2 * Common SMP CPU bringup/teardown functions
7 #include <linux/delay.h>
8 #include <linux/init.h>
9 #include <linux/list.h>
10 #include <linux/slab.h>
11 #include <linux/sched.h>
12 #include <linux/export.h>
13 #include <linux/percpu.h>
14 #include <linux/kthread.h>
15 #include <linux/smpboot.h>
21 #ifdef CONFIG_GENERIC_SMP_IDLE_THREAD
23 * For the hotplug case we keep the task structs around and reuse
26 static DEFINE_PER_CPU(struct task_struct *, idle_threads);
28 struct task_struct *idle_thread_get(unsigned int cpu)
30 struct task_struct *tsk = per_cpu(idle_threads, cpu);
33 return ERR_PTR(-ENOMEM);
38 void __init idle_thread_set_boot_cpu(void)
40 per_cpu(idle_threads, smp_processor_id()) = current;
44 * idle_init - Initialize the idle thread for a cpu
45 * @cpu: The cpu for which the idle thread should be initialized
47 * Creates the thread if it does not exist.
49 static inline void idle_init(unsigned int cpu)
51 struct task_struct *tsk = per_cpu(idle_threads, cpu);
56 pr_err("SMP: fork_idle() failed for CPU %u\n", cpu);
58 per_cpu(idle_threads, cpu) = tsk;
63 * idle_threads_init - Initialize idle threads for all cpus
65 void __init idle_threads_init(void)
67 unsigned int cpu, boot_cpu;
69 boot_cpu = smp_processor_id();
71 for_each_possible_cpu(cpu) {
78 #endif /* #ifdef CONFIG_SMP */
80 static LIST_HEAD(hotplug_threads);
81 static DEFINE_MUTEX(smpboot_threads_lock);
83 struct smpboot_thread_data {
86 struct smp_hotplug_thread *ht;
96 * smpboot_thread_fn - percpu hotplug thread loop function
97 * @data: thread data pointer
99 * Checks for thread stop and park conditions. Calls the necessary
100 * setup, cleanup, park and unpark functions for the registered
103 * Returns 1 when the thread should exit, 0 otherwise.
105 static int smpboot_thread_fn(void *data)
107 struct smpboot_thread_data *td = data;
108 struct smp_hotplug_thread *ht = td->ht;
111 set_current_state(TASK_INTERRUPTIBLE);
113 if (kthread_should_stop()) {
114 __set_current_state(TASK_RUNNING);
117 ht->cleanup(td->cpu, cpu_online(td->cpu));
122 if (kthread_should_park()) {
123 __set_current_state(TASK_RUNNING);
125 if (ht->park && td->status == HP_THREAD_ACTIVE) {
126 BUG_ON(td->cpu != smp_processor_id());
128 td->status = HP_THREAD_PARKED;
131 /* We might have been woken for stop */
135 BUG_ON(td->cpu != smp_processor_id());
137 /* Check for state change setup */
138 switch (td->status) {
140 __set_current_state(TASK_RUNNING);
144 td->status = HP_THREAD_ACTIVE;
147 case HP_THREAD_PARKED:
148 __set_current_state(TASK_RUNNING);
152 td->status = HP_THREAD_ACTIVE;
156 if (!ht->thread_should_run(td->cpu)) {
157 preempt_enable_no_resched();
160 __set_current_state(TASK_RUNNING);
162 ht->thread_fn(td->cpu);
168 __smpboot_create_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
170 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
171 struct smpboot_thread_data *td;
176 td = kzalloc_node(sizeof(*td), GFP_KERNEL, cpu_to_node(cpu));
182 tsk = kthread_create_on_cpu(smpboot_thread_fn, td, cpu,
188 get_task_struct(tsk);
189 *per_cpu_ptr(ht->store, cpu) = tsk;
192 * Make sure that the task has actually scheduled out
193 * into park position, before calling the create
194 * callback. At least the migration thread callback
195 * requires that the task is off the runqueue.
