2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
4 * This code is licenced under the GPL.
6 #include <linux/proc_fs.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>
19 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
20 static DEFINE_MUTEX(cpu_add_remove_lock
);
22 static __cpuinitdata
RAW_NOTIFIER_HEAD(cpu_chain
);
24 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
25 * Should always be manipulated under cpu_add_remove_lock
27 static int cpu_hotplug_disabled
;
30 struct task_struct
*active_writer
;
31 struct mutex lock
; /* Synchronizes accesses to refcount, */
33 * Also blocks the new readers during
34 * an ongoing cpu hotplug operation.
39 void __init
cpu_hotplug_init(void)
41 cpu_hotplug
.active_writer
= NULL
;
42 mutex_init(&cpu_hotplug
.lock
);
43 cpu_hotplug
.refcount
= 0;
46 #ifdef CONFIG_HOTPLUG_CPU
48 void get_online_cpus(void)
51 if (cpu_hotplug
.active_writer
== current
)
53 mutex_lock(&cpu_hotplug
.lock
);
54 cpu_hotplug
.refcount
++;
55 mutex_unlock(&cpu_hotplug
.lock
);
58 EXPORT_SYMBOL_GPL(get_online_cpus
);
60 void put_online_cpus(void)
62 if (cpu_hotplug
.active_writer
== current
)
64 mutex_lock(&cpu_hotplug
.lock
);
65 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
66 wake_up_process(cpu_hotplug
.active_writer
);
67 mutex_unlock(&cpu_hotplug
.lock
);
70 EXPORT_SYMBOL_GPL(put_online_cpus
);
72 #endif /* CONFIG_HOTPLUG_CPU */
75 * The following two API's must be used when attempting
76 * to serialize the updates to cpu_online_mask, cpu_present_mask.
78 void cpu_maps_update_begin(void)
80 mutex_lock(&cpu_add_remove_lock
);
83 void cpu_maps_update_done(void)
85 mutex_unlock(&cpu_add_remove_lock
);
89 * This ensures that the hotplug operation can begin only when the
90 * refcount goes to zero.
92 * Note that during a cpu-hotplug operation, the new readers, if any,
93 * will be blocked by the cpu_hotplug.lock
95 * Since cpu_hotplug_begin() is always called after invoking
96 * cpu_maps_update_begin(), we can be sure that only one writer is active.
98 * Note that theoretically, there is a possibility of a livelock:
99 * - Refcount goes to zero, last reader wakes up the sleeping
101 * - Last reader unlocks the cpu_hotplug.lock.
102 * - A new reader arrives at this moment, bumps up the refcount.
103 * - The writer acquires the cpu_hotplug.lock finds the refcount
104 * non zero and goes to sleep again.
106 * However, this is very difficult to achieve in practice since
107 * get_online_cpus() not an api which is called all that often.
110 static void cpu_hotplug_begin(void)
112 cpu_hotplug
.active_writer
= current
;
115 mutex_lock(&cpu_hotplug
.lock
);
116 if (likely(!cpu_hotplug
.refcount
))
118 __set_current_state(TASK_UNINTERRUPTIBLE
);
119 mutex_unlock(&cpu_hotplug
.lock
);
124 static void cpu_hotplug_done(void)
126 cpu_hotplug
.active_writer
= NULL
;
127 mutex_unlock(&cpu_hotplug
.lock
);
129 /* Need to know about CPUs going up/down? */
130 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
133 cpu_maps_update_begin();
134 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
135 cpu_maps_update_done();
139 #ifdef CONFIG_HOTPLUG_CPU
141 EXPORT_SYMBOL(register_cpu_notifier
);
143 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
145 cpu_maps_update_begin();
146 raw_notifier_chain_unregister(&cpu_chain
, nb
);
147 cpu_maps_update_done();
149 EXPORT_SYMBOL(unregister_cpu_notifier
);
151 static inline void check_for_tasks(int cpu
)
153 struct task_struct
*p
;
155 write_lock_irq(&tasklist_lock
);
156 for_each_process(p
) {
157 if (task_cpu(p
) == cpu
&&
158 (!cputime_eq(p
->utime
, cputime_zero
) ||
159 !cputime_eq(p
->stime
, cputime_zero
)))
160 printk(KERN_WARNING
"Task %s (pid = %d) is on cpu %d\
161 (state = %ld, flags = %x) \n",
162 p
->comm
, task_pid_nr(p
), cpu
,
165 write_unlock_irq(&tasklist_lock
);
168 struct take_cpu_down_param
{
173 /* Take this CPU down. */
174 static int __ref
take_cpu_down(void *_param
)
176 struct take_cpu_down_param
*param
= _param
;
179 /* Ensure this CPU doesn't handle any more interrupts. */
