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
);
37 hotplug_cfd(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
39 long cpu
= (long)hcpu
;
40 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
44 case CPU_UP_PREPARE_FROZEN
:
45 if (!zalloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
47 return notifier_from_errno(-ENOMEM
);
48 cfd
->csd
= alloc_percpu(struct call_single_data
);
50 free_cpumask_var(cfd
->cpumask
);
51 return notifier_from_errno(-ENOMEM
);
55 #ifdef CONFIG_HOTPLUG_CPU
57 case CPU_UP_CANCELED_FROZEN
:
58 /* Fall-through to the CPU_DEAD[_FROZEN] case. */
62 free_cpumask_var(cfd
->cpumask
);
63 free_percpu(cfd
->csd
);
67 case CPU_DYING_FROZEN
:
69 * The IPIs for the smp-call-function callbacks queued by other
70 * CPUs might arrive late, either due to hardware latencies or
71 * because this CPU disabled interrupts (inside stop-machine)
72 * before the IPIs were sent. So flush out any pending callbacks
73 * explicitly (without waiting for the IPIs to arrive), to
74 * ensure that the outgoing CPU doesn't go offline with work
77 flush_smp_call_function_queue(false);
85 static struct notifier_block hotplug_cfd_notifier
= {
86 .notifier_call
= hotplug_cfd
,
89 void __init
call_function_init(void)
91 void *cpu
= (void *)(long)smp_processor_id();
94 for_each_possible_cpu(i
)
95 init_llist_head(&per_cpu(call_single_queue
, i
));
97 hotplug_cfd(&hotplug_cfd_notifier
, CPU_UP_PREPARE
, cpu
);
98 register_cpu_notifier(&hotplug_cfd_notifier
);
102 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
104 * For non-synchronous ipi calls the csd can still be in use by the
105 * previous function call. For multi-cpu calls its even more interesting
106 * as we'll have to ensure no other cpu is observing our csd.
108 static void csd_lock_wait(struct call_single_data
*csd
)
110 while (smp_load_acquire(&csd
->flags
) & CSD_FLAG_LOCK
)
114 static void csd_lock(struct call_single_data
*csd
)
117 csd
->flags
|= CSD_FLAG_LOCK
;
120 * prevent CPU from reordering the above assignment
121 * to ->flags with any subsequent assignments to other
122 * fields of the specified call_single_data structure:
127 static void csd_unlock(struct call_single_data
*csd
)
129 WARN_ON(!(csd
->flags
& CSD_FLAG_LOCK
));
132 * ensure we're all done before releasing data:
134 smp_store_release(&csd
->flags
, 0);
137 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data
, csd_data
);
140 * Insert a previously allocated call_single_data element
141 * for execution on the given CPU. data must already have
142 * ->func, ->info, and ->flags set.
144 static int generic_exec_single(int cpu
, struct call_single_data
*csd
,
145 smp_call_func_t func
, void *info
)
147 if (cpu
== smp_processor_id()) {
151 * We can unlock early even for the synchronous on-stack case,
152 * since we're doing this from the same CPU..
155 local_irq_save(flags
);
157 local_irq_restore(flags
);
162 if ((unsigned)cpu
>= nr_cpu_ids
|| !cpu_online(cpu
)) {
171 * The list addition should be visible before sending the IPI
172 * handler locks the list to pull the entry off it because of
173 * normal cache coherency rules implied by spinlocks.
175 * If IPIs can go out of order to the cache coherency protocol
176 * in an architecture, sufficient synchronisation should be added
177 * to arch code to make it appear to obey cache coherency WRT
178 * locking and barrier primitives. Generic code isn't really
179 * equipped to do the right thing...
181 if (llist_add(&csd
->llist
, &per_cpu(call_single_queue
, cpu
)))
182 arch_send_call_function_single_ipi(cpu
);
188 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
190 * Invoked by arch to handle an IPI for call function single.
191 * Must be called with interrupts disabled.
193 void generic_smp_call_function_single_interrupt(void)
195 flush_smp_call_function_queue(true);
199 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
201 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
202 * offline CPU. Skip this check if set to 'false'.
204 * Flush any pending smp-call-function callbacks queued on this CPU. This is
205 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
206 * to ensure that all pending IPI callbacks are run before it goes completely
209 * Loop through the call_single_queue and run all the queued callbacks.
210 * Must be called with interrupts disabled.
