construct a good cpumask for its sched_domain to span. It |
9c1cfda2 JH |
6912 | * should be one that prevents unnecessary balancing, but also spreads tasks |
6913 | * out optimally. | |
6914 | */ | |
4bdbaad3 | 6915 | static void sched_domain_node_span(int node, cpumask_t *span) |
9c1cfda2 | 6916 | { |
c5f59f08 | 6917 | nodemask_t used_nodes; |
c5f59f08 | 6918 | node_to_cpumask_ptr(nodemask, node); |
48f24c4d | 6919 | int i; |
9c1cfda2 | 6920 | |
4bdbaad3 | 6921 | cpus_clear(*span); |
c5f59f08 | 6922 | nodes_clear(used_nodes); |
9c1cfda2 | 6923 | |
4bdbaad3 | 6924 | cpus_or(*span, *span, *nodemask); |
c5f59f08 | 6925 | node_set(node, used_nodes); |
9c1cfda2 JH |
6926 | |
6927 | for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { | |
c5f59f08 | 6928 | int next_node = find_next_best_node(node, &used_nodes); |
48f24c4d | 6929 | |
c5f59f08 | 6930 | node_to_cpumask_ptr_next(nodemask, next_node); |
4bdbaad3 | 6931 | cpus_or(*span, *span, *nodemask); |
9c1cfda2 | 6932 | } |
9c1cfda2 JH |
6933 | } |
6934 | #endif | |
6935 | ||
5c45bf27 | 6936 | int sched_smt_power_savings = 0, sched_mc_power_savings = 0; |
48f24c4d | 6937 | |
9c1cfda2 | 6938 | /* |
48f24c4d | 6939 | * SMT sched-domains: |
9c1cfda2 | 6940 | */ |
1da177e4 LT |
6941 | #ifdef CONFIG_SCHED_SMT |
6942 | static DEFINE_PER_CPU(struct sched_domain, cpu_domains); | |
6711cab4 | 6943 | static DEFINE_PER_CPU(struct sched_group, sched_group_cpus); |
48f24c4d | 6944 | |
41a2d6cf | 6945 | static int |
7c16ec58 MT |
6946 | cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6947 | cpumask_t *unused) | |
1da177e4 | 6948 | { |
6711cab4 SS |
6949 | if (sg) |
6950 | *sg = &per_cpu(sched_group_cpus, cpu); | |
1da177e4 LT |
6951 | return cpu; |
6952 | } | |
6953 | #endif | |
6954 | ||
48f24c4d IM |
6955 | /* |
6956 | * multi-core sched-domains: | |
6957 | */ | |
1e9f28fa SS |
6958 | #ifdef CONFIG_SCHED_MC |
6959 | static DEFINE_PER_CPU(struct sched_domain, core_domains); | |
6711cab4 | 6960 | static DEFINE_PER_CPU(struct sched_group, sched_group_core); |
1e9f28fa SS |
6961 | #endif |
6962 | ||
6963 | #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT) | |
41a2d6cf | 6964 | static int |
7c16ec58 MT |
6965 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6966 | cpumask_t *mask) | |
1e9f28fa | 6967 | { |
6711cab4 | 6968 | int group; |
7c16ec58 MT |
6969 | |
6970 | *mask = per_cpu(cpu_sibling_map, cpu); | |
6971 | cpus_and(*mask, *mask, *cpu_map); | |
6972 | group = first_cpu(*mask); | |
6711cab4 SS |
6973 | if (sg) |
6974 | *sg = &per_cpu(sched_group_core, group); | |
6975 | return group; | |
1e9f28fa SS |
6976 | } |
6977 | #elif defined(CONFIG_SCHED_MC) | |
41a2d6cf | 6978 | static int |
7c16ec58 MT |
6979 | cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6980 | cpumask_t *unused) | |
1e9f28fa | 6981 | { |
6711cab4 SS |
6982 | if (sg) |
6983 | *sg = &per_cpu(sched_group_core, cpu); | |
1e9f28fa SS |
6984 | return cpu; |
6985 | } | |
6986 | #endif | |
6987 | ||
1da177e4 | 6988 | static DEFINE_PER_CPU(struct sched_domain, phys_domains); |
6711cab4 | 6989 | static DEFINE_PER_CPU(struct sched_group, sched_group_phys); |
48f24c4d | 6990 | |
41a2d6cf | 6991 | static int |
7c16ec58 MT |
6992 | cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg, |
6993 | cpumask_t *mask) | |
1da177e4 | 6994 | { |
6711cab4 | 6995 | int group; |
48f24c4d | 6996 | #ifdef CONFIG_SCHED_MC |
7c16ec58 MT |
6997 | *mask = cpu_coregroup_map(cpu); |
6998 | cpus_and(*mask, *mask, *cpu_map); | |
6999 | group = first_cpu(*mask); | |
1e9f28fa | 7000 | #elif defined(CONFIG_SCHED_SMT) |
7c16ec58 MT |
7001 | *mask = per_cpu(cpu_sibling_map, cpu); |
7002 | cpus_and(*mask, *mask, *cpu_map); | |
7003 | group = first_cpu(*mask); | |
1da177e4 | 7004 | #else |
6711cab4 | 7005 | group = cpu; |
1da177e4 | 7006 | #endif |
6711cab4 SS |
7007 | if (sg) |
7008 | *sg = &per_cpu(sched_group_phys, group); | |
7009 | return group; | |
1da177e4 LT |
7010 | } |
7011 | ||
7012 | #ifdef CONFIG_NUMA | |
1da177e4 | 7013 | /* |
9c1cfda2 JH |
7014 | * The init_sched_build_groups can't handle what we want to do with node |
7015 | * groups, so roll our own. Now each node has its own list of groups which | |
7016 | * gets dynamically allocated. | |
1da177e4 | 7017 | */ |
9c1cfda2 | 7018 | static DEFINE_PER_CPU(struct sched_domain, node_domains); |
434d53b0 | 7019 | static struct sched_group ***sched_group_nodes_bycpu; |
1da177e4 | 7020 | |
9c1cfda2 | 7021 | static DEFINE_PER_CPU(struct sched_domain, allnodes_domains); |
6711cab4 | 7022 | static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes); |
9c1cfda2 | 7023 | |
6711cab4 | 7024 | static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map, |
7c16ec58 | 7025 | struct sched_group **sg, cpumask_t *nodemask) |
9c1cfda2 | 7026 | { |
6711cab4 SS |
7027 | int group; |
7028 | ||
7c16ec58 MT |
7029 | *nodemask = node_to_cpumask(cpu_to_node(cpu)); |
7030 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7031 | group = first_cpu(*nodemask); | |
6711cab4 SS |
7032 | |
7033 | if (sg) | |
7034 | *sg = &per_cpu(sched_group_allnodes, group); | |
7035 | return group; | |
1da177e4 | 7036 | } |
6711cab4 | 7037 | |
08069033 SS |
7038 | static void init_numa_sched_groups_power(struct sched_group *group_head) |
7039 | { | |
7040 | struct sched_group *sg = group_head; | |
7041 | int j; | |
7042 | ||
7043 | if (!sg) | |
7044 | return; | |
3a5c359a AK |
7045 | do { |
7046 | for_each_cpu_mask(j, sg->cpumask) { | |
7047 | struct sched_domain *sd; | |
08069033 | 7048 | |
3a5c359a AK |
7049 | sd = &per_cpu(phys_domains, j); |
7050 | if (j != first_cpu(sd->groups->cpumask)) { | |
7051 | /* | |
7052 | * Only add "power" once for each | |
7053 | * physical package. | |
7054 | */ | |
7055 | continue; | |
7056 | } | |
08069033 | 7057 | |
3a5c359a AK |
7058 | sg_inc_cpu_power(sg, sd->groups->__cpu_power); |
7059 | } | |
7060 | sg = sg->next; | |
7061 | } while (sg != group_head); | |
08069033 | 7062 | } |
1da177e4 LT |
7063 | #endif |
7064 | ||
a616058b | 7065 | #ifdef CONFIG_NUMA |
51888ca2 | 7066 | /* Free memory allocated for various sched_group structures */ |
7c16ec58 | 7067 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
51888ca2 | 7068 | { |
a616058b | 7069 | int cpu, i; |
51888ca2 SV |
7070 | |
7071 | for_each_cpu_mask(cpu, *cpu_map) { | |
51888ca2 SV |
7072 | struct sched_group **sched_group_nodes |
7073 | = sched_group_nodes_bycpu[cpu]; | |
7074 | ||
51888ca2 SV |
7075 | if (!sched_group_nodes) |
7076 | continue; | |
7077 | ||
7078 | for (i = 0; i < MAX_NUMNODES; i++) { | |
51888ca2 SV |
7079 | struct sched_group *oldsg, *sg = sched_group_nodes[i]; |
7080 | ||
7c16ec58 MT |
7081 | *nodemask = node_to_cpumask(i); |
7082 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7083 | if (cpus_empty(*nodemask)) | |
51888ca2 SV |
7084 | continue; |
7085 | ||
7086 | if (sg == NULL) | |
7087 | continue; | |
7088 | sg = sg->next; | |
7089 | next_sg: | |
7090 | oldsg = sg; | |
7091 | sg = sg->next; | |
7092 | kfree(oldsg); | |
7093 | if (oldsg != sched_group_nodes[i]) | |
7094 | goto next_sg; | |
7095 | } | |
7096 | kfree(sched_group_nodes); | |
7097 | sched_group_nodes_bycpu[cpu] = NULL; | |
7098 | } | |
51888ca2 | 7099 | } |
a616058b | 7100 | #else |
7c16ec58 | 7101 | static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask) |
a616058b SS |
7102 | { |
7103 | } | |
7104 | #endif | |
51888ca2 | 7105 | |
89c4710e SS |
7106 | /* |
7107 | * Initialize sched groups cpu_power. | |
7108 | * | |
7109 | * cpu_power indicates the capacity of sched group, which is used while | |
7110 | * distributing the load between different sched groups in a sched domain. | |
7111 | * Typically cpu_power for all the groups in a sched domain will be same unless | |
7112 | * there are asymmetries in the topology. If there are asymmetries, group | |
7113 | * having more cpu_power will pickup more load compared to the group having | |
7114 | * less cpu_power. | |
7115 | * | |
7116 | * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents | |
7117 | * the maximum number of tasks a group can handle in the presence of other idle | |
7118 | * or lightly loaded groups in the same sched domain. | |
7119 | */ | |
7120 | static void init_sched_groups_power(int cpu, struct sched_domain *sd) | |
7121 | { | |
7122 | struct sched_domain *child; | |
7123 | struct sched_group *group; | |
7124 | ||
7125 | WARN_ON(!sd || !sd->groups); | |
7126 | ||
7127 | if (cpu != first_cpu(sd->groups->cpumask)) | |
7128 | return; | |
7129 | ||
7130 | child = sd->child; | |
7131 | ||
5517d86b ED |
7132 | sd->groups->__cpu_power = 0; |
7133 | ||
89c4710e SS |
7134 | /* |
7135 | * For perf policy, if the groups in child domain share resources | |
7136 | * (for example cores sharing some portions of the cache hierarchy | |
7137 | * or SMT), then set this domain groups cpu_power such that each group | |
7138 | * can handle only one task, when there are other idle groups in the | |
7139 | * same sched domain. | |
7140 | */ | |
7141 | if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) && | |
7142 | (child->flags & | |
7143 | (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) { | |
5517d86b | 7144 | sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE); |
89c4710e SS |
7145 | return; |
7146 | } | |
7147 | ||
89c4710e SS |
7148 | /* |
7149 | * add cpu_power of each child group to this groups cpu_power | |
7150 | */ | |
7151 | group = child->groups; | |
7152 | do { | |
5517d86b | 7153 | sg_inc_cpu_power(sd->groups, group->__cpu_power); |
89c4710e SS |
7154 | group = group->next; |
7155 | } while (group != child->groups); | |
7156 | } | |
7157 | ||
7c16ec58 MT |
7158 | /* |
7159 | * Initializers for schedule domains | |
7160 | * Non-inlined to reduce accumulated stack pressure in build_sched_domains() | |
7161 | */ | |
7162 | ||
7163 | #define SD_INIT(sd, type) sd_init_##type(sd) | |
7164 | #define SD_INIT_FUNC(type) \ | |
7165 | static noinline void sd_init_##type(struct sched_domain *sd) \ | |
7166 | { \ | |
7167 | memset(sd, 0, sizeof(*sd)); \ | |
7168 | *sd = SD_##type##_INIT; \ | |
1d3504fc | 7169 | sd->level = SD_LV_##type; \ |
7c16ec58 MT |
7170 | } |
7171 | ||
7172 | SD_INIT_FUNC(CPU) | |
7173 | #ifdef CONFIG_NUMA | |
7174 | SD_INIT_FUNC(ALLNODES) | |
7175 | SD_INIT_FUNC(NODE) | |
7176 | #endif | |
7177 | #ifdef CONFIG_SCHED_SMT | |
7178 | SD_INIT_FUNC(SIBLING) | |
7179 | #endif | |
7180 | #ifdef CONFIG_SCHED_MC | |
7181 | SD_INIT_FUNC(MC) | |
7182 | #endif | |
7183 | ||
7184 | /* | |
7185 | * To minimize stack usage kmalloc room for cpumasks and share the | |
7186 | * space as the usage in build_sched_domains() dictates. Used only | |
7187 | * if the amount of space is significant. | |
7188 | */ | |
7189 | struct allmasks { | |
7190 | cpumask_t tmpmask; /* make this one first */ | |
