mem-hotplug: fix potential race while building zonelist for new populated zone
[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 static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
24
25 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
26  * Should always be manipulated under cpu_add_remove_lock
27  */
28 static int cpu_hotplug_disabled;
29
30 static struct {
31         struct task_struct *active_writer;
32         struct mutex lock; /* Synchronizes accesses to refcount, */
33         /*
34          * Also blocks the new readers during
35          * an ongoing cpu hotplug operation.
36          */
37         int refcount;
38 } cpu_hotplug = {
39         .active_writer = NULL,
40         .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
41         .refcount = 0,
42 };
43
44 #ifdef CONFIG_HOTPLUG_CPU
45
46 void get_online_cpus(void)
47 {
48         might_sleep();
49         if (cpu_hotplug.active_writer == current)
50                 return;
51         mutex_lock(&cpu_hotplug.lock);
52         cpu_hotplug.refcount++;
53         mutex_unlock(&cpu_hotplug.lock);
54
55 }
56 EXPORT_SYMBOL_GPL(get_online_cpus);
57
58 void put_online_cpus(void)
59 {
60         if (cpu_hotplug.active_writer == current)
61                 return;
62         mutex_lock(&cpu_hotplug.lock);
63         if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
64                 wake_up_process(cpu_hotplug.active_writer);
65         mutex_unlock(&cpu_hotplug.lock);
66
67 }
68 EXPORT_SYMBOL_GPL(put_online_cpus);
69
70 #endif  /* CONFIG_HOTPLUG_CPU */
71
72 /*
73  * The following two API's must be used when attempting
74  * to serialize the updates to cpu_online_mask, cpu_present_mask.
75  */
76 void cpu_maps_update_begin(void)
77 {
78         mutex_lock(&cpu_add_remove_lock);
79 }
80
81 void cpu_maps_update_done(void)
82 {
83         mutex_unlock(&cpu_add_remove_lock);
84 }
85
86 /*
87  * This ensures that the hotplug operation can begin only when the
88  * refcount goes to zero.
89  *
90  * Note that during a cpu-hotplug operation, the new readers, if any,
91  * will be blocked by the cpu_hotplug.lock
92  *
93  * Since cpu_hotplug_begin() is always called after invoking
94  * cpu_maps_update_begin(), we can be sure that only one writer is active.
95  *
96  * Note that theoretically, there is a possibility of a livelock:
97  * - Refcount goes to zero, last reader wakes up the sleeping
98  *   writer.
99  * - Last reader unlocks the cpu_hotplug.lock.
100  * - A new reader arrives at this moment, bumps up the refcount.
101  * - The writer acquires the cpu_hotplug.lock finds the refcount
102  *   non zero and goes to sleep again.
103  *
104  * However, this is very difficult to achieve in practice since
105  * get_online_cpus() not an api which is called all that often.
