2 * Generic helpers for smp ipi calls
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
6 #include <linux/rcupdate.h>
7 #include <linux/rculist.h>
8 #include <linux/kernel.h>
9 #include <linux/export.h>
10 #include <linux/percpu.h>
11 #include <linux/init.h>
12 #include <linux/gfp.h>
13 #include <linux/smp.h>
14 #include <linux/cpu.h>
18 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
23 struct call_function_data
{
24 struct call_single_data __percpu
*csd
;
25 cpumask_var_t cpumask
;
26 cpumask_var_t cpumask_ipi
;
29 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data
, cfd_data
);
31 struct call_single_queue
{
32 struct list_head list
;
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue
, call_single_queue
);
39 hotplug_cfd(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
41 long cpu
= (long)hcpu
;
42 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
46 case CPU_UP_PREPARE_FROZEN
:
47 if (!zalloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
49 return notifier_from_errno(-ENOMEM
);
50 if (!zalloc_cpumask_var_node(&cfd
->cpumask_ipi
, GFP_KERNEL
,
52 return notifier_from_errno(-ENOMEM
);
53 cfd
->csd
= alloc_percpu(struct call_single_data
);
55 free_cpumask_var(cfd
->cpumask
);
56 return notifier_from_errno(-ENOMEM
);
60 #ifdef CONFIG_HOTPLUG_CPU
62 case CPU_UP_CANCELED_FROZEN
:
66 free_cpumask_var(cfd
->cpumask
);
67 free_cpumask_var(cfd
->cpumask_ipi
);
68 free_percpu(cfd
->csd
);
76 static struct notifier_block __cpuinitdata hotplug_cfd_notifier
= {
77 .notifier_call
= hotplug_cfd
,
80 void __init
call_function_init(void)
82 void *cpu
= (void *)(long)smp_processor_id();
85 for_each_possible_cpu(i
) {
86 struct call_single_queue
*q
= &per_cpu(call_single_queue
, i
);
88 raw_spin_lock_init(&q
->lock
);
89 INIT_LIST_HEAD(&q
->list
);
92 hotplug_cfd(&hotplug_cfd_notifier
, CPU_UP_PREPARE
, cpu
);
93 register_cpu_notifier(&hotplug_cfd_notifier
);
97 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
99 * For non-synchronous ipi calls the csd can still be in use by the
100 * previous function call. For multi-cpu calls its even more interesting
101 * as we'll have to ensure no other cpu is observing our csd.
103 static void csd_lock_wait(struct call_single_data
*csd
)
105 while (csd
->flags
& CSD_FLAG_LOCK
)
109 static void csd_lock(struct call_single_data
*csd
)
112 csd
->flags
|= CSD_FLAG_LOCK
;
115 * prevent CPU from reordering the above assignment
116 * to ->flags with any subsequent assignments to other
117 * fields of the specified call_single_data structure:
122 static void csd_unlock(struct call_single_data
*csd
)
124 WARN_ON(!(csd
->flags
& CSD_FLAG_LOCK
));
127 * ensure we're all done before releasing data:
131 csd
->flags
&= ~CSD_FLAG_LOCK
;
135 * Insert a previously allocated call_single_data element
136 * for execution on the given CPU. data must already have
137 * ->func, ->info, and ->flags set.
140 void generic_exec_single(int cpu
, struct call_single_data
*csd
, int wait
)
142 struct call_single_queue
*dst
= &per_cpu(call_single_queue
, cpu
);
146 raw_spin_lock_irqsave(&dst
->lock
, flags
);
147 ipi
= list_empty(&dst
->list
);
148 list_add_tail(&csd
->list
, &dst
->list
);
149 raw_spin_unlock_irqrestore(&dst
->lock
, flags
);
152 * The list addition should be visible before sending the IPI
153 * handler locks the list to pull the entry off it because of
154 * normal cache coherency rules implied by spinlocks.
156 * If IPIs can go out of order to the cache coherency protocol
157 * in an architecture, sufficient synchronisation should be added
158 * to arch code to make it appear to obey cache coherency WRT
159 * locking and barrier primitives. Generic code isn't really
160 * equipped to do the right thing...
163 arch_send_call_function_single_ipi(cpu
);
170 * Invoked by arch to handle an IPI for call function single. Must be
171 * called from the arch with interrupts disabled.
173 void generic_smp_call_function_single_interrupt(void)
175 struct call_single_queue
*q
= &__get_cpu_var(call_single_queue
);
179 * Shouldn't receive this interrupt on a cpu that is not yet online.
