cpu hotplug, sched: Introduce cpu_active_map and redo sched domain managment (take 2)
[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 /*
19  * Represents all cpu's present in the system
20  * In systems capable of hotplug, this map could dynamically grow
21  * as new cpu's are detected in the system via any platform specific
22  * method, such as ACPI for e.g.
23  */
24 cpumask_t cpu_present_map __read_mostly;
25 EXPORT_SYMBOL(cpu_present_map);
26
27 #ifndef CONFIG_SMP
28
29 /*
30  * Represents all cpu's that are currently online.
31  */
32 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
33 EXPORT_SYMBOL(cpu_online_map);
34
35 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
36 EXPORT_SYMBOL(cpu_possible_map);
37
38 #else /* CONFIG_SMP */
39
40 /* Serializes the updates to cpu_online_map, cpu_present_map */
41 static DEFINE_MUTEX(cpu_add_remove_lock);
42
43 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
44
45 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
46  * Should always be manipulated under cpu_add_remove_lock
47  */
48 static int cpu_hotplug_disabled;
49
50 static struct {
51         struct task_struct *active_writer;
52         struct mutex lock; /* Synchronizes accesses to refcount, */
53         /*
54          * Also blocks the new readers during
55          * an ongoing cpu hotplug operation.
56          */
57         int refcount;
58 } cpu_hotplug;
59
60 void __init cpu_hotplug_init(void)
61 {
62         cpu_hotplug.active_writer = NULL;
63         mutex_init(&cpu_hotplug.lock);
64         cpu_hotplug.refcount = 0;
65 }
66
67 cpumask_t cpu_active_map;
68
69 #ifdef CONFIG_HOTPLUG_CPU
70
71 void get_online_cpus(void)
72 {
73         might_sleep();
74         if (cpu_hotplug.active_writer == current)
75                 return;
76         mutex_lock(&cpu_hotplug.lock);
77         cpu_hotplug.refcount++;
78         mutex_unlock(&cpu_hotplug.lock);
79
80 }
81 EXPORT_SYMBOL_GPL(get_online_cpus);
82
83 void put_online_cpus(void)
84 {
85         if (cpu_hotplug.active_writer == current)
86                 return;
87         mutex_lock(&cpu_hotplug.lock);
88         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
89                 wake_up_process(cpu_hotplug.active_writer);
90         mutex_unlock(&cpu_hotplug.lock);
91
92 }
93 EXPORT_SYMBOL_GPL(put_online_cpus);
94
95 #endif  /* CONFIG_HOTPLUG_CPU */
96
97 /*
98  * The following two API's must be used when attempting
99  * to serialize the updates to cpu_online_map, cpu_present_map.
100  */
101 void cpu_maps_update_begin(void)
102 {
103         mutex_lock(&cpu_add_remove_lock);
104 }
105
106 void cpu_maps_update_done(void)
107 {
108         mutex_unlock(&cpu_add_remove_lock);
109 }
110
111 /*
112  * This ensures that the hotplug operation can begin only when the
113  * refcount goes to zero.
114  *
115  * Note that during a cpu-hotplug operation, the new readers, if any,
116  * will be blocked by the cpu_hotplug.lock
117  *
118  * Since cpu_hotplug_begin() is always called after invoking
119  * cpu_maps_update_begin(), we can be sure that only one writer is active.
120  *
121  * Note that theoretically, there is a possibility of a livelock:
122  * - Refcount goes to zero, last reader wakes up the sleeping
123  *   writer.
124  * - Last reader unlocks the cpu_hotplug.lock.
125  * - A new reader arrives at this moment, bumps up the refcount.
126  * - The writer acquires the cpu_hotplug.lock finds the refcount
127  *   non zero and goes to sleep again.
128  *
129  * However, this is very difficult to achieve in practice since
130  * get_online_cpus() not an api which is called all that often.
