]> bbs.cooldavid.org Git - net-next-2.6.git/blame - kernel/cpu.c
cpu-hotplug: refcount based cpu hotplug
[net-next-2.6.git] / kernel / cpu.c
CommitLineData
1da177e4
LT
1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/proc_fs.h>
7#include <linux/smp.h>
8#include <linux/init.h>
9#include <linux/notifier.h>
10#include <linux/sched.h>
11#include <linux/unistd.h>
12#include <linux/cpu.h>
13#include <linux/module.h>
14#include <linux/kthread.h>
15#include <linux/stop_machine.h>
81615b62 16#include <linux/mutex.h>
1da177e4 17
d221938c 18/* Serializes the updates to cpu_online_map, cpu_present_map */
aa953877 19static DEFINE_MUTEX(cpu_add_remove_lock);
1da177e4 20
bd5349cf 21static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
1da177e4 22
e3920fb4
RW
23/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
24 * Should always be manipulated under cpu_add_remove_lock
25 */
26static int cpu_hotplug_disabled;
27
d221938c
GS
28static struct {
29 struct task_struct *active_writer;
30 struct mutex lock; /* Synchronizes accesses to refcount, */
31 /*
32 * Also blocks the new readers during
33 * an ongoing cpu hotplug operation.
34 */
35 int refcount;
36 wait_queue_head_t writer_queue;
37} cpu_hotplug;
90d45d17 38
d221938c
GS
39#define writer_exists() (cpu_hotplug.active_writer != NULL)
40
41void __init cpu_hotplug_init(void)
42{
43 cpu_hotplug.active_writer = NULL;
44 mutex_init(&cpu_hotplug.lock);
45 cpu_hotplug.refcount = 0;
46 init_waitqueue_head(&cpu_hotplug.writer_queue);
47}
48
49#ifdef CONFIG_HOTPLUG_CPU
90d45d17 50
a9d9baa1
AR
51void lock_cpu_hotplug(void)
52{
d221938c
GS
53 might_sleep();
54 if (cpu_hotplug.active_writer == current)
aa953877 55 return;
d221938c
GS
56 mutex_lock(&cpu_hotplug.lock);
57 cpu_hotplug.refcount++;
58 mutex_unlock(&cpu_hotplug.lock);
59
a9d9baa1
AR
60}
61EXPORT_SYMBOL_GPL(lock_cpu_hotplug);
90d45d17 62
a9d9baa1
AR
63void unlock_cpu_hotplug(void)
64{
d221938c 65 if (cpu_hotplug.active_writer == current)
aa953877 66 return;
d221938c
GS
67 mutex_lock(&cpu_hotplug.lock);
68 cpu_hotplug.refcount--;
69
70 if (unlikely(writer_exists()) && !cpu_hotplug.refcount)
71 wake_up(&cpu_hotplug.writer_queue);
72
73 mutex_unlock(&cpu_hotplug.lock);
74
a9d9baa1
AR
75}
76EXPORT_SYMBOL_GPL(unlock_cpu_hotplug);
77
a9d9baa1 78#endif /* CONFIG_HOTPLUG_CPU */
90d45d17 79
d221938c
GS
80/*
81 * The following two API's must be used when attempting
82 * to serialize the updates to cpu_online_map, cpu_present_map.
83 */
84void cpu_maps_update_begin(void)
85{
86 mutex_lock(&cpu_add_remove_lock);
87}
88
89void cpu_maps_update_done(void)
90{
91 mutex_unlock(&cpu_add_remove_lock);
92}
93
94/*
95 * This ensures that the hotplug operation can begin only when the
96 * refcount goes to zero.
97 *
98 * Note that during a cpu-hotplug operation, the new readers, if any,
99 * will be blocked by the cpu_hotplug.lock
100 *
101 * Since cpu_maps_update_begin is always called after invoking
102 * cpu_maps_update_begin, we can be sure that only one writer is active.
103 *
104 * Note that theoretically, there is a possibility of a livelock:
105 * - Refcount goes to zero, last reader wakes up the sleeping
106 * writer.