7191 | union { | |
7192 | cpumask_t nodemask; | |
7193 | cpumask_t this_sibling_map; | |
7194 | cpumask_t this_core_map; | |
7195 | }; | |
7196 | cpumask_t send_covered; | |
7197 | ||
7198 | #ifdef CONFIG_NUMA | |
7199 | cpumask_t domainspan; | |
7200 | cpumask_t covered; | |
7201 | cpumask_t notcovered; | |
7202 | #endif | |
7203 | }; | |
7204 | ||
7205 | #if NR_CPUS > 128 | |
7206 | #define SCHED_CPUMASK_ALLOC 1 | |
7207 | #define SCHED_CPUMASK_FREE(v) kfree(v) | |
7208 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v | |
7209 | #else | |
7210 | #define SCHED_CPUMASK_ALLOC 0 | |
7211 | #define SCHED_CPUMASK_FREE(v) | |
7212 | #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v | |
7213 | #endif | |
7214 | ||
7215 | #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \ | |
7216 | ((unsigned long)(a) + offsetof(struct allmasks, v)) | |
7217 | ||
1d3504fc HS |
7218 | static int default_relax_domain_level = -1; |
7219 | ||
7220 | static int __init setup_relax_domain_level(char *str) | |
7221 | { | |
7222 | default_relax_domain_level = simple_strtoul(str, NULL, 0); | |
7223 | return 1; | |
7224 | } | |
7225 | __setup("relax_domain_level=", setup_relax_domain_level); | |
7226 | ||
7227 | static void set_domain_attribute(struct sched_domain *sd, | |
7228 | struct sched_domain_attr *attr) | |
7229 | { | |
7230 | int request; | |
7231 | ||
7232 | if (!attr || attr->relax_domain_level < 0) { | |
7233 | if (default_relax_domain_level < 0) | |
7234 | return; | |
7235 | else | |
7236 | request = default_relax_domain_level; | |
7237 | } else | |
7238 | request = attr->relax_domain_level; | |
7239 | if (request < sd->level) { | |
7240 | /* turn off idle balance on this domain */ | |
7241 | sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE); | |
7242 | } else { | |
7243 | /* turn on idle balance on this domain */ | |
7244 | sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE); | |
7245 | } | |
7246 | } | |
7247 | ||
1da177e4 | 7248 | /* |
1a20ff27 DG |
7249 | * Build sched domains for a given set of cpus and attach the sched domains |
7250 | * to the individual cpus | |
1da177e4 | 7251 | */ |
1d3504fc HS |
7252 | static int __build_sched_domains(const cpumask_t *cpu_map, |
7253 | struct sched_domain_attr *attr) | |
1da177e4 LT |
7254 | { |
7255 | int i; | |
57d885fe | 7256 | struct root_domain *rd; |
7c16ec58 MT |
7257 | SCHED_CPUMASK_DECLARE(allmasks); |
7258 | cpumask_t *tmpmask; | |
d1b55138 JH |
7259 | #ifdef CONFIG_NUMA |
7260 | struct sched_group **sched_group_nodes = NULL; | |
6711cab4 | 7261 | int sd_allnodes = 0; |
d1b55138 JH |
7262 | |
7263 | /* | |
7264 | * Allocate the per-node list of sched groups | |
7265 | */ | |
5cf9f062 | 7266 | sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *), |
41a2d6cf | 7267 | GFP_KERNEL); |
d1b55138 JH |
7268 | if (!sched_group_nodes) { |
7269 | printk(KERN_WARNING "Can not alloc sched group node list\n"); | |
51888ca2 | 7270 | return -ENOMEM; |
d1b55138 | 7271 | } |
d1b55138 | 7272 | #endif |
1da177e4 | 7273 | |
dc938520 | 7274 | rd = alloc_rootdomain(); |
57d885fe GH |
7275 | if (!rd) { |
7276 | printk(KERN_WARNING "Cannot alloc root domain\n"); | |
7c16ec58 MT |
7277 | #ifdef CONFIG_NUMA |
7278 | kfree(sched_group_nodes); | |
7279 | #endif | |
57d885fe GH |
7280 | return -ENOMEM; |
7281 | } | |
7282 | ||
7c16ec58 MT |
7283 | #if SCHED_CPUMASK_ALLOC |
7284 | /* get space for all scratch cpumask variables */ | |
7285 | allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL); | |
7286 | if (!allmasks) { | |
7287 | printk(KERN_WARNING "Cannot alloc cpumask array\n"); | |
7288 | kfree(rd); | |
7289 | #ifdef CONFIG_NUMA | |
7290 | kfree(sched_group_nodes); | |
7291 | #endif | |
7292 | return -ENOMEM; | |
7293 | } | |
7294 | #endif | |
7295 | tmpmask = (cpumask_t *)allmasks; | |
7296 | ||
7297 | ||
7298 | #ifdef CONFIG_NUMA | |
7299 | sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes; | |
7300 | #endif | |
7301 | ||
1da177e4 | 7302 | /* |
1a20ff27 | 7303 | * Set up domains for cpus specified by the cpu_map. |
1da177e4 | 7304 | */ |
1a20ff27 | 7305 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 | 7306 | struct sched_domain *sd = NULL, *p; |
7c16ec58 | 7307 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
1da177e4 | 7308 | |
7c16ec58 MT |
7309 | *nodemask = node_to_cpumask(cpu_to_node(i)); |
7310 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
1da177e4 LT |
7311 | |
7312 | #ifdef CONFIG_NUMA | |
dd41f596 | 7313 | if (cpus_weight(*cpu_map) > |
7c16ec58 | 7314 | SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) { |
9c1cfda2 | 7315 | sd = &per_cpu(allnodes_domains, i); |
7c16ec58 | 7316 | SD_INIT(sd, ALLNODES); |
1d3504fc | 7317 | set_domain_attribute(sd, attr); |
9c1cfda2 | 7318 | sd->span = *cpu_map; |
18d95a28 | 7319 | sd->first_cpu = first_cpu(sd->span); |
7c16ec58 | 7320 | cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask); |
9c1cfda2 | 7321 | p = sd; |
6711cab4 | 7322 | sd_allnodes = 1; |
9c1cfda2 JH |
7323 | } else |
7324 | p = NULL; | |
7325 | ||
1da177e4 | 7326 | sd = &per_cpu(node_domains, i); |
7c16ec58 | 7327 | SD_INIT(sd, NODE); |
1d3504fc | 7328 | set_domain_attribute(sd, attr); |
4bdbaad3 | 7329 | sched_domain_node_span(cpu_to_node(i), &sd->span); |
18d95a28 | 7330 | sd->first_cpu = first_cpu(sd->span); |
9c1cfda2 | 7331 | sd->parent = p; |
1a848870 SS |
7332 | if (p) |
7333 | p->child = sd; | |
9c1cfda2 | 7334 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 LT |
7335 | #endif |
7336 | ||
7337 | p = sd; | |
7338 | sd = &per_cpu(phys_domains, i); | |
7c16ec58 | 7339 | SD_INIT(sd, CPU); |
1d3504fc | 7340 | set_domain_attribute(sd, attr); |
7c16ec58 | 7341 | sd->span = *nodemask; |
18d95a28 | 7342 | sd->first_cpu = first_cpu(sd->span); |
1da177e4 | 7343 | sd->parent = p; |
1a848870 SS |
7344 | if (p) |
7345 | p->child = sd; | |
7c16ec58 | 7346 | cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 | 7347 | |
1e9f28fa SS |
7348 | #ifdef CONFIG_SCHED_MC |
7349 | p = sd; | |
7350 | sd = &per_cpu(core_domains, i); | |
7c16ec58 | 7351 | SD_INIT(sd, MC); |
1d3504fc | 7352 | set_domain_attribute(sd, attr); |
1e9f28fa | 7353 | sd->span = cpu_coregroup_map(i); |
18d95a28 | 7354 | sd->first_cpu = first_cpu(sd->span); |
1e9f28fa SS |
7355 | cpus_and(sd->span, sd->span, *cpu_map); |
7356 | sd->parent = p; | |
1a848870 | 7357 | p->child = sd; |
7c16ec58 | 7358 | cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask); |
1e9f28fa SS |
7359 | #endif |
7360 | ||
1da177e4 LT |
7361 | #ifdef CONFIG_SCHED_SMT |
7362 | p = sd; | |
7363 | sd = &per_cpu(cpu_domains, i); | |
7c16ec58 | 7364 | SD_INIT(sd, SIBLING); |
1d3504fc | 7365 | set_domain_attribute(sd, attr); |
d5a7430d | 7366 | sd->span = per_cpu(cpu_sibling_map, i); |
18d95a28 | 7367 | sd->first_cpu = first_cpu(sd->span); |
1a20ff27 | 7368 | cpus_and(sd->span, sd->span, *cpu_map); |
1da177e4 | 7369 | sd->parent = p; |
1a848870 | 7370 | p->child = sd; |
7c16ec58 | 7371 | cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask); |
1da177e4 LT |
7372 | #endif |
7373 | } | |
7374 | ||
7375 | #ifdef CONFIG_SCHED_SMT | |
7376 | /* Set up CPU (sibling) groups */ | |
9c1cfda2 | 7377 | for_each_cpu_mask(i, *cpu_map) { |
7c16ec58 MT |
7378 | SCHED_CPUMASK_VAR(this_sibling_map, allmasks); |
7379 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7380 | ||
7381 | *this_sibling_map = per_cpu(cpu_sibling_map, i); | |
7382 | cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map); | |
7383 | if (i != first_cpu(*this_sibling_map)) | |
1da177e4 LT |
7384 | continue; |
7385 | ||
dd41f596 | 7386 | init_sched_build_groups(this_sibling_map, cpu_map, |
7c16ec58 MT |
7387 | &cpu_to_cpu_group, |
7388 | send_covered, tmpmask); | |
1da177e4 LT |
7389 | } |
7390 | #endif | |
7391 | ||
1e9f28fa SS |
7392 | #ifdef CONFIG_SCHED_MC |
7393 | /* Set up multi-core groups */ | |
7394 | for_each_cpu_mask(i, *cpu_map) { | |
7c16ec58 MT |
7395 | SCHED_CPUMASK_VAR(this_core_map, allmasks); |
7396 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7397 | ||
7398 | *this_core_map = cpu_coregroup_map(i); | |
7399 | cpus_and(*this_core_map, *this_core_map, *cpu_map); | |
7400 | if (i != first_cpu(*this_core_map)) | |
1e9f28fa | 7401 | continue; |
7c16ec58 | 7402 | |
dd41f596 | 7403 | init_sched_build_groups(this_core_map, cpu_map, |
7c16ec58 MT |
7404 | &cpu_to_core_group, |
7405 | send_covered, tmpmask); | |
1e9f28fa SS |
7406 | } |
7407 | #endif | |
7408 | ||
1da177e4 LT |
7409 | /* Set up physical groups */ |
7410 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7c16ec58 MT |
7411 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7412 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
1da177e4 | 7413 | |
7c16ec58 MT |
7414 | *nodemask = node_to_cpumask(i); |
7415 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7416 | if (cpus_empty(*nodemask)) | |
1da177e4 LT |
7417 | continue; |
7418 | ||
7c16ec58 MT |
7419 | init_sched_build_groups(nodemask, cpu_map, |
7420 | &cpu_to_phys_group, | |
7421 | send_covered, tmpmask); | |
1da177e4 LT |
7422 | } |
7423 | ||
7424 | #ifdef CONFIG_NUMA | |
7425 | /* Set up node groups */ | |
7c16ec58 MT |
7426 | if (sd_allnodes) { |
7427 | SCHED_CPUMASK_VAR(send_covered, allmasks); | |
7428 | ||
7429 | init_sched_build_groups(cpu_map, cpu_map, | |
7430 | &cpu_to_allnodes_group, | |
7431 | send_covered, tmpmask); | |
7432 | } | |
9c1cfda2 JH |
7433 | |
7434 | for (i = 0; i < MAX_NUMNODES; i++) { | |
7435 | /* Set up node groups */ | |
7436 | struct sched_group *sg, *prev; | |
7c16ec58 MT |
7437 | SCHED_CPUMASK_VAR(nodemask, allmasks); |
7438 | SCHED_CPUMASK_VAR(domainspan, allmasks); | |
7439 | SCHED_CPUMASK_VAR(covered, allmasks); | |
9c1cfda2 JH |
7440 | int j; |
7441 | ||
7c16ec58 MT |
7442 | *nodemask = node_to_cpumask(i); |
7443 | cpus_clear(*covered); | |
7444 | ||
7445 | cpus_and(*nodemask, *nodemask, *cpu_map); | |
7446 | if (cpus_empty(*nodemask)) { | |
d1b55138 | 7447 | sched_group_nodes[i] = NULL; |
9c1cfda2 | 7448 | continue; |
d1b55138 | 7449 | } |
9c1cfda2 | 7450 | |
4bdbaad3 | 7451 | sched_domain_node_span(i, domainspan); |
7c16ec58 | 7452 | cpus_and(*domainspan, *domainspan, *cpu_map); |
9c1cfda2 | 7453 | |
15f0b676 | 7454 | sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i); |
51888ca2 SV |
7455 | if (!sg) { |
7456 | printk(KERN_WARNING "Can not alloc domain group for " | |
7457 | "node %d\n", i); | |
7458 | goto error; | |
7459 | } | |
9c1cfda2 | 7460 | sched_group_nodes[i] = sg; |
7c16ec58 | 7461 | for_each_cpu_mask(j, *nodemask) { |
9c1cfda2 | 7462 | struct sched_domain *sd; |
9761eea8 | 7463 | |
9c1cfda2 JH |
7464 | sd = &per_cpu(node_domains, j); |
7465 | sd->groups = sg; | |
9c1cfda2 | 7466 | } |
5517d86b | 7467 | sg->__cpu_power = 0; |
7c16ec58 | 7468 | sg->cpumask = *nodemask; |