181 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
183 raw_spin_lock(&q
->lock
);
184 list_replace_init(&q
->list
, &list
);
185 raw_spin_unlock(&q
->lock
);
187 while (!list_empty(&list
)) {
188 struct call_single_data
*csd
;
189 unsigned int csd_flags
;
191 csd
= list_entry(list
.next
, struct call_single_data
, list
);
192 list_del(&csd
->list
);
195 * 'csd' can be invalid after this call if flags == 0
196 * (when called through generic_exec_single()),
197 * so save them away before making the call:
199 csd_flags
= csd
->flags
;
201 csd
->func(csd
->info
);
204 * Unlocked CSDs are valid through generic_exec_single():
206 if (csd_flags
& CSD_FLAG_LOCK
)
211 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data
, csd_data
);
214 * smp_call_function_single - Run a function on a specific CPU
215 * @func: The function to run. This must be fast and non-blocking.
216 * @info: An arbitrary pointer to pass to the function.
217 * @wait: If true, wait until function has completed on other CPUs.
219 * Returns 0 on success, else a negative status code.
221 int smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
224 struct call_single_data d
= {
232 * prevent preemption and reschedule on another processor,
233 * as well as CPU removal
235 this_cpu
= get_cpu();
238 * Can deadlock when called with interrupts disabled.
239 * We allow cpu's that are not yet online though, as no one else can
240 * send smp call function interrupt to this cpu and as such deadlocks
243 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
244 && !oops_in_progress
);
246 if (cpu
== this_cpu
) {
247 local_irq_save(flags
);
249 local_irq_restore(flags
);
251 if ((unsigned)cpu
< nr_cpu_ids
&& cpu_online(cpu
)) {
252 struct call_single_data
*csd
= &d
;
255 csd
= &__get_cpu_var(csd_data
);
261 generic_exec_single(cpu
, csd
, wait
);
263 err
= -ENXIO
; /* CPU not online */
271 EXPORT_SYMBOL(smp_call_function_single
);
273 /* This function can be used by MTK Monitor only */
274 /* Dont use this function directly */
275 int mtk_smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
278 struct call_single_data d
= {
286 * prevent preemption and reschedule on another processor,
287 * as well as CPU removal
289 this_cpu
= get_cpu();
291 if (cpu
== this_cpu
) {
292 local_irq_save(flags
);
294 local_irq_restore(flags
);
296 if ((unsigned)cpu
< nr_cpu_ids
&& cpu_online(cpu
)) {
297 struct call_single_data
*data
= &d
;
300 data
= &__get_cpu_var(csd_data
);
304 cpumask_set_cpu(cpu
, data
->cpumask
);
307 generic_exec_single(cpu
, data
, wait
);
309 err
= -ENXIO
; /* CPU not online */
319 * smp_call_function_any - Run a function on any of the given cpus
320 * @mask: The mask of cpus it can run on.
321 * @func: The function to run. This must be fast and non-blocking.
322 * @info: An arbitrary pointer to pass to the function.
323 * @wait: If true, wait until function has completed.
325 * Returns 0 on success, else a negative status code (if no cpus were online).
326 * Note that @wait will be implicitly turned on in case of allocation failures,
327 * since we fall back to on-stack allocation.
329 * Selection preference:
330 * 1) current cpu if in @mask
331 * 2) any cpu of current node if in @mask
332 * 3) any other online cpu in @mask
334 int smp_call_function_any(const struct cpumask
*mask
,
335 smp_call_func_t func
, void *info
, int wait
)
338 const struct cpumask
*nodemask
;
341 /* Try for same CPU (cheapest) */
343 if (cpumask_test_cpu(cpu
, mask
))
346 /* Try for same node. */
347 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
348 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
349 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
354 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
355 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
357 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
361 EXPORT_SYMBOL_GPL(smp_call_function_any
);
364 * __smp_call_function_single(): Run a function on a specific CPU
365 * @cpu: The CPU to run on.
366 * @data: Pre-allocated and setup data structure
367 * @wait: If true, wait until function has completed on specified CPU.
369 * Like smp_call_function_single(), but allow caller to pass in a
370 * pre-allocated data structure. Useful for embedding @data inside
371 * other structures, for instance.
373 void __smp_call_function_single(int cpu
, struct call_single_data
*csd
,
376 unsigned int this_cpu
;
379 this_cpu
= get_cpu();
381 * Can deadlock when called with interrupts disabled.
382 * We allow cpu's that are not yet online though, as no one else can
383 * send smp call function interrupt to this cpu and as such deadlocks
386 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait
&& irqs_disabled()
387 && !oops_in_progress
);
389 if (cpu
== this_cpu
) {
390 local_irq_save(flags
);
391 csd
->func(csd
->info
);
392 local_irq_restore(flags
);
395 generic_exec_single(cpu
, csd
, wait
);
401 * smp_call_function_many(): Run a function on a set of other CPUs.
