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
);
116 /* cleanup must mirror setup */
117 if (ht
->cleanup
&& td
->status
!= HP_THREAD_NONE
)
118 ht
->cleanup(td
->cpu
, cpu_online(td
->cpu
));
123 if (kthread_should_park()) {
124 __set_current_state(TASK_RUNNING
);
126 if (ht
->park
&& td
->status
== HP_THREAD_ACTIVE
) {
127 BUG_ON(td
->cpu
!= smp_processor_id());
129 td
->status
= HP_THREAD_PARKED
;
132 /* We might have been woken for stop */
136 BUG_ON(td
->cpu
!= smp_processor_id());
138 /* Check for state change setup */
139 switch (td
->status
) {
141 __set_current_state(TASK_RUNNING
);
145 td
->status
= HP_THREAD_ACTIVE
;
148 case HP_THREAD_PARKED
:
149 __set_current_state(TASK_RUNNING
);
153 td
->status
= HP_THREAD_ACTIVE
;
157 if (!ht
->thread_should_run(td
->cpu
)) {
158 preempt_enable_no_resched();
161 __set_current_state(TASK_RUNNING
);
163 ht
->thread_fn(td
->cpu
);
169 __smpboot_create_thread(struct smp_hotplug_thread
*ht
, unsigned int cpu
)
171 struct task_struct
*tsk
= *per_cpu_ptr(ht
->store
, cpu
);
172 struct smpboot_thread_data
*td
;
177 td
= kzalloc_node(sizeof(*td
), GFP_KERNEL
, cpu_to_node(cpu
));
183 tsk
= kthread_create_on_cpu(smpboot_thread_fn
, td
, cpu
,
189 get_task_struct(tsk
);
190 *per_cpu_ptr(ht
->store
, cpu
) = tsk
;
193 * Make sure that the task has actually scheduled out
194 * into park position, before calling the create
195 * callback. At least the migration thread callback
196 * requires that the task is off the runqueue.
198 if (!wait_task_inactive(tsk
, TASK_PARKED
))
206 int smpboot_create_threads(unsigned int cpu
)
208 struct smp_hotplug_thread
*cur
;
211 mutex_lock(&smpboot_threads_lock
);
212 list_for_each_entry(cur
, &hotplug_threads
, list
) {
213 ret
= __smpboot_create_thread(cur
, cpu
);
217 mutex_unlock(&smpboot_threads_lock
);
221 static void smpboot_unpark_thread(struct smp_hotplug_thread
*ht
, unsigned int cpu
)
223 struct task_struct
*tsk
= *per_cpu_ptr(ht
->store
, cpu
);
225 if (!ht
->selfparking
)
229 int 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
);
241 static void smpboot_park_thread(struct smp_hotplug_thread
*ht
, unsigned int cpu
)
243 struct task_struct
*tsk
= *per_cpu_ptr(ht
->store
, cpu
);
245 if (tsk
&& !ht
->selfparking
)
249 int smpboot_park_threads(unsigned int cpu
)
251 struct smp_hotplug_thread
*cur
;
253 mutex_lock(&smpboot_threads_lock
);
254 list_for_each_entry_reverse(cur
, &hotplug_threads
, list
)
255 smpboot_park_thread(cur
, cpu
);
256 mutex_unlock(&smpboot_threads_lock
);
260 static void smpboot_destroy_threads(struct smp_hotplug_thread
*ht
)
264 /* We need to destroy also the parked threads of offline cpus */
265 for_each_possible_cpu(cpu
) {
266 struct task_struct
*tsk
= *per_cpu_ptr(ht
->store
, cpu
);
270 put_task_struct(tsk
);
271 *per_cpu_ptr(ht
->store
, cpu
) = NULL
;
277 * smpboot_register_percpu_thread_cpumask - Register a per_cpu thread related
279 * @plug_thread: Hotplug thread descriptor
280 * @cpumask: The cpumask where threads run
282 * Creates and starts the threads on all online cpus.
