sched: adjust when cpu_active and cpuset configurations are updated during cpu on...
[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 #include <linux/gfp.h>
18
19 #ifdef CONFIG_SMP
20 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
21 static DEFINE_MUTEX(cpu_add_remove_lock);
22
23 /*
24  * The following two API's must be used when attempting
25  * to serialize the updates to cpu_online_mask, cpu_present_mask.
26  */
27 void cpu_maps_update_begin(void)
28 {
29         mutex_lock(&cpu_add_remove_lock);
30 }
31
32 void cpu_maps_update_done(void)
33 {
34         mutex_unlock(&cpu_add_remove_lock);
35 }
36
37 static RAW_NOTIFIER_HEAD(cpu_chain);
38
39 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
40  * Should always be manipulated under cpu_add_remove_lock
41  */
42 static int cpu_hotplug_disabled;
43
44 #ifdef CONFIG_HOTPLUG_CPU
45
46 static struct {
47         struct task_struct *active_writer;
48         struct mutex lock; /* Synchronizes accesses to refcount, */
49         /*
50          * Also blocks the new readers during
51          * an ongoing cpu hotplug operation.
52          */
53         int refcount;
54 } cpu_hotplug = {
55         .active_writer = NULL,
56         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
57         .refcount = 0,
58 };
59
60 void get_online_cpus(void)
61 {
62         might_sleep();
63         if (cpu_hotplug.active_writer == current)
64                 return;
65         mutex_lock(&cpu_hotplug.lock);
66         cpu_hotplug.refcount++;
67         mutex_unlock(&cpu_hotplug.lock);
68
69 }
70 EXPORT_SYMBOL_GPL(get_online_cpus);
71
72 void put_online_cpus(void)
73 {
74         if (cpu_hotplug.active_writer == current)
75                 return;
76         mutex_lock(&cpu_hotplug.lock);
77         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
78                 wake_up_process(cpu_hotplug.active_writer);
79         mutex_unlock(&cpu_hotplug.lock);
80
81 }
82 EXPORT_SYMBOL_GPL(put_online_cpus);
83
84 /*
85  * This ensures that the hotplug operation can begin only when the
86  * refcount goes to zero.
87  *
88  * Note that during a cpu-hotplug operation, the new readers, if any,
89  * will be blocked by the cpu_hotplug.lock
90  *
91  * Since cpu_hotplug_begin() is always called after invoking
92  * cpu_maps_update_begin(), we can be sure that only one writer is active.
93  *
94  * Note that theoretically, there is a possibility of a livelock:
95  * - Refcount goes to zero, last reader wakes up the sleeping
96  *   writer.
97  * - Last reader unlocks the cpu_hotplug.lock.
98  * - A new reader arrives at this moment, bumps up the refcount.
99  * - The writer acquires the cpu_hotplug.lock finds the refcount
100  *   non zero and goes to sleep again.
101  *
102  * However, this is very difficult to achieve in practice since
103  * get_online_cpus() not an api which is called all that often.