131  *
132  */
133 static void cpu_hotplug_begin(void)
134 {
135         cpu_hotplug.active_writer = current;
136
137         for (;;) {
138                 mutex_lock(&cpu_hotplug.lock);
139                 if (likely(!cpu_hotplug.refcount))
140                         break;
141                 __set_current_state(TASK_UNINTERRUPTIBLE);
142                 mutex_unlock(&cpu_hotplug.lock);
143                 schedule();
144         }
145 }
146
147 static void cpu_hotplug_done(void)
148 {
149         cpu_hotplug.active_writer = NULL;
150         mutex_unlock(&cpu_hotplug.lock);
151 }
152 /* Need to know about CPUs going up/down? */
153 int __ref register_cpu_notifier(struct notifier_block *nb)
154 {
155         int ret;
156         cpu_maps_update_begin();
157         ret = raw_notifier_chain_register(&cpu_chain, nb);
158         cpu_maps_update_done();
159         return ret;
160 }
161
162 #ifdef CONFIG_HOTPLUG_CPU
163
164 EXPORT_SYMBOL(register_cpu_notifier);
165
166 void __ref unregister_cpu_notifier(struct notifier_block *nb)
167 {
168         cpu_maps_update_begin();
169         raw_notifier_chain_unregister(&cpu_chain, nb);
170         cpu_maps_update_done();
171 }
172 EXPORT_SYMBOL(unregister_cpu_notifier);
173
174 static inline void check_for_tasks(int cpu)
175 {
176         struct task_struct *p;
177
178         write_lock_irq(&tasklist_lock);
179         for_each_process(p) {
180                 if (task_cpu(p) == cpu &&
181                     (!cputime_eq(p->utime, cputime_zero) ||
182                      !cputime_eq(p->stime, cputime_zero)))
183                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
184                                 (state = %ld, flags = %x) \n",
185                                  p->comm, task_pid_nr(p), cpu,
186                                  p->state, p->flags);
187         }
188         write_unlock_irq(&tasklist_lock);
189 }
190
191 struct take_cpu_down_param {
192         unsigned long mod;
193         void *hcpu;
194 };
195
196 /* Take this CPU down. */
197 static int __ref take_cpu_down(void *_param)
198 {
199         struct take_cpu_down_param *param = _param;
200         int err;
201
202         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
203                                 param->hcpu);
204         /* Ensure this CPU doesn't handle any more interrupts. */
205         err = __cpu_disable();
206         if (err < 0)
207                 return err;
208
209         /* Force idle task to run as soon as we yield: it should
210            immediately notice cpu is offline and die quickly. */
211         sched_idle_next();
212         return 0;
213 }
214
215 /* Requires cpu_add_remove_lock to be held */
216 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
217 {
218         int err, nr_calls = 0;
219         struct task_struct *p;
220         cpumask_t old_allowed, tmp;
221         void *hcpu = (void *)(long)cpu;
222         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
223         struct take_cpu_down_param tcd_param = {
224                 .mod = mod,
225                 .hcpu = hcpu,
226         };
227
228         if (num_online_cpus() == 1)
229                 return -EBUSY;
230
231         if (!cpu_online(cpu))
232                 return -EINVAL;
233
234         cpu_hotplug_begin();
235         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
236                                         hcpu, -1, &nr_calls);
237         if (err == NOTIFY_BAD) {
238                 nr_calls--;
239                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
240                                           hcpu, nr_calls, NULL);
241                 printk("%s: attempt to take down CPU %u failed\n",
242                                 __func__, cpu);
243                 err = -EINVAL;
244                 goto out_release;
245         }
246
247         /* Ensure that we are not runnable on dying cpu */
248         old_allowed = current->cpus_allowed;
249         cpus_setall(tmp);
250         cpu_clear(cpu, tmp);
251         set_cpus_allowed_ptr(current, &tmp);
252
253         p = __stop_machine_run(take_cpu_down, &tcd_param, cpu);
254
255         if (IS_ERR(p) || cpu_online(cpu)) {
256                 /* CPU didn't die: tell everyone.  Can't complain. */
257                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
258                                             hcpu) == NOTIFY_BAD)
259                         BUG();
260
261                 if (IS_ERR(p)) {
262                         err = PTR_ERR(p);
263                         goto out_allowed;
264                 }
265                 goto out_thread;
266         }
267
268         /* Wait for it to sleep (leaving idle task). */
269         while (!idle_cpu(cpu))
270                 yield();
271
272         /* This actually kills the CPU. */
273         __cpu_die(cpu);
274
275         /* CPU is completely dead: tell everyone.  Too late to complain. */
276         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
277                                     hcpu) == NOTIFY_BAD)
278                 BUG();
279
280         check_for_tasks(cpu);
281
282 out_thread:
283         err = kthread_stop(p);
