x86, pat/mtrr: Rendezvous all the cpus for MTRR/PAT init
[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                 nr_calls--;
216                 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
217                                           hcpu, nr_calls, NULL);
218                 printk("%s: attempt to take down CPU %u failed\n",
219                                 __func__, cpu);
220                 err = -EINVAL;
221                 goto out_release;
222         }
223
224         /* Ensure that we are not runnable on dying cpu */
225         cpumask_copy(old_allowed, &current->cpus_allowed);
226         set_cpus_allowed_ptr(current,
227                              cpumask_of(cpumask_any_but(cpu_online_mask, cpu)));
228
229         err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
230         if (err) {
231                 /* CPU didn't die: tell everyone.  Can't complain. */
232                 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
233                                             hcpu) == NOTIFY_BAD)
234                         BUG();
235
236                 goto out_allowed;
237         }
238         BUG_ON(cpu_online(cpu));
239
240         /* Wait for it to sleep (leaving idle task). */
241         while (!idle_cpu(cpu))
242                 yield();
243
244         /* This actually kills the CPU. */
245         __cpu_die(cpu);
246
247         /* CPU is completely dead: tell everyone.  Too late to complain. */
248         if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
249                                     hcpu) == NOTIFY_BAD)
250                 BUG();
251
252         check_for_tasks(cpu);
253
254 out_allowed:
255         set_cpus_allowed_ptr(current, old_allowed);
256 out_release:
257         cpu_hotplug_done();
258         if (!err) {
259                 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
260                                             hcpu) == NOTIFY_BAD)
261                         BUG();
262         }
263         free_cpumask_var(old_allowed);
264         return err;
265 }
266
267 int __ref cpu_down(unsigned int cpu)
268 {
269         int err;
270
271         err = stop_machine_create();
272         if (err)
273                 return err;
274         cpu_maps_update_begin();
275
276         if (cpu_hotplug_disabled) {
277                 err = -EBUSY;
278                 goto out;
279         }
280
281         set_cpu_active(cpu, false);
282
283         /*
284          * Make sure the all cpus did the reschedule and are not
285          * using stale version of the cpu_active_mask.
286          * This is not strictly necessary becuase stop_machine()
287          * that we run down the line already provides the required
288          * synchronization. But it's really a side effect and we do not
289          * want to depend on the innards of the stop_machine here.
290          */
291         synchronize_sched();
292
293         err = _cpu_down(cpu, 0);
294
295         if (cpu_online(cpu))
296                 set_cpu_active(cpu, true);
297
298 out:
299         cpu_maps_update_done();
300         stop_machine_destroy();
301         return err;
302 }
303 EXPORT_SYMBOL(cpu_down);
304 #endif /*CONFIG_HOTPLUG_CPU*/
305
306 /* Requires cpu_add_remove_lock to be held */
307 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
308 {
309         int ret, nr_calls = 0;
310         void *hcpu = (void *)(long)cpu;
311         unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
312
313         if (cpu_online(cpu) || !cpu_present(cpu))
314                 return -EINVAL;
315
316         cpu_hotplug_begin();
317         ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
318                                                         -1, &nr_calls);
319         if (ret == NOTIFY_BAD) {
320                 nr_calls--;
321                 printk("%s: attempt to bring up CPU %u failed\n",
322                                 __func__, cpu);
323                 ret = -EINVAL;
324                 goto out_notify;
325         }
326
327         /* Arch-specific enabling code. */
328         ret = __cpu_up(cpu);
329         if (ret != 0)
330                 goto out_notify;
331         BUG_ON(!cpu_online(cpu));
332
333         set_cpu_active(cpu, true);
334
335         /* Now call notifier in preparation. */
336         raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
337
338 out_notify:
339         if (ret != 0)
340                 __raw_notifier_call_chain(&cpu_chain,
341                                 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
342         cpu_hotplug_done();
343
344         return ret;
345 }
346
347 int __cpuinit cpu_up(unsigned int cpu)
348 {
349         int err = 0;
350         if (!cpu_possible(cpu)) {
351                 printk(KERN_ERR "can't online cpu %d because it is not "
352                         "configured as may-hotadd at boot time\n", cpu);
353 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
354                 printk(KERN_ERR "please check additional_cpus= boot "
355                                 "parameter\n");
356 #endif
357                 return -EINVAL;
358         }
359
360         cpu_maps_update_begin();
361
362         if (cpu_hotplug_disabled) {
363                 err = -EBUSY;
364                 goto out;
365         }
366
367         err = _cpu_up(cpu, 0);
368
369 out:
370         cpu_maps_update_done();
371         return err;
372 }
373
374 #ifdef CONFIG_PM_SLEEP_SMP
375 static cpumask_var_t frozen_cpus;
376
377 int disable_nonboot_cpus(void)
378 {
379         int cpu, first_cpu, error;
380
381         error = stop_machine_create();
382         if (error)
383                 return error;
384         cpu_maps_update_begin();
385         first_cpu = cpumask_first(cpu_online_mask);
386         /* We take down all of the non-boot CPUs in one shot to avoid races
387          * with the userspace trying to use the CPU hotplug at the same time
388          */
389         cpumask_clear(frozen_cpus);
390         printk("Disabling non-boot CPUs ...\n");
391         for_each_online_cpu(cpu) {
392                 if (cpu == first_cpu)
393                         continue;
394                 error = _cpu_down(cpu, 1);
