cpu-hotplug: refcount based cpu hotplug
[linux-2.6.git] / kernel / cpu.c
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>
16 #include <linux/mutex.h>
17
18 /* Serializes the updates to cpu_online_map, cpu_present_map */
19 static DEFINE_MUTEX(cpu_add_remove_lock);
20
21 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
22
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  */
26 static int cpu_hotplug_disabled;
27
28 static 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;
38
39 #define writer_exists() (cpu_hotplug.active_writer != NULL)
40
41 void __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
50
51 void lock_cpu_hotplug(void)
52 {
53         might_sleep();
54         if (cpu_hotplug.active_writer == current)
55                 return;
56         mutex_lock(&cpu_hotplug.lock);
57         cpu_hotplug.refcount++;
58         mutex_unlock(&cpu_hotplug.lock);
59
60 }
61 EXPORT_SYMBOL_GPL(lock_cpu_hotplug);
62
63 void unlock_cpu_hotplug(void)
64 {
65         if (cpu_hotplug.active_writer == current)
66                 return;
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
75 }
76 EXPORT_SYMBOL_GPL(unlock_cpu_hotplug);
77
78 #endif  /* CONFIG_HOTPLUG_CPU */
79
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  */
84 void cpu_maps_update_begin(void)
85 {
86         mutex_lock(&cpu_add_remove_lock);
87 }
88
89 void 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  */
116 static 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
133 static void cpu_hotplug_done(void)
134 {
135         cpu_hotplug.active_writer = NULL;
136         mutex_unlock(&cpu_hotplug.lock);
137 }
138 /* Need to know about CPUs going up/down? */
139 int __cpuinit register_cpu_notifier(struct notifier_block *nb)
140 {
141         int ret;
142         cpu_maps_update_begin();
143         ret = raw_notifier_chain_register(&cpu_chain, nb);
144         cpu_maps_update_done();
145         return ret;
146 }
147
148 #ifdef CONFIG_HOTPLUG_CPU
149
150 EXPORT_SYMBOL(register_cpu_notifier);
151
152 void unregister_cpu_notifier(struct notifier_block *nb)
153 {
154         cpu_maps_update_begin();
155         raw_notifier_chain_unregister(&cpu_chain, nb);
156         cpu_maps_update_done();
157 }
158 EXPORT_SYMBOL(unregister_cpu_notifier);
159
160 static 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\
170                                 (state = %ld, flags = %x) \n",
171                                  p->comm, task_pid_nr(p), cpu,
172                                  p->state, p->flags);
173         }
174         write_unlock_irq(&tasklist_lock);
175 }
176
177 struct take_cpu_down_param {
178         unsigned long mod;
179         void *hcpu;
180 };
181
182 /* Take this CPU down. */
183 static int take_cpu_down(void *_param)
184 {
185         struct take_cpu_down_param *param = _param;
186         int err;
187
188         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
189                                 param->hcpu);
190         /* Ensure this CPU doesn't handle any more interrupts. */
191         err = __cpu_disable();
192         if (err < 0)
193                 return err;
194
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;
199 }
200
201 /* Requires cpu_add_remove_lock to be held */
202 static int _cpu_down(unsigned int cpu, int tasks_frozen)
203 {
204         int err, nr_calls = 0;
205         struct task_struct *p;
206         cpumask_t old_allowed, tmp;
207         void *hcpu = (void *)(long)cpu;
208         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
209         struct take_cpu_down_param tcd_param = {
210                 .mod = mod,
211                 .hcpu = hcpu,
212         };
213
214         if (num_online_cpus() == 1)
215                 return -EBUSY;
216
217         if (!cpu_online(cpu))
218                 return -EINVAL;
219
220         cpu_hotplug_begin();
221         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
222         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
223                                         hcpu, -1, &nr_calls);
224         if (err == NOTIFY_BAD) {
225                 nr_calls--;
226                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
227                                           hcpu, nr_calls, NULL);
228                 printk("%s: attempt to take down CPU %u failed\n",
229                                 __FUNCTION__, cpu);
230                 err = -EINVAL;
231                 goto out_release;