212 static void flush_smp_call_function_queue(bool warn_cpu_offline
)
214 struct llist_head
*head
;
215 struct llist_node
*entry
;
216 struct call_single_data
*csd
, *csd_next
;
219 WARN_ON(!irqs_disabled());
221 head
= this_cpu_ptr(&call_single_queue
);
222 entry
= llist_del_all(head
);
223 entry
= llist_reverse_order(entry
);
225 /* There shouldn't be any pending callbacks on an offline CPU. */
226 if (unlikely(warn_cpu_offline
&& !cpu_online(smp_processor_id()) &&
227 !warned
&& !llist_empty(head
))) {
229 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
232 * We don't have to use the _safe() variant here
233 * because we are not invoking the IPI handlers yet.
235 llist_for_each_entry(csd
, entry
, llist
)
236 pr_warn("IPI callback %pS sent to offline CPU\n",
240 llist_for_each_entry_safe(csd
, csd_next
, entry
, llist
) {
241 smp_call_func_t func
= csd
->func
;
242 void *info
= csd
->info
;
244 /* Do we wait until *after* callback? */
245 if (csd
->flags
& CSD_FLAG_SYNCHRONOUS
) {
255 * Handle irq works queued remotely by irq_work_queue_on().
256 * Smp functions above are typically synchronous so they
257 * better run first since some other CPUs may be busy waiting
264 * smp_call_function_single - Run a function on a specific CPU
265 * @func: The function to run. This must be fast and non-blocking.
266 * @info: An arbitrary pointer to pass to the function.
267 * @wait: If true, wait until function has completed on other CPUs.
269 * Returns 0 on success, else a negative status code.
271 int smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
274 struct call_single_data
*csd
;
275 struct call_single_data csd_stack
= { .flags
= CSD_FLAG_LOCK
| CSD_FLAG_SYNCHRONOUS
};
280 * prevent preemption and reschedule on another processor,
281 * as well as CPU removal
283 this_cpu
= get_cpu();
286 * Can deadlock when called with interrupts disabled.
287 * We allow cpu's that are not yet online though, as no one else can
288 * send smp call function interrupt to this cpu and as such deadlocks
291 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
292 && !oops_in_progress
);
296 csd
= this_cpu_ptr(&csd_data
);
300 err
= generic_exec_single(cpu
, csd
, func
, info
);
309 EXPORT_SYMBOL(smp_call_function_single
);
312 * smp_call_function_single_async(): Run an asynchronous function on a
314 * @cpu: The CPU to run on.
315 * @csd: Pre-allocated and setup data structure
317 * Like smp_call_function_single(), but the call is asynchonous and
318 * can thus be done from contexts with disabled interrupts.
320 * The caller passes his own pre-allocated data structure
321 * (ie: embedded in an object) and is responsible for synchronizing it
322 * such that the IPIs performed on the @csd are strictly serialized.
324 * NOTE: Be careful, there is unfortunately no current debugging facility to
325 * validate the correctness of this serialization.
327 int smp_call_function_single_async(int cpu
, struct call_single_data
*csd
)
333 /* We could deadlock if we have to wait here with interrupts disabled! */
334 if (WARN_ON_ONCE(csd
->flags
& CSD_FLAG_LOCK
))
337 csd
->flags
= CSD_FLAG_LOCK
;
340 err
= generic_exec_single(cpu
, csd
, csd
->func
, csd
->info
);
345 EXPORT_SYMBOL_GPL(smp_call_function_single_async
);
348 * smp_call_function_any - Run a function on any of the given cpus
349 * @mask: The mask of cpus it can run on.
350 * @func: The function to run. This must be fast and non-blocking.
351 * @info: An arbitrary pointer to pass to the function.
352 * @wait: If true, wait until function has completed.
354 * Returns 0 on success, else a negative status code (if no cpus were online).
356 * Selection preference:
357 * 1) current cpu if in @mask
358 * 2) any cpu of current node if in @mask
359 * 3) any other online cpu in @mask
361 int smp_call_function_any(const struct cpumask
*mask
,
362 smp_call_func_t func
, void *info
, int wait
)
365 const struct cpumask
*nodemask
;
368 /* Try for same CPU (cheapest) */
370 if (cpumask_test_cpu(cpu
, mask
))
373 /* Try for same node. */
374 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
375 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
376 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
381 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
382 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
384 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
388 EXPORT_SYMBOL_GPL(smp_call_function_any
);
391 * smp_call_function_many(): Run a function on a set of other CPUs.
392 * @mask: The set of cpus to run on (only runs on online subset).
393 * @func: The function to run. This must be fast and non-blocking.
394 * @info: An arbitrary pointer to pass to the function.
395 * @wait: If true, wait (atomically) until function has completed
398 * If @wait is true, then returns once @func has returned.