180 err
= __cpu_disable();
184 raw_notifier_call_chain(&cpu_chain
, CPU_DYING
| param
->mod
,
187 /* Force idle task to run as soon as we yield: it should
188 immediately notice cpu is offline and die quickly. */
193 /* Requires cpu_add_remove_lock to be held */
194 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
196 int err
, nr_calls
= 0;
197 cpumask_var_t old_allowed
;
198 void *hcpu
= (void *)(long)cpu
;
199 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
200 struct take_cpu_down_param tcd_param
= {
205 if (num_online_cpus() == 1)
208 if (!cpu_online(cpu
))
211 if (!alloc_cpumask_var(&old_allowed
, GFP_KERNEL
))
215 err
= __raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_PREPARE
| mod
,
216 hcpu
, -1, &nr_calls
);
217 if (err
== NOTIFY_BAD
) {
219 __raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_FAILED
| mod
,
220 hcpu
, nr_calls
, NULL
);
221 printk("%s: attempt to take down CPU %u failed\n",
227 /* Ensure that we are not runnable on dying cpu */
228 cpumask_copy(old_allowed
, ¤t
->cpus_allowed
);
229 set_cpus_allowed_ptr(current
,
230 cpumask_of(cpumask_any_but(cpu_online_mask
, cpu
)));
232 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
234 /* CPU didn't die: tell everyone. Can't complain. */
235 if (raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_FAILED
| mod
,
241 BUG_ON(cpu_online(cpu
));
243 /* Wait for it to sleep (leaving idle task). */
244 while (!idle_cpu(cpu
))
247 /* This actually kills the CPU. */
250 /* CPU is completely dead: tell everyone. Too late to complain. */
251 if (raw_notifier_call_chain(&cpu_chain
, CPU_DEAD
| mod
,
255 check_for_tasks(cpu
);
258 set_cpus_allowed_ptr(current
, old_allowed
);
262 if (raw_notifier_call_chain(&cpu_chain
, CPU_POST_DEAD
| mod
,
266 free_cpumask_var(old_allowed
);
270 int __ref
cpu_down(unsigned int cpu
)
274 cpu_maps_update_begin();
276 if (cpu_hotplug_disabled
) {
281 cpu_clear(cpu
, cpu_active_map
);
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.
293 err
= _cpu_down(cpu
, 0);
296 cpu_set(cpu
, cpu_active_map
);
299 cpu_maps_update_done();
302 EXPORT_SYMBOL(cpu_down
);
303 #endif /*CONFIG_HOTPLUG_CPU*/
305 /* Requires cpu_add_remove_lock to be held */
306 static int __cpuinit
_cpu_up(unsigned int cpu
, int tasks_frozen
)
308 int ret
, nr_calls
= 0;
309 void *hcpu
= (void *)(long)cpu
;
310 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
312 if (cpu_online(cpu
) || !cpu_present(cpu
))
316 ret
= __raw_notifier_call_chain(&cpu_chain
, CPU_UP_PREPARE
| mod
, hcpu
,
318 if (ret
== NOTIFY_BAD
) {
320 printk("%s: attempt to bring up CPU %u failed\n",
326 /* Arch-specific enabling code. */
330 BUG_ON(!cpu_online(cpu
));
332 cpu_set(cpu
, cpu_active_map
);
334 /* Now call notifier in preparation. */
335 raw_notifier_call_chain(&cpu_chain
, CPU_ONLINE
| mod
, hcpu
);
339 __raw_notifier_call_chain(&cpu_chain
,
340 CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
346 int __cpuinit
cpu_up(unsigned int cpu
)
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 "
359 cpu_maps_update_begin();
361 if (cpu_hotplug_disabled
) {
366 err
= _cpu_up(cpu
, 0);
369 cpu_maps_update_done();
373 #ifdef CONFIG_PM_SLEEP_SMP
374 static cpumask_var_t frozen_cpus
;
376 int disable_nonboot_cpus(void)
378 int cpu
, first_cpu
, error
= 0;
380 cpu_maps_update_begin();
381 first_cpu
= cpumask_first(cpu_online_mask
);
382 /* We take down all of the non-boot CPUs in one shot to avoid races
383 * with the userspace trying to use the CPU hotplug at the same time
385 cpumask_clear(frozen_cpus
);
386 printk("Disabling non-boot CPUs ...\n");
387 for_each_online_cpu(cpu
) {
388 if (cpu
== first_cpu
)
390 error
= _cpu_down(cpu
, 1);
392 cpumask_set_cpu(cpu
, frozen_cpus
);
393 printk("CPU%d is down\n", cpu
);
395 printk(KERN_ERR
"Error taking CPU%d down: %d\n",
401 BUG_ON(num_online_cpus() > 1);
402 /* Make sure the CPUs won't be enabled by someone else */
403 cpu_hotplug_disabled
= 1;
405 printk(KERN_ERR
"Non-boot CPUs are not disabled\n");
407 cpu_maps_update_done();