106  *
107  */
108 static void cpu_hotplug_begin(void)
109 {
110         cpu_hotplug.active_writer = current;
111
112         for (;;) {
113                 mutex_lock(&cpu_hotplug.lock);
114                 if (likely(!cpu_hotplug.refcount))
115                         break;
116                 __set_current_state(TASK_UNINTERRUPTIBLE);
117                 mutex_unlock(&cpu_hotplug.lock);
118                 schedule();
119         }
120 }
121
122 static void cpu_hotplug_done(void)
123 {
124         cpu_hotplug.active_writer = NULL;
125         mutex_unlock(&cpu_hotplug.lock);
126 }
127 /* Need to know about CPUs going up/down? */
128 int __ref register_cpu_notifier(struct notifier_block *nb)
129 {
130         int ret;
131         cpu_maps_update_begin();
132         ret = raw_notifier_chain_register(&cpu_chain, nb);
133         cpu_maps_update_done();
134         return ret;
135 }
136
137 #ifdef CONFIG_HOTPLUG_CPU
138
139 EXPORT_SYMBOL(register_cpu_notifier);
140
141 void __ref unregister_cpu_notifier(struct notifier_block *nb)
142 {
143         cpu_maps_update_begin();
144         raw_notifier_chain_unregister(&cpu_chain, nb);
145         cpu_maps_update_done();
146 }
147 EXPORT_SYMBOL(unregister_cpu_notifier);
148
149 static inline void check_for_tasks(int cpu)
150 {
151         struct task_struct *p;
152
153         write_lock_irq(&tasklist_lock);
154         for_each_process(p) {
155                 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
156                     (!cputime_eq(p->utime, cputime_zero) ||
157                      !cputime_eq(p->stime, cputime_zero)))
158                         printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
159                                 "(state = %ld, flags = %x)\n",
160                                 p->comm, task_pid_nr(p), cpu,
161                                 p->state, p->flags);
162         }
163         write_unlock_irq(&tasklist_lock);
164 }
165
166 struct take_cpu_down_param {
167         struct task_struct *caller;
168         unsigned long mod;
169         void *hcpu;
170 };
171
172 /* Take this CPU down. */
173 static int __ref take_cpu_down(void *_param)
174 {
175         struct take_cpu_down_param *param = _param;
176         unsigned int cpu = (unsigned long)param->hcpu;
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         if (task_cpu(param->caller) == cpu)
188                 move_task_off_dead_cpu(cpu, param->caller);
189         /* Force idle task to run as soon as we yield: it should
190            immediately notice cpu is offline and die quickly. */
191         sched_idle_next();
192         return 0;
193 }
194
195 /* Requires cpu_add_remove_lock to be held */
196 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
197 {
198         int err, nr_calls = 0;
199         void *hcpu = (void *)(long)cpu;
200         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
201         struct take_cpu_down_param tcd_param = {
202                 .caller = current,
203                 .mod = mod,
204                 .hcpu = hcpu,
205         };
206
207         if (num_online_cpus() == 1)
208                 return -EBUSY;
209
210         if (!cpu_online(cpu))
211                 return -EINVAL;
212
213         cpu_hotplug_begin();
214         set_cpu_active(cpu, false);
215         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
216                                         hcpu, -1, &nr_calls);
217         if (err == NOTIFY_BAD) {
218                 set_cpu_active(cpu, true);
219
220                 nr_calls--;
221                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
222                                           hcpu, nr_calls, NULL);
223                 printk("%s: attempt to take down CPU %u failed\n",
224                                 __func__, cpu);
225                 err = -EINVAL;
226                 goto out_release;
227         }
228
229         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
230         if (err) {
231                 set_cpu_active(cpu, true);
232                 /* CPU didn't die: tell everyone.  Can't complain. */
233                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
234                                             hcpu) == NOTIFY_BAD)
235                         BUG();
236
237                 goto out_release;
238         }
239         BUG_ON(cpu_online(cpu));
240
241         /* Wait for it to sleep (leaving idle task). */
242         while (!idle_cpu(cpu))
243                 yield();
244
245         /* This actually kills the CPU. */
246         __cpu_die(cpu);
247
248         /* CPU is completely dead: tell everyone.  Too late to complain. */
249         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
250                                     hcpu) == NOTIFY_BAD)
251                 BUG();
252
253         check_for_tasks(cpu);
254
255 out_release:
256         cpu_hotplug_done();
257         if (!err) {
258                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
259                                             hcpu) == NOTIFY_BAD)