402 * @mask: The set of cpus to run on (only runs on online subset).
403 * @func: The function to run. This must be fast and non-blocking.
404 * @info: An arbitrary pointer to pass to the function.
405 * @wait: If true, wait (atomically) until function has completed
408 * If @wait is true, then returns once @func has returned.
410 * You must not call this function with disabled interrupts or from a
411 * hardware interrupt handler or from a bottom half handler. Preemption
412 * must be disabled when calling this function.
414 void smp_call_function_many(const struct cpumask
*mask
,
415 smp_call_func_t func
, void *info
, bool wait
)
417 struct call_function_data
*cfd
;
418 int cpu
, next_cpu
, this_cpu
= smp_processor_id();
421 * Can deadlock when called with interrupts disabled.
422 * We allow cpu's that are not yet online though, as no one else can
423 * send smp call function interrupt to this cpu and as such deadlocks
426 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
427 && !oops_in_progress
&& !early_boot_irqs_disabled
);
429 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
430 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
432 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
434 /* No online cpus? We're done. */
435 if (cpu
>= nr_cpu_ids
)
438 /* Do we have another CPU which isn't us? */
439 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
440 if (next_cpu
== this_cpu
)
441 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
443 /* Fastpath: do that cpu by itself. */
444 if (next_cpu
>= nr_cpu_ids
) {
445 smp_call_function_single(cpu
, func
, info
, wait
);
449 cfd
= &__get_cpu_var(cfd_data
);
451 cpumask_and(cfd
->cpumask
, mask
, cpu_online_mask
);
452 cpumask_clear_cpu(this_cpu
, cfd
->cpumask
);
454 /* Some callers race with other cpus changing the passed mask */
455 if (unlikely(!cpumask_weight(cfd
->cpumask
)))
459 * After we put an entry into the list, cfd->cpumask may be cleared
460 * again when another CPU sends another IPI for a SMP function call, so
461 * cfd->cpumask will be zero.
463 cpumask_copy(cfd
->cpumask_ipi
, cfd
->cpumask
);
465 for_each_cpu(cpu
, cfd
->cpumask
) {
466 struct call_single_data
*csd
= per_cpu_ptr(cfd
->csd
, cpu
);
467 struct call_single_queue
*dst
=
468 &per_cpu(call_single_queue
, cpu
);
475 raw_spin_lock_irqsave(&dst
->lock
, flags
);
476 list_add_tail(&csd
->list
, &dst
->list
);
477 raw_spin_unlock_irqrestore(&dst
->lock
, flags
);
480 /* Send a message to all CPUs in the map */
481 arch_send_call_function_ipi_mask(cfd
->cpumask_ipi
);
484 for_each_cpu(cpu
, cfd
->cpumask
) {
485 struct call_single_data
*csd
;
487 csd
= per_cpu_ptr(cfd
->csd
, cpu
);
492 EXPORT_SYMBOL(smp_call_function_many
);
495 * smp_call_function(): Run a function on all other CPUs.
496 * @func: The function to run. This must be fast and non-blocking.
497 * @info: An arbitrary pointer to pass to the function.
498 * @wait: If true, wait (atomically) until function has completed
503 * If @wait is true, then returns once @func has returned; otherwise
504 * it returns just before the target cpu calls @func.
506 * You must not call this function with disabled interrupts or from a
507 * hardware interrupt handler or from a bottom half handler.
509 int smp_call_function(smp_call_func_t func
, void *info
, int wait
)
512 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
517 EXPORT_SYMBOL(smp_call_function
);
518 #endif /* USE_GENERIC_SMP_HELPERS */
520 /* Setup configured maximum number of CPUs to activate */
521 unsigned int setup_max_cpus
= NR_CPUS
;
522 EXPORT_SYMBOL(setup_max_cpus
);
526 * Setup routine for controlling SMP activation
528 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
529 * activation entirely (the MPS table probe still happens, though).
