2 * Generic helpers for smp ipi calls
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
6 #include <linux/irq_work.h>
7 #include <linux/rcupdate.h>
8 #include <linux/rculist.h>
9 #include <linux/kernel.h>
10 #include <linux/export.h>
11 #include <linux/percpu.h>
12 #include <linux/init.h>
13 #include <linux/gfp.h>
14 #include <linux/smp.h>
15 #include <linux/cpu.h>
16 #include <linux/sched.h>
22 CSD_FLAG_SYNCHRONOUS
= 0x02,
25 struct call_function_data
{
26 struct call_single_data __percpu
*csd
;
27 cpumask_var_t cpumask
;
30 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data
, cfd_data
);
32 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head
, call_single_queue
);
34 static void flush_smp_call_function_queue(bool warn_cpu_offline
);
36 int smpcfd_prepare_cpu(unsigned int cpu
)
38 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
40 if (!zalloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
43 cfd
->csd
= alloc_percpu(struct call_single_data
);
45 free_cpumask_var(cfd
->cpumask
);
52 int smpcfd_dead_cpu(unsigned int cpu
)
54 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
56 free_cpumask_var(cfd
->cpumask
);
57 free_percpu(cfd
->csd
);
61 int smpcfd_dying_cpu(unsigned int cpu
)
64 * The IPIs for the smp-call-function callbacks queued by other
65 * CPUs might arrive late, either due to hardware latencies or
66 * because this CPU disabled interrupts (inside stop-machine)
67 * before the IPIs were sent. So flush out any pending callbacks
68 * explicitly (without waiting for the IPIs to arrive), to
69 * ensure that the outgoing CPU doesn't go offline with work
72 flush_smp_call_function_queue(false);
76 void __init
call_function_init(void)
80 for_each_possible_cpu(i
)
81 init_llist_head(&per_cpu(call_single_queue
, i
));
83 smpcfd_prepare_cpu(smp_processor_id());
87 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
89 * For non-synchronous ipi calls the csd can still be in use by the
90 * previous function call. For multi-cpu calls its even more interesting
91 * as we'll have to ensure no other cpu is observing our csd.
93 static __always_inline
void csd_lock_wait(struct call_single_data
*csd
)
95 smp_cond_load_acquire(&csd
->flags
, !(VAL
& CSD_FLAG_LOCK
));
98 static __always_inline
void csd_lock(struct call_single_data
*csd
)
101 csd
->flags
|= CSD_FLAG_LOCK
;
104 * prevent CPU from reordering the above assignment
105 * to ->flags with any subsequent assignments to other
106 * fields of the specified call_single_data structure:
111 static __always_inline
void csd_unlock(struct call_single_data
*csd
)
113 WARN_ON(!(csd
->flags
& CSD_FLAG_LOCK
));
116 * ensure we're all done before releasing data:
118 smp_store_release(&csd
->flags
, 0);
121 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data
, csd_data
);
124 * Insert a previously allocated call_single_data element
125 * for execution on the given CPU. data must already have
126 * ->func, ->info, and ->flags set.
128 static int generic_exec_single(int cpu
, struct call_single_data
*csd
,
129 smp_call_func_t func
, void *info
)
131 if (cpu
== smp_processor_id()) {
135 * We can unlock early even for the synchronous on-stack case,
136 * since we're doing this from the same CPU..
139 local_irq_save(flags
);
141 local_irq_restore(flags
);
146 if ((unsigned)cpu
>= nr_cpu_ids
|| !cpu_online(cpu
)) {
155 * The list addition should be visible before sending the IPI
156 * handler locks the list to pull the entry off it because of
157 * normal cache coherency rules implied by spinlocks.
159 * If IPIs can go out of order to the cache coherency protocol
160 * in an architecture, sufficient synchronisation should be added
161 * to arch code to make it appear to obey cache coherency WRT
162 * locking and barrier primitives. Generic code isn't really
163 * equipped to do the right thing...
165 if (llist_add(&csd
->llist
, &per_cpu(call_single_queue
, cpu
)))
166 arch_send_call_function_single_ipi(cpu
);
172 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
174 * Invoked by arch to handle an IPI for call function single.
175 * Must be called with interrupts disabled.
177 void generic_smp_call_function_single_interrupt(void)
179 flush_smp_call_function_queue(true);
183 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
185 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
186 * offline CPU. Skip this check if set to 'false'.
188 * Flush any pending smp-call-function callbacks queued on this CPU. This is
189 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
190 * to ensure that all pending IPI callbacks are run before it goes completely
193 * Loop through the call_single_queue and run all the queued callbacks.
194 * Must be called with interrupts disabled.