260                         BUG();
261         }
262         return err;
263 }
264
265 int __ref cpu_down(unsigned int cpu)
266 {
267         int err;
268
269         cpu_maps_update_begin();
270
271         if (cpu_hotplug_disabled) {
272                 err = -EBUSY;
273                 goto out;
274         }
275
276         err = _cpu_down(cpu, 0);
277
278 out:
279         cpu_maps_update_done();
280         return err;
281 }
282 EXPORT_SYMBOL(cpu_down);
283 #endif /*CONFIG_HOTPLUG_CPU*/
284
285 /* Requires cpu_add_remove_lock to be held */
286 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
287 {
288         int ret, nr_calls = 0;
289         void *hcpu = (void *)(long)cpu;
290         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
291
292         if (cpu_online(cpu) || !cpu_present(cpu))
293                 return -EINVAL;
294
295         cpu_hotplug_begin();
296         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
297                                                         -1, &nr_calls);
298         if (ret == NOTIFY_BAD) {
299                 nr_calls--;
300                 printk("%s: attempt to bring up CPU %u failed\n",
301                                 __func__, cpu);
302                 ret = -EINVAL;
303                 goto out_notify;
304         }
305
306         /* Arch-specific enabling code. */
307         ret = __cpu_up(cpu);
308         if (ret != 0)
309                 goto out_notify;
310         BUG_ON(!cpu_online(cpu));
311
312         set_cpu_active(cpu, true);
313
314         /* Now call notifier in preparation. */
315         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
316
317 out_notify:
318         if (ret != 0)
319                 __raw_notifier_call_chain(&cpu_chain,
320                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
321         cpu_hotplug_done();
322
323         return ret;
324 }
325
326 int __cpuinit cpu_up(unsigned int cpu)
327 {
328         int err = 0;
329
330 #ifdef  CONFIG_MEMORY_HOTPLUG
331         int nid;
332         pg_data_t       *pgdat;
333 #endif
334
335         if (!cpu_possible(cpu)) {
336                 printk(KERN_ERR "can't online cpu %d because it is not "
337                         "configured as may-hotadd at boot time\n", cpu);
338 #if defined(CONFIG_IA64)
339                 printk(KERN_ERR "please check additional_cpus= boot "
340                                 "parameter\n");
341 #endif
342                 return -EINVAL;
343         }
344
345 #ifdef  CONFIG_MEMORY_HOTPLUG
346         nid = cpu_to_node(cpu);
347         if (!node_online(nid)) {
348                 err = mem_online_node(nid);
349                 if (err)
350                         return err;
351         }
352
353         pgdat = NODE_DATA(nid);
354         if (!pgdat) {
355                 printk(KERN_ERR
356                         "Can't online cpu %d due to NULL pgdat\n", cpu);
357                 return -ENOMEM;
358         }
359
360         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
361                 mutex_lock(&zonelists_mutex);
362                 build_all_zonelists(NULL);
363                 mutex_unlock(&zonelists_mutex);
364         }
365 #endif
366
367         cpu_maps_update_begin();
368
369         if (cpu_hotplug_disabled) {
370                 err = -EBUSY;
371                 goto out;
372         }
373
374         err = _cpu_up(cpu, 0);
375
376 out:
377         cpu_maps_update_done();
378         return err;
379 }
380
381 #ifdef CONFIG_PM_SLEEP_SMP
382 static cpumask_var_t frozen_cpus;
383
384 int disable_nonboot_cpus(void)
385 {
386         int cpu, first_cpu, error;
387
388         cpu_maps_update_begin();
389         first_cpu = cpumask_first(cpu_online_mask);
390         /*
391          * We take down all of the non-boot CPUs in one shot to avoid races
392          * with the userspace trying to use the CPU hotplug at the same time
393          */
394         cpumask_clear(frozen_cpus);
395
396         printk("Disabling non-boot CPUs ...\n");
397         for_each_online_cpu(cpu) {
398                 if (cpu == first_cpu)
399                         continue;
400                 error = _cpu_down(cpu, 1);
401                 if (!error)
402                         cpumask_set_cpu(cpu, frozen_cpus);
403                 else {
404                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
405                                 cpu, error);
406                         break;
407                 }
408         }
409
410         if (!error) {
411                 BUG_ON(num_online_cpus() > 1);
412                 /* Make sure the CPUs won't be enabled by someone else */
413                 cpu_hotplug_disabled = 1;
414         } else {
415                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
416         }
417         cpu_maps_update_done();
418         return error;
419 }
420
421 void __weak arch_enable_nonboot_cpus_begin(void)
422 {
423 }
424
425 void __weak arch_enable_nonboot_cpus_end(void)
426 {
427 }
428
429 void __ref enable_nonboot_cpus(void)
430 {
431         int cpu, error;