104  *
105  */
106 static void cpu_hotplug_begin(void)
107 {
108         cpu_hotplug.active_writer = current;
109
110         for (;;) {
111                 mutex_lock(&cpu_hotplug.lock);
112                 if (likely(!cpu_hotplug.refcount))
113                         break;
114                 __set_current_state(TASK_UNINTERRUPTIBLE);
115                 mutex_unlock(&cpu_hotplug.lock);
116                 schedule();
117         }
118 }
119
120 static void cpu_hotplug_done(void)
121 {
122         cpu_hotplug.active_writer = NULL;
123         mutex_unlock(&cpu_hotplug.lock);
124 }
125
126 #else /* #if CONFIG_HOTPLUG_CPU */
127 static void cpu_hotplug_begin(void) {}
128 static void cpu_hotplug_done(void) {}
129 #endif  /* #esle #if CONFIG_HOTPLUG_CPU */
130
131 /* Need to know about CPUs going up/down? */
132 int __ref register_cpu_notifier(struct notifier_block *nb)
133 {
134         int ret;
135         cpu_maps_update_begin();
136         ret = raw_notifier_chain_register(&cpu_chain, nb);
137         cpu_maps_update_done();
138         return ret;
139 }
140
141 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
142                         int *nr_calls)
143 {
144         int ret;
145
146         ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
147                                         nr_calls);
148
149         return notifier_to_errno(ret);
150 }
151
152 static int cpu_notify(unsigned long val, void *v)
153 {
154         return __cpu_notify(val, v, -1, NULL);
155 }
156
157 #ifdef CONFIG_HOTPLUG_CPU
158
159 static void cpu_notify_nofail(unsigned long val, void *v)
160 {
161         BUG_ON(cpu_notify(val, v));
162 }
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 && p->state == TASK_RUNNING &&
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         struct task_struct *caller;
193         unsigned long mod;
194         void *hcpu;
195 };
196
197 /* Take this CPU down. */
198 static int __ref take_cpu_down(void *_param)
199 {
200         struct take_cpu_down_param *param = _param;
201         unsigned int cpu = (unsigned long)param->hcpu;
202         int err;
203
204         /* Ensure this CPU doesn't handle any more interrupts. */
205         err = __cpu_disable();
206         if (err < 0)
207                 return err;
208
209         cpu_notify(CPU_DYING | param->mod, param->hcpu);
210
211         if (task_cpu(param->caller) == cpu)
212                 move_task_off_dead_cpu(cpu, param->caller);
213         /* Force idle task to run as soon as we yield: it should
214            immediately notice cpu is offline and die quickly. */
215         sched_idle_next();
216         return 0;
217 }
218
219 /* Requires cpu_add_remove_lock to be held */
220 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
221 {
222         int err, nr_calls = 0;
223         void *hcpu = (void *)(long)cpu;
224         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
225         struct take_cpu_down_param tcd_param = {
226                 .caller = current,
227                 .mod = mod,
228                 .hcpu = hcpu,
229         };
230
231         if (num_online_cpus() == 1)
232                 return -EBUSY;
233
234         if (!cpu_online(cpu))
235                 return -EINVAL;
236
237         cpu_hotplug_begin();
238         err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
239         if (err) {
240                 nr_calls--;
241                 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
242                 printk("%s: attempt to take down CPU %u failed\n",
243                                 __func__, cpu);
244                 goto out_release;
245         }
246
247         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
248         if (err) {
249                 /* CPU didn't die: tell everyone.  Can't complain. */
250                 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
251
252                 goto out_release;
253         }
254         BUG_ON(cpu_online(cpu));
255
256         /* Wait for it to sleep (leaving idle task). */
257         while (!idle_cpu(cpu))
258                 yield();
259
260         /* This actually kills the CPU. */
261         __cpu_die(cpu);
262
263         /* CPU is completely dead: tell everyone.  Too late to complain. */
264         cpu_notify_nofail(CPU_DEAD | mod, hcpu);
265
266         check_for_tasks(cpu);
267
268 out_release:
269         cpu_hotplug_done();
270         if (!err)
271                 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
272         return err;
273 }
274
275 int __ref cpu_down(unsigned int cpu)
276 {
277         int err;
278
279         cpu_maps_update_begin();
280
281         if (cpu_hotplug_disabled) {
282                 err = -EBUSY;
283                 goto out;
284         }
285
286         err = _cpu_down(cpu, 0);
287
288 out:
289         cpu_maps_update_done();
290         return err;
291 }
292 EXPORT_SYMBOL(cpu_down);
293 #endif /*CONFIG_HOTPLUG_CPU*/
294
295 /* Requires cpu_add_remove_lock to be held */
296 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
297 {
298         int ret, nr_calls = 0;
299         void *hcpu = (void *)(long)cpu;
300         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
301
302         if (cpu_online(cpu) || !cpu_present(cpu))
