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 * Represents all cpu's present in the system
20 * In systems capable of hotplug, this map could dynamically grow
21 * as new cpu's are detected in the system via any platform specific
22 * method, such as ACPI for e.g.
24 cpumask_t cpu_present_map __read_mostly
;
25 EXPORT_SYMBOL(cpu_present_map
);
28 * Represents all cpu's that are currently online.
30 cpumask_t cpu_online_map __read_mostly
;
31 EXPORT_SYMBOL(cpu_online_map
);
33 #ifdef CONFIG_INIT_ALL_POSSIBLE
34 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
36 cpumask_t cpu_possible_map __read_mostly
;
38 EXPORT_SYMBOL(cpu_possible_map
);
41 /* Serializes the updates to cpu_online_map, cpu_present_map */
42 static DEFINE_MUTEX(cpu_add_remove_lock
);
44 static __cpuinitdata
RAW_NOTIFIER_HEAD(cpu_chain
);
46 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
47 * Should always be manipulated under cpu_add_remove_lock
49 static int cpu_hotplug_disabled
;
52 struct task_struct
*active_writer
;
53 struct mutex lock
; /* Synchronizes accesses to refcount, */
55 * Also blocks the new readers during
56 * an ongoing cpu hotplug operation.
61 void __init
cpu_hotplug_init(void)
63 cpu_hotplug
.active_writer
= NULL
;
64 mutex_init(&cpu_hotplug
.lock
);
65 cpu_hotplug
.refcount
= 0;
68 cpumask_t cpu_active_map
;
70 #ifdef CONFIG_HOTPLUG_CPU
72 void get_online_cpus(void)
75 if (cpu_hotplug
.active_writer
== current
)
77 mutex_lock(&cpu_hotplug
.lock
);
78 cpu_hotplug
.refcount
++;
79 mutex_unlock(&cpu_hotplug
.lock
);
82 EXPORT_SYMBOL_GPL(get_online_cpus
);
84 void put_online_cpus(void)
86 if (cpu_hotplug
.active_writer
== current
)
88 mutex_lock(&cpu_hotplug
.lock
);
89 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
90 wake_up_process(cpu_hotplug
.active_writer
);
91 mutex_unlock(&cpu_hotplug
.lock
);
94 EXPORT_SYMBOL_GPL(put_online_cpus
);
96 #endif /* CONFIG_HOTPLUG_CPU */
99 * The following two API's must be used when attempting
100 * to serialize the updates to cpu_online_map, cpu_present_map.
102 void cpu_maps_update_begin(void)
104 mutex_lock(&cpu_add_remove_lock
);
107 void cpu_maps_update_done(void)
109 mutex_unlock(&cpu_add_remove_lock
);
113 * This ensures that the hotplug operation can begin only when the
114 * refcount goes to zero.
116 * Note that during a cpu-hotplug operation, the new readers, if any,
117 * will be blocked by the cpu_hotplug.lock
119 * Since cpu_hotplug_begin() is always called after invoking
120 * cpu_maps_update_begin(), we can be sure that only one writer is active.
122 * Note that theoretically, there is a possibility of a livelock:
123 * - Refcount goes to zero, last reader wakes up the sleeping
125 * - Last reader unlocks the cpu_hotplug.lock.
126 * - A new reader arrives at this moment, bumps up the refcount.
127 * - The writer acquires the cpu_hotplug.lock finds the refcount
128 * non zero and goes to sleep again.
130 * However, this is very difficult to achieve in practice since
131 * get_online_cpus() not an api which is called all that often.
