cpumask: switch over to cpu_online/possible/active/present_mask: core
[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 #ifdef CONFIG_SMP
19 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
20 static DEFINE_MUTEX(cpu_add_remove_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 static struct {
30         struct task_struct *active_writer;
31         struct mutex lock; /* Synchronizes accesses to refcount, */
32         /*
33          * Also blocks the new readers during
34          * an ongoing cpu hotplug operation.
35          */
36         int refcount;
37 } cpu_hotplug;
38
39 void __init cpu_hotplug_init(void)
40 {
41         cpu_hotplug.active_writer = NULL;
42         mutex_init(&cpu_hotplug.lock);
43         cpu_hotplug.refcount = 0;
44 }
45
46 #ifdef CONFIG_HOTPLUG_CPU
47
48 void get_online_cpus(void)
49 {
50         might_sleep();
51         if (cpu_hotplug.active_writer == current)
52                 return;
53         mutex_lock(&cpu_hotplug.lock);
54         cpu_hotplug.refcount++;
55         mutex_unlock(&cpu_hotplug.lock);
56
57 }
58 EXPORT_SYMBOL_GPL(get_online_cpus);
59
60 void put_online_cpus(void)
61 {
62         if (cpu_hotplug.active_writer == current)
63                 return;
64         mutex_lock(&cpu_hotplug.lock);
65         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
66                 wake_up_process(cpu_hotplug.active_writer);
67         mutex_unlock(&cpu_hotplug.lock);
68
69 }
70 EXPORT_SYMBOL_GPL(put_online_cpus);
71
72 #endif  /* CONFIG_HOTPLUG_CPU */
73
74 /*
75  * The following two API's must be used when attempting
76  * to serialize the updates to cpu_online_mask, cpu_present_mask.
77  */
78 void cpu_maps_update_begin(void)
79 {
80         mutex_lock(&cpu_add_remove_lock);
81 }
82
83 void cpu_maps_update_done(void)
84 {
85         mutex_unlock(&cpu_add_remove_lock);
86 }
87
88 /*
89  * This ensures that the hotplug operation can begin only when the
90  * refcount goes to zero.
91  *
92  * Note that during a cpu-hotplug operation, the new readers, if any,
93  * will be blocked by the cpu_hotplug.lock
94  *
95  * Since cpu_hotplug_begin() is always called after invoking
96  * cpu_maps_update_begin(), we can be sure that only one writer is active.
97  *
98  * Note that theoretically, there is a possibility of a livelock:
99  * - Refcount goes to zero, last reader wakes up the sleeping
100  *   writer.
101  * - Last reader unlocks the cpu_hotplug.lock.
102  * - A new reader arrives at this moment, bumps up the refcount.
103  * - The writer acquires the cpu_hotplug.lock finds the refcount
104  *   non zero and goes to sleep again.
105  *
106  * However, this is very difficult to achieve in practice since
107  * get_online_cpus() not an api which is called all that often.
108  *
109  */
110 static void cpu_hotplug_begin(void)
111 {
112         cpu_hotplug.active_writer = current;
113
114         for (;;) {
115                 mutex_lock(&cpu_hotplug.lock);
116                 if (likely(!cpu_hotplug.refcount))
117                         break;
118                 __set_current_state(TASK_UNINTERRUPTIBLE);
119                 mutex_unlock(&cpu_hotplug.lock);
120                 schedule();
121         }
122 }
123
124 static void cpu_hotplug_done(void)
125 {
126         cpu_hotplug.active_writer = NULL;
127         mutex_unlock(&cpu_hotplug.lock);
128 }
129 /* Need to know about CPUs going up/down? */
130 int __ref register_cpu_notifier(struct notifier_block *nb)
131 {
132         int ret;
133         cpu_maps_update_begin();
134         ret = raw_notifier_chain_register(&cpu_chain, nb);
135         cpu_maps_update_done();
136         return ret;
137 }
138
139 #ifdef CONFIG_HOTPLUG_CPU
140
141 EXPORT_SYMBOL(register_cpu_notifier);
142
143 void __ref unregister_cpu_notifier(struct notifier_block *nb)
144 {
145         cpu_maps_update_begin();
146         raw_notifier_chain_unregister(&cpu_chain, nb);
147         cpu_maps_update_done();
148 }
149 EXPORT_SYMBOL(unregister_cpu_notifier);
150
151 static inline void check_for_tasks(int cpu)
152 {
153         struct task_struct *p;
154
155         write_lock_irq(&tasklist_lock);
156         for_each_process(p) {
157                 if (task_cpu(p) == cpu &&
158                     (!cputime_eq(p->utime, cputime_zero) ||
159                      !cputime_eq(p->stime, cputime_zero)))
160                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