395                 if (!error) {
396                         cpumask_set_cpu(cpu, frozen_cpus);
397                         printk("CPU%d is down\n", cpu);
398                 } else {
399                         printk(KERN_ERR "Error taking CPU%d down: %d\n",
400                                 cpu, error);
401                         break;
402                 }
403         }
404         if (!error) {
405                 BUG_ON(num_online_cpus() > 1);
406                 /* Make sure the CPUs won't be enabled by someone else */
407                 cpu_hotplug_disabled = 1;
408         } else {
409                 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
410         }
411         cpu_maps_update_done();
412         stop_machine_destroy();
413         return error;
414 }
415
416 void __weak arch_enable_nonboot_cpus_begin(void)
417 {
418 }
419
420 void __weak arch_enable_nonboot_cpus_end(void)
421 {
422 }
423
424 void __ref enable_nonboot_cpus(void)
425 {
426         int cpu, error;
427
428         /* Allow everyone to use the CPU hotplug again */
429         cpu_maps_update_begin();
430         cpu_hotplug_disabled = 0;
431         if (cpumask_empty(frozen_cpus))
432                 goto out;
433
434         printk("Enabling non-boot CPUs ...\n");
435
436         arch_enable_nonboot_cpus_begin();
437
438         for_each_cpu(cpu, frozen_cpus) {
439                 error = _cpu_up(cpu, 1);
440                 if (!error) {
441                         printk("CPU%d is up\n", cpu);
442                         continue;
443                 }
444                 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
445         }
446
447         arch_enable_nonboot_cpus_end();
448
449         cpumask_clear(frozen_cpus);
450 out:
451         cpu_maps_update_done();
452 }
453
454 static int alloc_frozen_cpus(void)
455 {
456         if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
457                 return -ENOMEM;
458         return 0;
459 }
460 core_initcall(alloc_frozen_cpus);
461 #endif /* CONFIG_PM_SLEEP_SMP */
462
463 /**
464  * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
465  * @cpu: cpu that just started
466  *
467  * This function calls the cpu_chain notifiers with CPU_STARTING.
468  * It must be called by the arch code on the new cpu, before the new cpu
469  * enables interrupts and before the "boot" cpu returns from __cpu_up().
470  */
471 void __cpuinit notify_cpu_starting(unsigned int cpu)
472 {
473         unsigned long val = CPU_STARTING;
474
475 #ifdef CONFIG_PM_SLEEP_SMP
476         if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
477                 val = CPU_STARTING_FROZEN;
478 #endif /* CONFIG_PM_SLEEP_SMP */
479         raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
480 }
481
482 #endif /* CONFIG_SMP */
483
484 /*
485  * cpu_bit_bitmap[] is a special, "compressed" data structure that
486  * represents all NR_CPUS bits binary values of 1<<nr.
487  *
488  * It is used by cpumask_of() to get a constant address to a CPU
489  * mask value that has a single bit set only.
490  */
491
492 /* cpu_bit_bitmap[0] is empty - so we can back into it */
493 #define MASK_DECLARE_1(x)       [x+1][0] = 1UL << (x)
494 #define MASK_DECLARE_2(x)       MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
495 #define MASK_DECLARE_4(x)       MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
496 #define MASK_DECLARE_8(x)       MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
497
498 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
499
500         MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
501         MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
502 #if BITS_PER_LONG > 32
503         MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
504         MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
505 #endif
506 };
507 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
508
509 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
510 EXPORT_SYMBOL(cpu_all_bits);
511
512 #ifdef CONFIG_INIT_ALL_POSSIBLE
513 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
514         = CPU_BITS_ALL;
515 #else
516 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
517 #endif
518 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
519 EXPORT_SYMBOL(cpu_possible_mask);
520
521 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
522 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
523 EXPORT_SYMBOL(cpu_online_mask);
524
525 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
526 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
527 EXPORT_SYMBOL(cpu_present_mask);
528
529 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
530 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
531 EXPORT_SYMBOL(cpu_active_mask);
532
533 void set_cpu_possible(unsigned int cpu, bool possible)
534 {
535         if (possible)
536                 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
537         else
538                 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
539 }
540
541 void set_cpu_present(unsigned int cpu, bool present)
542 {
543         if (present)
544                 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
545         else
546                 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
547 }
548
549 void set_cpu_online(unsigned int cpu, bool online)
550 {
551         if (online)
552                 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
553         else
554                 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
555 }
556
557 void set_cpu_active(unsigned int cpu, bool active)
558 {
559         if (active)
560                 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
561         else
562                 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
563 }
564
565 void init_cpu_present(const struct cpumask *src)
566 {
567         cpumask_copy(to_cpumask(cpu_present_bits), src);
568 }
569
570 void init_cpu_possible(const struct cpumask *src)
571 {
572         cpumask_copy(to_cpumask(cpu_possible_bits), src);
573 }
574
575 void init_cpu_online(const struct cpumask *src)
576 {
577         cpumask_copy(to_cpumask(cpu_online_bits), src);
578 }