197 if (!wait_task_inactive(tsk, TASK_PARKED))
205 int smpboot_create_threads(unsigned int cpu)
207 struct smp_hotplug_thread *cur;
210 mutex_lock(&smpboot_threads_lock);
211 list_for_each_entry(cur, &hotplug_threads, list) {
212 ret = __smpboot_create_thread(cur, cpu);
216 mutex_unlock(&smpboot_threads_lock);
220 static void smpboot_unpark_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
222 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
229 void smpboot_unpark_threads(unsigned int cpu)
231 struct smp_hotplug_thread *cur;
233 mutex_lock(&smpboot_threads_lock);
234 list_for_each_entry(cur, &hotplug_threads, list)
235 if (cpumask_test_cpu(cpu, cur->cpumask))
236 smpboot_unpark_thread(cur, cpu);
237 mutex_unlock(&smpboot_threads_lock);
240 static void smpboot_park_thread(struct smp_hotplug_thread *ht, unsigned int cpu)
242 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
244 if (tsk && !ht->selfparking)
248 void smpboot_park_threads(unsigned int cpu)
250 struct smp_hotplug_thread *cur;
252 mutex_lock(&smpboot_threads_lock);
253 list_for_each_entry_reverse(cur, &hotplug_threads, list)
254 smpboot_park_thread(cur, cpu);
255 mutex_unlock(&smpboot_threads_lock);
258 static void smpboot_destroy_threads(struct smp_hotplug_thread *ht)
262 /* Unpark any threads that were voluntarily parked. */
263 for_each_cpu_not(cpu, ht->cpumask) {
264 if (cpu_online(cpu)) {
265 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
271 /* We need to destroy also the parked threads of offline cpus */
272 for_each_possible_cpu(cpu) {
273 struct task_struct *tsk = *per_cpu_ptr(ht->store, cpu);
277 put_task_struct(tsk);
278 *per_cpu_ptr(ht->store, cpu) = NULL;
284 * smpboot_register_percpu_thread - Register a per_cpu thread related to hotplug
285 * @plug_thread: Hotplug thread descriptor
287 * Creates and starts the threads on all online cpus.
289 int smpboot_register_percpu_thread(struct smp_hotplug_thread *plug_thread)
294 if (!alloc_cpumask_var(&plug_thread->cpumask, GFP_KERNEL))
296 cpumask_copy(plug_thread->cpumask, cpu_possible_mask);
299 mutex_lock(&smpboot_threads_lock);
300 for_each_online_cpu(cpu) {
301 ret = __smpboot_create_thread(plug_thread, cpu);
303 smpboot_destroy_threads(plug_thread);
306 smpboot_unpark_thread(plug_thread, cpu);
308 list_add(&plug_thread->list, &hotplug_threads);
310 mutex_unlock(&smpboot_threads_lock);
314 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread);
317 * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
318 * @plug_thread: Hotplug thread descriptor
320 * Stops all threads on all possible cpus.
322 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread *plug_thread)
325 mutex_lock(&smpboot_threads_lock);
326 list_del(&plug_thread->list);
327 smpboot_destroy_threads(plug_thread);
328 mutex_unlock(&smpboot_threads_lock);
330 free_cpumask_var(plug_thread->cpumask);
332 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread);
335 * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
336 * @plug_thread: Hotplug thread descriptor
337 * @new: Revised mask to use
339 * The cpumask field in the smp_hotplug_thread must not be updated directly
340 * by the client, but only by calling this function.
342 int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread *plug_thread,
343 const struct cpumask *new)
345 struct cpumask *old = plug_thread->cpumask;
349 if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
353 mutex_lock(&smpboot_threads_lock);
355 /* Park threads that were exclusively enabled on the old mask. */
356 cpumask_andnot(tmp, old, new);
357 for_each_cpu_and(cpu, tmp, cpu_online_mask)
358 smpboot_park_thread(plug_thread, cpu);
360 /* Unpark threads that are exclusively enabled on the new mask. */
361 cpumask_andnot(tmp, new, old);
362 for_each_cpu_and(cpu, tmp, cpu_online_mask)
363 smpboot_unpark_thread(plug_thread, cpu);
365 cpumask_copy(old, new);
367 mutex_unlock(&smpboot_threads_lock);
370 free_cpumask_var(tmp);
374 EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread);
376 static DEFINE_PER_CPU(atomic_t, cpu_hotplug_state) = ATOMIC_INIT(CPU_POST_DEAD);
379 * Called to poll specified CPU's state, for example, when waiting for
380 * a CPU to come online.
