i386: fix a hang on stuck nmi watchdog
[linux-2.6] / 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 /* This protects CPUs going up and down... */
19 static DEFINE_MUTEX(cpu_add_remove_lock);
20 static DEFINE_MUTEX(cpu_bitmask_lock);
21
22 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
23
24 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
25  * Should always be manipulated under cpu_add_remove_lock
26  */
27 static int cpu_hotplug_disabled;
28
29 #ifdef CONFIG_HOTPLUG_CPU
30
31 /* Crappy recursive lock-takers in cpufreq! Complain loudly about idiots */
32 static struct task_struct *recursive;
33 static int recursive_depth;
34
35 void lock_cpu_hotplug(void)
36 {
37         struct task_struct *tsk = current;
38
39         if (tsk == recursive) {
40                 static int warnings = 10;
41                 if (warnings) {
42                         printk(KERN_ERR "Lukewarm IQ detected in hotplug locking\n");
43                         WARN_ON(1);
44                         warnings--;
45                 }
46                 recursive_depth++;
47                 return;
48         }
49         mutex_lock(&cpu_bitmask_lock);
50         recursive = tsk;
51 }
52 EXPORT_SYMBOL_GPL(lock_cpu_hotplug);
53
54 void unlock_cpu_hotplug(void)
55 {
56         WARN_ON(recursive != current);
57         if (recursive_depth) {
58                 recursive_depth--;
59                 return;
60         }
61         recursive = NULL;
62         mutex_unlock(&cpu_bitmask_lock);
63 }
64 EXPORT_SYMBOL_GPL(unlock_cpu_hotplug);
65
66 #endif  /* CONFIG_HOTPLUG_CPU */
67
68 /* Need to know about CPUs going up/down? */
69 int __cpuinit register_cpu_notifier(struct notifier_block *nb)
70 {
71         int ret;
72         mutex_lock(&cpu_add_remove_lock);
73         ret = raw_notifier_chain_register(&cpu_chain, nb);
74         mutex_unlock(&cpu_add_remove_lock);
75         return ret;
76 }
77
78 #ifdef CONFIG_HOTPLUG_CPU
79
80 EXPORT_SYMBOL(register_cpu_notifier);
81
82 void unregister_cpu_notifier(struct notifier_block *nb)
83 {
84         mutex_lock(&cpu_add_remove_lock);
85         raw_notifier_chain_unregister(&cpu_chain, nb);
86         mutex_unlock(&cpu_add_remove_lock);
87 }
88 EXPORT_SYMBOL(unregister_cpu_notifier);
89
90 static inline void check_for_tasks(int cpu)
91 {
92         struct task_struct *p;
93
94         write_lock_irq(&tasklist_lock);
95         for_each_process(p) {
96                 if (task_cpu(p) == cpu &&
97                     (!cputime_eq(p->utime, cputime_zero) ||
98                      !cputime_eq(p->stime, cputime_zero)))
99                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
100                                 (state = %ld, flags = %x) \n",
101                                  p->comm, p->pid, cpu, p->state, p->flags);
102         }
103         write_unlock_irq(&tasklist_lock);
104 }
105
106 struct take_cpu_down_param {
107         unsigned long mod;
108         void *hcpu;
109 };
110
111 /* Take this CPU down. */
112 static int take_cpu_down(void *_param)
113 {
114         struct take_cpu_down_param *param = _param;
115         int err;
116
117         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
118                                 param->hcpu);
119         /* Ensure this CPU doesn't handle any more interrupts. */
120         err = __cpu_disable();
121         if (err < 0)
122                 return err;
123
124         /* Force idle task to run as soon as we yield: it should
125            immediately notice cpu is offline and die quickly. */
126         sched_idle_next();
127         return 0;
128 }
129
130 /* Requires cpu_add_remove_lock to be held */
131 static int _cpu_down(unsigned int cpu, int tasks_frozen)
132 {
133         int err, nr_calls = 0;
134         struct task_struct *p;
135         cpumask_t old_allowed, tmp;
136         void *hcpu = (void *)(long)cpu;
137         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
138         struct take_cpu_down_param tcd_param = {
139                 .mod = mod,
140                 .hcpu = hcpu,
141         };
142
143         if (num_online_cpus() == 1)
144                 return -EBUSY;
145
146         if (!cpu_online(cpu))
147                 return -EINVAL;
148
149         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
150         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
151                                         hcpu, -1, &nr_calls);
152         if (err == NOTIFY_BAD) {
153                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
154                                           hcpu, nr_calls, NULL);
155                 printk("%s: attempt to take down CPU %u failed\n",
156                                 __FUNCTION__, cpu);
157                 err = -EINVAL;
158                 goto out_release;
159         }
160
161         /* Ensure that we are not runnable on dying cpu */
162         old_allowed = current->cpus_allowed;
163         tmp = CPU_MASK_ALL;
164         cpu_clear(cpu, tmp);
165         set_cpus_allowed(current, tmp);
166
167         mutex_lock(&cpu_bitmask_lock);
168         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
169         mutex_unlock(&cpu_bitmask_lock);
170
171         if (IS_ERR(p) || cpu_online(cpu)) {
172                 /* CPU didn't die: tell everyone.  Can't complain. */
173                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
174                                             hcpu) == NOTIFY_BAD)