107 * - Last reader unlocks the cpu_hotplug.lock.
108 * - A new reader arrives at this moment, bumps up the refcount.
109 * - The writer acquires the cpu_hotplug.lock finds the refcount
110 * non zero and goes to sleep again.
111 *
112 * However, this is very difficult to achieve in practice since
113 * lock_cpu_hotplug() not an api which is called all that often.
114 *
115 */
116static void cpu_hotplug_begin(void)
117{
118 DECLARE_WAITQUEUE(wait, current);
119
120 mutex_lock(&cpu_hotplug.lock);
121
122 cpu_hotplug.active_writer = current;
123 add_wait_queue_exclusive(&cpu_hotplug.writer_queue, &wait);
124 while (cpu_hotplug.refcount) {
125 set_current_state(TASK_UNINTERRUPTIBLE);
126 mutex_unlock(&cpu_hotplug.lock);
127 schedule();
128 mutex_lock(&cpu_hotplug.lock);
129 }
130 remove_wait_queue_locked(&cpu_hotplug.writer_queue, &wait);
131}
132
133static void cpu_hotplug_done(void)
134{
135 cpu_hotplug.active_writer = NULL;
136 mutex_unlock(&cpu_hotplug.lock);
137}
1da177e4 138/* Need to know about CPUs going up/down? */
65edc68c 139int __cpuinit register_cpu_notifier(struct notifier_block *nb)
1da177e4 140{
bd5349cf 141 int ret;
d221938c 142 cpu_maps_update_begin();
bd5349cf 143 ret = raw_notifier_chain_register(&cpu_chain, nb);
d221938c 144 cpu_maps_update_done();
bd5349cf 145 return ret;
1da177e4 146}
65edc68c
CS
147
148#ifdef CONFIG_HOTPLUG_CPU
149
1da177e4
LT
150EXPORT_SYMBOL(register_cpu_notifier);
151
152void unregister_cpu_notifier(struct notifier_block *nb)
153{
d221938c 154 cpu_maps_update_begin();
bd5349cf 155 raw_notifier_chain_unregister(&cpu_chain, nb);
d221938c 156 cpu_maps_update_done();
1da177e4
LT
157}
158EXPORT_SYMBOL(unregister_cpu_notifier);
159
1da177e4
LT
160static inline void check_for_tasks(int cpu)
161{
162 struct task_struct *p;
163
164 write_lock_irq(&tasklist_lock);
165 for_each_process(p) {
166 if (task_cpu(p) == cpu &&
167 (!cputime_eq(p->utime, cputime_zero) ||
168 !cputime_eq(p->stime, cputime_zero)))
169 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
e7407dcc 170 (state = %ld, flags = %x) \n",
ba25f9dc
PE
171 p->comm, task_pid_nr(p), cpu,
172 p->state, p->flags);
1da177e4
LT
173 }
174 write_unlock_irq(&tasklist_lock);
175}
176
db912f96
AK
177struct take_cpu_down_param {
178 unsigned long mod;
179 void *hcpu;
180};
181
1da177e4 182/* Take this CPU down. */
db912f96 183static int take_cpu_down(void *_param)
1da177e4 184{
db912f96 185 struct take_cpu_down_param *param = _param;
1da177e4
LT
186 int err;
187
db912f96
AK
188 raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
189 param->hcpu);
1da177e4
LT
190 /* Ensure this CPU doesn't handle any more interrupts. */
191 err = __cpu_disable();
192 if (err < 0)
f3705136 193 return err;
1da177e4 194
f3705136
ZM
195 /* Force idle task to run as soon as we yield: it should
196 immediately notice cpu is offline and die quickly. */
197 sched_idle_next();
198 return 0;
1da177e4
LT
199}
200
e3920fb4 201/* Requires cpu_add_remove_lock to be held */
8bb78442 202static int _cpu_down(unsigned int cpu, int tasks_frozen)
1da177e4 203{
e7407dcc 204 int err, nr_calls = 0;
1da177e4
LT
205 struct task_struct *p;
206 cpumask_t old_allowed, tmp;
e7407dcc 207 void *hcpu = (void *)(long)cpu;