382 int cpu_report_state(int cpu)
384 return atomic_read(&per_cpu(cpu_hotplug_state, cpu));
388 * If CPU has died properly, set its state to CPU_UP_PREPARE and
389 * return success. Otherwise, return -EBUSY if the CPU died after
390 * cpu_wait_death() timed out. And yet otherwise again, return -EAGAIN
391 * if cpu_wait_death() timed out and the CPU still hasn't gotten around
392 * to dying. In the latter two cases, the CPU might not be set up
393 * properly, but it is up to the arch-specific code to decide.
394 * Finally, -EIO indicates an unanticipated problem.
396 * Note that it is permissible to omit this call entirely, as is
397 * done in architectures that do no CPU-hotplug error checking.
399 int cpu_check_up_prepare(int cpu)
401 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU)) {
402 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
406 switch (atomic_read(&per_cpu(cpu_hotplug_state, cpu))) {
410 /* The CPU died properly, so just start it up again. */
411 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_UP_PREPARE);
414 case CPU_DEAD_FROZEN:
417 * Timeout during CPU death, so let caller know.
418 * The outgoing CPU completed its processing, but after
419 * cpu_wait_death() timed out and reported the error. The
420 * caller is free to proceed, in which case the state
421 * will be reset properly by cpu_set_state_online().
422 * Proceeding despite this -EBUSY return makes sense
423 * for systems where the outgoing CPUs take themselves
424 * offline, with no post-death manipulation required from
432 * The most likely reason we got here is that there was
433 * a timeout during CPU death, and the outgoing CPU never
434 * did complete its processing. This could happen on
435 * a virtualized system if the outgoing VCPU gets preempted
436 * for more than five seconds, and the user attempts to
437 * immediately online that same CPU. Trying again later
438 * might return -EBUSY above, hence -EAGAIN.
444 /* Should not happen. Famous last words. */
450 * Mark the specified CPU online.
452 * Note that it is permissible to omit this call entirely, as is
453 * done in architectures that do no CPU-hotplug error checking.
455 void cpu_set_state_online(int cpu)
457 (void)atomic_xchg(&per_cpu(cpu_hotplug_state, cpu), CPU_ONLINE);
460 #ifdef CONFIG_HOTPLUG_CPU
463 * Wait for the specified CPU to exit the idle loop and die.
465 bool cpu_wait_death(unsigned int cpu, int seconds)
467 int jf_left = seconds * HZ;
474 /* The outgoing CPU will normally get done quite quickly. */
475 if (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) == CPU_DEAD)
479 /* But if the outgoing CPU dawdles, wait increasingly long times. */
480 while (atomic_read(&per_cpu(cpu_hotplug_state, cpu)) != CPU_DEAD) {
481 schedule_timeout_uninterruptible(sleep_jf);
485 sleep_jf = DIV_ROUND_UP(sleep_jf * 11, 10);
488 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
489 if (oldstate == CPU_DEAD) {
490 /* Outgoing CPU died normally, update state. */
491 smp_mb(); /* atomic_read() before update. */
492 atomic_set(&per_cpu(cpu_hotplug_state, cpu), CPU_POST_DEAD);
494 /* Outgoing CPU still hasn't died, set state accordingly. */
495 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
496 oldstate, CPU_BROKEN) != oldstate)
504 * Called by the outgoing CPU to report its successful death. Return
505 * false if this report follows the surviving CPU's timing out.
507 * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
508 * timed out. This approach allows architectures to omit calls to
509 * cpu_check_up_prepare() and cpu_set_state_online() without defeating
510 * the next cpu_wait_death()'s polling loop.
512 bool cpu_report_death(void)
516 int cpu = smp_processor_id();
519 oldstate = atomic_read(&per_cpu(cpu_hotplug_state, cpu));
520 if (oldstate != CPU_BROKEN)
523 newstate = CPU_DEAD_FROZEN;
524 } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state, cpu),
525 oldstate, newstate) != oldstate);
526 return newstate == CPU_DEAD;
529 #endif /* #ifdef CONFIG_HOTPLUG_CPU */