51888ca2 | 7469 | sg->next = sg; |
7c16ec58 | 7470 | cpus_or(*covered, *covered, *nodemask); |
9c1cfda2 JH |
7471 | prev = sg; |
7472 | ||
7473 | for (j = 0; j < MAX_NUMNODES; j++) { | |
7c16ec58 | 7474 | SCHED_CPUMASK_VAR(notcovered, allmasks); |
9c1cfda2 | 7475 | int n = (i + j) % MAX_NUMNODES; |
c5f59f08 | 7476 | node_to_cpumask_ptr(pnodemask, n); |
9c1cfda2 | 7477 | |
7c16ec58 MT |
7478 | cpus_complement(*notcovered, *covered); |
7479 | cpus_and(*tmpmask, *notcovered, *cpu_map); | |
7480 | cpus_and(*tmpmask, *tmpmask, *domainspan); | |
7481 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7482 | break; |
7483 | ||
7c16ec58 MT |
7484 | cpus_and(*tmpmask, *tmpmask, *pnodemask); |
7485 | if (cpus_empty(*tmpmask)) | |
9c1cfda2 JH |
7486 | continue; |
7487 | ||
15f0b676 SV |
7488 | sg = kmalloc_node(sizeof(struct sched_group), |
7489 | GFP_KERNEL, i); | |
9c1cfda2 JH |
7490 | if (!sg) { |
7491 | printk(KERN_WARNING | |
7492 | "Can not alloc domain group for node %d\n", j); | |
51888ca2 | 7493 | goto error; |
9c1cfda2 | 7494 | } |
5517d86b | 7495 | sg->__cpu_power = 0; |
7c16ec58 | 7496 | sg->cpumask = *tmpmask; |
51888ca2 | 7497 | sg->next = prev->next; |
7c16ec58 | 7498 | cpus_or(*covered, *covered, *tmpmask); |
9c1cfda2 JH |
7499 | prev->next = sg; |
7500 | prev = sg; | |
7501 | } | |
9c1cfda2 | 7502 | } |
1da177e4 LT |
7503 | #endif |
7504 | ||
7505 | /* Calculate CPU power for physical packages and nodes */ | |
5c45bf27 | 7506 | #ifdef CONFIG_SCHED_SMT |
1a20ff27 | 7507 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7508 | struct sched_domain *sd = &per_cpu(cpu_domains, i); |
7509 | ||
89c4710e | 7510 | init_sched_groups_power(i, sd); |
5c45bf27 | 7511 | } |
1da177e4 | 7512 | #endif |
1e9f28fa | 7513 | #ifdef CONFIG_SCHED_MC |
5c45bf27 | 7514 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7515 | struct sched_domain *sd = &per_cpu(core_domains, i); |
7516 | ||
89c4710e | 7517 | init_sched_groups_power(i, sd); |
5c45bf27 SS |
7518 | } |
7519 | #endif | |
1e9f28fa | 7520 | |
5c45bf27 | 7521 | for_each_cpu_mask(i, *cpu_map) { |
dd41f596 IM |
7522 | struct sched_domain *sd = &per_cpu(phys_domains, i); |
7523 | ||
89c4710e | 7524 | init_sched_groups_power(i, sd); |
1da177e4 LT |
7525 | } |
7526 | ||
9c1cfda2 | 7527 | #ifdef CONFIG_NUMA |
08069033 SS |
7528 | for (i = 0; i < MAX_NUMNODES; i++) |
7529 | init_numa_sched_groups_power(sched_group_nodes[i]); | |
9c1cfda2 | 7530 | |
6711cab4 SS |
7531 | if (sd_allnodes) { |
7532 | struct sched_group *sg; | |
f712c0c7 | 7533 | |
7c16ec58 MT |
7534 | cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg, |
7535 | tmpmask); | |
f712c0c7 SS |
7536 | init_numa_sched_groups_power(sg); |
7537 | } | |
9c1cfda2 JH |
7538 | #endif |
7539 | ||
1da177e4 | 7540 | /* Attach the domains */ |
1a20ff27 | 7541 | for_each_cpu_mask(i, *cpu_map) { |
1da177e4 LT |
7542 | struct sched_domain *sd; |
7543 | #ifdef CONFIG_SCHED_SMT | |
7544 | sd = &per_cpu(cpu_domains, i); | |
1e9f28fa SS |
7545 | #elif defined(CONFIG_SCHED_MC) |
7546 | sd = &per_cpu(core_domains, i); | |
1da177e4 LT |
7547 | #else |
7548 | sd = &per_cpu(phys_domains, i); | |
7549 | #endif | |
57d885fe | 7550 | cpu_attach_domain(sd, rd, i); |
1da177e4 | 7551 | } |
51888ca2 | 7552 | |
7c16ec58 | 7553 | SCHED_CPUMASK_FREE((void *)allmasks); |
51888ca2 SV |
7554 | return 0; |
7555 | ||
a616058b | 7556 | #ifdef CONFIG_NUMA |
51888ca2 | 7557 | error: |
7c16ec58 MT |
7558 | free_sched_groups(cpu_map, tmpmask); |
7559 | SCHED_CPUMASK_FREE((void *)allmasks); | |
51888ca2 | 7560 | return -ENOMEM; |
a616058b | 7561 | #endif |
1da177e4 | 7562 | } |
029190c5 | 7563 | |
1d3504fc HS |
7564 | static int build_sched_domains(const cpumask_t *cpu_map) |
7565 | { | |
7566 | return __build_sched_domains(cpu_map, NULL); | |
7567 | } | |
7568 | ||
029190c5 PJ |
7569 | static cpumask_t *doms_cur; /* current sched domains */ |
7570 | static int ndoms_cur; /* number of sched domains in 'doms_cur' */ | |
1d3504fc HS |
7571 | static struct sched_domain_attr *dattr_cur; /* attribues of custom domains |
7572 | in 'doms_cur' */ | |
029190c5 PJ |
7573 | |
7574 | /* | |
7575 | * Special case: If a kmalloc of a doms_cur partition (array of | |
7576 | * cpumask_t) fails, then fallback to a single sched domain, | |
7577 | * as determined by the single cpumask_t fallback_doms. | |
7578 | */ | |
7579 | static cpumask_t fallback_doms; | |
7580 | ||
22e52b07 HC |
7581 | void __attribute__((weak)) arch_update_cpu_topology(void) |
7582 | { | |
7583 | } | |
7584 | ||
1a20ff27 | 7585 | /* |
41a2d6cf | 7586 | * Set up scheduler domains and groups. Callers must hold the hotplug lock. |
029190c5 PJ |
7587 | * For now this just excludes isolated cpus, but could be used to |
7588 | * exclude other special cases in the future. | |
1a20ff27 | 7589 | */ |
51888ca2 | 7590 | static int arch_init_sched_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7591 | { |
7378547f MM |
7592 | int err; |
7593 | ||
22e52b07 | 7594 | arch_update_cpu_topology(); |
029190c5 PJ |
7595 | ndoms_cur = 1; |
7596 | doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
7597 | if (!doms_cur) | |
7598 | doms_cur = &fallback_doms; | |
7599 | cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map); | |
1d3504fc | 7600 | dattr_cur = NULL; |
7378547f | 7601 | err = build_sched_domains(doms_cur); |
6382bc90 | 7602 | register_sched_domain_sysctl(); |
7378547f MM |
7603 | |
7604 | return err; | |
1a20ff27 DG |
7605 | } |
7606 | ||
7c16ec58 MT |
7607 | static void arch_destroy_sched_domains(const cpumask_t *cpu_map, |
7608 | cpumask_t *tmpmask) | |
1da177e4 | 7609 | { |
7c16ec58 | 7610 | free_sched_groups(cpu_map, tmpmask); |
9c1cfda2 | 7611 | } |
1da177e4 | 7612 | |
1a20ff27 DG |
7613 | /* |
7614 | * Detach sched domains from a group of cpus specified in cpu_map | |
7615 | * These cpus will now be attached to the NULL domain | |
7616 | */ | |
858119e1 | 7617 | static void detach_destroy_domains(const cpumask_t *cpu_map) |
1a20ff27 | 7618 | { |
7c16ec58 | 7619 | cpumask_t tmpmask; |
1a20ff27 DG |
7620 | int i; |
7621 | ||
6382bc90 MM |
7622 | unregister_sched_domain_sysctl(); |
7623 | ||
1a20ff27 | 7624 | for_each_cpu_mask(i, *cpu_map) |
57d885fe | 7625 | cpu_attach_domain(NULL, &def_root_domain, i); |
1a20ff27 | 7626 | synchronize_sched(); |
7c16ec58 | 7627 | arch_destroy_sched_domains(cpu_map, &tmpmask); |
1a20ff27 DG |
7628 | } |
7629 | ||
1d3504fc HS |
7630 | /* handle null as "default" */ |
7631 | static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, | |
7632 | struct sched_domain_attr *new, int idx_new) | |
7633 | { | |
7634 | struct sched_domain_attr tmp; | |
7635 | ||
7636 | /* fast path */ | |
7637 | if (!new && !cur) | |
7638 | return 1; | |
7639 | ||
7640 | tmp = SD_ATTR_INIT; | |
7641 | return !memcmp(cur ? (cur + idx_cur) : &tmp, | |
7642 | new ? (new + idx_new) : &tmp, | |
7643 | sizeof(struct sched_domain_attr)); | |
7644 | } | |
7645 | ||
029190c5 PJ |
7646 | /* |
7647 | * Partition sched domains as specified by the 'ndoms_new' | |
41a2d6cf | 7648 | * cpumasks in the array doms_new[] of cpumasks. This compares |
029190c5 PJ |
7649 | * doms_new[] to the current sched domain partitioning, doms_cur[]. |
7650 | * It destroys each deleted domain and builds each new domain. | |
7651 | * | |
7652 | * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'. | |
41a2d6cf IM |
7653 | * The masks don't intersect (don't overlap.) We should setup one |
7654 | * sched domain for each mask. CPUs not in any of the cpumasks will | |
7655 | * not be load balanced. If the same cpumask appears both in the | |
029190c5 PJ |
7656 | * current 'doms_cur' domains and in the new 'doms_new', we can leave |
7657 | * it as it is. | |
7658 | * | |
41a2d6cf IM |
7659 | * The passed in 'doms_new' should be kmalloc'd. This routine takes |
7660 | * ownership of it and will kfree it when done with it. If the caller | |
029190c5 PJ |
7661 | * failed the kmalloc call, then it can pass in doms_new == NULL, |
7662 | * and partition_sched_domains() will fallback to the single partition | |
7663 | * 'fallback_doms'. | |
7664 | * | |
7665 | * Call with hotplug lock held | |
7666 | */ | |
1d3504fc HS |
7667 | void partition_sched_domains(int ndoms_new, cpumask_t *doms_new, |
7668 | struct sched_domain_attr *dattr_new) | |
029190c5 PJ |
7669 | { |
7670 | int i, j; | |
7671 | ||
712555ee | 7672 | mutex_lock(&sched_domains_mutex); |
a1835615 | 7673 | |
7378547f MM |
7674 | /* always unregister in case we don't destroy any domains */ |
7675 | unregister_sched_domain_sysctl(); | |
7676 | ||
029190c5 PJ |
7677 | if (doms_new == NULL) { |
7678 | ndoms_new = 1; | |
7679 | doms_new = &fallback_doms; | |
7680 | cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map); | |
1d3504fc | 7681 | dattr_new = NULL; |
029190c5 PJ |
7682 | } |
7683 | ||
7684 | /* Destroy deleted domains */ | |
7685 | for (i = 0; i < ndoms_cur; i++) { | |
7686 | for (j = 0; j < ndoms_new; j++) { | |
1d3504fc HS |
7687 | if (cpus_equal(doms_cur[i], doms_new[j]) |
7688 | && dattrs_equal(dattr_cur, i, dattr_new, j)) | |
029190c5 PJ |
7689 | goto match1; |
7690 | } | |
7691 | /* no match - a current sched domain not in new doms_new[] */ | |
7692 | detach_destroy_domains(doms_cur + i); | |
7693 | match1: | |
7694 | ; | |
7695 | } | |
7696 | ||
7697 | /* Build new domains */ | |
7698 | for (i = 0; i < ndoms_new; i++) { | |
7699 | for (j = 0; j < ndoms_cur; j++) { | |
1d3504fc HS |
7700 | if (cpus_equal(doms_new[i], doms_cur[j]) |
7701 | && dattrs_equal(dattr_new, i, dattr_cur, j)) | |
029190c5 PJ |
7702 | goto match2; |
7703 | } | |
7704 | /* no match - add a new doms_new */ | |
1d3504fc HS |
7705 | __build_sched_domains(doms_new + i, |
7706 | dattr_new ? dattr_new + i : NULL); | |
029190c5 PJ |
7707 | match2: |
7708 | ; | |
7709 | } | |
7710 | ||
7711 | /* Remember the new sched domains */ | |
7712 | if (doms_cur != &fallback_doms) | |
7713 | kfree(doms_cur); | |
1d3504fc | 7714 | kfree(dattr_cur); /* kfree(NULL) is safe */ |
029190c5 | 7715 | doms_cur = doms_new; |
1d3504fc | 7716 | dattr_cur = dattr_new; |
029190c5 | 7717 | ndoms_cur = ndoms_new; |
7378547f MM |
7718 | |
7719 | register_sched_domain_sysctl(); | |
a1835615 | 7720 | |
712555ee | 7721 | mutex_unlock(&sched_domains_mutex); |
029190c5 PJ |
7722 | } |
7723 | ||
5c45bf27 | 7724 | #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) |
9aefd0ab | 7725 | int arch_reinit_sched_domains(void) |
5c45bf27 SS |
7726 | { |
7727 | int err; | |
7728 | ||
95402b38 | 7729 | get_online_cpus(); |
712555ee | 7730 | mutex_lock(&sched_domains_mutex); |
5c45bf27 SS |
7731 | detach_destroy_domains(&cpu_online_map); |
7732 | err = arch_init_sched_domains(&cpu_online_map); | |
712555ee | 7733 | mutex_unlock(&sched_domains_mutex); |
95402b38 | 7734 | put_online_cpus(); |
5c45bf27 SS |