196 static void flush_smp_call_function_queue(bool warn_cpu_offline
)
198 struct llist_head
*head
;
199 struct llist_node
*entry
;
200 struct call_single_data
*csd
, *csd_next
;
203 WARN_ON(!irqs_disabled());
205 head
= this_cpu_ptr(&call_single_queue
);
206 entry
= llist_del_all(head
);
207 entry
= llist_reverse_order(entry
);
209 /* There shouldn't be any pending callbacks on an offline CPU. */
210 if (unlikely(warn_cpu_offline
&& !cpu_online(smp_processor_id()) &&
211 !warned
&& !llist_empty(head
))) {
213 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
216 * We don't have to use the _safe() variant here
217 * because we are not invoking the IPI handlers yet.
219 llist_for_each_entry(csd
, entry
, llist
)
220 pr_warn("IPI callback %pS sent to offline CPU\n",
224 llist_for_each_entry_safe(csd
, csd_next
, entry
, llist
) {
225 smp_call_func_t func
= csd
->func
;
226 void *info
= csd
->info
;
228 /* Do we wait until *after* callback? */
229 if (csd
->flags
& CSD_FLAG_SYNCHRONOUS
) {
239 * Handle irq works queued remotely by irq_work_queue_on().
240 * Smp functions above are typically synchronous so they
241 * better run first since some other CPUs may be busy waiting
248 * smp_call_function_single - Run a function on a specific CPU
249 * @func: The function to run. This must be fast and non-blocking.
250 * @info: An arbitrary pointer to pass to the function.
251 * @wait: If true, wait until function has completed on other CPUs.
253 * Returns 0 on success, else a negative status code.
255 int smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
258 struct call_single_data
*csd
;
259 struct call_single_data csd_stack
= { .flags
= CSD_FLAG_LOCK
| CSD_FLAG_SYNCHRONOUS
};
264 * prevent preemption and reschedule on another processor,
265 * as well as CPU removal
267 this_cpu
= get_cpu();
270 * Can deadlock when called with interrupts disabled.
271 * We allow cpu's that are not yet online though, as no one else can
272 * send smp call function interrupt to this cpu and as such deadlocks
275 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
276 && !oops_in_progress
);
280 csd
= this_cpu_ptr(&csd_data
);
284 err
= generic_exec_single(cpu
, csd
, func
, info
);
293 EXPORT_SYMBOL(smp_call_function_single
);
296 * smp_call_function_single_async(): Run an asynchronous function on a
298 * @cpu: The CPU to run on.
299 * @csd: Pre-allocated and setup data structure
301 * Like smp_call_function_single(), but the call is asynchonous and
302 * can thus be done from contexts with disabled interrupts.
304 * The caller passes his own pre-allocated data structure
305 * (ie: embedded in an object) and is responsible for synchronizing it
306 * such that the IPIs performed on the @csd are strictly serialized.
308 * NOTE: Be careful, there is unfortunately no current debugging facility to
309 * validate the correctness of this serialization.
311 int smp_call_function_single_async(int cpu
, struct call_single_data
*csd
)
317 /* We could deadlock if we have to wait here with interrupts disabled! */
318 if (WARN_ON_ONCE(csd
->flags
& CSD_FLAG_LOCK
))
321 csd
->flags
= CSD_FLAG_LOCK
;
324 err
= generic_exec_single(cpu
, csd
, csd
->func
, csd
->info
);
329 EXPORT_SYMBOL_GPL(smp_call_function_single_async
);
332 * smp_call_function_any - Run a function on any of the given cpus
333 * @mask: The mask of cpus it can run on.
334 * @func: The function to run. This must be fast and non-blocking.
335 * @info: An arbitrary pointer to pass to the function.
336 * @wait: If true, wait until function has completed.
338 * Returns 0 on success, else a negative status code (if no cpus were online).
340 * Selection preference:
341 * 1) current cpu if in @mask
342 * 2) any cpu of current node if in @mask
343 * 3) any other online cpu in @mask
345 int smp_call_function_any(const struct cpumask
*mask
,
346 smp_call_func_t func
, void *info
, int wait
)
349 const struct cpumask
*nodemask
;
352 /* Try for same CPU (cheapest) */
354 if (cpumask_test_cpu(cpu
, mask
))
357 /* Try for same node. */
358 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
359 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
360 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
365 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
366 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
368 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
372 EXPORT_SYMBOL_GPL(smp_call_function_any
);
375 * smp_call_function_many(): Run a function on a set of other CPUs.
376 * @mask: The set of cpus to run on (only runs on online subset).
377 * @func: The function to run. This must be fast and non-blocking.
378 * @info: An arbitrary pointer to pass to the function.