134 static void cpu_hotplug_begin(void)
136 cpu_hotplug
.active_writer
= current
;
139 mutex_lock(&cpu_hotplug
.lock
);
140 if (likely(!cpu_hotplug
.refcount
))
142 __set_current_state(TASK_UNINTERRUPTIBLE
);
143 mutex_unlock(&cpu_hotplug
.lock
);
148 static void cpu_hotplug_done(void)
150 cpu_hotplug
.active_writer
= NULL
;
151 mutex_unlock(&cpu_hotplug
.lock
);
153 /* Need to know about CPUs going up/down? */
154 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
157 cpu_maps_update_begin();
158 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
159 cpu_maps_update_done();
163 #ifdef CONFIG_HOTPLUG_CPU
165 EXPORT_SYMBOL(register_cpu_notifier
);
167 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
169 cpu_maps_update_begin();
170 raw_notifier_chain_unregister(&cpu_chain
, nb
);
171 cpu_maps_update_done();
173 EXPORT_SYMBOL(unregister_cpu_notifier
);
175 static inline void check_for_tasks(int cpu
)
177 struct task_struct
*p
;
179 write_lock_irq(&tasklist_lock
);
180 for_each_process(p
) {
181 if (task_cpu(p
) == cpu
&&
182 (!cputime_eq(p
->utime
, cputime_zero
) ||
183 !cputime_eq(p
->stime
, cputime_zero
)))
184 printk(KERN_WARNING
"Task %s (pid = %d) is on cpu %d\
185 (state = %ld, flags = %x) \n",
186 p
->comm
, task_pid_nr(p
), cpu
,
189 write_unlock_irq(&tasklist_lock
);
192 struct take_cpu_down_param
{
197 /* Take this CPU down. */
198 static int __ref
take_cpu_down(void *_param
)
200 struct take_cpu_down_param
*param
= _param
;
203 /* Ensure this CPU doesn't handle any more interrupts. */
204 err
= __cpu_disable();
208 raw_notifier_call_chain(&cpu_chain
, CPU_DYING
| param
->mod
,
211 /* Force idle task to run as soon as we yield: it should
212 immediately notice cpu is offline and die quickly. */
217 /* Requires cpu_add_remove_lock to be held */
218 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
220 int err
, nr_calls
= 0;
221 cpumask_t old_allowed
, tmp
;
222 void *hcpu
= (void *)(long)cpu
;
223 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
224 struct take_cpu_down_param tcd_param
= {
229 if (num_online_cpus() == 1)
232 if (!cpu_online(cpu
))
236 err
= __raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_PREPARE
| mod
,
237 hcpu
, -1, &nr_calls
);
238 if (err
== NOTIFY_BAD
) {
240 __raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_FAILED
| mod
,
241 hcpu
, nr_calls
, NULL
);
242 printk("%s: attempt to take down CPU %u failed\n",
248 /* Ensure that we are not runnable on dying cpu */
249 old_allowed
= current
->cpus_allowed
;
252 set_cpus_allowed_ptr(current
, &tmp
);
253 tmp
= cpumask_of_cpu(cpu
);
255 err
= __stop_machine(take_cpu_down
, &tcd_param
, &tmp
);
257 /* CPU didn't die: tell everyone. Can't complain. */
258 if (raw_notifier_call_chain(&cpu_chain
, CPU_DOWN_FAILED
| mod
,
264 BUG_ON(cpu_online(cpu
));
266 /* Wait for it to sleep (leaving idle task). */
267 while (!idle_cpu(cpu
))
270 /* This actually kills the CPU. */
273 /* CPU is completely dead: tell everyone. Too late to complain. */
274 if (raw_notifier_call_chain(&cpu_chain
, CPU_DEAD
| mod
,
278 check_for_tasks(cpu
);
281 set_cpus_allowed_ptr(current
, &old_allowed
);
285 if (raw_notifier_call_chain(&cpu_chain
, CPU_POST_DEAD
| mod
,
292 int __ref
cpu_down(unsigned int cpu
)
296 cpu_maps_update_begin();
298 if (cpu_hotplug_disabled
) {
303 cpu_clear(cpu
, cpu_active_map
);
306 * Make sure the all cpus did the reschedule and are not
307 * using stale version of the cpu_active_map.
308 * This is not strictly necessary becuase stop_machine()
309 * that we run down the line already provides the required
310 * synchronization. But it's really a side effect and we do not
311 * want to depend on the innards of the stop_machine here.