7735 | |
7736 | return err; | |
7737 | } | |
7738 | ||
7739 | static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) | |
7740 | { | |
7741 | int ret; | |
7742 | ||
7743 | if (buf[0] != '0' && buf[0] != '1') | |
7744 | return -EINVAL; | |
7745 | ||
7746 | if (smt) | |
7747 | sched_smt_power_savings = (buf[0] == '1'); | |
7748 | else | |
7749 | sched_mc_power_savings = (buf[0] == '1'); | |
7750 | ||
7751 | ret = arch_reinit_sched_domains(); | |
7752 | ||
7753 | return ret ? ret : count; | |
7754 | } | |
7755 | ||
5c45bf27 SS |
7756 | #ifdef CONFIG_SCHED_MC |
7757 | static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page) | |
7758 | { | |
7759 | return sprintf(page, "%u\n", sched_mc_power_savings); | |
7760 | } | |
48f24c4d IM |
7761 | static ssize_t sched_mc_power_savings_store(struct sys_device *dev, |
7762 | const char *buf, size_t count) | |
5c45bf27 SS |
7763 | { |
7764 | return sched_power_savings_store(buf, count, 0); | |
7765 | } | |
6707de00 AB |
7766 | static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show, |
7767 | sched_mc_power_savings_store); | |
5c45bf27 SS |
7768 | #endif |
7769 | ||
7770 | #ifdef CONFIG_SCHED_SMT | |
7771 | static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page) | |
7772 | { | |
7773 | return sprintf(page, "%u\n", sched_smt_power_savings); | |
7774 | } | |
48f24c4d IM |
7775 | static ssize_t sched_smt_power_savings_store(struct sys_device *dev, |
7776 | const char *buf, size_t count) | |
5c45bf27 SS |
7777 | { |
7778 | return sched_power_savings_store(buf, count, 1); | |
7779 | } | |
6707de00 AB |
7780 | static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show, |
7781 | sched_smt_power_savings_store); | |
7782 | #endif | |
7783 | ||
7784 | int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls) | |
7785 | { | |
7786 | int err = 0; | |
7787 | ||
7788 | #ifdef CONFIG_SCHED_SMT | |
7789 | if (smt_capable()) | |
7790 | err = sysfs_create_file(&cls->kset.kobj, | |
7791 | &attr_sched_smt_power_savings.attr); | |
7792 | #endif | |
7793 | #ifdef CONFIG_SCHED_MC | |
7794 | if (!err && mc_capable()) | |
7795 | err = sysfs_create_file(&cls->kset.kobj, | |
7796 | &attr_sched_mc_power_savings.attr); | |
7797 | #endif | |
7798 | return err; | |
7799 | } | |
5c45bf27 SS |
7800 | #endif |
7801 | ||
1da177e4 | 7802 | /* |
41a2d6cf | 7803 | * Force a reinitialization of the sched domains hierarchy. The domains |
1da177e4 | 7804 | * and groups cannot be updated in place without racing with the balancing |
41c7ce9a | 7805 | * code, so we temporarily attach all running cpus to the NULL domain |
1da177e4 LT |
7806 | * which will prevent rebalancing while the sched domains are recalculated. |
7807 | */ | |
7808 | static int update_sched_domains(struct notifier_block *nfb, | |
7809 | unsigned long action, void *hcpu) | |
7810 | { | |
1da177e4 LT |
7811 | switch (action) { |
7812 | case CPU_UP_PREPARE: | |
8bb78442 | 7813 | case CPU_UP_PREPARE_FROZEN: |
1da177e4 | 7814 | case CPU_DOWN_PREPARE: |
8bb78442 | 7815 | case CPU_DOWN_PREPARE_FROZEN: |
1a20ff27 | 7816 | detach_destroy_domains(&cpu_online_map); |
1da177e4 LT |
7817 | return NOTIFY_OK; |
7818 | ||
7819 | case CPU_UP_CANCELED: | |
8bb78442 | 7820 | case CPU_UP_CANCELED_FROZEN: |
1da177e4 | 7821 | case CPU_DOWN_FAILED: |
8bb78442 | 7822 | case CPU_DOWN_FAILED_FROZEN: |
1da177e4 | 7823 | case CPU_ONLINE: |
8bb78442 | 7824 | case CPU_ONLINE_FROZEN: |
1da177e4 | 7825 | case CPU_DEAD: |
8bb78442 | 7826 | case CPU_DEAD_FROZEN: |
1da177e4 LT |
7827 | /* |
7828 | * Fall through and re-initialise the domains. | |
7829 | */ | |
7830 | break; | |
7831 | default: | |
7832 | return NOTIFY_DONE; | |
7833 | } | |
7834 | ||
7835 | /* The hotplug lock is already held by cpu_up/cpu_down */ | |
1a20ff27 | 7836 | arch_init_sched_domains(&cpu_online_map); |
1da177e4 LT |
7837 | |
7838 | return NOTIFY_OK; | |
7839 | } | |
1da177e4 LT |
7840 | |
7841 | void __init sched_init_smp(void) | |
7842 | { | |
5c1e1767 NP |
7843 | cpumask_t non_isolated_cpus; |
7844 | ||
434d53b0 MT |
7845 | #if defined(CONFIG_NUMA) |
7846 | sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **), | |
7847 | GFP_KERNEL); | |
7848 | BUG_ON(sched_group_nodes_bycpu == NULL); | |
7849 | #endif | |
95402b38 | 7850 | get_online_cpus(); |
712555ee | 7851 | mutex_lock(&sched_domains_mutex); |
1a20ff27 | 7852 | arch_init_sched_domains(&cpu_online_map); |
e5e5673f | 7853 | cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map); |
5c1e1767 NP |
7854 | if (cpus_empty(non_isolated_cpus)) |
7855 | cpu_set(smp_processor_id(), non_isolated_cpus); | |
712555ee | 7856 | mutex_unlock(&sched_domains_mutex); |
95402b38 | 7857 | put_online_cpus(); |
1da177e4 LT |
7858 | /* XXX: Theoretical race here - CPU may be hotplugged now */ |
7859 | hotcpu_notifier(update_sched_domains, 0); | |
b328ca18 | 7860 | init_hrtick(); |
5c1e1767 NP |
7861 | |
7862 | /* Move init over to a non-isolated CPU */ | |
7c16ec58 | 7863 | if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0) |
5c1e1767 | 7864 | BUG(); |
19978ca6 | 7865 | sched_init_granularity(); |
1da177e4 LT |
7866 | } |
7867 | #else | |
7868 | void __init sched_init_smp(void) | |
7869 | { | |
19978ca6 | 7870 | sched_init_granularity(); |
1da177e4 LT |
7871 | } |
7872 | #endif /* CONFIG_SMP */ | |
7873 | ||
7874 | int in_sched_functions(unsigned long addr) | |
7875 | { | |
1da177e4 LT |
7876 | return in_lock_functions(addr) || |
7877 | (addr >= (unsigned long)__sched_text_start | |
7878 | && addr < (unsigned long)__sched_text_end); | |
7879 | } | |
7880 | ||
a9957449 | 7881 | static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq) |
dd41f596 IM |
7882 | { |
7883 | cfs_rq->tasks_timeline = RB_ROOT; | |
4a55bd5e | 7884 | INIT_LIST_HEAD(&cfs_rq->tasks); |
dd41f596 IM |
7885 | #ifdef CONFIG_FAIR_GROUP_SCHED |
7886 | cfs_rq->rq = rq; | |
7887 | #endif | |
67e9fb2a | 7888 | cfs_rq->min_vruntime = (u64)(-(1LL << 20)); |
dd41f596 IM |
7889 | } |
7890 | ||
fa85ae24 PZ |
7891 | static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq) |
7892 | { | |
7893 | struct rt_prio_array *array; | |
7894 | int i; | |
7895 | ||
7896 | array = &rt_rq->active; | |
7897 | for (i = 0; i < MAX_RT_PRIO; i++) { | |
7898 | INIT_LIST_HEAD(array->queue + i); | |
7899 | __clear_bit(i, array->bitmap); | |
7900 | } | |
7901 | /* delimiter for bitsearch: */ | |
7902 | __set_bit(MAX_RT_PRIO, array->bitmap); | |
7903 | ||
052f1dc7 | 7904 | #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED |
48d5e258 PZ |
7905 | rt_rq->highest_prio = MAX_RT_PRIO; |
7906 | #endif | |
fa85ae24 PZ |
7907 | #ifdef CONFIG_SMP |
7908 | rt_rq->rt_nr_migratory = 0; | |
fa85ae24 PZ |
7909 | rt_rq->overloaded = 0; |
7910 | #endif | |
7911 | ||
7912 | rt_rq->rt_time = 0; | |
7913 | rt_rq->rt_throttled = 0; | |
ac086bc2 PZ |
7914 | rt_rq->rt_runtime = 0; |
7915 | spin_lock_init(&rt_rq->rt_runtime_lock); | |
6f505b16 | 7916 | |
052f1dc7 | 7917 | #ifdef CONFIG_RT_GROUP_SCHED |
23b0fdfc | 7918 | rt_rq->rt_nr_boosted = 0; |
6f505b16 PZ |
7919 | rt_rq->rq = rq; |
7920 | #endif | |
fa85ae24 PZ |
7921 | } |
7922 | ||
6f505b16 | 7923 | #ifdef CONFIG_FAIR_GROUP_SCHED |
ec7dc8ac DG |
7924 | static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq, |
7925 | struct sched_entity *se, int cpu, int add, | |
7926 | struct sched_entity *parent) | |
6f505b16 | 7927 | { |
ec7dc8ac | 7928 | struct rq *rq = cpu_rq(cpu); |
6f505b16 PZ |
7929 | tg->cfs_rq[cpu] = cfs_rq; |
7930 | init_cfs_rq(cfs_rq, rq); | |
7931 | cfs_rq->tg = tg; | |
7932 | if (add) | |
7933 | list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list); | |
7934 | ||
7935 | tg->se[cpu] = se; | |
354d60c2 DG |
7936 | /* se could be NULL for init_task_group */ |
7937 | if (!se) | |
7938 | return; | |
7939 | ||
ec7dc8ac DG |
7940 | if (!parent) |
7941 | se->cfs_rq = &rq->cfs; | |
7942 | else | |
7943 | se->cfs_rq = parent->my_q; | |
7944 | ||
6f505b16 PZ |
7945 | se->my_q = cfs_rq; |
7946 | se->load.weight = tg->shares; | |
e05510d0 | 7947 | se->load.inv_weight = 0; |
ec7dc8ac | 7948 | se->parent = parent; |
6f505b16 | 7949 | } |
052f1dc7 | 7950 | #endif |
6f505b16 | 7951 | |
052f1dc7 | 7952 | #ifdef CONFIG_RT_GROUP_SCHED |
ec7dc8ac DG |
7953 | static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq, |
7954 | struct sched_rt_entity *rt_se, int cpu, int add, | |
7955 | struct sched_rt_entity *parent) | |
6f505b16 | 7956 | { |
ec7dc8ac DG |
7957 | struct rq *rq = cpu_rq(cpu); |
7958 | ||
6f505b16 PZ |
7959 | tg->rt_rq[cpu] = rt_rq; |
7960 | init_rt_rq(rt_rq, rq); | |
7961 | rt_rq->tg = tg; | |
7962 | rt_rq->rt_se = rt_se; | |
ac086bc2 | 7963 | rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime; |
6f505b16 PZ |
7964 | if (add) |
7965 | list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list); | |
7966 | ||
7967 | tg->rt_se[cpu] = rt_se; | |
354d60c2 DG |
7968 | if (!rt_se) |
7969 | return; | |
7970 | ||
ec7dc8ac DG |
7971 | if (!parent) |
7972 | rt_se->rt_rq = &rq->rt; | |
7973 | else | |
7974 | rt_se->rt_rq = parent->my_q; | |
7975 | ||
6f505b16 PZ |
7976 | rt_se->rt_rq = &rq->rt; |
7977 | rt_se->my_q = rt_rq; | |
ec7dc8ac | 7978 | rt_se->parent = parent; |
6f505b16 PZ |
7979 | INIT_LIST_HEAD(&rt_se->run_list); |
7980 | } | |
7981 | #endif | |
7982 | ||
1da177e4 LT |
7983 | void __init sched_init(void) |
7984 | { | |
dd41f596 | 7985 | int i, j; |
434d53b0 MT |
7986 | unsigned long alloc_size = 0, ptr; |
7987 | ||
7988 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
7989 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
7990 | #endif | |
7991 | #ifdef CONFIG_RT_GROUP_SCHED | |
7992 | alloc_size += 2 * nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
7993 | #endif |
7994 | #ifdef CONFIG_USER_SCHED | |
7995 | alloc_size *= 2; | |
434d53b0 MT |
7996 | #endif |
7997 | /* | |
7998 | * As sched_init() is called before page_alloc is setup, | |
7999 | * we use alloc_bootmem(). | |
8000 | */ | |
8001 | if (alloc_size) { | |
5a9d3225 | 8002 | ptr = (unsigned long)alloc_bootmem(alloc_size); |
434d53b0 MT |
8003 | |
8004 | #ifdef CONFIG_FAIR_GROUP_SCHED | |
8005 | init_task_group.se = (struct sched_entity **)ptr; | |
8006 | ptr += nr_cpu_ids * sizeof(void **); | |
8007 | ||
8008 | init_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8009 | ptr += nr_cpu_ids * sizeof(void **); | |
eff766a6 PZ |
8010 | |
8011 | #ifdef CONFIG_USER_SCHED | |
8012 | root_task_group.se = (struct sched_entity **)ptr; | |
8013 | ptr += nr_cpu_ids * sizeof(void **); | |
8014 | ||
8015 | root_task_group.cfs_rq = (struct cfs_rq **)ptr; | |
8016 | ptr += nr_cpu_ids * sizeof(void **); | |
8017 | #endif | |
434d53b0 MT |
8018 | #endif |
8019 | #ifdef CONFIG_RT_GROUP_SCHED | |
8020 | init_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8021 | ptr += nr_cpu_ids * sizeof(void **); | |
8022 | ||