379 * @wait: If true, wait (atomically) until function has completed
382 * If @wait is true, then returns once @func has returned.
384 * You must not call this function with disabled interrupts or from a
385 * hardware interrupt handler or from a bottom half handler. Preemption
386 * must be disabled when calling this function.
388 void smp_call_function_many(const struct cpumask
*mask
,
389 smp_call_func_t func
, void *info
, bool wait
)
391 struct call_function_data
*cfd
;
392 int cpu
, next_cpu
, this_cpu
= smp_processor_id();
395 * Can deadlock when called with interrupts disabled.
396 * We allow cpu's that are not yet online though, as no one else can
397 * send smp call function interrupt to this cpu and as such deadlocks
400 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
401 && !oops_in_progress
&& !early_boot_irqs_disabled
);
403 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
404 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
406 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
408 /* No online cpus? We're done. */
409 if (cpu
>= nr_cpu_ids
)
412 /* Do we have another CPU which isn't us? */
413 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
414 if (next_cpu
== this_cpu
)
415 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
417 /* Fastpath: do that cpu by itself. */
418 if (next_cpu
>= nr_cpu_ids
) {
419 smp_call_function_single(cpu
, func
, info
, wait
);
423 cfd
= this_cpu_ptr(&cfd_data
);
425 cpumask_and(cfd
->cpumask
, mask
, cpu_online_mask
);
426 cpumask_clear_cpu(this_cpu
, cfd
->cpumask
);
428 /* Some callers race with other cpus changing the passed mask */
429 if (unlikely(!cpumask_weight(cfd
->cpumask
)))
432 for_each_cpu(cpu
, cfd
->cpumask
) {
433 struct call_single_data
*csd
= per_cpu_ptr(cfd
->csd
, cpu
);
437 csd
->flags
|= CSD_FLAG_SYNCHRONOUS
;
440 llist_add(&csd
->llist
, &per_cpu(call_single_queue
, cpu
));
443 /* Send a message to all CPUs in the map */
444 arch_send_call_function_ipi_mask(cfd
->cpumask
);
447 for_each_cpu(cpu
, cfd
->cpumask
) {
448 struct call_single_data
*csd
;
450 csd
= per_cpu_ptr(cfd
->csd
, cpu
);
455 EXPORT_SYMBOL(smp_call_function_many
);
458 * smp_call_function(): Run a function on all other CPUs.
459 * @func: The function to run. This must be fast and non-blocking.
460 * @info: An arbitrary pointer to pass to the function.
461 * @wait: If true, wait (atomically) until function has completed
466 * If @wait is true, then returns once @func has returned; otherwise
467 * it returns just before the target cpu calls @func.
469 * You must not call this function with disabled interrupts or from a
470 * hardware interrupt handler or from a bottom half handler.
472 int smp_call_function(smp_call_func_t func
, void *info
, int wait
)
475 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
480 EXPORT_SYMBOL(smp_call_function
);
482 /* Setup configured maximum number of CPUs to activate */
483 unsigned int setup_max_cpus
= NR_CPUS
;
484 EXPORT_SYMBOL(setup_max_cpus
);
488 * Setup routine for controlling SMP activation
490 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
491 * activation entirely (the MPS table probe still happens, though).
493 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
494 * greater than 0, limits the maximum number of CPUs activated in
498 void __weak
arch_disable_smp_support(void) { }
500 static int __init
nosmp(char *str
)
503 arch_disable_smp_support();
508 early_param("nosmp", nosmp
);
510 /* this is hard limit */
511 static int __init
nrcpus(char *str
)
515 get_option(&str
, &nr_cpus
);
516 if (nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
517 nr_cpu_ids
= nr_cpus
;
522 early_param("nr_cpus", nrcpus
);
524 static int __init
maxcpus(char *str
)
526 get_option(&str
, &setup_max_cpus
);
527 if (setup_max_cpus
== 0)
528 arch_disable_smp_support();
533 early_param("maxcpus", maxcpus
);
535 /* Setup number of possible processor ids */
536 int nr_cpu_ids __read_mostly
= NR_CPUS
;
537 EXPORT_SYMBOL(nr_cpu_ids
);
539 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
540 void __init
setup_nr_cpu_ids(void)
542 nr_cpu_ids
= find_last_bit(cpumask_bits(cpu_possible_mask
),NR_CPUS
) + 1;
545 void __weak
smp_announce(void)
547 printk(KERN_INFO
"Brought up %d CPUs\n", num_online_cpus());
550 /* Called by boot processor to activate the rest. */
551 void __init
smp_init(void)
556 cpuhp_threads_init();
558 /* FIXME: This should be done in userspace --RR */
559 for_each_present_cpu(cpu
) {
560 if (num_online_cpus() >= setup_max_cpus
)
562 if (!cpu_online(cpu
))
566 /* Any cleanup work */
568 smp_cpus_done(setup_max_cpus
);
572 * Call a function on all processors. May be used during early boot while
573 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
574 * of local_irq_disable/enable().