161                                 (state = %ld, flags = %x) \n",
162                                  p->comm, task_pid_nr(p), cpu,
163                                  p->state, p->flags);
164         }
165         write_unlock_irq(&tasklist_lock);
166 }
167
168 struct take_cpu_down_param {
169         unsigned long mod;
170         void *hcpu;
171 };
172
173 /* Take this CPU down. */
174 static int __ref take_cpu_down(void *_param)
175 {
176         struct take_cpu_down_param *param = _param;
177         int err;
178
179         /* Ensure this CPU doesn't handle any more interrupts. */
180         err = __cpu_disable();
181         if (err < 0)
182                 return err;
183
184         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
185                                 param->hcpu);
186
187         /* Force idle task to run as soon as we yield: it should
188            immediately notice cpu is offline and die quickly. */
189         sched_idle_next();
190         return 0;
191 }
192
193 /* Requires cpu_add_remove_lock to be held */
194 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
195 {
196         int err, nr_calls = 0;
197         cpumask_t old_allowed, tmp;
198         void *hcpu = (void *)(long)cpu;
199         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
200         struct take_cpu_down_param tcd_param = {
201                 .mod = mod,
202                 .hcpu = hcpu,
203         };
204
205         if (num_online_cpus() == 1)
206                 return -EBUSY;
207
208         if (!cpu_online(cpu))
209                 return -EINVAL;
210
211         cpu_hotplug_begin();
212         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
213                                         hcpu, -1, &nr_calls);
214         if (err == NOTIFY_BAD) {
215                 nr_calls--;
216                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
217                                           hcpu, nr_calls, NULL);
218                 printk("%s: attempt to take down CPU %u failed\n",
219                                 __func__, cpu);
220                 err = -EINVAL;
221                 goto out_release;
222         }
223
224         /* Ensure that we are not runnable on dying cpu */
225         old_allowed = current->cpus_allowed;
226         cpus_setall(tmp);
227         cpu_clear(cpu, tmp);
228         set_cpus_allowed_ptr(current, &tmp);
229         tmp = cpumask_of_cpu(cpu);
230
231         err = __stop_machine(take_cpu_down, &tcd_param, &tmp);
232         if (err) {
233                 /* CPU didn't die: tell everyone.  Can't complain. */
234                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
235                                             hcpu) == NOTIFY_BAD)
236                         BUG();
237
238                 goto out_allowed;
239         }
240         BUG_ON(cpu_online(cpu));
241
242         /* Wait for it to sleep (leaving idle task). */
243         while (!idle_cpu(cpu))
244                 yield();
245
246         /* This actually kills the CPU. */
247         __cpu_die(cpu);
248
249         /* CPU is completely dead: tell everyone.  Too late to complain. */
250         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
251                                     hcpu) == NOTIFY_BAD)
252                 BUG();
253
254         check_for_tasks(cpu);
255
256 out_allowed:
257         set_cpus_allowed_ptr(current, &old_allowed);
258 out_release:
259         cpu_hotplug_done();
260         if (!err) {
261                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
262                                             hcpu) == NOTIFY_BAD)
263                         BUG();
264         }
265         return err;
266 }
267
268 int __ref cpu_down(unsigned int cpu)
269 {
270         int err = 0;
271
272         cpu_maps_update_begin();
273
274         if (cpu_hotplug_disabled) {
275                 err = -EBUSY;
276                 goto out;
277         }
278
279         cpu_clear(cpu, cpu_active_map);
280
281         /*
282          * Make sure the all cpus did the reschedule and are not
283          * using stale version of the cpu_active_map.
284          * This is not strictly necessary becuase stop_machine()
285          * that we run down the line already provides the required
286          * synchronization. But it's really a side effect and we do not
287          * want to depend on the innards of the stop_machine here.