400 * You must not call this function with disabled interrupts or from a
401 * hardware interrupt handler or from a bottom half handler. Preemption
402 * must be disabled when calling this function.
404 void smp_call_function_many(const struct cpumask
*mask
,
405 smp_call_func_t func
, void *info
, bool wait
)
407 struct call_function_data
*cfd
;
408 int cpu
, next_cpu
, this_cpu
= smp_processor_id();
411 * Can deadlock when called with interrupts disabled.
412 * We allow cpu's that are not yet online though, as no one else can
413 * send smp call function interrupt to this cpu and as such deadlocks
416 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
417 && !oops_in_progress
&& !early_boot_irqs_disabled
);
419 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
420 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
422 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
424 /* No online cpus? We're done. */
425 if (cpu
>= nr_cpu_ids
)
428 /* Do we have another CPU which isn't us? */
429 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
430 if (next_cpu
== this_cpu
)
431 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
433 /* Fastpath: do that cpu by itself. */
434 if (next_cpu
>= nr_cpu_ids
) {
435 smp_call_function_single(cpu
, func
, info
, wait
);
439 cfd
= this_cpu_ptr(&cfd_data
);
441 cpumask_and(cfd
->cpumask
, mask
, cpu_online_mask
);
442 cpumask_clear_cpu(this_cpu
, cfd
->cpumask
);
444 /* Some callers race with other cpus changing the passed mask */
445 if (unlikely(!cpumask_weight(cfd
->cpumask
)))
448 for_each_cpu(cpu
, cfd
->cpumask
) {
449 struct call_single_data
*csd
= per_cpu_ptr(cfd
->csd
, cpu
);
453 csd
->flags
|= CSD_FLAG_SYNCHRONOUS
;
456 llist_add(&csd
->llist
, &per_cpu(call_single_queue
, cpu
));
459 /* Send a message to all CPUs in the map */
460 arch_send_call_function_ipi_mask(cfd
->cpumask
);
463 for_each_cpu(cpu
, cfd
->cpumask
) {
464 struct call_single_data
*csd
;
466 csd
= per_cpu_ptr(cfd
->csd
, cpu
);
471 EXPORT_SYMBOL(smp_call_function_many
);
474 * smp_call_function(): Run a function on all other CPUs.
475 * @func: The function to run. This must be fast and non-blocking.
476 * @info: An arbitrary pointer to pass to the function.
477 * @wait: If true, wait (atomically) until function has completed
482 * If @wait is true, then returns once @func has returned; otherwise
483 * it returns just before the target cpu calls @func.
485 * You must not call this function with disabled interrupts or from a
486 * hardware interrupt handler or from a bottom half handler.
488 int smp_call_function(smp_call_func_t func
, void *info
, int wait
)
491 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
496 EXPORT_SYMBOL(smp_call_function
);
498 /* Setup configured maximum number of CPUs to activate */
499 unsigned int setup_max_cpus
= NR_CPUS
;
500 EXPORT_SYMBOL(setup_max_cpus
);
504 * Setup routine for controlling SMP activation
506 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
507 * activation entirely (the MPS table probe still happens, though).
509 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
510 * greater than 0, limits the maximum number of CPUs activated in
514 void __weak
arch_disable_smp_support(void) { }
516 static int __init
nosmp(char *str
)
519 arch_disable_smp_support();
524 early_param("nosmp", nosmp
);
526 /* this is hard limit */
527 static int __init
nrcpus(char *str
)
531 get_option(&str
, &nr_cpus
);
532 if (nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
533 nr_cpu_ids
= nr_cpus
;
538 early_param("nr_cpus", nrcpus
);
540 static int __init
maxcpus(char *str
)
542 get_option(&str
, &setup_max_cpus
);
543 if (setup_max_cpus
== 0)
544 arch_disable_smp_support();
549 early_param("maxcpus", maxcpus
);
551 /* Setup number of possible processor ids */
552 int nr_cpu_ids __read_mostly
= NR_CPUS
;
553 EXPORT_SYMBOL(nr_cpu_ids
);
555 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
556 void __init
setup_nr_cpu_ids(void)
558 nr_cpu_ids
= find_last_bit(cpumask_bits(cpu_possible_mask
),NR_CPUS
) + 1;
561 void __weak
smp_announce(void)
563 printk(KERN_INFO
"Brought up %d CPUs\n", num_online_cpus());
566 /* Called by boot processor to activate the rest. */
567 void __init
smp_init(void)
573 /* FIXME: This should be done in userspace --RR */
574 for_each_present_cpu(cpu
) {
575 if (num_online_cpus() >= setup_max_cpus
)
577 if (!cpu_online(cpu
))
581 /* Any cleanup work */
583 smp_cpus_done(setup_max_cpus
);
587 * Call a function on all processors. May be used during early boot while
588 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
589 * of local_irq_disable/enable().