284 out_allowed:
285         set_cpus_allowed_ptr(current, &old_allowed);
286 out_release:
287         cpu_hotplug_done();
288         return err;
289 }
290
291 int __ref cpu_down(unsigned int cpu)
292 {
293         int err = 0;
294
295         cpu_maps_update_begin();
296
297         if (cpu_hotplug_disabled) {
298                 err = -EBUSY;
299                 goto out;
300         }
301
302         cpu_clear(cpu, cpu_active_map);
303
304         err = _cpu_down(cpu, 0);
305
306         if (cpu_online(cpu))
307                 cpu_set(cpu, cpu_active_map);
308
309 out:
310         cpu_maps_update_done();
311         return err;
312 }
313 EXPORT_SYMBOL(cpu_down);
314 #endif /*CONFIG_HOTPLUG_CPU*/
315
316 /* Requires cpu_add_remove_lock to be held */
317 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
318 {
319         int ret, nr_calls = 0;
320         void *hcpu = (void *)(long)cpu;
321         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
322
323         if (cpu_online(cpu) || !cpu_present(cpu))
324                 return -EINVAL;
325
326         cpu_hotplug_begin();
327         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
328                                                         -1, &nr_calls);
329         if (ret == NOTIFY_BAD) {
330                 nr_calls--;
331                 printk("%s: attempt to bring up CPU %u failed\n",
332                                 __func__, cpu);
333                 ret = -EINVAL;
334                 goto out_notify;
335         }
336
337         /* Arch-specific enabling code. */
338         ret = __cpu_up(cpu);
339         if (ret != 0)
340                 goto out_notify;
341         BUG_ON(!cpu_online(cpu));
342
343         /* Now call notifier in preparation. */
344         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
345
346 out_notify:
347         if (ret != 0)
348                 __raw_notifier_call_chain(&cpu_chain,
349                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
350         cpu_hotplug_done();
351
352         return ret;
353 }
354
355 int __cpuinit cpu_up(unsigned int cpu)
356 {
357         int err = 0;
358         if (!cpu_isset(cpu, cpu_possible_map)) {
359                 printk(KERN_ERR "can't online cpu %d because it is not "
360                         "configured as may-hotadd at boot time\n", cpu);
361 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
362                 printk(KERN_ERR "please check additional_cpus= boot "
363                                 "parameter\n");
364 #endif
365                 return -EINVAL;
366         }
367
368         cpu_maps_update_begin();
369
370         if (cpu_hotplug_disabled) {
371                 err = -EBUSY;
372                 goto out;
373         }
374
375         err = _cpu_up(cpu, 0);
376
377         if (cpu_online(cpu))
378                 cpu_set(cpu, cpu_active_map);
379
380 out:
381         cpu_maps_update_done();
382         return err;
383 }
384
385 #ifdef CONFIG_PM_SLEEP_SMP
386 static cpumask_t frozen_cpus;
387
388 int disable_nonboot_cpus(void)
389 {
390         int cpu, first_cpu, error = 0;
391
392         cpu_maps_update_begin();
393         first_cpu = first_cpu(cpu_online_map);
394         /* We take down all of the non-boot CPUs in one shot to avoid races
395          * with the userspace trying to use the CPU hotplug at the same time
396          */
397         cpus_clear(frozen_cpus);
398         printk("Disabling non-boot CPUs ...\n");
399         for_each_online_cpu(cpu) {
400                 if (cpu == first_cpu)
401                         continue;
402                 error = _cpu_down(cpu, 1);
403                 if (!error) {
404                         cpu_set(cpu, frozen_cpus);
405                         printk("CPU%d is down\n", cpu);
406                 } else {
407                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
408                                 cpu, error);
409                         break;
410                 }
411         }
412         if (!error) {
413                 BUG_ON(num_online_cpus() > 1);
414                 /* Make sure the CPUs won't be enabled by someone else */
415                 cpu_hotplug_disabled = 1;
416         } else {
417                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
418         }
419         cpu_maps_update_done();
420         return error;
421 }
422
423 void __ref enable_nonboot_cpus(void)
424 {
425         int cpu, error;
426
427         /* Allow everyone to use the CPU hotplug again */
428         cpu_maps_update_begin();
429         cpu_hotplug_disabled = 0;
430         if (cpus_empty(frozen_cpus))
431                 goto out;
432
433         printk("Enabling non-boot CPUs ...\n");
434         for_each_cpu_mask(cpu, frozen_cpus) {
435                 error = _cpu_up(cpu, 1);
436                 if (!error) {
437                         printk("CPU%d is up\n", cpu);
438                         continue;
439                 }
440                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
441         }
442         cpus_clear(frozen_cpus);
443 out:
444         cpu_maps_update_done();
445 }
446 #endif /* CONFIG_PM_SLEEP_SMP */
447
448 #endif /* CONFIG_SMP */