411 void __ref
enable_nonboot_cpus(void)
415 /* Allow everyone to use the CPU hotplug again */
416 cpu_maps_update_begin();
417 cpu_hotplug_disabled
= 0;
418 if (cpumask_empty(frozen_cpus
))
421 printk("Enabling non-boot CPUs ...\n");
422 for_each_cpu(cpu
, frozen_cpus
) {
423 error
= _cpu_up(cpu
, 1);
425 printk("CPU%d is up\n", cpu
);
428 printk(KERN_WARNING
"Error taking CPU%d up: %d\n", cpu
, error
);
430 cpumask_clear(frozen_cpus
);
432 cpu_maps_update_done();
435 static int alloc_frozen_cpus(void)
437 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
441 core_initcall(alloc_frozen_cpus
);
442 #endif /* CONFIG_PM_SLEEP_SMP */
445 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
446 * @cpu: cpu that just started
448 * This function calls the cpu_chain notifiers with CPU_STARTING.
449 * It must be called by the arch code on the new cpu, before the new cpu
450 * enables interrupts and before the "boot" cpu returns from __cpu_up().
452 void __cpuinit
notify_cpu_starting(unsigned int cpu
)
454 unsigned long val
= CPU_STARTING
;
456 #ifdef CONFIG_PM_SLEEP_SMP
457 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
458 val
= CPU_STARTING_FROZEN
;
459 #endif /* CONFIG_PM_SLEEP_SMP */
460 raw_notifier_call_chain(&cpu_chain
, val
, (void *)(long)cpu
);
463 #endif /* CONFIG_SMP */
466 * cpu_bit_bitmap[] is a special, "compressed" data structure that
467 * represents all NR_CPUS bits binary values of 1<<nr.
469 * It is used by cpumask_of() to get a constant address to a CPU
470 * mask value that has a single bit set only.
473 /* cpu_bit_bitmap[0] is empty - so we can back into it */
474 #define MASK_DECLARE_1(x) [x+1][0] = 1UL << (x)
475 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
476 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
477 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
479 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
481 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
482 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
483 #if BITS_PER_LONG > 32
484 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
485 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
488 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
490 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
491 EXPORT_SYMBOL(cpu_all_bits
);
493 #ifdef CONFIG_INIT_ALL_POSSIBLE
494 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
497 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
499 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
500 EXPORT_SYMBOL(cpu_possible_mask
);
502 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
503 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
504 EXPORT_SYMBOL(cpu_online_mask
);
506 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
507 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
508 EXPORT_SYMBOL(cpu_present_mask
);
510 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
511 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
512 EXPORT_SYMBOL(cpu_active_mask
);
514 void set_cpu_possible(unsigned int cpu
, bool possible
)
517 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
519 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
522 void set_cpu_present(unsigned int cpu
, bool present
)
525 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
527 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
530 void set_cpu_online(unsigned int cpu
, bool online
)
533 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
535 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
538 void set_cpu_active(unsigned int cpu
, bool active
)
541 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
543 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
546 void init_cpu_present(const struct cpumask
*src
)
548 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
551 void init_cpu_possible(const struct cpumask
*src
)
553 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
556 void init_cpu_online(const struct cpumask
*src
)
558 cpumask_copy(to_cpumask(cpu_online_bits
), src
);