8023 | init_task_group.rt_rq = (struct rt_rq **)ptr; | |
eff766a6 PZ |
8024 | ptr += nr_cpu_ids * sizeof(void **); |
8025 | ||
8026 | #ifdef CONFIG_USER_SCHED | |
8027 | root_task_group.rt_se = (struct sched_rt_entity **)ptr; | |
8028 | ptr += nr_cpu_ids * sizeof(void **); | |
8029 | ||
8030 | root_task_group.rt_rq = (struct rt_rq **)ptr; | |
8031 | ptr += nr_cpu_ids * sizeof(void **); | |
8032 | #endif | |
434d53b0 MT |
8033 | #endif |
8034 | } | |
dd41f596 | 8035 | |
57d885fe | 8036 | #ifdef CONFIG_SMP |
18d95a28 | 8037 | init_aggregate(); |
57d885fe GH |
8038 | init_defrootdomain(); |
8039 | #endif | |
8040 | ||
d0b27fa7 PZ |
8041 | init_rt_bandwidth(&def_rt_bandwidth, |
8042 | global_rt_period(), global_rt_runtime()); | |
8043 | ||
8044 | #ifdef CONFIG_RT_GROUP_SCHED | |
8045 | init_rt_bandwidth(&init_task_group.rt_bandwidth, | |
8046 | global_rt_period(), global_rt_runtime()); | |
eff766a6 PZ |
8047 | #ifdef CONFIG_USER_SCHED |
8048 | init_rt_bandwidth(&root_task_group.rt_bandwidth, | |
8049 | global_rt_period(), RUNTIME_INF); | |
8050 | #endif | |
d0b27fa7 PZ |
8051 | #endif |
8052 | ||
052f1dc7 | 8053 | #ifdef CONFIG_GROUP_SCHED |
6f505b16 | 8054 | list_add(&init_task_group.list, &task_groups); |
f473aa5e PZ |
8055 | INIT_LIST_HEAD(&init_task_group.children); |
8056 | ||
8057 | #ifdef CONFIG_USER_SCHED | |
8058 | INIT_LIST_HEAD(&root_task_group.children); | |
8059 | init_task_group.parent = &root_task_group; | |
8060 | list_add(&init_task_group.siblings, &root_task_group.children); | |
8061 | #endif | |
6f505b16 PZ |
8062 | #endif |
8063 | ||
0a945022 | 8064 | for_each_possible_cpu(i) { |
70b97a7f | 8065 | struct rq *rq; |
1da177e4 LT |
8066 | |
8067 | rq = cpu_rq(i); | |
8068 | spin_lock_init(&rq->lock); | |
fcb99371 | 8069 | lockdep_set_class(&rq->lock, &rq->rq_lock_key); |
7897986b | 8070 | rq->nr_running = 0; |
dd41f596 | 8071 | init_cfs_rq(&rq->cfs, rq); |
6f505b16 | 8072 | init_rt_rq(&rq->rt, rq); |
dd41f596 | 8073 | #ifdef CONFIG_FAIR_GROUP_SCHED |
4cf86d77 | 8074 | init_task_group.shares = init_task_group_load; |
6f505b16 | 8075 | INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); |
354d60c2 DG |
8076 | #ifdef CONFIG_CGROUP_SCHED |
8077 | /* | |
8078 | * How much cpu bandwidth does init_task_group get? | |
8079 | * | |
8080 | * In case of task-groups formed thr' the cgroup filesystem, it | |
8081 | * gets 100% of the cpu resources in the system. This overall | |
8082 | * system cpu resource is divided among the tasks of | |
8083 | * init_task_group and its child task-groups in a fair manner, | |
8084 | * based on each entity's (task or task-group's) weight | |
8085 | * (se->load.weight). | |
8086 | * | |
8087 | * In other words, if init_task_group has 10 tasks of weight | |
8088 | * 1024) and two child groups A0 and A1 (of weight 1024 each), | |
8089 | * then A0's share of the cpu resource is: | |
8090 | * | |
8091 | * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% | |
8092 | * | |
8093 | * We achieve this by letting init_task_group's tasks sit | |
8094 | * directly in rq->cfs (i.e init_task_group->se[] = NULL). | |
8095 | */ | |
ec7dc8ac | 8096 | init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL); |
354d60c2 | 8097 | #elif defined CONFIG_USER_SCHED |
eff766a6 PZ |
8098 | root_task_group.shares = NICE_0_LOAD; |
8099 | init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL); | |
354d60c2 DG |
8100 | /* |
8101 | * In case of task-groups formed thr' the user id of tasks, | |
8102 | * init_task_group represents tasks belonging to root user. | |
8103 | * Hence it forms a sibling of all subsequent groups formed. | |
8104 | * In this case, init_task_group gets only a fraction of overall | |
8105 | * system cpu resource, based on the weight assigned to root | |
8106 | * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished | |
8107 | * by letting tasks of init_task_group sit in a separate cfs_rq | |
8108 | * (init_cfs_rq) and having one entity represent this group of | |
8109 | * tasks in rq->cfs (i.e init_task_group->se[] != NULL). | |
8110 | */ | |
ec7dc8ac | 8111 | init_tg_cfs_entry(&init_task_group, |
6f505b16 | 8112 | &per_cpu(init_cfs_rq, i), |
eff766a6 PZ |
8113 | &per_cpu(init_sched_entity, i), i, 1, |
8114 | root_task_group.se[i]); | |
6f505b16 | 8115 | |
052f1dc7 | 8116 | #endif |
354d60c2 DG |
8117 | #endif /* CONFIG_FAIR_GROUP_SCHED */ |
8118 | ||
8119 | rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; | |
052f1dc7 | 8120 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8121 | INIT_LIST_HEAD(&rq->leaf_rt_rq_list); |
354d60c2 | 8122 | #ifdef CONFIG_CGROUP_SCHED |
ec7dc8ac | 8123 | init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL); |
354d60c2 | 8124 | #elif defined CONFIG_USER_SCHED |
eff766a6 | 8125 | init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL); |
ec7dc8ac | 8126 | init_tg_rt_entry(&init_task_group, |
6f505b16 | 8127 | &per_cpu(init_rt_rq, i), |
eff766a6 PZ |
8128 | &per_cpu(init_sched_rt_entity, i), i, 1, |
8129 | root_task_group.rt_se[i]); | |
354d60c2 | 8130 | #endif |
dd41f596 | 8131 | #endif |
1da177e4 | 8132 | |
dd41f596 IM |
8133 | for (j = 0; j < CPU_LOAD_IDX_MAX; j++) |
8134 | rq->cpu_load[j] = 0; | |
1da177e4 | 8135 | #ifdef CONFIG_SMP |
41c7ce9a | 8136 | rq->sd = NULL; |
57d885fe | 8137 | rq->rd = NULL; |
1da177e4 | 8138 | rq->active_balance = 0; |
dd41f596 | 8139 | rq->next_balance = jiffies; |
1da177e4 | 8140 | rq->push_cpu = 0; |
0a2966b4 | 8141 | rq->cpu = i; |
1da177e4 LT |
8142 | rq->migration_thread = NULL; |
8143 | INIT_LIST_HEAD(&rq->migration_queue); | |
dc938520 | 8144 | rq_attach_root(rq, &def_root_domain); |
1da177e4 | 8145 | #endif |
8f4d37ec | 8146 | init_rq_hrtick(rq); |
1da177e4 | 8147 | atomic_set(&rq->nr_iowait, 0); |
1da177e4 LT |
8148 | } |
8149 | ||
2dd73a4f | 8150 | set_load_weight(&init_task); |
b50f60ce | 8151 | |
e107be36 AK |
8152 | #ifdef CONFIG_PREEMPT_NOTIFIERS |
8153 | INIT_HLIST_HEAD(&init_task.preempt_notifiers); | |
8154 | #endif | |
8155 | ||
c9819f45 CL |
8156 | #ifdef CONFIG_SMP |
8157 | open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL); | |
8158 | #endif | |
8159 | ||
b50f60ce HC |
8160 | #ifdef CONFIG_RT_MUTEXES |
8161 | plist_head_init(&init_task.pi_waiters, &init_task.pi_lock); | |
8162 | #endif | |
8163 | ||
1da177e4 LT |
8164 | /* |
8165 | * The boot idle thread does lazy MMU switching as well: | |
8166 | */ | |
8167 | atomic_inc(&init_mm.mm_count); | |
8168 | enter_lazy_tlb(&init_mm, current); | |
8169 | ||
8170 | /* | |
8171 | * Make us the idle thread. Technically, schedule() should not be | |
8172 | * called from this thread, however somewhere below it might be, | |
8173 | * but because we are the idle thread, we just pick up running again | |
8174 | * when this runqueue becomes "idle". | |
8175 | */ | |
8176 | init_idle(current, smp_processor_id()); | |
dd41f596 IM |
8177 | /* |
8178 | * During early bootup we pretend to be a normal task: | |
8179 | */ | |
8180 | current->sched_class = &fair_sched_class; | |
6892b75e IM |
8181 | |
8182 | scheduler_running = 1; | |
1da177e4 LT |
8183 | } |
8184 | ||
8185 | #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP | |
8186 | void __might_sleep(char *file, int line) | |
8187 | { | |
48f24c4d | 8188 | #ifdef in_atomic |
1da177e4 LT |
8189 | static unsigned long prev_jiffy; /* ratelimiting */ |
8190 | ||
8191 | if ((in_atomic() || irqs_disabled()) && | |
8192 | system_state == SYSTEM_RUNNING && !oops_in_progress) { | |
8193 | if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) | |
8194 | return; | |
8195 | prev_jiffy = jiffies; | |
91368d73 | 8196 | printk(KERN_ERR "BUG: sleeping function called from invalid" |
1da177e4 LT |
8197 | " context at %s:%d\n", file, line); |
8198 | printk("in_atomic():%d, irqs_disabled():%d\n", | |
8199 | in_atomic(), irqs_disabled()); | |
a4c410f0 | 8200 | debug_show_held_locks(current); |
3117df04 IM |
8201 | if (irqs_disabled()) |
8202 | print_irqtrace_events(current); | |
1da177e4 LT |
8203 | dump_stack(); |
8204 | } | |
8205 | #endif | |
8206 | } | |
8207 | EXPORT_SYMBOL(__might_sleep); | |
8208 | #endif | |
8209 | ||
8210 | #ifdef CONFIG_MAGIC_SYSRQ | |
3a5e4dc1 AK |
8211 | static void normalize_task(struct rq *rq, struct task_struct *p) |
8212 | { | |
8213 | int on_rq; | |
3e51f33f | 8214 | |
3a5e4dc1 AK |
8215 | update_rq_clock(rq); |
8216 | on_rq = p->se.on_rq; | |
8217 | if (on_rq) | |
8218 | deactivate_task(rq, p, 0); | |
8219 | __setscheduler(rq, p, SCHED_NORMAL, 0); | |
8220 | if (on_rq) { | |
8221 | activate_task(rq, p, 0); | |
8222 | resched_task(rq->curr); | |
8223 | } | |
8224 | } | |
8225 | ||
1da177e4 LT |
8226 | void normalize_rt_tasks(void) |
8227 | { | |
a0f98a1c | 8228 | struct task_struct *g, *p; |
1da177e4 | 8229 | unsigned long flags; |
70b97a7f | 8230 | struct rq *rq; |
1da177e4 | 8231 | |
4cf5d77a | 8232 | read_lock_irqsave(&tasklist_lock, flags); |
a0f98a1c | 8233 | do_each_thread(g, p) { |
178be793 IM |
8234 | /* |
8235 | * Only normalize user tasks: | |
8236 | */ | |
8237 | if (!p->mm) | |
8238 | continue; | |
8239 | ||
6cfb0d5d | 8240 | p->se.exec_start = 0; |
6cfb0d5d | 8241 | #ifdef CONFIG_SCHEDSTATS |
dd41f596 | 8242 | p->se.wait_start = 0; |
dd41f596 | 8243 | p->se.sleep_start = 0; |
dd41f596 | 8244 | p->se.block_start = 0; |
6cfb0d5d | 8245 | #endif |
dd41f596 IM |
8246 | |
8247 | if (!rt_task(p)) { | |
8248 | /* | |
8249 | * Renice negative nice level userspace | |
8250 | * tasks back to 0: | |
8251 | */ | |
8252 | if (TASK_NICE(p) < 0 && p->mm) | |
8253 | set_user_nice(p, 0); | |
1da177e4 | 8254 | continue; |
dd41f596 | 8255 | } |
1da177e4 | 8256 | |
4cf5d77a | 8257 | spin_lock(&p->pi_lock); |
b29739f9 | 8258 | rq = __task_rq_lock(p); |
1da177e4 | 8259 | |
178be793 | 8260 | normalize_task(rq, p); |
3a5e4dc1 | 8261 | |
b29739f9 | 8262 | __task_rq_unlock(rq); |
4cf5d77a | 8263 | spin_unlock(&p->pi_lock); |
a0f98a1c IM |
8264 | } while_each_thread(g, p); |
8265 | ||
4cf5d77a | 8266 | read_unlock_irqrestore(&tasklist_lock, flags); |
1da177e4 LT |
8267 | } |
8268 | ||
8269 | #endif /* CONFIG_MAGIC_SYSRQ */ | |
1df5c10a LT |
8270 | |
8271 | #ifdef CONFIG_IA64 | |
8272 | /* | |
8273 | * These functions are only useful for the IA64 MCA handling. | |
8274 | * | |
8275 | * They can only be called when the whole system has been | |
8276 | * stopped - every CPU needs to be quiescent, and no scheduling | |
8277 | * activity can take place. Using them for anything else would | |
8278 | * be a serious bug, and as a result, they aren't even visible | |
8279 | * under any other configuration. | |
8280 | */ | |
8281 | ||
8282 | /** | |
8283 | * curr_task - return the current task for a given cpu. | |
8284 | * @cpu: the processor in question. | |
8285 | * | |
8286 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8287 | */ | |