8bb78442 208 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
db912f96
AK
209 struct take_cpu_down_param tcd_param = {
210 .mod = mod,
211 .hcpu = hcpu,
212 };
1da177e4 213
e3920fb4
RW
214 if (num_online_cpus() == 1)
215 return -EBUSY;
1da177e4 216
e3920fb4
RW
217 if (!cpu_online(cpu))
218 return -EINVAL;
1da177e4 219
d221938c 220 cpu_hotplug_begin();
e7407dcc 221 raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
8bb78442 222 err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
e7407dcc 223 hcpu, -1, &nr_calls);
1da177e4 224 if (err == NOTIFY_BAD) {
a0d8cdb6 225 nr_calls--;
8bb78442
RW
226 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
227 hcpu, nr_calls, NULL);
1da177e4
LT
228 printk("%s: attempt to take down CPU %u failed\n",
229 __FUNCTION__, cpu);
baaca49f
GS
230 err = -EINVAL;
231 goto out_release;
1da177e4
LT
232 }
233
234 /* Ensure that we are not runnable on dying cpu */
235 old_allowed = current->cpus_allowed;
236 tmp = CPU_MASK_ALL;
237 cpu_clear(cpu, tmp);
238 set_cpus_allowed(current, tmp);
239
db912f96 240 p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
aa953877 241
8fa1d7d3 242 if (IS_ERR(p) || cpu_online(cpu)) {
1da177e4 243 /* CPU didn't die: tell everyone. Can't complain. */
8bb78442 244 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
e7407dcc 245 hcpu) == NOTIFY_BAD)
1da177e4
LT
246 BUG();
247
8fa1d7d3
ST
248 if (IS_ERR(p)) {
249 err = PTR_ERR(p);
250 goto out_allowed;
251 }
1da177e4 252 goto out_thread;
8fa1d7d3 253 }
1da177e4
LT
254
255 /* Wait for it to sleep (leaving idle task). */
256 while (!idle_cpu(cpu))
257 yield();
258
259 /* This actually kills the CPU. */
260 __cpu_die(cpu);
261
1da177e4 262 /* CPU is completely dead: tell everyone. Too late to complain. */
8bb78442
RW
263 if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
264 hcpu) == NOTIFY_BAD)
1da177e4
LT
265 BUG();
266
267 check_for_tasks(cpu);
268
269out_thread:
270 err = kthread_stop(p);
271out_allowed:
272 set_cpus_allowed(current, old_allowed);
baaca49f 273out_release:
8bb78442 274 raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
d221938c 275 cpu_hotplug_done();
e3920fb4
RW
276 return err;
277}
278
279int cpu_down(unsigned int cpu)
280{
281 int err = 0;
282
d221938c 283 cpu_maps_update_begin();
e3920fb4
RW
284 if (cpu_hotplug_disabled)
285 err = -EBUSY;
286 else
8bb78442 287 err = _cpu_down(cpu, 0);
e3920fb4 288
d221938c 289 cpu_maps_update_done();
1da177e4
LT
290 return err;
291}
292#endif /*CONFIG_HOTPLUG_CPU*/
293
e3920fb4 294/* Requires cpu_add_remove_lock to be held */
8bb78442 295static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
1da177e4 296{
baaca49f 297 int ret, nr_calls = 0;
1da177e4 298 void *hcpu = (void *)(long)cpu;
8bb78442 299 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
1da177e4 300
e3920fb4
RW
301 if (cpu_online(cpu) || !cpu_present(cpu))
302 return -EINVAL;
90d45d17 303
d221938c 304 cpu_hotplug_begin();
baaca49f 305 raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
8bb78442 306 ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
baaca49f 307 -1, &nr_calls);
1da177e4 308 if (ret == NOTIFY_BAD) {
a0d8cdb6 309 nr_calls--;
1da177e4
LT
310 printk("%s: attempt to bring up CPU %u failed\n",