303                 return -EINVAL;
304
305         cpu_hotplug_begin();
306         ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
307         if (ret) {
308                 nr_calls--;
309                 printk("%s: attempt to bring up CPU %u failed\n",
310                                 __func__, cpu);
311                 goto out_notify;
312         }
313
314         /* Arch-specific enabling code. */
315         ret = __cpu_up(cpu);
316         if (ret != 0)
317                 goto out_notify;
318         BUG_ON(!cpu_online(cpu));
319
320         /* Now call notifier in preparation. */
321         cpu_notify(CPU_ONLINE | mod, hcpu);
322
323 out_notify:
324         if (ret != 0)
325                 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
326         cpu_hotplug_done();
327
328         return ret;
329 }
330
331 int __cpuinit cpu_up(unsigned int cpu)
332 {
333         int err = 0;
334
335 #ifdef  CONFIG_MEMORY_HOTPLUG
336         int nid;
337         pg_data_t       *pgdat;
338 #endif
339
340         if (!cpu_possible(cpu)) {
341                 printk(KERN_ERR "can't online cpu %d because it is not "
342                         "configured as may-hotadd at boot time\n", cpu);
343 #if defined(CONFIG_IA64)
344                 printk(KERN_ERR "please check additional_cpus= boot "
345                                 "parameter\n");
346 #endif
347                 return -EINVAL;
348         }
349
350 #ifdef  CONFIG_MEMORY_HOTPLUG
351         nid = cpu_to_node(cpu);
352         if (!node_online(nid)) {
353                 err = mem_online_node(nid);
354                 if (err)
355                         return err;
356         }
357
358         pgdat = NODE_DATA(nid);
359         if (!pgdat) {
360                 printk(KERN_ERR
361                         "Can't online cpu %d due to NULL pgdat\n", cpu);
362                 return -ENOMEM;
363         }
364
365         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
366                 mutex_lock(&zonelists_mutex);
367                 build_all_zonelists(NULL);
368                 mutex_unlock(&zonelists_mutex);
369         }
370 #endif
371
372         cpu_maps_update_begin();
373
374         if (cpu_hotplug_disabled) {
375                 err = -EBUSY;
376                 goto out;
377         }
378
379         err = _cpu_up(cpu, 0);
380
381 out:
382         cpu_maps_update_done();
383         return err;
384 }
385
386 #ifdef CONFIG_PM_SLEEP_SMP
387 static cpumask_var_t frozen_cpus;
388
389 int disable_nonboot_cpus(void)
390 {
391         int cpu, first_cpu, error = 0;
392
393         cpu_maps_update_begin();
394         first_cpu = cpumask_first(cpu_online_mask);
395         /*
396          * We take down all of the non-boot CPUs in one shot to avoid races
397          * with the userspace trying to use the CPU hotplug at the same time
398          */
399         cpumask_clear(frozen_cpus);
400
401         printk("Disabling non-boot CPUs ...\n");
402         for_each_online_cpu(cpu) {
403                 if (cpu == first_cpu)
404                         continue;
405                 error = _cpu_down(cpu, 1);
406                 if (!error)
407                         cpumask_set_cpu(cpu, frozen_cpus);
408                 else {
409                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
410                                 cpu, error);
411                         break;
412                 }
413         }
414
415         if (!error) {
416                 BUG_ON(num_online_cpus() > 1);
417                 /* Make sure the CPUs won't be enabled by someone else */
418                 cpu_hotplug_disabled = 1;
419         } else {
420                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
421         }
422         cpu_maps_update_done();
423         return error;
424 }
425
426 void __weak arch_enable_nonboot_cpus_begin(void)
427 {
428 }
429
430 void __weak arch_enable_nonboot_cpus_end(void)
431 {
432 }
433
434 void __ref enable_nonboot_cpus(void)
435 {
436         int cpu, error;
437
438         /* Allow everyone to use the CPU hotplug again */
439         cpu_maps_update_begin();
440         cpu_hotplug_disabled = 0;
441         if (cpumask_empty(frozen_cpus))
442                 goto out;
443
444         printk("Enabling non-boot CPUs ...\n");
445
446         arch_enable_nonboot_cpus_begin();
447
448         for_each_cpu(cpu, frozen_cpus) {
449                 error = _cpu_up(cpu, 1);
450                 if (!error) {
451                         printk("CPU%d is up\n", cpu);
452                         continue;
453                 }
454                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
455         }
456
457         arch_enable_nonboot_cpus_end();
458
459         cpumask_clear(frozen_cpus);
460 out:
461         cpu_maps_update_done();
462 }
463
464 static int alloc_frozen_cpus(void)
465 {
466         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
467                 return -ENOMEM;
468         return 0;
469 }
470 core_initcall(alloc_frozen_cpus);
471 #endif /* CONFIG_PM_SLEEP_SMP */
472
473 /**
474  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
475  * @cpu: cpu that just started
476  *
477  * This function calls the cpu_chain notifiers with CPU_STARTING.
