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>
16 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
18 struct list_head queue
;
20 } call_function __cacheline_aligned_in_smp
=
22 .queue
= LIST_HEAD_INIT(call_function
.queue
),
23 .lock
= __RAW_SPIN_LOCK_UNLOCKED(call_function
.lock
),
30 struct call_function_data
{
31 struct call_single_data csd
;
33 cpumask_var_t cpumask
;
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data
, cfd_data
);
38 struct call_single_queue
{
39 struct list_head list
;
43 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue
, call_single_queue
);
46 hotplug_cfd(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
48 long cpu
= (long)hcpu
;
49 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
53 case CPU_UP_PREPARE_FROZEN
:
54 if (!zalloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
56 return notifier_from_errno(-ENOMEM
);
59 #ifdef CONFIG_HOTPLUG_CPU
61 case CPU_UP_CANCELED_FROZEN
:
65 free_cpumask_var(cfd
->cpumask
);
73 static struct notifier_block __cpuinitdata hotplug_cfd_notifier
= {
74 .notifier_call
= hotplug_cfd
,
77 void __init
call_function_init(void)
79 void *cpu
= (void *)(long)smp_processor_id();
82 for_each_possible_cpu(i
) {
83 struct call_single_queue
*q
= &per_cpu(call_single_queue
, i
);
85 raw_spin_lock_init(&q
->lock
);
86 INIT_LIST_HEAD(&q
->list
);
89 hotplug_cfd(&hotplug_cfd_notifier
, CPU_UP_PREPARE
, cpu
);
90 register_cpu_notifier(&hotplug_cfd_notifier
);
94 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
96 * For non-synchronous ipi calls the csd can still be in use by the
97 * previous function call. For multi-cpu calls its even more interesting
98 * as we'll have to ensure no other cpu is observing our csd.
100 static void csd_lock_wait(struct call_single_data
*data
)
102 while (data
->flags
& CSD_FLAG_LOCK
)
106 static void csd_lock(struct call_single_data
*data
)
109 data
->flags
= CSD_FLAG_LOCK
;
112 * prevent CPU from reordering the above assignment
113 * to ->flags with any subsequent assignments to other
114 * fields of the specified call_single_data structure:
119 static void csd_unlock(struct call_single_data
*data
)
121 WARN_ON(!(data
->flags
& CSD_FLAG_LOCK
));
124 * ensure we're all done before releasing data:
128 data
->flags
&= ~CSD_FLAG_LOCK
;
132 * Insert a previously allocated call_single_data element
133 * for execution on the given CPU. data must already have
134 * ->func, ->info, and ->flags set.
137 void generic_exec_single(int cpu
, struct call_single_data
*data
, int wait
)
139 struct call_single_queue
*dst
= &per_cpu(call_single_queue
, cpu
);
143 raw_spin_lock_irqsave(&dst
->lock
, flags
);
144 ipi
= list_empty(&dst
->list
);
145 list_add_tail(&data
->list
, &dst
->list
);
146 raw_spin_unlock_irqrestore(&dst
->lock
, flags
);
149 * The list addition should be visible before sending the IPI
150 * handler locks the list to pull the entry off it because of
151 * normal cache coherency rules implied by spinlocks.
153 * If IPIs can go out of order to the cache coherency protocol
154 * in an architecture, sufficient synchronisation should be added
155 * to arch code to make it appear to obey cache coherency WRT
156 * locking and barrier primitives. Generic code isn't really
157 * equipped to do the right thing...
160 arch_send_call_function_single_ipi(cpu
);
167 * Invoked by arch to handle an IPI for call function. Must be called with
168 * interrupts disabled.
170 void generic_smp_call_function_interrupt(void)
172 struct call_function_data
*data
;
173 int cpu
= smp_processor_id();
176 * Shouldn't receive this interrupt on a cpu that is not yet online.
178 WARN_ON_ONCE(!cpu_online(cpu
));
181 * Ensure entry is visible on call_function_queue after we have
182 * entered the IPI. See comment in smp_call_function_many.
183 * If we don't have this, then we may miss an entry on the list
184 * and never get another IPI to process it.
189 * It's ok to use list_for_each_rcu() here even though we may
190 * delete 'pos', since list_del_rcu() doesn't clear ->next
192 list_for_each_entry_rcu(data
, &call_function
.queue
, csd
.list
) {
194 smp_call_func_t func
;
197 * Since we walk the list without any locks, we might
198 * see an entry that was completed, removed from the
199 * list and is in the process of being reused.