315 err
= _cpu_down(cpu
, 0);
318 cpu_set(cpu
, cpu_active_map
);
321 cpu_maps_update_done();
324 EXPORT_SYMBOL(cpu_down
);
325 #endif /*CONFIG_HOTPLUG_CPU*/
327 /* Requires cpu_add_remove_lock to be held */
328 static int __cpuinit
_cpu_up(unsigned int cpu
, int tasks_frozen
)
330 int ret
, nr_calls
= 0;
331 void *hcpu
= (void *)(long)cpu
;
332 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
334 if (cpu_online(cpu
) || !cpu_present(cpu
))
338 ret
= __raw_notifier_call_chain(&cpu_chain
, CPU_UP_PREPARE
| mod
, hcpu
,
340 if (ret
== NOTIFY_BAD
) {
342 printk("%s: attempt to bring up CPU %u failed\n",
348 /* Arch-specific enabling code. */
352 BUG_ON(!cpu_online(cpu
));
354 cpu_set(cpu
, cpu_active_map
);
356 /* Now call notifier in preparation. */
357 raw_notifier_call_chain(&cpu_chain
, CPU_ONLINE
| mod
, hcpu
);
361 __raw_notifier_call_chain(&cpu_chain
,
362 CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
368 int __cpuinit
cpu_up(unsigned int cpu
)
371 if (!cpu_isset(cpu
, cpu_possible_map
)) {
372 printk(KERN_ERR
"can't online cpu %d because it is not "
373 "configured as may-hotadd at boot time\n", cpu
);
374 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
375 printk(KERN_ERR
"please check additional_cpus= boot "
381 cpu_maps_update_begin();
383 if (cpu_hotplug_disabled
) {
388 err
= _cpu_up(cpu
, 0);
391 cpu_maps_update_done();
395 #ifdef CONFIG_PM_SLEEP_SMP
396 static cpumask_t frozen_cpus
;
398 int disable_nonboot_cpus(void)
400 int cpu
, first_cpu
, error
= 0;
402 cpu_maps_update_begin();
403 first_cpu
= first_cpu(cpu_online_map
);
404 /* We take down all of the non-boot CPUs in one shot to avoid races
405 * with the userspace trying to use the CPU hotplug at the same time
407 cpus_clear(frozen_cpus
);
408 printk("Disabling non-boot CPUs ...\n");
409 for_each_online_cpu(cpu
) {
410 if (cpu
== first_cpu
)
412 error
= _cpu_down(cpu
, 1);
414 cpu_set(cpu
, frozen_cpus
);
415 printk("CPU%d is down\n", cpu
);
417 printk(KERN_ERR
"Error taking CPU%d down: %d\n",
423 BUG_ON(num_online_cpus() > 1);
424 /* Make sure the CPUs won't be enabled by someone else */
425 cpu_hotplug_disabled
= 1;
427 printk(KERN_ERR
"Non-boot CPUs are not disabled\n");
429 cpu_maps_update_done();
433 void __ref
enable_nonboot_cpus(void)
437 /* Allow everyone to use the CPU hotplug again */
438 cpu_maps_update_begin();
439 cpu_hotplug_disabled
= 0;
440 if (cpus_empty(frozen_cpus
))
443 printk("Enabling non-boot CPUs ...\n");
444 for_each_cpu_mask_nr(cpu
, frozen_cpus
) {
445 error
= _cpu_up(cpu
, 1);
447 printk("CPU%d is up\n", cpu
);
450 printk(KERN_WARNING
"Error taking CPU%d up: %d\n", cpu
, error
);
452 cpus_clear(frozen_cpus
);
454 cpu_maps_update_done();
456 #endif /* CONFIG_PM_SLEEP_SMP */
459 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
460 * @cpu: cpu that just started
462 * This function calls the cpu_chain notifiers with CPU_STARTING.
463 * It must be called by the arch code on the new cpu, before the new cpu
464 * enables interrupts and before the "boot" cpu returns from __cpu_up().
466 void __cpuinit
notify_cpu_starting(unsigned int cpu
)
468 unsigned long val
= CPU_STARTING
;
470 #ifdef CONFIG_PM_SLEEP_SMP
471 if (cpu_isset(cpu
, frozen_cpus
))
472 val
= CPU_STARTING_FROZEN
;
473 #endif /* CONFIG_PM_SLEEP_SMP */
474 raw_notifier_call_chain(&cpu_chain
, val
, (void *)(long)cpu
);
477 #endif /* CONFIG_SMP */
480 * cpu_bit_bitmap[] is a special, "compressed" data structure that
481 * represents all NR_CPUS bits binary values of 1<<nr.
483 * It is used by cpumask_of_cpu() to get a constant address to a CPU
484 * mask value that has a single bit set only.
487 /* cpu_bit_bitmap[0] is empty - so we can back into it */
488 #define MASK_DECLARE_1(x) [x+1][0] = 1UL << (x)
489 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
490 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
491 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
493 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
495 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
496 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
497 #if BITS_PER_LONG > 32
498 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
499 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
502 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
504 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
505 EXPORT_SYMBOL(cpu_all_bits
);