sched: Simplify cpu-hot-unplug task migration
[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         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         unsigned int cpu = (unsigned long)param->hcpu;
201         int err;
202
203         /* Ensure this CPU doesn't handle any more interrupts. */
204         err = __cpu_disable();
205         if (err < 0)
206                 return err;
207
208         cpu_notify(CPU_DYING | param->mod, param->hcpu);
209
210         return 0;
211 }
212
213 /* Requires cpu_add_remove_lock to be held */
214 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
215 {
216         int err, nr_calls = 0;
217         void *hcpu = (void *)(long)cpu;
218         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
219         struct take_cpu_down_param tcd_param = {
220                 .mod = mod,
221                 .hcpu = hcpu,
222         };
223
224         if (num_online_cpus() == 1)
225                 return -EBUSY;
226
227         if (!cpu_online(cpu))
228                 return -EINVAL;
229
230         cpu_hotplug_begin();
231         err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
232         if (err) {
233                 nr_calls--;
234                 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
235                 printk("%s: attempt to take down CPU %u failed\n",
236                                 __func__, cpu);
237                 goto out_release;
238         }
239
240         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
241         if (err) {
242                 /* CPU didn't die: tell everyone.  Can't complain. */
243                 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
244
245                 goto out_release;
246         }
247         BUG_ON(cpu_online(cpu));
248
249         /*
250          * The migration_call() CPU_DYING callback will have removed all
251          * runnable tasks from the cpu, there's only the idle task left now
252          * that the migration thread is done doing the stop_machine thing.
253          */
254         BUG_ON(!idle_cpu(cpu));
255
256         /* This actually kills the CPU. */
257         __cpu_die(cpu);
258
259         /* CPU is completely dead: tell everyone.  Too late to complain. */
260         cpu_notify_nofail(CPU_DEAD | mod, hcpu);
261
262         check_for_tasks(cpu);
263
264 out_release:
265         cpu_hotplug_done();
266         if (!err)
267                 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
268         return err;
269 }
270
271 int __ref cpu_down(unsigned int cpu)
272 {
273         int err;
274
275         cpu_maps_update_begin();
276
277         if (cpu_hotplug_disabled) {
278                 err = -EBUSY;
279                 goto out;
280         }
281
282         err = _cpu_down(cpu, 0);
283
284 out:
285         cpu_maps_update_done();
286         return err;
287 }
288 EXPORT_SYMBOL(cpu_down);
289 #endif /*CONFIG_HOTPLUG_CPU*/
290
291 /* Requires cpu_add_remove_lock to be held */
292 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
293 {
294         int ret, nr_calls = 0;
295         void *hcpu = (void *)(long)cpu;
296         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
297
298         if (cpu_online(cpu) || !cpu_present(cpu))
299                 return -EINVAL;
300
301         cpu_hotplug_begin();
302         ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
303         if (ret) {
304                 nr_calls--;
305                 printk("%s: attempt to bring up CPU %u failed\n",
306                                 __func__, cpu);
307                 goto out_notify;
308         }
309
310         /* Arch-specific enabling code. */
311         ret = __cpu_up(cpu);
312         if (ret != 0)
313                 goto out_notify;
314         BUG_ON(!cpu_online(cpu));
315
316         /* Now call notifier in preparation. */
317         cpu_notify(CPU_ONLINE | mod, hcpu);
318
319 out_notify:
320         if (ret != 0)
321                 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
322         cpu_hotplug_done();
323
324         return ret;
325 }
326
327 int __cpuinit cpu_up(unsigned int cpu)
328 {
329         int err = 0;
330
331 #ifdef  CONFIG_MEMORY_HOTPLUG
332         int nid;
333         pg_data_t       *pgdat;
334 #endif
335
336         if (!cpu_possible(cpu)) {
337                 printk(KERN_ERR "can't online cpu %d because it is not "
338                         "configured as may-hotadd at boot time\n", cpu);
339 #if defined(CONFIG_IA64)
340                 printk(KERN_ERR "please check additional_cpus= boot "
341                                 "parameter\n");
342 #endif
343                 return -EINVAL;
344         }
345
346 #ifdef  CONFIG_MEMORY_HOTPLUG
347         nid = cpu_to_node(cpu);
348         if (!node_online(nid)) {
349                 err = mem_online_node(nid);
350                 if (err)
351                         return err;
352         }
353
354         pgdat = NODE_DATA(nid);
355         if (!pgdat) {
356                 printk(KERN_ERR
357                         "Can't online cpu %d due to NULL pgdat\n", cpu);
358                 return -ENOMEM;
359         }
360
361         if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
362                 mutex_lock(&zonelists_mutex);
363                 build_all_zonelists(NULL);
364                 mutex_unlock(&zonelists_mutex);
365         }
366 #endif
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 out:
378         cpu_maps_update_done();
379         return err;
380 }
381
382 #ifdef CONFIG_PM_SLEEP_SMP
383 static cpumask_var_t frozen_cpus;
384
385 int disable_nonboot_cpus(void)
386 {
387         int cpu, first_cpu, error = 0;
388
389         cpu_maps_update_begin();
390         first_cpu = cpumask_first(cpu_online_mask);
391         /*
392          * We take down all of the non-boot CPUs in one shot to avoid races
393          * with the userspace trying to use the CPU hotplug at the same time
394          */
395         cpumask_clear(frozen_cpus);
396
397         printk("Disabling non-boot CPUs ...\n");
398         for_each_online_cpu(cpu) {