284 int smpboot_register_percpu_thread_cpumask(struct smp_hotplug_thread
*plug_thread
,
285 const struct cpumask
*cpumask
)
290 if (!alloc_cpumask_var(&plug_thread
->cpumask
, GFP_KERNEL
))
292 cpumask_copy(plug_thread
->cpumask
, cpumask
);
295 mutex_lock(&smpboot_threads_lock
);
296 for_each_online_cpu(cpu
) {
297 ret
= __smpboot_create_thread(plug_thread
, cpu
);
299 smpboot_destroy_threads(plug_thread
);
300 free_cpumask_var(plug_thread
->cpumask
);
303 if (cpumask_test_cpu(cpu
, cpumask
))
304 smpboot_unpark_thread(plug_thread
, cpu
);
306 list_add(&plug_thread
->list
, &hotplug_threads
);
308 mutex_unlock(&smpboot_threads_lock
);
312 EXPORT_SYMBOL_GPL(smpboot_register_percpu_thread_cpumask
);
315 * smpboot_unregister_percpu_thread - Unregister a per_cpu thread related to hotplug
316 * @plug_thread: Hotplug thread descriptor
318 * Stops all threads on all possible cpus.
320 void smpboot_unregister_percpu_thread(struct smp_hotplug_thread
*plug_thread
)
323 mutex_lock(&smpboot_threads_lock
);
324 list_del(&plug_thread
->list
);
325 smpboot_destroy_threads(plug_thread
);
326 mutex_unlock(&smpboot_threads_lock
);
328 free_cpumask_var(plug_thread
->cpumask
);
330 EXPORT_SYMBOL_GPL(smpboot_unregister_percpu_thread
);
333 * smpboot_update_cpumask_percpu_thread - Adjust which per_cpu hotplug threads stay parked
334 * @plug_thread: Hotplug thread descriptor
335 * @new: Revised mask to use
337 * The cpumask field in the smp_hotplug_thread must not be updated directly
338 * by the client, but only by calling this function.
339 * This function can only be called on a registered smp_hotplug_thread.
341 int smpboot_update_cpumask_percpu_thread(struct smp_hotplug_thread
*plug_thread
,
342 const struct cpumask
*new)
344 struct cpumask
*old
= plug_thread
->cpumask
;
348 if (!alloc_cpumask_var(&tmp
, GFP_KERNEL
))
352 mutex_lock(&smpboot_threads_lock
);
354 /* Park threads that were exclusively enabled on the old mask. */
355 cpumask_andnot(tmp
, old
, new);
356 for_each_cpu_and(cpu
, tmp
, cpu_online_mask
)
357 smpboot_park_thread(plug_thread
, cpu
);
359 /* Unpark threads that are exclusively enabled on the new mask. */
360 cpumask_andnot(tmp
, new, old
);
361 for_each_cpu_and(cpu
, tmp
, cpu_online_mask
)
362 smpboot_unpark_thread(plug_thread
, cpu
);
364 cpumask_copy(old
, new);
366 mutex_unlock(&smpboot_threads_lock
);
369 free_cpumask_var(tmp
);
373 EXPORT_SYMBOL_GPL(smpboot_update_cpumask_percpu_thread
);
375 static DEFINE_PER_CPU(atomic_t
, cpu_hotplug_state
) = ATOMIC_INIT(CPU_POST_DEAD
);
378 * Called to poll specified CPU's state, for example, when waiting for
379 * a CPU to come online.
381 int cpu_report_state(int cpu
)
383 return atomic_read(&per_cpu(cpu_hotplug_state
, cpu
));
387 * If CPU has died properly, set its state to CPU_UP_PREPARE and
388 * return success. Otherwise, return -EBUSY if the CPU died after
389 * cpu_wait_death() timed out. And yet otherwise again, return -EAGAIN
390 * if cpu_wait_death() timed out and the CPU still hasn't gotten around
391 * to dying. In the latter two cases, the CPU might not be set up
392 * properly, but it is up to the arch-specific code to decide.
393 * Finally, -EIO indicates an unanticipated problem.
395 * Note that it is permissible to omit this call entirely, as is
396 * done in architectures that do no CPU-hotplug error checking.
