sched: Fix balance vs hotplug race
[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         .active_writer = NULL,
39         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
40         .refcount = 0,
41 };
42
43 #ifdef CONFIG_HOTPLUG_CPU
44
45 void get_online_cpus(void)
46 {
47         might_sleep();
48         if (cpu_hotplug.active_writer == current)
49                 return;
50         mutex_lock(&cpu_hotplug.lock);
51         cpu_hotplug.refcount++;
52         mutex_unlock(&cpu_hotplug.lock);
53
54 }
55 EXPORT_SYMBOL_GPL(get_online_cpus);
56
57 void put_online_cpus(void)
58 {
59         if (cpu_hotplug.active_writer == current)
60                 return;
61         mutex_lock(&cpu_hotplug.lock);
62         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
63                 wake_up_process(cpu_hotplug.active_writer);
64         mutex_unlock(&cpu_hotplug.lock);
65
66 }
67 EXPORT_SYMBOL_GPL(put_online_cpus);
68
69 #endif  /* CONFIG_HOTPLUG_CPU */
70
71 /*
72  * The following two API's must be used when attempting
73  * to serialize the updates to cpu_online_mask, cpu_present_mask.
74  */
75 void cpu_maps_update_begin(void)
76 {
77         mutex_lock(&cpu_add_remove_lock);
78 }
79
80 void cpu_maps_update_done(void)
81 {
82         mutex_unlock(&cpu_add_remove_lock);
83 }
84
85 /*
86  * This ensures that the hotplug operation can begin only when the
87  * refcount goes to zero.
88  *
89  * Note that during a cpu-hotplug operation, the new readers, if any,
90  * will be blocked by the cpu_hotplug.lock
91  *
92  * Since cpu_hotplug_begin() is always called after invoking
93  * cpu_maps_update_begin(), we can be sure that only one writer is active.
94  *
95  * Note that theoretically, there is a possibility of a livelock:
96  * - Refcount goes to zero, last reader wakes up the sleeping
97  *   writer.
98  * - Last reader unlocks the cpu_hotplug.lock.
99  * - A new reader arrives at this moment, bumps up the refcount.
100  * - The writer acquires the cpu_hotplug.lock finds the refcount
101  *   non zero and goes to sleep again.
102  *
103  * However, this is very difficult to achieve in practice since
104  * get_online_cpus() not an api which is called all that often.
105  *
106  */
107 static void cpu_hotplug_begin(void)
108 {
109         cpu_hotplug.active_writer = current;
110
111         for (;;) {
112                 mutex_lock(&cpu_hotplug.lock);
113                 if (likely(!cpu_hotplug.refcount))
114                         break;
115                 __set_current_state(TASK_UNINTERRUPTIBLE);
116                 mutex_unlock(&cpu_hotplug.lock);
117                 schedule();
118         }
119 }
120
121 static void cpu_hotplug_done(void)
122 {
123         cpu_hotplug.active_writer = NULL;
124         mutex_unlock(&cpu_hotplug.lock);
125 }
126 /* Need to know about CPUs going up/down? */
127 int __ref register_cpu_notifier(struct notifier_block *nb)
128 {
129         int ret;
130         cpu_maps_update_begin();
131         ret = raw_notifier_chain_register(&cpu_chain, nb);
132         cpu_maps_update_done();
133         return ret;
134 }
135
136 #ifdef CONFIG_HOTPLUG_CPU
137
138 EXPORT_SYMBOL(register_cpu_notifier);
139
140 void __ref unregister_cpu_notifier(struct notifier_block *nb)
141 {
142         cpu_maps_update_begin();
143         raw_notifier_chain_unregister(&cpu_chain, nb);
144         cpu_maps_update_done();
145 }
146 EXPORT_SYMBOL(unregister_cpu_notifier);
147
148 static inline void check_for_tasks(int cpu)
149 {
150         struct task_struct *p;
151
152         write_lock_irq(&tasklist_lock);
153         for_each_process(p) {
154                 if (task_cpu(p) == cpu &&
155                     (!cputime_eq(p->utime, cputime_zero) ||
156                      !cputime_eq(p->stime, cputime_zero)))
157                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
158                                 (state = %ld, flags = %x) \n",
159                                  p->comm, task_pid_nr(p), cpu,
160                                  p->state, p->flags);
161         }
162         write_unlock_irq(&tasklist_lock);
163 }
164
165 struct take_cpu_down_param {
166         unsigned long mod;
167         void *hcpu;
168 };
169
170 /* Take this CPU down. */
171 static int __ref take_cpu_down(void *_param)
172 {
173         struct take_cpu_down_param *param = _param;
174         int err;
175
176         /* Ensure this CPU doesn't handle any more interrupts. */
177         err = __cpu_disable();
178         if (err < 0)
179                 return err;
180
181         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
182                                 param->hcpu);
183
184         /* Force idle task to run as soon as we yield: it should
185            immediately notice cpu is offline and die quickly. */
186         sched_idle_next();
187         return 0;
188 }
189
190 /* Requires cpu_add_remove_lock to be held */
191 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
192 {
193         int err, nr_calls = 0;
194         cpumask_var_t old_allowed;
195         void *hcpu = (void *)(long)cpu;
196         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
197         struct take_cpu_down_param tcd_param = {
198                 .mod = mod,
199                 .hcpu = hcpu,
200         };
201
202         if (num_online_cpus() == 1)
203                 return -EBUSY;
204
205         if (!cpu_online(cpu))
206                 return -EINVAL;
207
