PM QOS update
[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         unsigned long mod;
168         void *hcpu;
169 };
170
171 /* Take this CPU down. */
172 static int __ref take_cpu_down(void *_param)
173 {
174         struct take_cpu_down_param *param = _param;
175         int err;
176
177         /* Ensure this CPU doesn't handle any more interrupts. */
178         err = __cpu_disable();
179         if (err < 0)
180                 return err;
181
182         raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
183                                 param->hcpu);
184
185         /* Force idle task to run as soon as we yield: it should
186            immediately notice cpu is offline and die quickly. */
187         sched_idle_next();
188         return 0;
189 }
190
191 /* Requires cpu_add_remove_lock to be held */
192 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
193 {
194         int err, nr_calls = 0;
195         cpumask_var_t old_allowed;
196         void *hcpu = (void *)(long)cpu;
197         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
198         struct take_cpu_down_param tcd_param = {
199                 .mod = mod,
200                 .hcpu = hcpu,
201         };
202
203         if (num_online_cpus() == 1)
204                 return -EBUSY;
205
206         if (!cpu_online(cpu))
207                 return -EINVAL;
208
209         if (!alloc_cpumask_var(&old_allowed, GFP_KERNEL))
210                 return -ENOMEM;
211
212         cpu_hotplug_begin();
213         set_cpu_active(cpu, false);
214         err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
215                                         hcpu, -1, &nr_calls);
216         if (err == NOTIFY_BAD) {
217                 set_cpu_active(cpu, true);
218
219                 nr_calls--;
220                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
221                                           hcpu, nr_calls, NULL);
222                 printk("%s: attempt to take down CPU %u failed\n",
223                                 __func__, cpu);
224                 err = -EINVAL;
225                 goto out_release;
226         }
227
228         /* Ensure that we are not runnable on dying cpu */
229         cpumask_copy(old_allowed, &current->cpus_allowed);
230         set_cpus_allowed_ptr(current, cpu_active_mask);
231
232         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
233         if (err) {
234                 set_cpu_active(cpu, true);
235                 /* CPU didn't die: tell everyone.  Can't complain. */
236                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
237                                             hcpu) == NOTIFY_BAD)
238                         BUG();
239
240                 goto out_allowed;
241         }
242         BUG_ON(cpu_online(cpu));
243
244         /* Wait for it to sleep (leaving idle task). */
245         while (!idle_cpu(cpu))
246                 yield();
247
248         /* This actually kills the CPU. */
249         __cpu_die(cpu);
250
251         /* CPU is completely dead: tell everyone.  Too late to complain. */
252         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
253                                     hcpu) == NOTIFY_BAD)
254                 BUG();
255
256         check_for_tasks(cpu);
257
258 out_allowed:
259         set_cpus_allowed_ptr(current, old_allowed);
260 out_release:
261         cpu_hotplug_done();
262         if (!err) {
263                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
264                                             hcpu) == NOTIFY_BAD)
265                         BUG();
266         }
267         free_cpumask_var(old_allowed);
268         return err;
269 }
270
271 int __ref cpu_down(unsigned int cpu)
272 {
273         int err;
274
275         err = stop_machine_create();
276         if (err)
277                 return err;
278         cpu_maps_update_begin();
279
280         if (cpu_hotplug_disabled) {
281                 err = -EBUSY;
282                 goto out;
283         }
284
285         err = _cpu_down(cpu, 0);
286
287 out:
288         cpu_maps_update_done();
289         stop_machine_destroy();
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 = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
307                                                         -1, &nr_calls);
308         if (ret == NOTIFY_BAD) {
309                 nr_calls--;
310                 printk("%s: attempt to bring up CPU %u failed\n",
311                                 __func__, cpu);
312                 ret = -EINVAL;
313                 goto out_notify;
314         }
315
316         /* Arch-specific enabling code. */
317         ret = __cpu_up(cpu);
318         if (ret != 0)
319                 goto out_notify;
320         BUG_ON(!cpu_online(cpu));
321
322         set_cpu_active(cpu, true);
323
324         /* Now call notifier in preparation. */
325         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
326
327 out_notify:
328         if (ret != 0)
329                 __raw_notifier_call_chain(&cpu_chain,
330                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
331         cpu_hotplug_done();
332
333         return ret;
334 }
335
336 int __cpuinit cpu_up(unsigned int cpu)
337 {
338         int err = 0;
339         if (!cpu_possible(cpu)) {
340                 printk(KERN_ERR "can't online cpu %d because it is not "
341                         "configured as may-hotadd at boot time\n", cpu);
342 #if defined(CONFIG_IA64)
343                 printk(KERN_ERR "please check additional_cpus= boot "
344                                 "parameter\n");
345 #endif
346                 return -EINVAL;
347         }
348
349         cpu_maps_update_begin();
350
351         if (cpu_hotplug_disabled) {
352                 err = -EBUSY;
353                 goto out;
354         }
355
356         err = _cpu_up(cpu, 0);
357
358 out:
359         cpu_maps_update_done();
360         return err;
361 }
362
363 #ifdef CONFIG_PM_SLEEP_SMP
364 static cpumask_var_t frozen_cpus;
365
366 int disable_nonboot_cpus(void)
367 {
368         int cpu, first_cpu, error;
369
370         error = stop_machine_create();
371         if (error)
372                 return error;
373         cpu_maps_update_begin();
374         first_cpu = cpumask_first(cpu_online_mask);
375         /*
376          * We take down all of the non-boot CPUs in one shot to avoid races
