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.
38 .active_writer = NULL,
39 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
43 #ifdef CONFIG_HOTPLUG_CPU
45 void get_online_cpus(void)
48 if (cpu_hotplug.active_writer == current)
50 mutex_lock(&cpu_hotplug.lock);
51 cpu_hotplug.refcount++;
52 mutex_unlock(&cpu_hotplug.lock);
55 EXPORT_SYMBOL_GPL(get_online_cpus);
57 void put_online_cpus(void)
59 if (cpu_hotplug.active_writer == current)
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);
67 EXPORT_SYMBOL_GPL(put_online_cpus);
69 #endif /* CONFIG_HOTPLUG_CPU */
72 * The following two API's must be used when attempting
73 * to serialize the updates to cpu_online_mask, cpu_present_mask.
75 void cpu_maps_update_begin(void)
77 mutex_lock(&cpu_add_remove_lock);
80 void cpu_maps_update_done(void)
82 mutex_unlock(&cpu_add_remove_lock);
86 * This ensures that the hotplug operation can begin only when the
87 * refcount goes to zero.
89 * Note that during a cpu-hotplug operation, the new readers, if any,
90 * will be blocked by the cpu_hotplug.lock
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.
95 * Note that theoretically, there is a possibility of a livelock:
96 * - Refcount goes to zero, last reader wakes up the sleeping
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.
103 * However, this is very difficult to achieve in practice since
104 * get_online_cpus() not an api which is called all that often.
107 static void cpu_hotplug_begin(void)
109 cpu_hotplug.active_writer = current;
112 mutex_lock(&cpu_hotplug.lock);
113 if (likely(!cpu_hotplug.refcount))
115 __set_current_state(TASK_UNINTERRUPTIBLE);
116 mutex_unlock(&cpu_hotplug.lock);
121 static void cpu_hotplug_done(void)
123 cpu_hotplug.active_writer = NULL;
124 mutex_unlock(&cpu_hotplug.lock);
126 /* Need to know about CPUs going up/down? */
127 int __ref register_cpu_notifier(struct notifier_block *nb)
130 cpu_maps_update_begin();
131 ret = raw_notifier_chain_register(&cpu_chain, nb);
132 cpu_maps_update_done();
136 #ifdef CONFIG_HOTPLUG_CPU
138 EXPORT_SYMBOL(register_cpu_notifier);
140 void __ref unregister_cpu_notifier(struct notifier_block *nb)
142 cpu_maps_update_begin();
143 raw_notifier_chain_unregister(&cpu_chain, nb);
144 cpu_maps_update_done();
146 EXPORT_SYMBOL(unregister_cpu_notifier);
148 static inline void check_for_tasks(int cpu)
150 struct task_struct *p;
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,
162 write_unlock_irq(&tasklist_lock);
165 struct take_cpu_down_param {
170 /* Take this CPU down. */
171 static int __ref take_cpu_down(void *_param)
173 struct take_cpu_down_param *param = _param;
176 /* Ensure this CPU doesn't handle any more interrupts. */
177 err = __cpu_disable();
181 raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
184 /* Force idle task to run as soon as we yield: it should
185 immediately notice cpu is offline and die quickly. */
190 /* Requires cpu_add_remove_lock to be held */
191 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
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 = {
202 if (num_online_cpus() == 1)
205 if (!cpu_online(cpu))
208 if (!alloc_cpumask_var(&old_allowed, GFP_KERNEL))
212 err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
213 hcpu, -1, &nr_calls);
214 if (err == NOTIFY_BAD) {
215 set_cpu_active(cpu, true);
218 __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
219 hcpu, nr_calls, NULL);
220 printk("%s: attempt to take down CPU %u failed\n",
226 /* Ensure that we are not runnable on dying cpu */
227 cpumask_copy(old_allowed, ¤t->cpus_allowed);
228 set_cpus_allowed_ptr(current, cpu_active_mask);
230 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
232 set_cpu_active(cpu, true);
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,
265 free_cpumask_var(old_allowed);
269 int __ref cpu_down(unsigned int cpu)
273 err = stop_machine_create();
276 cpu_maps_update_begin();
278 if (cpu_hotplug_disabled) {
283 set_cpu_active(cpu, false);
286 * Make sure the all cpus did the reschedule and are not
287 * using stale version of the cpu_active_mask.
288 * This is not strictly necessary becuase stop_machine()
289 * that we run down the line already provides the required
290 * synchronization. But it's really a side effect and we do not
291 * want to depend on the innards of the stop_machine here.