36c8b586 | 8288 | struct task_struct *curr_task(int cpu) |
1df5c10a LT |
8289 | { |
8290 | return cpu_curr(cpu); | |
8291 | } | |
8292 | ||
8293 | /** | |
8294 | * set_curr_task - set the current task for a given cpu. | |
8295 | * @cpu: the processor in question. | |
8296 | * @p: the task pointer to set. | |
8297 | * | |
8298 | * Description: This function must only be used when non-maskable interrupts | |
41a2d6cf IM |
8299 | * are serviced on a separate stack. It allows the architecture to switch the |
8300 | * notion of the current task on a cpu in a non-blocking manner. This function | |
1df5c10a LT |
8301 | * must be called with all CPU's synchronized, and interrupts disabled, the |
8302 | * and caller must save the original value of the current task (see | |
8303 | * curr_task() above) and restore that value before reenabling interrupts and | |
8304 | * re-starting the system. | |
8305 | * | |
8306 | * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! | |
8307 | */ | |
36c8b586 | 8308 | void set_curr_task(int cpu, struct task_struct *p) |
1df5c10a LT |
8309 | { |
8310 | cpu_curr(cpu) = p; | |
8311 | } | |
8312 | ||
8313 | #endif | |
29f59db3 | 8314 | |
bccbe08a PZ |
8315 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8316 | static void free_fair_sched_group(struct task_group *tg) | |
6f505b16 PZ |
8317 | { |
8318 | int i; | |
8319 | ||
8320 | for_each_possible_cpu(i) { | |
8321 | if (tg->cfs_rq) | |
8322 | kfree(tg->cfs_rq[i]); | |
8323 | if (tg->se) | |
8324 | kfree(tg->se[i]); | |
6f505b16 PZ |
8325 | } |
8326 | ||
8327 | kfree(tg->cfs_rq); | |
8328 | kfree(tg->se); | |
6f505b16 PZ |
8329 | } |
8330 | ||
ec7dc8ac DG |
8331 | static |
8332 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
29f59db3 | 8333 | { |
29f59db3 | 8334 | struct cfs_rq *cfs_rq; |
ec7dc8ac | 8335 | struct sched_entity *se, *parent_se; |
9b5b7751 | 8336 | struct rq *rq; |
29f59db3 SV |
8337 | int i; |
8338 | ||
434d53b0 | 8339 | tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8340 | if (!tg->cfs_rq) |
8341 | goto err; | |
434d53b0 | 8342 | tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL); |
29f59db3 SV |
8343 | if (!tg->se) |
8344 | goto err; | |
052f1dc7 PZ |
8345 | |
8346 | tg->shares = NICE_0_LOAD; | |
29f59db3 SV |
8347 | |
8348 | for_each_possible_cpu(i) { | |
9b5b7751 | 8349 | rq = cpu_rq(i); |
29f59db3 | 8350 | |
6f505b16 PZ |
8351 | cfs_rq = kmalloc_node(sizeof(struct cfs_rq), |
8352 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8353 | if (!cfs_rq) |
8354 | goto err; | |
8355 | ||
6f505b16 PZ |
8356 | se = kmalloc_node(sizeof(struct sched_entity), |
8357 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
29f59db3 SV |
8358 | if (!se) |
8359 | goto err; | |
8360 | ||
ec7dc8ac DG |
8361 | parent_se = parent ? parent->se[i] : NULL; |
8362 | init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se); | |
bccbe08a PZ |
8363 | } |
8364 | ||
8365 | return 1; | |
8366 | ||
8367 | err: | |
8368 | return 0; | |
8369 | } | |
8370 | ||
8371 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8372 | { | |
8373 | list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list, | |
8374 | &cpu_rq(cpu)->leaf_cfs_rq_list); | |
8375 | } | |
8376 | ||
8377 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8378 | { | |
8379 | list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list); | |
8380 | } | |
8381 | #else | |
8382 | static inline void free_fair_sched_group(struct task_group *tg) | |
8383 | { | |
8384 | } | |
8385 | ||
ec7dc8ac DG |
8386 | static inline |
8387 | int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8388 | { |
8389 | return 1; | |
8390 | } | |
8391 | ||
8392 | static inline void register_fair_sched_group(struct task_group *tg, int cpu) | |
8393 | { | |
8394 | } | |
8395 | ||
8396 | static inline void unregister_fair_sched_group(struct task_group *tg, int cpu) | |
8397 | { | |
8398 | } | |
052f1dc7 PZ |
8399 | #endif |
8400 | ||
8401 | #ifdef CONFIG_RT_GROUP_SCHED | |
bccbe08a PZ |
8402 | static void free_rt_sched_group(struct task_group *tg) |
8403 | { | |
8404 | int i; | |
8405 | ||
d0b27fa7 PZ |
8406 | destroy_rt_bandwidth(&tg->rt_bandwidth); |
8407 | ||
bccbe08a PZ |
8408 | for_each_possible_cpu(i) { |
8409 | if (tg->rt_rq) | |
8410 | kfree(tg->rt_rq[i]); | |
8411 | if (tg->rt_se) | |
8412 | kfree(tg->rt_se[i]); | |
8413 | } | |
8414 | ||
8415 | kfree(tg->rt_rq); | |
8416 | kfree(tg->rt_se); | |
8417 | } | |
8418 | ||
ec7dc8ac DG |
8419 | static |
8420 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8421 | { |
8422 | struct rt_rq *rt_rq; | |
ec7dc8ac | 8423 | struct sched_rt_entity *rt_se, *parent_se; |
bccbe08a PZ |
8424 | struct rq *rq; |
8425 | int i; | |
8426 | ||
434d53b0 | 8427 | tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8428 | if (!tg->rt_rq) |
8429 | goto err; | |
434d53b0 | 8430 | tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL); |
bccbe08a PZ |
8431 | if (!tg->rt_se) |
8432 | goto err; | |
8433 | ||
d0b27fa7 PZ |
8434 | init_rt_bandwidth(&tg->rt_bandwidth, |
8435 | ktime_to_ns(def_rt_bandwidth.rt_period), 0); | |
bccbe08a PZ |
8436 | |
8437 | for_each_possible_cpu(i) { | |
8438 | rq = cpu_rq(i); | |
8439 | ||
6f505b16 PZ |
8440 | rt_rq = kmalloc_node(sizeof(struct rt_rq), |
8441 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8442 | if (!rt_rq) | |
8443 | goto err; | |
29f59db3 | 8444 | |
6f505b16 PZ |
8445 | rt_se = kmalloc_node(sizeof(struct sched_rt_entity), |
8446 | GFP_KERNEL|__GFP_ZERO, cpu_to_node(i)); | |
8447 | if (!rt_se) | |
8448 | goto err; | |
29f59db3 | 8449 | |
ec7dc8ac DG |
8450 | parent_se = parent ? parent->rt_se[i] : NULL; |
8451 | init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se); | |
29f59db3 SV |
8452 | } |
8453 | ||
bccbe08a PZ |
8454 | return 1; |
8455 | ||
8456 | err: | |
8457 | return 0; | |
8458 | } | |
8459 | ||
8460 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8461 | { | |
8462 | list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list, | |
8463 | &cpu_rq(cpu)->leaf_rt_rq_list); | |
8464 | } | |
8465 | ||
8466 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8467 | { | |
8468 | list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list); | |
8469 | } | |
8470 | #else | |
8471 | static inline void free_rt_sched_group(struct task_group *tg) | |
8472 | { | |
8473 | } | |
8474 | ||
ec7dc8ac DG |
8475 | static inline |
8476 | int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent) | |
bccbe08a PZ |
8477 | { |
8478 | return 1; | |
8479 | } | |
8480 | ||
8481 | static inline void register_rt_sched_group(struct task_group *tg, int cpu) | |
8482 | { | |
8483 | } | |
8484 | ||
8485 | static inline void unregister_rt_sched_group(struct task_group *tg, int cpu) | |
8486 | { | |
8487 | } | |
8488 | #endif | |
8489 | ||
d0b27fa7 | 8490 | #ifdef CONFIG_GROUP_SCHED |
bccbe08a PZ |
8491 | static void free_sched_group(struct task_group *tg) |
8492 | { | |
8493 | free_fair_sched_group(tg); | |
8494 | free_rt_sched_group(tg); | |
8495 | kfree(tg); | |
8496 | } | |
8497 | ||
8498 | /* allocate runqueue etc for a new task group */ | |
ec7dc8ac | 8499 | struct task_group *sched_create_group(struct task_group *parent) |
bccbe08a PZ |
8500 | { |
8501 | struct task_group *tg; | |
8502 | unsigned long flags; | |
8503 | int i; | |
8504 | ||
8505 | tg = kzalloc(sizeof(*tg), GFP_KERNEL); | |
8506 | if (!tg) | |
8507 | return ERR_PTR(-ENOMEM); | |
8508 | ||
ec7dc8ac | 8509 | if (!alloc_fair_sched_group(tg, parent)) |
bccbe08a PZ |
8510 | goto err; |
8511 | ||
ec7dc8ac | 8512 | if (!alloc_rt_sched_group(tg, parent)) |
bccbe08a PZ |
8513 | goto err; |
8514 | ||
8ed36996 | 8515 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8516 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8517 | register_fair_sched_group(tg, i); |
8518 | register_rt_sched_group(tg, i); | |
9b5b7751 | 8519 | } |
6f505b16 | 8520 | list_add_rcu(&tg->list, &task_groups); |
f473aa5e PZ |
8521 | |
8522 | WARN_ON(!parent); /* root should already exist */ | |
8523 | ||
8524 | tg->parent = parent; | |
8525 | list_add_rcu(&tg->siblings, &parent->children); | |
8526 | INIT_LIST_HEAD(&tg->children); | |
8ed36996 | 8527 | spin_unlock_irqrestore(&task_group_lock, flags); |
29f59db3 | 8528 | |
9b5b7751 | 8529 | return tg; |
29f59db3 SV |
8530 | |
8531 | err: | |
6f505b16 | 8532 | free_sched_group(tg); |
29f59db3 SV |
8533 | return ERR_PTR(-ENOMEM); |
8534 | } | |
8535 | ||
9b5b7751 | 8536 | /* rcu callback to free various structures associated with a task group */ |
6f505b16 | 8537 | static void free_sched_group_rcu(struct rcu_head *rhp) |
29f59db3 | 8538 | { |
29f59db3 | 8539 | /* now it should be safe to free those cfs_rqs */ |
6f505b16 | 8540 | free_sched_group(container_of(rhp, struct task_group, rcu)); |
29f59db3 SV |
8541 | } |
8542 | ||
9b5b7751 | 8543 | /* Destroy runqueue etc associated with a task group */ |
4cf86d77 | 8544 | void sched_destroy_group(struct task_group *tg) |
29f59db3 | 8545 | { |
8ed36996 | 8546 | unsigned long flags; |
9b5b7751 | 8547 | int i; |
29f59db3 | 8548 | |
8ed36996 | 8549 | spin_lock_irqsave(&task_group_lock, flags); |
9b5b7751 | 8550 | for_each_possible_cpu(i) { |
bccbe08a PZ |
8551 | unregister_fair_sched_group(tg, i); |
8552 | unregister_rt_sched_group(tg, i); | |
9b5b7751 | 8553 | } |
6f505b16 | 8554 | list_del_rcu(&tg->list); |
f473aa5e | 8555 | list_del_rcu(&tg->siblings); |
8ed36996 | 8556 | spin_unlock_irqrestore(&task_group_lock, flags); |
9b5b7751 | 8557 | |
9b5b7751 | 8558 | /* wait for possible concurrent references to cfs_rqs complete */ |
6f505b16 | 8559 | call_rcu(&tg->rcu, free_sched_group_rcu); |
29f59db3 SV |
8560 | } |
8561 | ||
9b5b7751 | 8562 | /* change task's runqueue when it moves between groups. |
3a252015 IM |
8563 | * The caller of this function should have put the task in its new group |
8564 | * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to | |
8565 | * reflect its new group. | |
9b5b7751 SV |
8566 | */ |
8567 | void sched_move_task(struct task_struct *tsk) | |
29f59db3 SV |
8568 | { |
8569 | int on_rq, running; | |
8570 | unsigned long flags; | |
8571 | struct rq *rq; | |
8572 | ||
8573 | rq = task_rq_lock(tsk, &flags); | |
8574 | ||
29f59db3 SV |
8575 | update_rq_clock(rq); |
8576 | ||
051a1d1a | 8577 | running = task_current(rq, tsk); |
29f59db3 SV |
8578 | on_rq = tsk->se.on_rq; |
8579 | ||
0e1f3483 | 8580 | if (on_rq) |
29f59db3 | 8581 | dequeue_task(rq, tsk, 0); |
0e1f3483 HS |
8582 | if (unlikely(running)) |
8583 | tsk->sched_class->put_prev_task(rq, tsk); | |
29f59db3 | 8584 | |
6f505b16 | 8585 | set_task_rq(tsk, task_cpu(tsk)); |
29f59db3 | 8586 | |
810b3817 PZ |
8587 | #ifdef CONFIG_FAIR_GROUP_SCHED |
8588 | if (tsk->sched_class->moved_group) | |
8589 | tsk->sched_class->moved_group(tsk); | |
8590 | #endif | |
8591 | ||
