Add suspend-related notifications for 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 /* 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 /* Take this CPU down. */
107 static int take_cpu_down(void *unused)
108 {
109         int err;
110
111         /* Ensure this CPU doesn't handle any more interrupts. */
112         err = __cpu_disable();
113         if (err < 0)
114                 return err;
115
116         /* Force idle task to run as soon as we yield: it should
117            immediately notice cpu is offline and die quickly. */
118         sched_idle_next();
119         return 0;
120 }
121
122 /* Requires cpu_add_remove_lock to be held */
123 static int _cpu_down(unsigned int cpu, int tasks_frozen)
124 {
125         int err, nr_calls = 0;
126         struct task_struct *p;
127         cpumask_t old_allowed, tmp;
128         void *hcpu = (void *)(long)cpu;
129         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
130
131         if (num_online_cpus() == 1)
132                 return -EBUSY;
133
134         if (!cpu_online(cpu))
135                 return -EINVAL;
136
137         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
138         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
139                                         hcpu, -1, &nr_calls);
140         if (err == NOTIFY_BAD) {
141                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
142                                           hcpu, nr_calls, NULL);
143                 printk("%s: attempt to take down CPU %u failed\n",
144                                 __FUNCTION__, cpu);
145                 err = -EINVAL;
146                 goto out_release;
147         }
148
149         /* Ensure that we are not runnable on dying cpu */
150         old_allowed = current->cpus_allowed;
151         tmp = CPU_MASK_ALL;
152         cpu_clear(cpu, tmp);
153         set_cpus_allowed(current, tmp);
154
155         mutex_lock(&cpu_bitmask_lock);
156         p = __stop_machine_run(take_cpu_down, NULL, cpu);
157         mutex_unlock(&cpu_bitmask_lock);
158
159         if (IS_ERR(p) || cpu_online(cpu)) {
160                 /* CPU didn't die: tell everyone.  Can't complain. */
161                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
162                                             hcpu) == NOTIFY_BAD)
163                         BUG();
164
165                 if (IS_ERR(p)) {
166                         err = PTR_ERR(p);
167                         goto out_allowed;
168                 }
169                 goto out_thread;
170         }
171
172         /* Wait for it to sleep (leaving idle task). */
173         while (!idle_cpu(cpu))
174                 yield();
175
176         /* This actually kills the CPU. */
177         __cpu_die(cpu);
178
179         /* CPU is completely dead: tell everyone.  Too late to complain. */
180         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
181                                     hcpu) == NOTIFY_BAD)
182                 BUG();
183
184         check_for_tasks(cpu);
185
186 out_thread:
187         err = kthread_stop(p);
188 out_allowed:
189         set_cpus_allowed(current, old_allowed);
190 out_release:
191         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
192         return err;
193 }
194
195 int cpu_down(unsigned int cpu)
196 {
197         int err = 0;
198
199         mutex_lock(&cpu_add_remove_lock);
200         if (cpu_hotplug_disabled)
201                 err = -EBUSY;
202         else
203                 err = _cpu_down(cpu, 0);
204
205         mutex_unlock(&cpu_add_remove_lock);
206         return err;
207 }
208 #endif /*CONFIG_HOTPLUG_CPU*/
209
210 /* Requires cpu_add_remove_lock to be held */
211 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
212 {
213         int ret, nr_calls = 0;
214         void *hcpu = (void *)(long)cpu;
215         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
216
217         if (cpu_online(cpu) || !cpu_present(cpu))
218                 return -EINVAL;
219
220         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_ACQUIRE, hcpu);
221         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
222                                                         -1, &nr_calls);
223         if (ret == NOTIFY_BAD) {
224                 printk("%s: attempt to bring up CPU %u failed\n",
225                                 __FUNCTION__, cpu);
226                 ret = -EINVAL;
227                 goto out_notify;
228         }
229
230         /* Arch-specific enabling code. */
231         mutex_lock(&cpu_bitmask_lock);
232         ret = __cpu_up(cpu);
233         mutex_unlock(&cpu_bitmask_lock);
234         if (ret != 0)
235                 goto out_notify;
236         BUG_ON(!cpu_online(cpu));
237
238         /* Now call notifier in preparation. */
239         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
240
241 out_notify:
242         if (ret != 0)
243                 __raw_notifier_call_chain(&cpu_chain,
244                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
245         raw_notifier_call_chain(&cpu_chain, CPU_LOCK_RELEASE, hcpu);
246
247         return ret;
248 }
249
250 int __cpuinit cpu_up(unsigned int cpu)
251 {
252         int err = 0;
253
254         mutex_lock(&cpu_add_remove_lock);
255         if (cpu_hotplug_disabled)
256                 err = -EBUSY;
257         else
258                 err = _cpu_up(cpu, 0);
259
260         mutex_unlock(&cpu_add_remove_lock);
261         return err;
262 }
263
264 #ifdef CONFIG_SUSPEND_SMP
265 /* Needed to prevent the microcode driver from requesting firmware in its CPU
266  * hotplug notifier during the suspend/resume.
267  */
268 int suspend_cpu_hotplug;
269 EXPORT_SYMBOL(suspend_cpu_hotplug);
270
271 static cpumask_t frozen_cpus;
272
273 int disable_nonboot_cpus(void)
274 {
275         int cpu, first_cpu, error = 0;
276
277         mutex_lock(&cpu_add_remove_lock);
278         suspend_cpu_hotplug = 1;
279         first_cpu = first_cpu(cpu_online_map);
280         /* We take down all of the non-boot CPUs in one shot to avoid races
281          * with the userspace trying to use the CPU hotplug at the same time
282          */
283         cpus_clear(frozen_cpus);
284         printk("Disabling non-boot CPUs ...\n");
285         for_each_online_cpu(cpu) {
286                 if (cpu == first_cpu)
287                         continue;
288                 error = _cpu_down(cpu, 1);
289                 if (!error) {
290                         cpu_set(cpu, frozen_cpus);
291                         printk("CPU%d is down\n", cpu);
292                 } else {
293                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
294                                 cpu, error);
295                         break;
296                 }
297         }
298         if (!error) {
299                 BUG_ON(num_online_cpus() > 1);
300                 /* Make sure the CPUs won't be enabled by someone else */
301                 cpu_hotplug_disabled = 1;
302         } else {
303                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
304         }
305         suspend_cpu_hotplug = 0;
306         mutex_unlock(&cpu_add_remove_lock);
307         return error;
308 }
309
310 void enable_nonboot_cpus(void)
311 {
312         int cpu, error;
313
314         /* Allow everyone to use the CPU hotplug again */
315         mutex_lock(&cpu_add_remove_lock);
316         cpu_hotplug_disabled = 0;
317         if (cpus_empty(frozen_cpus))
318                 goto out;
319
320         suspend_cpu_hotplug = 1;
321         printk("Enabling non-boot CPUs ...\n");
322         for_each_cpu_mask(cpu, frozen_cpus) {
323                 error = _cpu_up(cpu, 1);
324                 if (!error) {
325                         printk("CPU%d is up\n", cpu);
326                         continue;
327                 }
328                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
329         }
330         cpus_clear(frozen_cpus);
331         suspend_cpu_hotplug = 0;
332 out:
333         mutex_unlock(&cpu_add_remove_lock);
334 }
335 #endif