208         if (!alloc_cpumask_var(&old_allowed, GFP_KERNEL))
209                 return -ENOMEM;
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                 set_cpu_active(cpu, true);
216
217                 nr_calls--;
218                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
219                                           hcpu, nr_calls, NULL);
220                 printk("%s: attempt to take down CPU %u failed\n",
221                                 __func__, cpu);
222                 err = -EINVAL;
223                 goto out_release;
224         }
225
226         /* Ensure that we are not runnable on dying cpu */
227         cpumask_copy(old_allowed, &current->cpus_allowed);
228         set_cpus_allowed_ptr(current, cpu_active_mask);
229
230         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
231         if (err) {
232                 set_cpu_active(cpu, true);
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         free_cpumask_var(old_allowed);
266         return err;
267 }
268
269 int __ref cpu_down(unsigned int cpu)
270 {
271         int err;
272
273         err = stop_machine_create();
274         if (err)
275                 return err;
276         cpu_maps_update_begin();
277
278         if (cpu_hotplug_disabled) {
279                 err = -EBUSY;
280                 goto out;
281         }
282
283         set_cpu_active(cpu, false);
284
285         /*
286          * Make sure the all cpus did the reschedule and are not
287          * using stale version of the cpu_active_mask.
288          * This is not strictly necessary becuase stop_machine()
289          * that we run down the line already provides the required
290          * synchronization. But it's really a side effect and we do not
291          * want to depend on the innards of the stop_machine here.
292          */
293         synchronize_sched();
294
295         err = _cpu_down(cpu, 0);
296
297 out:
298         cpu_maps_update_done();
299         stop_machine_destroy();
300         return err;
301 }
302 EXPORT_SYMBOL(cpu_down);
303 #endif /*CONFIG_HOTPLUG_CPU*/
304
305 /* Requires cpu_add_remove_lock to be held */
306 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
307 {
308         int ret, nr_calls = 0;
309         void *hcpu = (void *)(long)cpu;
310         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
311
312         if (cpu_online(cpu) || !cpu_present(cpu))
313                 return -EINVAL;
314
315         cpu_hotplug_begin();
316         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
317                                                         -1, &nr_calls);
318         if (ret == NOTIFY_BAD) {
319                 nr_calls--;
320                 printk("%s: attempt to bring up CPU %u failed\n",
321                                 __func__, cpu);
322                 ret = -EINVAL;
323                 goto out_notify;
324         }
325
326         /* Arch-specific enabling code. */
327         ret = __cpu_up(cpu);
328         if (ret != 0)
329                 goto out_notify;
330         BUG_ON(!cpu_online(cpu));
331
332         set_cpu_active(cpu, true);
333
334         /* Now call notifier in preparation. */
335         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
336
337 out_notify:
338         if (ret != 0)
339                 __raw_notifier_call_chain(&cpu_chain,
340                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
341         cpu_hotplug_done();
342
343         return ret;
344 }
345
346 int __cpuinit cpu_up(unsigned int cpu)
347 {
348         int err = 0;
349         if (!cpu_possible(cpu)) {
350                 printk(KERN_ERR "can't online cpu %d because it is not "
351                         "configured as may-hotadd at boot time\n", cpu);
352 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
353                 printk(KERN_ERR "please check additional_cpus= boot "
354                                 "parameter\n");
355 #endif
356                 return -EINVAL;
357         }
358
359         cpu_maps_update_begin();
360
361         if (cpu_hotplug_disabled) {
362                 err = -EBUSY;
363                 goto out;
364         }
365
366         err = _cpu_up(cpu, 0);
367
368 out:
369         cpu_maps_update_done();
370         return err;
371 }
372
373 #ifdef CONFIG_PM_SLEEP_SMP
374 static cpumask_var_t frozen_cpus;
375
376 int disable_nonboot_cpus(void)
377 {
378         int cpu, first_cpu, error;
379
380         error = stop_machine_create();
381         if (error)
382                 return error;
383         cpu_maps_update_begin();
384         first_cpu = cpumask_first(cpu_online_mask);
385         /* We take down all of the non-boot CPUs in one shot to avoid races
386          * with the userspace trying to use the CPU hotplug at the same time
387          */
388         cpumask_clear(frozen_cpus);
389
390         for_each_online_cpu(cpu) {
391                 if (cpu == first_cpu)
392                         continue;
393                 set_cpu_active(cpu, false);
394         }
395
396         synchronize_sched();
397
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                         cpumask_set_cpu(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
413         if (!error) {
414                 BUG_ON(num_online_cpus() > 1);
415                 /* Make sure the CPUs won't be enabled by someone else */
416                 cpu_hotplug_disabled = 1;
417         } else {