0e1f3483 HS |
8592 | if (unlikely(running)) |
8593 | tsk->sched_class->set_curr_task(rq); | |
8594 | if (on_rq) | |
7074badb | 8595 | enqueue_task(rq, tsk, 0); |
29f59db3 | 8596 | |
29f59db3 SV |
8597 | task_rq_unlock(rq, &flags); |
8598 | } | |
d0b27fa7 | 8599 | #endif |
29f59db3 | 8600 | |
052f1dc7 | 8601 | #ifdef CONFIG_FAIR_GROUP_SCHED |
18d95a28 | 8602 | static void __set_se_shares(struct sched_entity *se, unsigned long shares) |
29f59db3 SV |
8603 | { |
8604 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
29f59db3 SV |
8605 | int on_rq; |
8606 | ||
29f59db3 | 8607 | on_rq = se->on_rq; |
62fb1851 | 8608 | if (on_rq) |
29f59db3 SV |
8609 | dequeue_entity(cfs_rq, se, 0); |
8610 | ||
8611 | se->load.weight = shares; | |
e05510d0 | 8612 | se->load.inv_weight = 0; |
29f59db3 | 8613 | |
62fb1851 | 8614 | if (on_rq) |
29f59db3 | 8615 | enqueue_entity(cfs_rq, se, 0); |
18d95a28 | 8616 | } |
62fb1851 | 8617 | |
18d95a28 PZ |
8618 | static void set_se_shares(struct sched_entity *se, unsigned long shares) |
8619 | { | |
8620 | struct cfs_rq *cfs_rq = se->cfs_rq; | |
8621 | struct rq *rq = cfs_rq->rq; | |
8622 | unsigned long flags; | |
8623 | ||
8624 | spin_lock_irqsave(&rq->lock, flags); | |
8625 | __set_se_shares(se, shares); | |
8626 | spin_unlock_irqrestore(&rq->lock, flags); | |
29f59db3 SV |
8627 | } |
8628 | ||
8ed36996 PZ |
8629 | static DEFINE_MUTEX(shares_mutex); |
8630 | ||
4cf86d77 | 8631 | int sched_group_set_shares(struct task_group *tg, unsigned long shares) |
29f59db3 SV |
8632 | { |
8633 | int i; | |
8ed36996 | 8634 | unsigned long flags; |
c61935fd | 8635 | |
ec7dc8ac DG |
8636 | /* |
8637 | * We can't change the weight of the root cgroup. | |
8638 | */ | |
8639 | if (!tg->se[0]) | |
8640 | return -EINVAL; | |
8641 | ||
18d95a28 PZ |
8642 | if (shares < MIN_SHARES) |
8643 | shares = MIN_SHARES; | |
cb4ad1ff MX |
8644 | else if (shares > MAX_SHARES) |
8645 | shares = MAX_SHARES; | |
62fb1851 | 8646 | |
8ed36996 | 8647 | mutex_lock(&shares_mutex); |
9b5b7751 | 8648 | if (tg->shares == shares) |
5cb350ba | 8649 | goto done; |
29f59db3 | 8650 | |
8ed36996 | 8651 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8652 | for_each_possible_cpu(i) |
8653 | unregister_fair_sched_group(tg, i); | |
f473aa5e | 8654 | list_del_rcu(&tg->siblings); |
8ed36996 | 8655 | spin_unlock_irqrestore(&task_group_lock, flags); |
6b2d7700 SV |
8656 | |
8657 | /* wait for any ongoing reference to this group to finish */ | |
8658 | synchronize_sched(); | |
8659 | ||
8660 | /* | |
8661 | * Now we are free to modify the group's share on each cpu | |
8662 | * w/o tripping rebalance_share or load_balance_fair. | |
8663 | */ | |
9b5b7751 | 8664 | tg->shares = shares; |
18d95a28 PZ |
8665 | for_each_possible_cpu(i) { |
8666 | /* | |
8667 | * force a rebalance | |
8668 | */ | |
8669 | cfs_rq_set_shares(tg->cfs_rq[i], 0); | |
cb4ad1ff | 8670 | set_se_shares(tg->se[i], shares); |
18d95a28 | 8671 | } |
29f59db3 | 8672 | |
6b2d7700 SV |
8673 | /* |
8674 | * Enable load balance activity on this group, by inserting it back on | |
8675 | * each cpu's rq->leaf_cfs_rq_list. | |
8676 | */ | |
8ed36996 | 8677 | spin_lock_irqsave(&task_group_lock, flags); |
bccbe08a PZ |
8678 | for_each_possible_cpu(i) |
8679 | register_fair_sched_group(tg, i); | |
f473aa5e | 8680 | list_add_rcu(&tg->siblings, &tg->parent->children); |
8ed36996 | 8681 | spin_unlock_irqrestore(&task_group_lock, flags); |
5cb350ba | 8682 | done: |
8ed36996 | 8683 | mutex_unlock(&shares_mutex); |
9b5b7751 | 8684 | return 0; |
29f59db3 SV |
8685 | } |
8686 | ||
5cb350ba DG |
8687 | unsigned long sched_group_shares(struct task_group *tg) |
8688 | { | |
8689 | return tg->shares; | |
8690 | } | |
052f1dc7 | 8691 | #endif |
5cb350ba | 8692 | |
052f1dc7 | 8693 | #ifdef CONFIG_RT_GROUP_SCHED |
6f505b16 | 8694 | /* |
9f0c1e56 | 8695 | * Ensure that the real time constraints are schedulable. |
6f505b16 | 8696 | */ |
9f0c1e56 PZ |
8697 | static DEFINE_MUTEX(rt_constraints_mutex); |
8698 | ||
8699 | static unsigned long to_ratio(u64 period, u64 runtime) | |
8700 | { | |
8701 | if (runtime == RUNTIME_INF) | |
8702 | return 1ULL << 16; | |
8703 | ||
6f6d6a1a | 8704 | return div64_u64(runtime << 16, period); |
9f0c1e56 PZ |
8705 | } |
8706 | ||
b40b2e8e PZ |
8707 | #ifdef CONFIG_CGROUP_SCHED |
8708 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) | |
8709 | { | |
8710 | struct task_group *tgi, *parent = tg->parent; | |
8711 | unsigned long total = 0; | |
8712 | ||
8713 | if (!parent) { | |
8714 | if (global_rt_period() < period) | |
8715 | return 0; | |
8716 | ||
8717 | return to_ratio(period, runtime) < | |
8718 | to_ratio(global_rt_period(), global_rt_runtime()); | |
8719 | } | |
8720 | ||
8721 | if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period) | |
8722 | return 0; | |
8723 | ||
8724 | rcu_read_lock(); | |
8725 | list_for_each_entry_rcu(tgi, &parent->children, siblings) { | |
8726 | if (tgi == tg) | |
8727 | continue; | |
8728 | ||
8729 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), | |
8730 | tgi->rt_bandwidth.rt_runtime); | |
8731 | } | |
8732 | rcu_read_unlock(); | |
8733 | ||
8734 | return total + to_ratio(period, runtime) < | |
8735 | to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period), | |
8736 | parent->rt_bandwidth.rt_runtime); | |
8737 | } | |
8738 | #elif defined CONFIG_USER_SCHED | |
9f0c1e56 | 8739 | static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) |
6f505b16 PZ |
8740 | { |
8741 | struct task_group *tgi; | |
8742 | unsigned long total = 0; | |
9f0c1e56 | 8743 | unsigned long global_ratio = |
d0b27fa7 | 8744 | to_ratio(global_rt_period(), global_rt_runtime()); |
6f505b16 PZ |
8745 | |
8746 | rcu_read_lock(); | |
9f0c1e56 PZ |
8747 | list_for_each_entry_rcu(tgi, &task_groups, list) { |
8748 | if (tgi == tg) | |
8749 | continue; | |
6f505b16 | 8750 | |
d0b27fa7 PZ |
8751 | total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period), |
8752 | tgi->rt_bandwidth.rt_runtime); | |
9f0c1e56 PZ |
8753 | } |
8754 | rcu_read_unlock(); | |
6f505b16 | 8755 | |
9f0c1e56 | 8756 | return total + to_ratio(period, runtime) < global_ratio; |
6f505b16 | 8757 | } |
b40b2e8e | 8758 | #endif |
6f505b16 | 8759 | |
521f1a24 DG |
8760 | /* Must be called with tasklist_lock held */ |
8761 | static inline int tg_has_rt_tasks(struct task_group *tg) | |
8762 | { | |
8763 | struct task_struct *g, *p; | |
8764 | do_each_thread(g, p) { | |
8765 | if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg) | |
8766 | return 1; | |
8767 | } while_each_thread(g, p); | |
8768 | return 0; | |
8769 | } | |
8770 | ||
d0b27fa7 PZ |
8771 | static int tg_set_bandwidth(struct task_group *tg, |
8772 | u64 rt_period, u64 rt_runtime) | |
6f505b16 | 8773 | { |
ac086bc2 | 8774 | int i, err = 0; |
9f0c1e56 | 8775 | |
9f0c1e56 | 8776 | mutex_lock(&rt_constraints_mutex); |
521f1a24 | 8777 | read_lock(&tasklist_lock); |
ac086bc2 | 8778 | if (rt_runtime == 0 && tg_has_rt_tasks(tg)) { |
521f1a24 DG |
8779 | err = -EBUSY; |
8780 | goto unlock; | |
8781 | } | |
9f0c1e56 PZ |
8782 | if (!__rt_schedulable(tg, rt_period, rt_runtime)) { |
8783 | err = -EINVAL; | |
8784 | goto unlock; | |
8785 | } | |
ac086bc2 PZ |
8786 | |
8787 | spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
d0b27fa7 PZ |
8788 | tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); |
8789 | tg->rt_bandwidth.rt_runtime = rt_runtime; | |
ac086bc2 PZ |
8790 | |
8791 | for_each_possible_cpu(i) { | |
8792 | struct rt_rq *rt_rq = tg->rt_rq[i]; | |
8793 | ||
8794 | spin_lock(&rt_rq->rt_runtime_lock); | |
8795 | rt_rq->rt_runtime = rt_runtime; | |
8796 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8797 | } | |
8798 | spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); | |
9f0c1e56 | 8799 | unlock: |
521f1a24 | 8800 | read_unlock(&tasklist_lock); |
9f0c1e56 PZ |
8801 | mutex_unlock(&rt_constraints_mutex); |
8802 | ||
8803 | return err; | |
6f505b16 PZ |
8804 | } |
8805 | ||
d0b27fa7 PZ |
8806 | int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) |
8807 | { | |
8808 | u64 rt_runtime, rt_period; | |
8809 | ||
8810 | rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8811 | rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; | |
8812 | if (rt_runtime_us < 0) | |
8813 | rt_runtime = RUNTIME_INF; | |
8814 | ||
8815 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8816 | } | |
8817 | ||
9f0c1e56 PZ |
8818 | long sched_group_rt_runtime(struct task_group *tg) |
8819 | { | |
8820 | u64 rt_runtime_us; | |
8821 | ||
d0b27fa7 | 8822 | if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) |
9f0c1e56 PZ |
8823 | return -1; |
8824 | ||
d0b27fa7 | 8825 | rt_runtime_us = tg->rt_bandwidth.rt_runtime; |
9f0c1e56 PZ |
8826 | do_div(rt_runtime_us, NSEC_PER_USEC); |
8827 | return rt_runtime_us; | |
8828 | } | |
d0b27fa7 PZ |
8829 | |
8830 | int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) | |
8831 | { | |
8832 | u64 rt_runtime, rt_period; | |
8833 | ||
8834 | rt_period = (u64)rt_period_us * NSEC_PER_USEC; | |
8835 | rt_runtime = tg->rt_bandwidth.rt_runtime; | |
8836 | ||
8837 | return tg_set_bandwidth(tg, rt_period, rt_runtime); | |
8838 | } | |
8839 | ||
8840 | long sched_group_rt_period(struct task_group *tg) | |
8841 | { | |
8842 | u64 rt_period_us; | |
8843 | ||
8844 | rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); | |
8845 | do_div(rt_period_us, NSEC_PER_USEC); | |
8846 | return rt_period_us; | |
8847 | } | |
8848 | ||
8849 | static int sched_rt_global_constraints(void) | |
8850 | { | |
8851 | int ret = 0; | |
8852 | ||
8853 | mutex_lock(&rt_constraints_mutex); | |
8854 | if (!__rt_schedulable(NULL, 1, 0)) | |
8855 | ret = -EINVAL; | |
8856 | mutex_unlock(&rt_constraints_mutex); | |
8857 | ||
8858 | return ret; | |
8859 | } | |
8860 | #else | |
8861 | static int sched_rt_global_constraints(void) | |
8862 | { | |
ac086bc2 PZ |
8863 | unsigned long flags; |
8864 | int i; | |
8865 | ||
8866 | spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8867 | for_each_possible_cpu(i) { | |
8868 | struct rt_rq *rt_rq = &cpu_rq(i)->rt; | |
8869 | ||
8870 | spin_lock(&rt_rq->rt_runtime_lock); | |
8871 | rt_rq->rt_runtime = global_rt_runtime(); | |
8872 | spin_unlock(&rt_rq->rt_runtime_lock); | |
8873 | } | |
8874 | spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); | |
8875 | ||
d0b27fa7 PZ |
8876 | return 0; |
8877 | } | |
052f1dc7 | 8878 | #endif |
d0b27fa7 PZ |
8879 | |
8880 | int sched_rt_handler(struct ctl_table *table, int write, | |
8881 | struct file *filp, void __user *buffer, size_t *lenp, | |
8882 | loff_t *ppos) | |
8883 | { | |
8884 | int ret; | |
8885 | int old_period, old_runtime; | |
8886 | static DEFINE_MUTEX(mutex); | |
8887 | ||
8888 | mutex_lock(&mutex); | |
8889 | old_period = sysctl_sched_rt_period; | |
8890 | old_runtime = sysctl_sched_rt_runtime; | |