576 int on_each_cpu(void (*func
) (void *info
), void *info
, int wait
)
582 ret
= smp_call_function(func
, info
, wait
);
583 local_irq_save(flags
);
585 local_irq_restore(flags
);
589 EXPORT_SYMBOL(on_each_cpu
);
592 * on_each_cpu_mask(): Run a function on processors specified by
593 * cpumask, which may include the local processor.
594 * @mask: The set of cpus to run on (only runs on online subset).
595 * @func: The function to run. This must be fast and non-blocking.
596 * @info: An arbitrary pointer to pass to the function.
597 * @wait: If true, wait (atomically) until function has completed
600 * If @wait is true, then returns once @func has returned.
602 * You must not call this function with disabled interrupts or from a
603 * hardware interrupt handler or from a bottom half handler. The
604 * exception is that it may be used during early boot while
605 * early_boot_irqs_disabled is set.
607 void on_each_cpu_mask(const struct cpumask
*mask
, smp_call_func_t func
,
608 void *info
, bool wait
)
612 smp_call_function_many(mask
, func
, info
, wait
);
613 if (cpumask_test_cpu(cpu
, mask
)) {
615 local_irq_save(flags
);
617 local_irq_restore(flags
);
621 EXPORT_SYMBOL(on_each_cpu_mask
);
624 * on_each_cpu_cond(): Call a function on each processor for which
625 * the supplied function cond_func returns true, optionally waiting
626 * for all the required CPUs to finish. This may include the local
628 * @cond_func: A callback function that is passed a cpu id and
629 * the the info parameter. The function is called
630 * with preemption disabled. The function should
631 * return a blooean value indicating whether to IPI
633 * @func: The function to run on all applicable CPUs.
634 * This must be fast and non-blocking.
635 * @info: An arbitrary pointer to pass to both functions.
636 * @wait: If true, wait (atomically) until function has
637 * completed on other CPUs.
638 * @gfp_flags: GFP flags to use when allocating the cpumask
639 * used internally by the function.
641 * The function might sleep if the GFP flags indicates a non
642 * atomic allocation is allowed.
644 * Preemption is disabled to protect against CPUs going offline but not online.
645 * CPUs going online during the call will not be seen or sent an IPI.
647 * You must not call this function with disabled interrupts or
648 * from a hardware interrupt handler or from a bottom half handler.
650 void on_each_cpu_cond(bool (*cond_func
)(int cpu
, void *info
),
651 smp_call_func_t func
, void *info
, bool wait
,
657 might_sleep_if(gfpflags_allow_blocking(gfp_flags
));
659 if (likely(zalloc_cpumask_var(&cpus
, (gfp_flags
|__GFP_NOWARN
)))) {
661 for_each_online_cpu(cpu
)
662 if (cond_func(cpu
, info
))
663 cpumask_set_cpu(cpu
, cpus
);
664 on_each_cpu_mask(cpus
, func
, info
, wait
);
666 free_cpumask_var(cpus
);
669 * No free cpumask, bother. No matter, we'll
670 * just have to IPI them one by one.
673 for_each_online_cpu(cpu
)
674 if (cond_func(cpu
, info
)) {
675 ret
= smp_call_function_single(cpu
, func
,
682 EXPORT_SYMBOL(on_each_cpu_cond
);
684 static void do_nothing(void *unused
)
689 * kick_all_cpus_sync - Force all cpus out of idle
691 * Used to synchronize the update of pm_idle function pointer. It's
692 * called after the pointer is updated and returns after the dummy
693 * callback function has been executed on all cpus. The execution of
694 * the function can only happen on the remote cpus after they have
695 * left the idle function which had been called via pm_idle function
696 * pointer. So it's guaranteed that nothing uses the previous pointer
699 void kick_all_cpus_sync(void)
701 /* Make sure the change is visible before we kick the cpus */
703 smp_call_function(do_nothing
, NULL
, 1);
705 EXPORT_SYMBOL_GPL(kick_all_cpus_sync
);
708 * wake_up_all_idle_cpus - break all cpus out of idle
709 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
710 * including idle polling cpus, for non-idle cpus, we will do nothing
713 void wake_up_all_idle_cpus(void)
718 for_each_online_cpu(cpu
) {
719 if (cpu
== smp_processor_id())
722 wake_up_if_idle(cpu
);
726 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus
);