418                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
419         }
420         cpu_maps_update_done();
421         stop_machine_destroy();
422         return error;
423 }
424
425 void __weak arch_enable_nonboot_cpus_begin(void)
426 {
427 }
428
429 void __weak arch_enable_nonboot_cpus_end(void)
430 {
431 }
432
433 void __ref enable_nonboot_cpus(void)
434 {
435         int cpu, error;
436
437         /* Allow everyone to use the CPU hotplug again */
438         cpu_maps_update_begin();
439         cpu_hotplug_disabled = 0;
440         if (cpumask_empty(frozen_cpus))
441                 goto out;
442
443         printk("Enabling non-boot CPUs ...\n");
444
445         arch_enable_nonboot_cpus_begin();
446
447         for_each_cpu(cpu, frozen_cpus) {
448                 error = _cpu_up(cpu, 1);
449                 if (!error) {
450                         printk("CPU%d is up\n", cpu);
451                         continue;
452                 }
453                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
454         }
455
456         arch_enable_nonboot_cpus_end();
457
458         cpumask_clear(frozen_cpus);
459 out:
460         cpu_maps_update_done();
461 }
462
463 static int alloc_frozen_cpus(void)
464 {
465         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
466                 return -ENOMEM;
467         return 0;
468 }
469 core_initcall(alloc_frozen_cpus);
470 #endif /* CONFIG_PM_SLEEP_SMP */
471
472 /**
473  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
474  * @cpu: cpu that just started
475  *
476  * This function calls the cpu_chain notifiers with CPU_STARTING.
477  * It must be called by the arch code on the new cpu, before the new cpu
478  * enables interrupts and before the "boot" cpu returns from __cpu_up().
479  */
480 void __cpuinit notify_cpu_starting(unsigned int cpu)
481 {
482         unsigned long val = CPU_STARTING;
483
484 #ifdef CONFIG_PM_SLEEP_SMP
485         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
486                 val = CPU_STARTING_FROZEN;
487 #endif /* CONFIG_PM_SLEEP_SMP */
488         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
489 }
490
491 #endif /* CONFIG_SMP */
492
493 /*
494  * cpu_bit_bitmap[] is a special, "compressed" data structure that
495  * represents all NR_CPUS bits binary values of 1<<nr.
496  *
497  * It is used by cpumask_of() to get a constant address to a CPU
498  * mask value that has a single bit set only.
499  */
500
501 /* cpu_bit_bitmap[0] is empty - so we can back into it */
502 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
503 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
504 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
505 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
506
507 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
508
509         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
510         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
511 #if BITS_PER_LONG > 32
512         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
513         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
514 #endif
515 };
516 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
517
518 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
519 EXPORT_SYMBOL(cpu_all_bits);
520
521 #ifdef CONFIG_INIT_ALL_POSSIBLE
522 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
523         = CPU_BITS_ALL;
524 #else
525 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
526 #endif
527 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
528 EXPORT_SYMBOL(cpu_possible_mask);
529
530 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
531 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
532 EXPORT_SYMBOL(cpu_online_mask);
533
534 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
535 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
536 EXPORT_SYMBOL(cpu_present_mask);
537
538 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
539 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
540 EXPORT_SYMBOL(cpu_active_mask);
541
542 void set_cpu_possible(unsigned int cpu, bool possible)
543 {
544         if (possible)
545                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
546         else
547                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
548 }
549
550 void set_cpu_present(unsigned int cpu, bool present)
551 {
552         if (present)
553                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
554         else
555                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
556 }
557
558 void set_cpu_online(unsigned int cpu, bool online)
559 {
560         if (online)
561                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
562         else
563                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
564 }
565
566 void set_cpu_active(unsigned int cpu, bool active)
567 {
568         if (active)
569                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
570         else
571                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
572 }
573
574 void init_cpu_present(const struct cpumask *src)
575 {
576         cpumask_copy(to_cpumask(cpu_present_bits), src);
577 }
578
579 void init_cpu_possible(const struct cpumask *src)
580 {
581         cpumask_copy(to_cpumask(cpu_possible_bits), src);
582 }
583
584 void init_cpu_online(const struct cpumask *src)
585 {
586         cpumask_copy(to_cpumask(cpu_online_bits), src);
587 }