8891 | ||
8892 | ret = proc_dointvec(table, write, filp, buffer, lenp, ppos); | |
8893 | ||
8894 | if (!ret && write) { | |
8895 | ret = sched_rt_global_constraints(); | |
8896 | if (ret) { | |
8897 | sysctl_sched_rt_period = old_period; | |
8898 | sysctl_sched_rt_runtime = old_runtime; | |
8899 | } else { | |
8900 | def_rt_bandwidth.rt_runtime = global_rt_runtime(); | |
8901 | def_rt_bandwidth.rt_period = | |
8902 | ns_to_ktime(global_rt_period()); | |
8903 | } | |
8904 | } | |
8905 | mutex_unlock(&mutex); | |
8906 | ||
8907 | return ret; | |
8908 | } | |
68318b8e | 8909 | |
052f1dc7 | 8910 | #ifdef CONFIG_CGROUP_SCHED |
68318b8e SV |
8911 | |
8912 | /* return corresponding task_group object of a cgroup */ | |
2b01dfe3 | 8913 | static inline struct task_group *cgroup_tg(struct cgroup *cgrp) |
68318b8e | 8914 | { |
2b01dfe3 PM |
8915 | return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), |
8916 | struct task_group, css); | |
68318b8e SV |
8917 | } |
8918 | ||
8919 | static struct cgroup_subsys_state * | |
2b01dfe3 | 8920 | cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) |
68318b8e | 8921 | { |
ec7dc8ac | 8922 | struct task_group *tg, *parent; |
68318b8e | 8923 | |
2b01dfe3 | 8924 | if (!cgrp->parent) { |
68318b8e | 8925 | /* This is early initialization for the top cgroup */ |
2b01dfe3 | 8926 | init_task_group.css.cgroup = cgrp; |
68318b8e SV |
8927 | return &init_task_group.css; |
8928 | } | |
8929 | ||
ec7dc8ac DG |
8930 | parent = cgroup_tg(cgrp->parent); |
8931 | tg = sched_create_group(parent); | |
68318b8e SV |
8932 | if (IS_ERR(tg)) |
8933 | return ERR_PTR(-ENOMEM); | |
8934 | ||
8935 | /* Bind the cgroup to task_group object we just created */ | |
2b01dfe3 | 8936 | tg->css.cgroup = cgrp; |
68318b8e SV |
8937 | |
8938 | return &tg->css; | |
8939 | } | |
8940 | ||
41a2d6cf IM |
8941 | static void |
8942 | cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) | |
68318b8e | 8943 | { |
2b01dfe3 | 8944 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
8945 | |
8946 | sched_destroy_group(tg); | |
8947 | } | |
8948 | ||
41a2d6cf IM |
8949 | static int |
8950 | cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, | |
8951 | struct task_struct *tsk) | |
68318b8e | 8952 | { |
b68aa230 PZ |
8953 | #ifdef CONFIG_RT_GROUP_SCHED |
8954 | /* Don't accept realtime tasks when there is no way for them to run */ | |
d0b27fa7 | 8955 | if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0) |
b68aa230 PZ |
8956 | return -EINVAL; |
8957 | #else | |
68318b8e SV |
8958 | /* We don't support RT-tasks being in separate groups */ |
8959 | if (tsk->sched_class != &fair_sched_class) | |
8960 | return -EINVAL; | |
b68aa230 | 8961 | #endif |
68318b8e SV |
8962 | |
8963 | return 0; | |
8964 | } | |
8965 | ||
8966 | static void | |
2b01dfe3 | 8967 | cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, |
68318b8e SV |
8968 | struct cgroup *old_cont, struct task_struct *tsk) |
8969 | { | |
8970 | sched_move_task(tsk); | |
8971 | } | |
8972 | ||
052f1dc7 | 8973 | #ifdef CONFIG_FAIR_GROUP_SCHED |
f4c753b7 | 8974 | static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, |
2b01dfe3 | 8975 | u64 shareval) |
68318b8e | 8976 | { |
2b01dfe3 | 8977 | return sched_group_set_shares(cgroup_tg(cgrp), shareval); |
68318b8e SV |
8978 | } |
8979 | ||
f4c753b7 | 8980 | static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) |
68318b8e | 8981 | { |
2b01dfe3 | 8982 | struct task_group *tg = cgroup_tg(cgrp); |
68318b8e SV |
8983 | |
8984 | return (u64) tg->shares; | |
8985 | } | |
052f1dc7 | 8986 | #endif |
68318b8e | 8987 | |
052f1dc7 | 8988 | #ifdef CONFIG_RT_GROUP_SCHED |
0c70814c | 8989 | static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, |
06ecb27c | 8990 | s64 val) |
6f505b16 | 8991 | { |
06ecb27c | 8992 | return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); |
6f505b16 PZ |
8993 | } |
8994 | ||
06ecb27c | 8995 | static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) |
6f505b16 | 8996 | { |
06ecb27c | 8997 | return sched_group_rt_runtime(cgroup_tg(cgrp)); |
6f505b16 | 8998 | } |
d0b27fa7 PZ |
8999 | |
9000 | static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, | |
9001 | u64 rt_period_us) | |
9002 | { | |
9003 | return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); | |
9004 | } | |
9005 | ||
9006 | static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) | |
9007 | { | |
9008 | return sched_group_rt_period(cgroup_tg(cgrp)); | |
9009 | } | |
052f1dc7 | 9010 | #endif |
6f505b16 | 9011 | |
fe5c7cc2 | 9012 | static struct cftype cpu_files[] = { |
052f1dc7 | 9013 | #ifdef CONFIG_FAIR_GROUP_SCHED |
fe5c7cc2 PM |
9014 | { |
9015 | .name = "shares", | |
f4c753b7 PM |
9016 | .read_u64 = cpu_shares_read_u64, |
9017 | .write_u64 = cpu_shares_write_u64, | |
fe5c7cc2 | 9018 | }, |
052f1dc7 PZ |
9019 | #endif |
9020 | #ifdef CONFIG_RT_GROUP_SCHED | |
6f505b16 | 9021 | { |
9f0c1e56 | 9022 | .name = "rt_runtime_us", |
06ecb27c PM |
9023 | .read_s64 = cpu_rt_runtime_read, |
9024 | .write_s64 = cpu_rt_runtime_write, | |
6f505b16 | 9025 | }, |
d0b27fa7 PZ |
9026 | { |
9027 | .name = "rt_period_us", | |
f4c753b7 PM |
9028 | .read_u64 = cpu_rt_period_read_uint, |
9029 | .write_u64 = cpu_rt_period_write_uint, | |
d0b27fa7 | 9030 | }, |
052f1dc7 | 9031 | #endif |
68318b8e SV |
9032 | }; |
9033 | ||
9034 | static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
9035 | { | |
fe5c7cc2 | 9036 | return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); |
68318b8e SV |
9037 | } |
9038 | ||
9039 | struct cgroup_subsys cpu_cgroup_subsys = { | |
38605cae IM |
9040 | .name = "cpu", |
9041 | .create = cpu_cgroup_create, | |
9042 | .destroy = cpu_cgroup_destroy, | |
9043 | .can_attach = cpu_cgroup_can_attach, | |
9044 | .attach = cpu_cgroup_attach, | |
9045 | .populate = cpu_cgroup_populate, | |
9046 | .subsys_id = cpu_cgroup_subsys_id, | |
68318b8e SV |
9047 | .early_init = 1, |
9048 | }; | |
9049 | ||
052f1dc7 | 9050 | #endif /* CONFIG_CGROUP_SCHED */ |
d842de87 SV |
9051 | |
9052 | #ifdef CONFIG_CGROUP_CPUACCT | |
9053 | ||
9054 | /* | |
9055 | * CPU accounting code for task groups. | |
9056 | * | |
9057 | * Based on the work by Paul Menage (menage@google.com) and Balbir Singh | |
9058 | * (balbir@in.ibm.com). | |
9059 | */ | |
9060 | ||
9061 | /* track cpu usage of a group of tasks */ | |
9062 | struct cpuacct { | |
9063 | struct cgroup_subsys_state css; | |
9064 | /* cpuusage holds pointer to a u64-type object on every cpu */ | |
9065 | u64 *cpuusage; | |
9066 | }; | |
9067 | ||
9068 | struct cgroup_subsys cpuacct_subsys; | |
9069 | ||
9070 | /* return cpu accounting group corresponding to this container */ | |
32cd756a | 9071 | static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp) |
d842de87 | 9072 | { |
32cd756a | 9073 | return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id), |
d842de87 SV |
9074 | struct cpuacct, css); |
9075 | } | |
9076 | ||
9077 | /* return cpu accounting group to which this task belongs */ | |
9078 | static inline struct cpuacct *task_ca(struct task_struct *tsk) | |
9079 | { | |
9080 | return container_of(task_subsys_state(tsk, cpuacct_subsys_id), | |
9081 | struct cpuacct, css); | |
9082 | } | |
9083 | ||
9084 | /* create a new cpu accounting group */ | |
9085 | static struct cgroup_subsys_state *cpuacct_create( | |
32cd756a | 9086 | struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 SV |
9087 | { |
9088 | struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL); | |
9089 | ||
9090 | if (!ca) | |
9091 | return ERR_PTR(-ENOMEM); | |
9092 | ||
9093 | ca->cpuusage = alloc_percpu(u64); | |
9094 | if (!ca->cpuusage) { | |
9095 | kfree(ca); | |
9096 | return ERR_PTR(-ENOMEM); | |
9097 | } | |
9098 | ||
9099 | return &ca->css; | |
9100 | } | |
9101 | ||
9102 | /* destroy an existing cpu accounting group */ | |
41a2d6cf | 9103 | static void |
32cd756a | 9104 | cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9105 | { |
32cd756a | 9106 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9107 | |
9108 | free_percpu(ca->cpuusage); | |
9109 | kfree(ca); | |
9110 | } | |
9111 | ||
9112 | /* return total cpu usage (in nanoseconds) of a group */ | |
32cd756a | 9113 | static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) |
d842de87 | 9114 | { |
32cd756a | 9115 | struct cpuacct *ca = cgroup_ca(cgrp); |
d842de87 SV |
9116 | u64 totalcpuusage = 0; |
9117 | int i; | |
9118 | ||
9119 | for_each_possible_cpu(i) { | |
9120 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9121 | ||
9122 | /* | |
9123 | * Take rq->lock to make 64-bit addition safe on 32-bit | |
9124 | * platforms. | |
9125 | */ | |
9126 | spin_lock_irq(&cpu_rq(i)->lock); | |
9127 | totalcpuusage += *cpuusage; | |
9128 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9129 | } | |
9130 | ||
9131 | return totalcpuusage; | |
9132 | } | |
9133 | ||
0297b803 DG |
9134 | static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, |
9135 | u64 reset) | |
9136 | { | |
9137 | struct cpuacct *ca = cgroup_ca(cgrp); | |
9138 | int err = 0; | |
9139 | int i; | |
9140 | ||
9141 | if (reset) { | |
9142 | err = -EINVAL; | |
9143 | goto out; | |
9144 | } | |
9145 | ||
9146 | for_each_possible_cpu(i) { | |
9147 | u64 *cpuusage = percpu_ptr(ca->cpuusage, i); | |
9148 | ||
9149 | spin_lock_irq(&cpu_rq(i)->lock); | |
9150 | *cpuusage = 0; | |
9151 | spin_unlock_irq(&cpu_rq(i)->lock); | |
9152 | } | |
9153 | out: | |
9154 | return err; | |
9155 | } | |
9156 | ||
d842de87 SV |
9157 | static struct cftype files[] = { |
9158 | { | |
9159 | .name = "usage", | |
f4c753b7 PM |
9160 | .read_u64 = cpuusage_read, |
9161 | .write_u64 = cpuusage_write, | |
d842de87 SV |
9162 | }, |
9163 | }; | |
9164 | ||
32cd756a | 9165 | static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) |
d842de87 | 9166 | { |
32cd756a | 9167 | return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); |
d842de87 SV |
9168 | } |
9169 | ||
9170 | /* | |
9171 | * charge this task's execution time to its accounting group. | |
9172 | * | |
9173 | * called with rq->lock held. | |
9174 | */ | |
9175 | static void cpuacct_charge(struct task_struct *tsk, u64 cputime) | |
9176 | { | |
9177 | struct cpuacct *ca; | |
9178 | ||
9179 | if (!cpuacct_subsys.active) | |
9180 | return; | |
9181 | ||
9182 | ca = task_ca(tsk); | |
9183 | if (ca) { | |
9184 | u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk)); | |
9185 | ||
9186 | *cpuusage += cputime; | |
9187 | } | |
9188 | } | |
9189 | ||
9190 | struct cgroup_subsys cpuacct_subsys = { | |
9191 | .name = "cpuacct", | |
9192 | .create = cpuacct_create, | |
9193 | .destroy = cpuacct_destroy, | |
9194 | .populate = cpuacct_populate, | |
9195 | .subsys_id = cpuacct_subsys_id, | |
9196 | }; | |
9197 | #endif /* CONFIG_CGROUP_CPUACCT */ |