377          * with the userspace trying to use the CPU hotplug at the same time
378          */
379         cpumask_clear(frozen_cpus);
380
381         printk("Disabling non-boot CPUs ...\n");
382         for_each_online_cpu(cpu) {
383                 if (cpu == first_cpu)
384                         continue;
385                 error = _cpu_down(cpu, 1);
386                 if (!error)
387                         cpumask_set_cpu(cpu, frozen_cpus);
388                 else {
389                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
390                                 cpu, error);
391                         break;
392                 }
393         }
394
395         if (!error) {
396                 BUG_ON(num_online_cpus() > 1);
397                 /* Make sure the CPUs won't be enabled by someone else */
398                 cpu_hotplug_disabled = 1;
399         } else {
400                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
401         }
402         cpu_maps_update_done();
403         stop_machine_destroy();
404         return error;
405 }
406
407 void __weak arch_enable_nonboot_cpus_begin(void)
408 {
409 }
410
411 void __weak arch_enable_nonboot_cpus_end(void)
412 {
413 }
414
415 void __ref enable_nonboot_cpus(void)
416 {
417         int cpu, error;
418
419         /* Allow everyone to use the CPU hotplug again */
420         cpu_maps_update_begin();
421         cpu_hotplug_disabled = 0;
422         if (cpumask_empty(frozen_cpus))
423                 goto out;
424
425         printk("Enabling non-boot CPUs ...\n");
426
427         arch_enable_nonboot_cpus_begin();
428
429         for_each_cpu(cpu, frozen_cpus) {
430                 error = _cpu_up(cpu, 1);
431                 if (!error) {
432                         printk("CPU%d is up\n", cpu);
433                         continue;
434                 }
435                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
436         }
437
438         arch_enable_nonboot_cpus_end();
439
440         cpumask_clear(frozen_cpus);
441 out:
442         cpu_maps_update_done();
443 }
444
445 static int alloc_frozen_cpus(void)
446 {
447         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
448                 return -ENOMEM;
449         return 0;
450 }
451 core_initcall(alloc_frozen_cpus);
452 #endif /* CONFIG_PM_SLEEP_SMP */
453
454 /**
455  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
456  * @cpu: cpu that just started
457  *
458  * This function calls the cpu_chain notifiers with CPU_STARTING.
459  * It must be called by the arch code on the new cpu, before the new cpu
460  * enables interrupts and before the "boot" cpu returns from __cpu_up().
461  */
462 void __cpuinit notify_cpu_starting(unsigned int cpu)
463 {
464         unsigned long val = CPU_STARTING;
465
466 #ifdef CONFIG_PM_SLEEP_SMP
467         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
468                 val = CPU_STARTING_FROZEN;
469 #endif /* CONFIG_PM_SLEEP_SMP */
470         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
471 }
472
473 #endif /* CONFIG_SMP */
474
475 /*
476  * cpu_bit_bitmap[] is a special, "compressed" data structure that
477  * represents all NR_CPUS bits binary values of 1<<nr.
478  *
479  * It is used by cpumask_of() to get a constant address to a CPU
480  * mask value that has a single bit set only.
481  */
482
483 /* cpu_bit_bitmap[0] is empty - so we can back into it */
484 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
485 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
486 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
487 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
488
489 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
490
491         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
492         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
493 #if BITS_PER_LONG > 32
494         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
495         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
496 #endif
497 };
498 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
499
500 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
501 EXPORT_SYMBOL(cpu_all_bits);
502
503 #ifdef CONFIG_INIT_ALL_POSSIBLE
504 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
505         = CPU_BITS_ALL;
506 #else
507 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
508 #endif
509 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
510 EXPORT_SYMBOL(cpu_possible_mask);
511
512 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
513 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
514 EXPORT_SYMBOL(cpu_online_mask);
515
516 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
517 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
518 EXPORT_SYMBOL(cpu_present_mask);
519
520 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
521 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
522 EXPORT_SYMBOL(cpu_active_mask);
523
524 void set_cpu_possible(unsigned int cpu, bool possible)
525 {
526         if (possible)
527                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
528         else
529                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
530 }
531
532 void set_cpu_present(unsigned int cpu, bool present)
533 {
534         if (present)
535                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
536         else
537                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
538 }
539
540 void set_cpu_online(unsigned int cpu, bool online)
541 {
542         if (online)
543                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
544         else
545                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
546 }
547
548 void set_cpu_active(unsigned int cpu, bool active)
549 {
550         if (active)
551                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
552         else
553                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
554 }
555
556 void init_cpu_present(const struct cpumask *src)
557 {
558         cpumask_copy(to_cpumask(cpu_present_bits), src);
559 }
560
561 void init_cpu_possible(const struct cpumask *src)
562 {
563         cpumask_copy(to_cpumask(cpu_possible_bits), src);
564 }
565
566 void init_cpu_online(const struct cpumask *src)
567 {
568         cpumask_copy(to_cpumask(cpu_online_bits), src);
569 }