399                 if (cpu == first_cpu)
400                         continue;
401                 error = _cpu_down(cpu, 1);
402                 if (!error)
403                         cpumask_set_cpu(cpu, frozen_cpus);
404                 else {
405                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
406                                 cpu, error);
407                         break;
408                 }
409         }
410
411         if (!error) {
412                 BUG_ON(num_online_cpus() > 1);
413                 /* Make sure the CPUs won't be enabled by someone else */
414                 cpu_hotplug_disabled = 1;
415         } else {
416                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
417         }
418         cpu_maps_update_done();
419         return error;
420 }
421
422 void __weak arch_enable_nonboot_cpus_begin(void)
423 {
424 }
425
426 void __weak arch_enable_nonboot_cpus_end(void)
427 {
428 }
429
430 void __ref enable_nonboot_cpus(void)
431 {
432         int cpu, error;
433
434         /* Allow everyone to use the CPU hotplug again */
435         cpu_maps_update_begin();
436         cpu_hotplug_disabled = 0;
437         if (cpumask_empty(frozen_cpus))
438                 goto out;
439
440         printk("Enabling non-boot CPUs ...\n");
441
442         arch_enable_nonboot_cpus_begin();
443
444         for_each_cpu(cpu, frozen_cpus) {
445                 error = _cpu_up(cpu, 1);
446                 if (!error) {
447                         printk("CPU%d is up\n", cpu);
448                         continue;
449                 }
450                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
451         }
452
453         arch_enable_nonboot_cpus_end();
454
455         cpumask_clear(frozen_cpus);
456 out:
457         cpu_maps_update_done();
458 }
459
460 static int alloc_frozen_cpus(void)
461 {
462         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
463                 return -ENOMEM;
464         return 0;
465 }
466 core_initcall(alloc_frozen_cpus);
467 #endif /* CONFIG_PM_SLEEP_SMP */
468
469 /**
470  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
471  * @cpu: cpu that just started
472  *
473  * This function calls the cpu_chain notifiers with CPU_STARTING.
474  * It must be called by the arch code on the new cpu, before the new cpu
475  * enables interrupts and before the "boot" cpu returns from __cpu_up().
476  */
477 void __cpuinit notify_cpu_starting(unsigned int cpu)
478 {
479         unsigned long val = CPU_STARTING;
480
481 #ifdef CONFIG_PM_SLEEP_SMP
482         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
483                 val = CPU_STARTING_FROZEN;
484 #endif /* CONFIG_PM_SLEEP_SMP */
485         cpu_notify(val, (void *)(long)cpu);
486 }
487
488 #endif /* CONFIG_SMP */
489
490 /*
491  * cpu_bit_bitmap[] is a special, "compressed" data structure that
492  * represents all NR_CPUS bits binary values of 1<<nr.
493  *
494  * It is used by cpumask_of() to get a constant address to a CPU
495  * mask value that has a single bit set only.
496  */
497
498 /* cpu_bit_bitmap[0] is empty - so we can back into it */
499 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
500 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
501 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
502 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
503
504 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
505
506         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
507         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
508 #if BITS_PER_LONG > 32
509         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
510         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
511 #endif
512 };
513 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
514
515 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
516 EXPORT_SYMBOL(cpu_all_bits);
517
518 #ifdef CONFIG_INIT_ALL_POSSIBLE
519 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
520         = CPU_BITS_ALL;
521 #else
522 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
523 #endif
524 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
525 EXPORT_SYMBOL(cpu_possible_mask);
526
527 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
528 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
529 EXPORT_SYMBOL(cpu_online_mask);
530
531 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
532 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
533 EXPORT_SYMBOL(cpu_present_mask);
534
535 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
536 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
537 EXPORT_SYMBOL(cpu_active_mask);
538
539 void set_cpu_possible(unsigned int cpu, bool possible)
540 {
541         if (possible)
542                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
543         else
544                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
545 }
546
547 void set_cpu_present(unsigned int cpu, bool present)
548 {
549         if (present)
550                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
551         else
552                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
553 }
554
555 void set_cpu_online(unsigned int cpu, bool online)
556 {
557         if (online)
558                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
559         else
560                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
561 }
562
563 void set_cpu_active(unsigned int cpu, bool active)
564 {
565         if (active)
566                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
567         else
568                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
569 }
570
571 void init_cpu_present(const struct cpumask *src)
572 {
573         cpumask_copy(to_cpumask(cpu_present_bits), src);
574 }
575
576 void init_cpu_possible(const struct cpumask *src)
577 {
578         cpumask_copy(to_cpumask(cpu_possible_bits), src);
579 }
580
581 void init_cpu_online(const struct cpumask *src)
582 {
583         cpumask_copy(to_cpumask(cpu_online_bits), src);
584 }