398 int cpu_check_up_prepare(int cpu
)
400 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU
)) {
401 atomic_set(&per_cpu(cpu_hotplug_state
, cpu
), CPU_UP_PREPARE
);
405 switch (atomic_read(&per_cpu(cpu_hotplug_state
, cpu
))) {
409 /* The CPU died properly, so just start it up again. */
410 atomic_set(&per_cpu(cpu_hotplug_state
, cpu
), CPU_UP_PREPARE
);
413 case CPU_DEAD_FROZEN
:
416 * Timeout during CPU death, so let caller know.
417 * The outgoing CPU completed its processing, but after
418 * cpu_wait_death() timed out and reported the error. The
419 * caller is free to proceed, in which case the state
420 * will be reset properly by cpu_set_state_online().
421 * Proceeding despite this -EBUSY return makes sense
422 * for systems where the outgoing CPUs take themselves
423 * offline, with no post-death manipulation required from
431 * The most likely reason we got here is that there was
432 * a timeout during CPU death, and the outgoing CPU never
433 * did complete its processing. This could happen on
434 * a virtualized system if the outgoing VCPU gets preempted
435 * for more than five seconds, and the user attempts to
436 * immediately online that same CPU. Trying again later
437 * might return -EBUSY above, hence -EAGAIN.
443 /* Should not happen. Famous last words. */
449 * Mark the specified CPU online.
451 * Note that it is permissible to omit this call entirely, as is
452 * done in architectures that do no CPU-hotplug error checking.
454 void cpu_set_state_online(int cpu
)
456 (void)atomic_xchg(&per_cpu(cpu_hotplug_state
, cpu
), CPU_ONLINE
);
459 #ifdef CONFIG_HOTPLUG_CPU
462 * Wait for the specified CPU to exit the idle loop and die.
464 bool cpu_wait_death(unsigned int cpu
, int seconds
)
466 int jf_left
= seconds
* HZ
;
473 /* The outgoing CPU will normally get done quite quickly. */
474 if (atomic_read(&per_cpu(cpu_hotplug_state
, cpu
)) == CPU_DEAD
)
478 /* But if the outgoing CPU dawdles, wait increasingly long times. */
479 while (atomic_read(&per_cpu(cpu_hotplug_state
, cpu
)) != CPU_DEAD
) {
480 schedule_timeout_uninterruptible(sleep_jf
);
484 sleep_jf
= DIV_ROUND_UP(sleep_jf
* 11, 10);
487 oldstate
= atomic_read(&per_cpu(cpu_hotplug_state
, cpu
));
488 if (oldstate
== CPU_DEAD
) {
489 /* Outgoing CPU died normally, update state. */
490 smp_mb(); /* atomic_read() before update. */
491 atomic_set(&per_cpu(cpu_hotplug_state
, cpu
), CPU_POST_DEAD
);
493 /* Outgoing CPU still hasn't died, set state accordingly. */
494 if (atomic_cmpxchg(&per_cpu(cpu_hotplug_state
, cpu
),
495 oldstate
, CPU_BROKEN
) != oldstate
)
503 * Called by the outgoing CPU to report its successful death. Return
504 * false if this report follows the surviving CPU's timing out.
506 * A separate "CPU_DEAD_FROZEN" is used when the surviving CPU
507 * timed out. This approach allows architectures to omit calls to
508 * cpu_check_up_prepare() and cpu_set_state_online() without defeating
509 * the next cpu_wait_death()'s polling loop.
511 bool cpu_report_death(void)
515 int cpu
= smp_processor_id();
518 oldstate
= atomic_read(&per_cpu(cpu_hotplug_state
, cpu
));
519 if (oldstate
!= CPU_BROKEN
)
522 newstate
= CPU_DEAD_FROZEN
;
523 } while (atomic_cmpxchg(&per_cpu(cpu_hotplug_state
, cpu
),
524 oldstate
, newstate
) != oldstate
);
525 return newstate
== CPU_DEAD
;
528 #endif /* #ifdef CONFIG_HOTPLUG_CPU */