591 int on_each_cpu(void (*func
) (void *info
), void *info
, int wait
)
597 ret
= smp_call_function(func
, info
, wait
);
598 local_irq_save(flags
);
600 local_irq_restore(flags
);
604 EXPORT_SYMBOL(on_each_cpu
);
607 * on_each_cpu_mask(): Run a function on processors specified by
608 * cpumask, which may include the local processor.
609 * @mask: The set of cpus to run on (only runs on online subset).
610 * @func: The function to run. This must be fast and non-blocking.
611 * @info: An arbitrary pointer to pass to the function.
612 * @wait: If true, wait (atomically) until function has completed
615 * If @wait is true, then returns once @func has returned.
617 * You must not call this function with disabled interrupts or from a
618 * hardware interrupt handler or from a bottom half handler. The
619 * exception is that it may be used during early boot while
620 * early_boot_irqs_disabled is set.
622 void on_each_cpu_mask(const struct cpumask
*mask
, smp_call_func_t func
,
623 void *info
, bool wait
)
627 smp_call_function_many(mask
, func
, info
, wait
);
628 if (cpumask_test_cpu(cpu
, mask
)) {
630 local_irq_save(flags
);
632 local_irq_restore(flags
);
636 EXPORT_SYMBOL(on_each_cpu_mask
);
639 * on_each_cpu_cond(): Call a function on each processor for which
640 * the supplied function cond_func returns true, optionally waiting
641 * for all the required CPUs to finish. This may include the local
643 * @cond_func: A callback function that is passed a cpu id and
644 * the the info parameter. The function is called
645 * with preemption disabled. The function should
646 * return a blooean value indicating whether to IPI
648 * @func: The function to run on all applicable CPUs.
649 * This must be fast and non-blocking.
650 * @info: An arbitrary pointer to pass to both functions.
651 * @wait: If true, wait (atomically) until function has
652 * completed on other CPUs.
653 * @gfp_flags: GFP flags to use when allocating the cpumask
654 * used internally by the function.
656 * The function might sleep if the GFP flags indicates a non
657 * atomic allocation is allowed.
659 * Preemption is disabled to protect against CPUs going offline but not online.
660 * CPUs going online during the call will not be seen or sent an IPI.
662 * You must not call this function with disabled interrupts or
663 * from a hardware interrupt handler or from a bottom half handler.
665 void on_each_cpu_cond(bool (*cond_func
)(int cpu
, void *info
),
666 smp_call_func_t func
, void *info
, bool wait
,
672 might_sleep_if(gfpflags_allow_blocking(gfp_flags
));
674 if (likely(zalloc_cpumask_var(&cpus
, (gfp_flags
|__GFP_NOWARN
)))) {
676 for_each_online_cpu(cpu
)
677 if (cond_func(cpu
, info
))
678 cpumask_set_cpu(cpu
, cpus
);
679 on_each_cpu_mask(cpus
, func
, info
, wait
);
681 free_cpumask_var(cpus
);
684 * No free cpumask, bother. No matter, we'll
685 * just have to IPI them one by one.
688 for_each_online_cpu(cpu
)
689 if (cond_func(cpu
, info
)) {
690 ret
= smp_call_function_single(cpu
, func
,
697 EXPORT_SYMBOL(on_each_cpu_cond
);
699 static void do_nothing(void *unused
)
704 * kick_all_cpus_sync - Force all cpus out of idle
706 * Used to synchronize the update of pm_idle function pointer. It's
707 * called after the pointer is updated and returns after the dummy
708 * callback function has been executed on all cpus. The execution of
709 * the function can only happen on the remote cpus after they have
710 * left the idle function which had been called via pm_idle function
711 * pointer. So it's guaranteed that nothing uses the previous pointer
714 void kick_all_cpus_sync(void)
716 /* Make sure the change is visible before we kick the cpus */
718 smp_call_function(do_nothing
, NULL
, 1);
720 EXPORT_SYMBOL_GPL(kick_all_cpus_sync
);
723 * wake_up_all_idle_cpus - break all cpus out of idle
724 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
725 * including idle polling cpus, for non-idle cpus, we will do nothing
728 void wake_up_all_idle_cpus(void)
733 for_each_online_cpu(cpu
) {
734 if (cpu
== smp_processor_id())
737 wake_up_if_idle(cpu
);
741 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus
);