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_t old_allowed, tmp;
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))
212 err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
213 hcpu, -1, &nr_calls);
214 if (err == NOTIFY_BAD) {
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",
224 /* Ensure that we are not runnable on dying cpu */
225 old_allowed = current->cpus_allowed;
228 set_cpus_allowed_ptr(current, &tmp);
229 tmp = cpumask_of_cpu(cpu);
231 err = __stop_machine(take_cpu_down, &tcd_param, &tmp);
233 /* CPU didn't die: tell everyone. Can't complain. */
234 if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
240 BUG_ON(cpu_online(cpu));
242 /* Wait for it to sleep (leaving idle task). */
243 while (!idle_cpu(cpu))
246 /* This actually kills the CPU. */
249 /* CPU is completely dead: tell everyone. Too late to complain. */
250 if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
254 check_for_tasks(cpu);
257 set_cpus_allowed_ptr(current, &old_allowed);
261 if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
268 int __ref cpu_down(unsigned int cpu)
272 cpu_maps_update_begin();
274 if (cpu_hotplug_disabled) {
279 cpu_clear(cpu, cpu_active_map);
282 * Make sure the all cpus did the reschedule and are not
283 * using stale version of the cpu_active_map.
284 * This is not strictly necessary becuase stop_machine()
285 * that we run down the line already provides the required
286 * synchronization. But it's really a side effect and we do not
287 * want to depend on the innards of the stop_machine here.
291 err = _cpu_down(cpu, 0);
294 cpu_set(cpu, cpu_active_map);
297 cpu_maps_update_done();
300 EXPORT_SYMBOL(cpu_down);
301 #endif /*CONFIG_HOTPLUG_CPU*/
303 /* Requires cpu_add_remove_lock to be held */
304 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
306 int ret, nr_calls = 0;
307 void *hcpu = (void *)(long)cpu;
308 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
310 if (cpu_online(cpu) || !cpu_present(cpu))
314 ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
316 if (ret == NOTIFY_BAD) {
318 printk("%s: attempt to bring up CPU %u failed\n",
324 /* Arch-specific enabling code. */
328 BUG_ON(!cpu_online(cpu));
330 cpu_set(cpu, cpu_active_map);
332 /* Now call notifier in preparation. */
333 raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
337 __raw_notifier_call_chain(&cpu_chain,
338 CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
344 int __cpuinit cpu_up(unsigned int cpu)
347 if (!cpu_isset(cpu, cpu_possible_map)) {
348 printk(KERN_ERR "can't online cpu %d because it is not "
349 "configured as may-hotadd at boot time\n", cpu);
350 #if defined(CONFIG_IA64) || defined(CONFIG_X86_64)
351 printk(KERN_ERR "please check additional_cpus= boot "
357 cpu_maps_update_begin();
359 if (cpu_hotplug_disabled) {
364 err = _cpu_up(cpu, 0);
367 cpu_maps_update_done();
371 #ifdef CONFIG_PM_SLEEP_SMP
372 static cpumask_t frozen_cpus;
374 int disable_nonboot_cpus(void)
376 int cpu, first_cpu, error = 0;
378 cpu_maps_update_begin();
379 first_cpu = first_cpu(cpu_online_map);
380 /* We take down all of the non-boot CPUs in one shot to avoid races
381 * with the userspace trying to use the CPU hotplug at the same time
383 cpus_clear(frozen_cpus);
384 printk("Disabling non-boot CPUs ...\n");
385 for_each_online_cpu(cpu) {
386 if (cpu == first_cpu)
388 error = _cpu_down(cpu, 1);
390 cpu_set(cpu, frozen_cpus);
391 printk("CPU%d is down\n", cpu);
393 printk(KERN_ERR "Error taking CPU%d down: %d\n",
399 BUG_ON(num_online_cpus() > 1);
400 /* Make sure the CPUs won't be enabled by someone else */
401 cpu_hotplug_disabled = 1;
403 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
405 cpu_maps_update_done();
409 void __ref enable_nonboot_cpus(void)
413 /* Allow everyone to use the CPU hotplug again */
414 cpu_maps_update_begin();
415 cpu_hotplug_disabled = 0;
416 if (cpus_empty(frozen_cpus))
419 printk("Enabling non-boot CPUs ...\n");
420 for_each_cpu_mask_nr(cpu, frozen_cpus) {
421 error = _cpu_up(cpu, 1);
423 printk("CPU%d is up\n", cpu);
426 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
428 cpus_clear(frozen_cpus);
430 cpu_maps_update_done();
432 #endif /* CONFIG_PM_SLEEP_SMP */
435 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
436 * @cpu: cpu that just started
438 * This function calls the cpu_chain notifiers with CPU_STARTING.
439 * It must be called by the arch code on the new cpu, before the new cpu
440 * enables interrupts and before the "boot" cpu returns from __cpu_up().
442 void __cpuinit notify_cpu_starting(unsigned int cpu)
444 unsigned long val = CPU_STARTING;
446 #ifdef CONFIG_PM_SLEEP_SMP
447 if (cpu_isset(cpu, frozen_cpus))
448 val = CPU_STARTING_FROZEN;
449 #endif /* CONFIG_PM_SLEEP_SMP */
450 raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
453 #endif /* CONFIG_SMP */
456 * cpu_bit_bitmap[] is a special, "compressed" data structure that
457 * represents all NR_CPUS bits binary values of 1<<nr.
459 * It is used by cpumask_of_cpu() to get a constant address to a CPU
460 * mask value that has a single bit set only.
463 /* cpu_bit_bitmap[0] is empty - so we can back into it */
464 #define MASK_DECLARE_1(x) [x+1][0] = 1UL << (x)
465 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
466 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
467 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
469 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
471 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
472 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
473 #if BITS_PER_LONG > 32
474 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
475 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
478 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
480 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
481 EXPORT_SYMBOL(cpu_all_bits);
483 #ifdef CONFIG_INIT_ALL_POSSIBLE
484 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
487 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
489 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
490 EXPORT_SYMBOL(cpu_possible_mask);
492 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
493 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
494 EXPORT_SYMBOL(cpu_online_mask);
496 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
497 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
498 EXPORT_SYMBOL(cpu_present_mask);
500 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
501 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
502 EXPORT_SYMBOL(cpu_active_mask);
504 void set_cpu_possible(unsigned int cpu, bool possible)
507 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
509 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
512 void set_cpu_present(unsigned int cpu, bool present)
515 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
517 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
520 void set_cpu_online(unsigned int cpu, bool online)
523 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
525 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
528 void set_cpu_active(unsigned int cpu, bool active)
531 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
533 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
536 void init_cpu_present(const struct cpumask *src)
538 cpumask_copy(to_cpumask(cpu_present_bits), src);
541 void init_cpu_possible(const struct cpumask *src)
543 cpumask_copy(to_cpumask(cpu_possible_bits), src);
546 void init_cpu_online(const struct cpumask *src)
548 cpumask_copy(to_cpumask(cpu_online_bits), src);