288          */
289         synchronize_sched();
290
291         err = _cpu_down(cpu, 0);
292
293         if (cpu_online(cpu))
294                 cpu_set(cpu, cpu_active_map);
295
296 out:
297         cpu_maps_update_done();
298         return err;
299 }
300 EXPORT_SYMBOL(cpu_down);
301 #endif /*CONFIG_HOTPLUG_CPU*/
302
303 /* Requires cpu_add_remove_lock to be held */
304 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
305 {
306         int ret, nr_calls = 0;
307         void *hcpu = (void *)(long)cpu;
308         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
309
310         if (cpu_online(cpu) || !cpu_present(cpu))
311                 return -EINVAL;
312
313         cpu_hotplug_begin();
314         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
315                                                         -1, &nr_calls);
316         if (ret == NOTIFY_BAD) {
317                 nr_calls--;
318                 printk("%s: attempt to bring up CPU %u failed\n",
319                                 __func__, cpu);
320                 ret = -EINVAL;
321                 goto out_notify;
322         }
323
324         /* Arch-specific enabling code. */
325         ret = __cpu_up(cpu);
326         if (ret != 0)
327                 goto out_notify;
328         BUG_ON(!cpu_online(cpu));
329
330         cpu_set(cpu, cpu_active_map);
331
332         /* Now call notifier in preparation. */
333         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
334
335 out_notify:
336         if (ret != 0)
337                 __raw_notifier_call_chain(&cpu_chain,
338                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
339         cpu_hotplug_done();
340
341         return ret;
342 }
343
344 int __cpuinit cpu_up(unsigned int cpu)
345 {
346         int err = 0;
347         if (!cpu_isset(cpu, cpu_possible_map)) {
348                 printk(KERN_ERR "can't online cpu %d because it is not "
349                         "configured as may-hotadd at boot time\n", cpu);
350 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
351                 printk(KERN_ERR "please check additional_cpus= boot "
352                                 "parameter\n");
353 #endif
354                 return -EINVAL;
355         }
356
357         cpu_maps_update_begin();
358
359         if (cpu_hotplug_disabled) {
360                 err = -EBUSY;
361                 goto out;
362         }
363
364         err = _cpu_up(cpu, 0);
365
366 out:
367         cpu_maps_update_done();
368         return err;
369 }
370
371 #ifdef CONFIG_PM_SLEEP_SMP
372 static cpumask_t frozen_cpus;
373
374 int disable_nonboot_cpus(void)
375 {
376         int cpu, first_cpu, error = 0;
377
378         cpu_maps_update_begin();
379         first_cpu = first_cpu(cpu_online_map);
380         /* We take down all of the non-boot CPUs in one shot to avoid races
381          * with the userspace trying to use the CPU hotplug at the same time
382          */
383         cpus_clear(frozen_cpus);
384         printk("Disabling non-boot CPUs ...\n");
385         for_each_online_cpu(cpu) {
386                 if (cpu == first_cpu)
387                         continue;
388                 error = _cpu_down(cpu, 1);
389                 if (!error) {
390                         cpu_set(cpu, frozen_cpus);
391                         printk("CPU%d is down\n", cpu);
392                 } else {
393                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
394                                 cpu, error);
395                         break;
396                 }
397         }
398         if (!error) {
399                 BUG_ON(num_online_cpus() > 1);
400                 /* Make sure the CPUs won't be enabled by someone else */
401                 cpu_hotplug_disabled = 1;
402         } else {
403                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
404         }
405         cpu_maps_update_done();
406         return error;
407 }
408
409 void __ref enable_nonboot_cpus(void)
410 {
411         int cpu, error;
412
413         /* Allow everyone to use the CPU hotplug again */
414         cpu_maps_update_begin();
415         cpu_hotplug_disabled = 0;
416         if (cpus_empty(frozen_cpus))
417                 goto out;
418
419         printk("Enabling non-boot CPUs ...\n");
420         for_each_cpu_mask_nr(cpu, frozen_cpus) {
421                 error = _cpu_up(cpu, 1);
422                 if (!error) {
423                         printk("CPU%d is up\n", cpu);
424                         continue;
425                 }
426                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
427         }
428         cpus_clear(frozen_cpus);
429 out:
430         cpu_maps_update_done();
431 }
432 #endif /* CONFIG_PM_SLEEP_SMP */
433
434 /**
435  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
436  * @cpu: cpu that just started
437  *
438  * This function calls the cpu_chain notifiers with CPU_STARTING.
439  * It must be called by the arch code on the new cpu, before the new cpu
440  * enables interrupts and before the "boot" cpu returns from __cpu_up().
441  */
442 void __cpuinit notify_cpu_starting(unsigned int cpu)
443 {
444         unsigned long val = CPU_STARTING;
445
446 #ifdef CONFIG_PM_SLEEP_SMP
447         if (cpu_isset(cpu, frozen_cpus))
448                 val = CPU_STARTING_FROZEN;
449 #endif /* CONFIG_PM_SLEEP_SMP */
450         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
451 }
452
453 #endif /* CONFIG_SMP */
454
455 /*
456  * cpu_bit_bitmap[] is a special, "compressed" data structure that
457  * represents all NR_CPUS bits binary values of 1<<nr.
458  *
459  * It is used by cpumask_of_cpu() to get a constant address to a CPU
460  * mask value that has a single bit set only.
461  */
462
463 /* cpu_bit_bitmap[0] is empty - so we can back into it */
464 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
465 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
466 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
467 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
468
469 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
470
471         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
472         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
473 #if BITS_PER_LONG > 32
474         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
475         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
476 #endif
477 };
478 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
479
480 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
481 EXPORT_SYMBOL(cpu_all_bits);
482
483 #ifdef CONFIG_INIT_ALL_POSSIBLE
484 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
485         = CPU_BITS_ALL;
486 #else
487 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
488 #endif
489 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
490 EXPORT_SYMBOL(cpu_possible_mask);
491
492 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
493 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
494 EXPORT_SYMBOL(cpu_online_mask);
495
496 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
497 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
498 EXPORT_SYMBOL(cpu_present_mask);
499
500 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
501 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
502 EXPORT_SYMBOL(cpu_active_mask);