478  * It must be called by the arch code on the new cpu, before the new cpu
479  * enables interrupts and before the "boot" cpu returns from __cpu_up().
480  */
481 void __cpuinit notify_cpu_starting(unsigned int cpu)
482 {
483         unsigned long val = CPU_STARTING;
484
485 #ifdef CONFIG_PM_SLEEP_SMP
486         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
487                 val = CPU_STARTING_FROZEN;
488 #endif /* CONFIG_PM_SLEEP_SMP */
489         cpu_notify(val, (void *)(long)cpu);
490 }
491
492 #endif /* CONFIG_SMP */
493
494 /*
495  * cpu_bit_bitmap[] is a special, "compressed" data structure that
496  * represents all NR_CPUS bits binary values of 1<<nr.
497  *
498  * It is used by cpumask_of() to get a constant address to a CPU
499  * mask value that has a single bit set only.
500  */
501
502 /* cpu_bit_bitmap[0] is empty - so we can back into it */
503 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
504 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
505 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
506 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
507
508 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
509
510         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
511         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
512 #if BITS_PER_LONG > 32
513         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
514         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
515 #endif
516 };
517 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
518
519 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
520 EXPORT_SYMBOL(cpu_all_bits);
521
522 #ifdef CONFIG_INIT_ALL_POSSIBLE
523 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
524         = CPU_BITS_ALL;
525 #else
526 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
527 #endif
528 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
529 EXPORT_SYMBOL(cpu_possible_mask);
530
531 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
532 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
533 EXPORT_SYMBOL(cpu_online_mask);
534
535 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
536 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
537 EXPORT_SYMBOL(cpu_present_mask);
538
539 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
540 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
541 EXPORT_SYMBOL(cpu_active_mask);
542
543 void set_cpu_possible(unsigned int cpu, bool possible)
544 {
545         if (possible)
546                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
547         else
548                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
549 }
550
551 void set_cpu_present(unsigned int cpu, bool present)
552 {
553         if (present)
554                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
555         else
556                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
557 }
558
559 void set_cpu_online(unsigned int cpu, bool online)
560 {
561         if (online)
562                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
563         else
564                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
565 }
566
567 void set_cpu_active(unsigned int cpu, bool active)
568 {
569         if (active)
570                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
571         else
572                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
573 }
574
575 void init_cpu_present(const struct cpumask *src)
576 {
577         cpumask_copy(to_cpumask(cpu_present_bits), src);
578 }
579
580 void init_cpu_possible(const struct cpumask *src)
581 {
582         cpumask_copy(to_cpumask(cpu_possible_bits), src);
583 }
584
585 void init_cpu_online(const struct cpumask *src)
586 {
587         cpumask_copy(to_cpumask(cpu_online_bits), src);
588 }