175                         BUG();
176
177                 if (IS_ERR(p)) {
178                         err = PTR_ERR(p);
179                         goto out_allowed;
180                 }
181                 goto out_thread;
182         }
183
184         /* Wait for it to sleep (leaving idle task). */
185         while (!idle_cpu(cpu))
186                 yield();
187
188         /* This actually kills the CPU. */
189         __cpu_die(cpu);
190
191         /* CPU is completely dead: tell everyone.  Too late to complain. */
192         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
193                                     hcpu) == NOTIFY_BAD)
194                 BUG();
195
196         check_for_tasks(cpu);
197
198 out_thread:
199         err = kthread_stop(p);
200 out_allowed:
201         set_cpus_allowed(current, old_allowed);
202 out_release:
203         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
204         return err;
205 }
206
207 int cpu_down(unsigned int cpu)
208 {
209         int err = 0;
210
211         mutex_lock(&cpu_add_remove_lock);
212         if (cpu_hotplug_disabled)
213                 err = -EBUSY;
214         else
215                 err = _cpu_down(cpu, 0);
216
217         mutex_unlock(&cpu_add_remove_lock);
218         return err;
219 }
220 #endif /*CONFIG_HOTPLUG_CPU*/
221
222 /* Requires cpu_add_remove_lock to be held */
223 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
224 {
225         int ret, nr_calls = 0;
226         void *hcpu = (void *)(long)cpu;
227         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
228
229         if (cpu_online(cpu) || !cpu_present(cpu))
230                 return -EINVAL;
231
232         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
233         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
234                                                         -1, &nr_calls);
235         if (ret == NOTIFY_BAD) {
236                 printk("%s: attempt to bring up CPU %u failed\n",
237                                 __FUNCTION__, cpu);
238                 ret = -EINVAL;
239                 goto out_notify;
240         }
241
242         /* Arch-specific enabling code. */
243         mutex_lock(&cpu_bitmask_lock);
244         ret = __cpu_up(cpu);
245         mutex_unlock(&cpu_bitmask_lock);
246         if (ret != 0)
247                 goto out_notify;
248         BUG_ON(!cpu_online(cpu));
249
250         /* Now call notifier in preparation. */
251         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
252
253 out_notify:
254         if (ret != 0)
255                 __raw_notifier_call_chain(&cpu_chain,
256                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
257         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
258
259         return ret;
260 }
261
262 int __cpuinit cpu_up(unsigned int cpu)
263 {
264         int err = 0;
265
266         mutex_lock(&cpu_add_remove_lock);
267         if (cpu_hotplug_disabled)
268                 err = -EBUSY;
269         else
270                 err = _cpu_up(cpu, 0);
271
272         mutex_unlock(&cpu_add_remove_lock);
273         return err;
274 }
275
276 #ifdef CONFIG_PM_SLEEP_SMP
277 static cpumask_t frozen_cpus;
278
279 int disable_nonboot_cpus(void)
280 {
281         int cpu, first_cpu, error = 0;
282
283         mutex_lock(&cpu_add_remove_lock);
284         first_cpu = first_cpu(cpu_online_map);
285         /* We take down all of the non-boot CPUs in one shot to avoid races
286          * with the userspace trying to use the CPU hotplug at the same time
287          */
288         cpus_clear(frozen_cpus);
289         printk("Disabling non-boot CPUs ...\n");
290         for_each_online_cpu(cpu) {
291                 if (cpu == first_cpu)
292                         continue;
293                 error = _cpu_down(cpu, 1);
294                 if (!error) {
295                         cpu_set(cpu, frozen_cpus);
296                         printk("CPU%d is down\n", cpu);
297                 } else {
298                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
299                                 cpu, error);
300                         break;
301                 }
302         }
303         if (!error) {
304                 BUG_ON(num_online_cpus() > 1);
305                 /* Make sure the CPUs won't be enabled by someone else */
306                 cpu_hotplug_disabled = 1;
307         } else {
308                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
309         }
310         mutex_unlock(&cpu_add_remove_lock);
311         return error;
312 }
313
314 void enable_nonboot_cpus(void)
315 {
316         int cpu, error;
317
318         /* Allow everyone to use the CPU hotplug again */
319         mutex_lock(&cpu_add_remove_lock);
320         cpu_hotplug_disabled = 0;
321         if (cpus_empty(frozen_cpus))
322                 goto out;
323
324         printk("Enabling non-boot CPUs ...\n");
325         for_each_cpu_mask(cpu, frozen_cpus) {
326                 error = _cpu_up(cpu, 1);
327                 if (!error) {
328                         printk("CPU%d is up\n", cpu);
329                         continue;
330                 }
331                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
332         }
333         cpus_clear(frozen_cpus);
334 out:
335         mutex_unlock(&cpu_add_remove_lock);
336 }
337 #endif /* CONFIG_PM_SLEEP_SMP */