432
433         /* Allow everyone to use the CPU hotplug again */
434         cpu_maps_update_begin();
435         cpu_hotplug_disabled = 0;
436         if (cpumask_empty(frozen_cpus))
437                 goto out;
438
439         printk("Enabling non-boot CPUs ...\n");
440
441         arch_enable_nonboot_cpus_begin();
442
443         for_each_cpu(cpu, frozen_cpus) {
444                 error = _cpu_up(cpu, 1);
445                 if (!error) {
446                         printk("CPU%d is up\n", cpu);
447                         continue;
448                 }
449                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
450         }
451
452         arch_enable_nonboot_cpus_end();
453
454         cpumask_clear(frozen_cpus);
455 out:
456         cpu_maps_update_done();
457 }
458
459 static int alloc_frozen_cpus(void)
460 {
461         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
462                 return -ENOMEM;
463         return 0;
464 }
465 core_initcall(alloc_frozen_cpus);
466 #endif /* CONFIG_PM_SLEEP_SMP */
467
468 /**
469  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
470  * @cpu: cpu that just started
471  *
472  * This function calls the cpu_chain notifiers with CPU_STARTING.
473  * It must be called by the arch code on the new cpu, before the new cpu
474  * enables interrupts and before the "boot" cpu returns from __cpu_up().
475  */
476 void __cpuinit notify_cpu_starting(unsigned int cpu)
477 {
478         unsigned long val = CPU_STARTING;
479
480 #ifdef CONFIG_PM_SLEEP_SMP
481         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
482                 val = CPU_STARTING_FROZEN;
483 #endif /* CONFIG_PM_SLEEP_SMP */
484         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
485 }
486
487 #endif /* CONFIG_SMP */
488
489 /*
490  * cpu_bit_bitmap[] is a special, "compressed" data structure that
491  * represents all NR_CPUS bits binary values of 1<<nr.
492  *
493  * It is used by cpumask_of() to get a constant address to a CPU
494  * mask value that has a single bit set only.
495  */
496
497 /* cpu_bit_bitmap[0] is empty - so we can back into it */
498 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
499 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
500 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
501 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
502
503 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
504
505         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
506         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
507 #if BITS_PER_LONG > 32
508         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
509         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
510 #endif
511 };
512 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
513
514 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
515 EXPORT_SYMBOL(cpu_all_bits);
516
517 #ifdef CONFIG_INIT_ALL_POSSIBLE
518 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
519         = CPU_BITS_ALL;
520 #else
521 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
522 #endif
523 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
524 EXPORT_SYMBOL(cpu_possible_mask);
525
526 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
527 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
528 EXPORT_SYMBOL(cpu_online_mask);
529
530 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
531 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
532 EXPORT_SYMBOL(cpu_present_mask);
533
534 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
535 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
536 EXPORT_SYMBOL(cpu_active_mask);
537
538 void set_cpu_possible(unsigned int cpu, bool possible)
539 {
540         if (possible)
541                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
542         else
543                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
544 }
545
546 void set_cpu_present(unsigned int cpu, bool present)
547 {
548         if (present)
549                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
550         else
551                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
552 }
553
554 void set_cpu_online(unsigned int cpu, bool online)
555 {
556         if (online)
557                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
558         else
559                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
560 }
561
562 void set_cpu_active(unsigned int cpu, bool active)
563 {
564         if (active)
565                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
566         else
567                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
568 }
569
570 void init_cpu_present(const struct cpumask *src)
571 {
572         cpumask_copy(to_cpumask(cpu_present_bits), src);
573 }
574
575 void init_cpu_possible(const struct cpumask *src)
576 {
577         cpumask_copy(to_cpumask(cpu_possible_bits), src);
578 }
579
580 void init_cpu_online(const struct cpumask *src)
581 {
582         cpumask_copy(to_cpumask(cpu_online_bits), src);
583 }