295 err = _cpu_down(cpu, 0);
298 cpu_maps_update_done();
299 stop_machine_destroy();
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 set_cpu_active(cpu, true);
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;
380 error = stop_machine_create();
383 cpu_maps_update_begin();
384 first_cpu = cpumask_first(cpu_online_mask);
385 /* We take down all of the non-boot CPUs in one shot to avoid races
386 * with the userspace trying to use the CPU hotplug at the same time
388 cpumask_clear(frozen_cpus);
390 for_each_online_cpu(cpu) {
391 if (cpu == first_cpu)
393 set_cpu_active(cpu, false);
398 printk("Disabling non-boot CPUs ...\n");
399 for_each_online_cpu(cpu) {
400 if (cpu == first_cpu)
402 error = _cpu_down(cpu, 1);
404 cpumask_set_cpu(cpu, frozen_cpus);
405 printk("CPU%d is down\n", cpu);
407 printk(KERN_ERR "Error taking CPU%d down: %d\n",
414 BUG_ON(num_online_cpus() > 1);
415 /* Make sure the CPUs won't be enabled by someone else */
416 cpu_hotplug_disabled = 1;
418 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
420 cpu_maps_update_done();
421 stop_machine_destroy();
425 void __weak arch_enable_nonboot_cpus_begin(void)
429 void __weak arch_enable_nonboot_cpus_end(void)
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 (cpumask_empty(frozen_cpus))
443 printk("Enabling non-boot CPUs ...\n");
445 arch_enable_nonboot_cpus_begin();
447 for_each_cpu(cpu, frozen_cpus) {
448 error = _cpu_up(cpu, 1);
450 printk("CPU%d is up\n", cpu);
453 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
456 arch_enable_nonboot_cpus_end();
458 cpumask_clear(frozen_cpus);
460 cpu_maps_update_done();
463 static int alloc_frozen_cpus(void)
465 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
469 core_initcall(alloc_frozen_cpus);
470 #endif /* CONFIG_PM_SLEEP_SMP */
473 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
474 * @cpu: cpu that just started
476 * This function calls the cpu_chain notifiers with CPU_STARTING.
477 * It must be called by the arch code on the new cpu, before the new cpu
478 * enables interrupts and before the "boot" cpu returns from __cpu_up().
480 void __cpuinit notify_cpu_starting(unsigned int cpu)
482 unsigned long val = CPU_STARTING;
484 #ifdef CONFIG_PM_SLEEP_SMP
485 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
486 val = CPU_STARTING_FROZEN;
487 #endif /* CONFIG_PM_SLEEP_SMP */
488 raw_notifier_call_chain(&cpu_chain, val, (void *)(long)cpu);
491 #endif /* CONFIG_SMP */
494 * cpu_bit_bitmap[] is a special, "compressed" data structure that
495 * represents all NR_CPUS bits binary values of 1<<nr.
497 * It is used by cpumask_of() to get a constant address to a CPU
498 * mask value that has a single bit set only.
501 /* cpu_bit_bitmap[0] is empty - so we can back into it */
502 #define MASK_DECLARE_1(x) [x+1][0] = 1UL << (x)
503 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
504 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
505 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
507 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
509 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
510 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
511 #if BITS_PER_LONG > 32
512 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
513 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
516 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
518 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
519 EXPORT_SYMBOL(cpu_all_bits);
521 #ifdef CONFIG_INIT_ALL_POSSIBLE
522 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
525 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
527 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
528 EXPORT_SYMBOL(cpu_possible_mask);
530 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
531 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
532 EXPORT_SYMBOL(cpu_online_mask);
534 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
535 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
536 EXPORT_SYMBOL(cpu_present_mask);
538 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
539 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
540 EXPORT_SYMBOL(cpu_active_mask);
542 void set_cpu_possible(unsigned int cpu, bool possible)
545 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
547 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
550 void set_cpu_present(unsigned int cpu, bool present)
553 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
555 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
558 void set_cpu_online(unsigned int cpu, bool online)
561 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
563 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
566 void set_cpu_active(unsigned int cpu, bool active)
569 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
571 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
574 void init_cpu_present(const struct cpumask *src)
576 cpumask_copy(to_cpumask(cpu_present_bits), src);
579 void init_cpu_possible(const struct cpumask *src)
581 cpumask_copy(to_cpumask(cpu_possible_bits), src);
584 void init_cpu_online(const struct cpumask *src)
586 cpumask_copy(to_cpumask(cpu_online_bits), src);