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
240         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
241
242         if (IS_ERR(p) || cpu_online(cpu)) {
243                 /* CPU didn't die: tell everyone.  Can't complain. */
244                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
245                                             hcpu) == NOTIFY_BAD)
246                         BUG();
247
248                 if (IS_ERR(p)) {
249                         err = PTR_ERR(p);
250                         goto out_allowed;
251                 }
252                 goto out_thread;
253         }
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
262         /* CPU is completely dead: tell everyone.  Too late to complain. */
263         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
264                                     hcpu) == NOTIFY_BAD)
265                 BUG();
266
267         check_for_tasks(cpu);
268
269 out_thread:
270         err = kthread_stop(p);
271 out_allowed:
272         set_cpus_allowed(current, old_allowed);
273 out_release:
274         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
275         cpu_hotplug_done();
276         return err;
277 }
278
279 int cpu_down(unsigned int cpu)
280 {
281         int err = 0;
282
283         cpu_maps_update_begin();
284         if (cpu_hotplug_disabled)
285                 err = -EBUSY;
286         else
287                 err = _cpu_down(cpu, 0);
288
289         cpu_maps_update_done();
290         return err;
291 }
292 #endif /*CONFIG_HOTPLUG_CPU*/
293
294 /* Requires cpu_add_remove_lock to be held */
295 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
296 {
297         int ret, nr_calls = 0;
298         void *hcpu = (void *)(long)cpu;
299         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
300
301         if (cpu_online(cpu) || !cpu_present(cpu))
302                 return -EINVAL;
303
304         cpu_hotplug_begin();
305         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
306         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
307                                                         -1, &nr_calls);
308         if (ret == NOTIFY_BAD) {
309                 nr_calls--;
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;
320         BUG_ON(!cpu_online(cpu));
321
322         /* Now call notifier in preparation. */
323         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
324
325 out_notify:
326         if (ret != 0)
327                 __raw_notifier_call_chain(&cpu_chain,
328                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
329         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
330         cpu_hotplug_done();
331
332         return ret;
333 }
334
335 int __cpuinit cpu_up(unsigned int cpu)
336 {
337         int err = 0;
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         }
347
348         cpu_maps_update_begin();
349         if (cpu_hotplug_disabled)
350                 err = -EBUSY;
351         else
352                 err = _cpu_up(cpu, 0);
353
354         cpu_maps_update_done();
355         return err;
356 }
357
358 #ifdef CONFIG_PM_SLEEP_SMP
359 static cpumask_t frozen_cpus;
360
361 int disable_nonboot_cpus(void)
362 {
363         int cpu, first_cpu, error = 0;
364
365         cpu_maps_update_begin();
366         first_cpu = first_cpu(cpu_online_map);
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;
375                 error = _cpu_down(cpu, 1);
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 {
390                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
391         }
392         cpu_maps_update_done();
393         return error;
394 }
395
396 void enable_nonboot_cpus(void)
397 {
398         int cpu, error;
399
400         /* Allow everyone to use the CPU hotplug again */
401         cpu_maps_update_begin();
402         cpu_hotplug_disabled = 0;
403         if (cpus_empty(frozen_cpus))
404                 goto out;
405
406         printk("Enabling non-boot CPUs ...\n");
407         for_each_cpu_mask(cpu, frozen_cpus) {
408                 error = _cpu_up(cpu, 1);
409                 if (!error) {
410                         printk("CPU%d is up\n", cpu);
411                         continue;
412                 }
413                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
414         }
415         cpus_clear(frozen_cpus);
416 out:
417         cpu_maps_update_done();
418 }
419 #endif /* CONFIG_PM_SLEEP_SMP */