311 __FUNCTION__, cpu);
312 ret = -EINVAL;
313 goto out_notify;
314 }
315
316 /* Arch-specific enabling code. */
317 ret = __cpu_up(cpu);
318 if (ret != 0)
319 goto out_notify;
6978c705 320 BUG_ON(!cpu_online(cpu));
1da177e4
LT
321
322 /* Now call notifier in preparation. */
8bb78442 323 raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
1da177e4
LT
324
325out_notify:
326 if (ret != 0)
baaca49f 327 __raw_notifier_call_chain(&cpu_chain,
8bb78442 328 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
baaca49f 329 raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
d221938c 330 cpu_hotplug_done();
e3920fb4
RW
331
332 return ret;
333}
334
b282b6f8 335int __cpuinit cpu_up(unsigned int cpu)
e3920fb4
RW
336{
337 int err = 0;
73e753a5
KH
338 if (!cpu_isset(cpu, cpu_possible_map)) {
339 printk(KERN_ERR "can't online cpu %d because it is not "
340 "configured as may-hotadd at boot time\n", cpu);
341#if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
342 printk(KERN_ERR "please check additional_cpus= boot "
343 "parameter\n");
344#endif
345 return -EINVAL;
346 }
e3920fb4 347
d221938c 348 cpu_maps_update_begin();
e3920fb4
RW
349 if (cpu_hotplug_disabled)
350 err = -EBUSY;
351 else
8bb78442 352 err = _cpu_up(cpu, 0);
e3920fb4 353
d221938c 354 cpu_maps_update_done();
e3920fb4
RW
355 return err;
356}
357
f3de4be9 358#ifdef CONFIG_PM_SLEEP_SMP
e3920fb4
RW
359static cpumask_t frozen_cpus;
360
361int disable_nonboot_cpus(void)
362{
e1d9fd2e 363 int cpu, first_cpu, error = 0;
e3920fb4 364
d221938c 365 cpu_maps_update_begin();
1d64b9cb 366 first_cpu = first_cpu(cpu_online_map);
e3920fb4
RW
367 /* We take down all of the non-boot CPUs in one shot to avoid races
368 * with the userspace trying to use the CPU hotplug at the same time
369 */
370 cpus_clear(frozen_cpus);
371 printk("Disabling non-boot CPUs ...\n");
372 for_each_online_cpu(cpu) {
373 if (cpu == first_cpu)
374 continue;
8bb78442 375 error = _cpu_down(cpu, 1);
e3920fb4
RW
376 if (!error) {
377 cpu_set(cpu, frozen_cpus);
378 printk("CPU%d is down\n", cpu);
379 } else {
380 printk(KERN_ERR "Error taking CPU%d down: %d\n",
381 cpu, error);
382 break;
383 }
384 }
385 if (!error) {
386 BUG_ON(num_online_cpus() > 1);
387 /* Make sure the CPUs won't be enabled by someone else */
388 cpu_hotplug_disabled = 1;
389 } else {
e1d9fd2e 390 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
e3920fb4 391 }
d221938c 392 cpu_maps_update_done();
e3920fb4
RW
393 return error;
394}
395
396void enable_nonboot_cpus(void)
397{
398 int cpu, error;
399
400 /* Allow everyone to use the CPU hotplug again */
d221938c 401 cpu_maps_update_begin();
e3920fb4 402 cpu_hotplug_disabled = 0;
ed746e3b 403 if (cpus_empty(frozen_cpus))
1d64b9cb 404 goto out;
e3920fb4
RW
405
406 printk("Enabling non-boot CPUs ...\n");
407 for_each_cpu_mask(cpu, frozen_cpus) {
8bb78442 408 error = _cpu_up(cpu, 1);
e3920fb4
RW
409 if (!error) {
410 printk("CPU%d is up\n", cpu);
411 continue;
412 }
1d64b9cb 413 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
e3920fb4
RW
414 }
415 cpus_clear(frozen_cpus);
1d64b9cb 416out:
d221938c 417 cpu_maps_update_done();
1da177e4 418}
f3de4be9 419#endif /* CONFIG_PM_SLEEP_SMP */