531 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
532 * greater than 0, limits the maximum number of CPUs activated in
536 void __weak
arch_disable_smp_support(void) { }
538 static int __init
nosmp(char *str
)
541 arch_disable_smp_support();
546 early_param("nosmp", nosmp
);
548 /* this is hard limit */
549 static int __init
nrcpus(char *str
)
553 get_option(&str
, &nr_cpus
);
554 if (nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
555 nr_cpu_ids
= nr_cpus
;
560 early_param("nr_cpus", nrcpus
);
562 static int __init
maxcpus(char *str
)
564 get_option(&str
, &setup_max_cpus
);
565 if (setup_max_cpus
== 0)
566 arch_disable_smp_support();
571 early_param("maxcpus", maxcpus
);
573 /* Setup number of possible processor ids */
574 int nr_cpu_ids __read_mostly
= NR_CPUS
;
575 EXPORT_SYMBOL(nr_cpu_ids
);
577 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
578 void __init
setup_nr_cpu_ids(void)
580 nr_cpu_ids
= find_last_bit(cpumask_bits(cpu_possible_mask
),NR_CPUS
) + 1;
583 /* Called by boot processor to activate the rest. */
584 void __init
smp_init(void)
590 /* FIXME: This should be done in userspace --RR */
591 for_each_present_cpu(cpu
) {
592 if (num_online_cpus() >= setup_max_cpus
)
594 if (!cpu_online(cpu
))
598 /* Any cleanup work */
599 printk(KERN_INFO
"Brought up %ld CPUs\n", (long)num_online_cpus());
600 smp_cpus_done(setup_max_cpus
);
604 * Call a function on all processors. May be used during early boot while
605 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
606 * of local_irq_disable/enable().
608 int on_each_cpu(void (*func
) (void *info
), void *info
, int wait
)
614 ret
= smp_call_function(func
, info
, wait
);
615 local_irq_save(flags
);
617 local_irq_restore(flags
);
621 EXPORT_SYMBOL(on_each_cpu
);
624 * on_each_cpu_mask(): Run a function on processors specified by
625 * cpumask, which may include the local processor.
626 * @mask: The set of cpus to run on (only runs on online subset).
627 * @func: The function to run. This must be fast and non-blocking.
628 * @info: An arbitrary pointer to pass to the function.
629 * @wait: If true, wait (atomically) until function has completed
632 * If @wait is true, then returns once @func has returned.
634 * You must not call this function with disabled interrupts or
635 * from a hardware interrupt handler or from a bottom half handler.
637 void on_each_cpu_mask(const struct cpumask
*mask
, smp_call_func_t func
,
638 void *info
, bool wait
)
642 smp_call_function_many(mask
, func
, info
, wait
);
643 if (cpumask_test_cpu(cpu
, mask
)) {
650 EXPORT_SYMBOL(on_each_cpu_mask
);
653 * on_each_cpu_cond(): Call a function on each processor for which
654 * the supplied function cond_func returns true, optionally waiting
655 * for all the required CPUs to finish. This may include the local
657 * @cond_func: A callback function that is passed a cpu id and
658 * the the info parameter. The function is called
659 * with preemption disabled. The function should
660 * return a blooean value indicating whether to IPI
662 * @func: The function to run on all applicable CPUs.
663 * This must be fast and non-blocking.
664 * @info: An arbitrary pointer to pass to both functions.
665 * @wait: If true, wait (atomically) until function has
666 * completed on other CPUs.
667 * @gfp_flags: GFP flags to use when allocating the cpumask
668 * used internally by the function.
670 * The function might sleep if the GFP flags indicates a non
671 * atomic allocation is allowed.
673 * Preemption is disabled to protect against CPUs going offline but not online.
674 * CPUs going online during the call will not be seen or sent an IPI.
676 * You must not call this function with disabled interrupts or
677 * from a hardware interrupt handler or from a bottom half handler.
679 void on_each_cpu_cond(bool (*cond_func
)(int cpu
, void *info
),
680 smp_call_func_t func
, void *info
, bool wait
,
686 might_sleep_if(gfp_flags
& __GFP_WAIT
);
688 if (likely(zalloc_cpumask_var(&cpus
, (gfp_flags
|__GFP_NOWARN
)))) {
690 for_each_online_cpu(cpu
)
691 if (cond_func(cpu
, info
))
692 cpumask_set_cpu(cpu
, cpus
);
693 on_each_cpu_mask(cpus
, func
, info
, wait
);
695 free_cpumask_var(cpus
);
698 * No free cpumask, bother. No matter, we'll
699 * just have to IPI them one by one.
702 for_each_online_cpu(cpu
)
703 if (cond_func(cpu
, info
)) {
704 ret
= smp_call_function_single(cpu
, func
,
711 EXPORT_SYMBOL(on_each_cpu_cond
);
713 static void do_nothing(void *unused
)
718 * kick_all_cpus_sync - Force all cpus out of idle
720 * Used to synchronize the update of pm_idle function pointer. It's
721 * called after the pointer is updated and returns after the dummy
722 * callback function has been executed on all cpus. The execution of
723 * the function can only happen on the remote cpus after they have
724 * left the idle function which had been called via pm_idle function
725 * pointer. So it's guaranteed that nothing uses the previous pointer
728 void kick_all_cpus_sync(void)
730 /* Make sure the change is visible before we kick the cpus */
732 smp_call_function(do_nothing
, NULL
, 1);
734 EXPORT_SYMBOL_GPL(kick_all_cpus_sync
);