201 * We must check that the cpu is in the cpumask before
202 * checking the refs, and both must be set before
203 * executing the callback on this cpu.
206 if (!cpumask_test_cpu(cpu
, data
->cpumask
))
211 if (atomic_read(&data
->refs
) == 0)
214 func
= data
->csd
.func
; /* save for later warn */
215 func(data
->csd
.info
);
218 * If the cpu mask is not still set then func enabled
219 * interrupts (BUG), and this cpu took another smp call
220 * function interrupt and executed func(info) twice
221 * on this cpu. That nested execution decremented refs.
223 if (!cpumask_test_and_clear_cpu(cpu
, data
->cpumask
)) {
224 WARN(1, "%pf enabled interrupts and double executed\n", func
);
228 refs
= atomic_dec_return(&data
->refs
);
234 WARN_ON(!cpumask_empty(data
->cpumask
));
236 raw_spin_lock(&call_function
.lock
);
237 list_del_rcu(&data
->csd
.list
);
238 raw_spin_unlock(&call_function
.lock
);
240 csd_unlock(&data
->csd
);
246 * Invoked by arch to handle an IPI for call function single. Must be
247 * called from the arch with interrupts disabled.
249 void generic_smp_call_function_single_interrupt(void)
251 struct call_single_queue
*q
= &__get_cpu_var(call_single_queue
);
252 unsigned int data_flags
;
256 * Shouldn't receive this interrupt on a cpu that is not yet online.
258 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
260 raw_spin_lock(&q
->lock
);
261 list_replace_init(&q
->list
, &list
);
262 raw_spin_unlock(&q
->lock
);
264 while (!list_empty(&list
)) {
265 struct call_single_data
*data
;
267 data
= list_entry(list
.next
, struct call_single_data
, list
);
268 list_del(&data
->list
);
271 * 'data' can be invalid after this call if flags == 0
272 * (when called through generic_exec_single()),
273 * so save them away before making the call:
275 data_flags
= data
->flags
;
277 data
->func(data
->info
);
280 * Unlocked CSDs are valid through generic_exec_single():
282 if (data_flags
& CSD_FLAG_LOCK
)
287 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data
, csd_data
);
290 * smp_call_function_single - Run a function on a specific CPU
291 * @func: The function to run. This must be fast and non-blocking.
292 * @info: An arbitrary pointer to pass to the function.
293 * @wait: If true, wait until function has completed on other CPUs.
295 * Returns 0 on success, else a negative status code.
297 int smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
300 struct call_single_data d
= {
308 * prevent preemption and reschedule on another processor,
309 * as well as CPU removal
311 this_cpu
= get_cpu();
314 * Can deadlock when called with interrupts disabled.
315 * We allow cpu's that are not yet online though, as no one else can
316 * send smp call function interrupt to this cpu and as such deadlocks
319 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
320 && !oops_in_progress
);
322 if (cpu
== this_cpu
) {
323 local_irq_save(flags
);
325 local_irq_restore(flags
);
327 if ((unsigned)cpu
< nr_cpu_ids
&& cpu_online(cpu
)) {
328 struct call_single_data
*data
= &d
;
331 data
= &__get_cpu_var(csd_data
);
337 generic_exec_single(cpu
, data
, wait
);
339 err
= -ENXIO
; /* CPU not online */
347 EXPORT_SYMBOL(smp_call_function_single
);
350 * smp_call_function_any - Run a function on any of the given cpus
351 * @mask: The mask of cpus it can run on.
352 * @func: The function to run. This must be fast and non-blocking.
353 * @info: An arbitrary pointer to pass to the function.
354 * @wait: If true, wait until function has completed.
356 * Returns 0 on success, else a negative status code (if no cpus were online).
357 * Note that @wait will be implicitly turned on in case of allocation failures,
358 * since we fall back to on-stack allocation.
360 * Selection preference:
361 * 1) current cpu if in @mask
362 * 2) any cpu of current node if in @mask
363 * 3) any other online cpu in @mask
365 int smp_call_function_any(const struct cpumask
*mask
,
366 smp_call_func_t func
, void *info
, int wait
)
369 const struct cpumask
*nodemask
;
372 /* Try for same CPU (cheapest) */
374 if (cpumask_test_cpu(cpu
, mask
))
377 /* Try for same node. */
378 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
379 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
380 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
385 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
386 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
388 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
392 EXPORT_SYMBOL_GPL(smp_call_function_any
);
395 * __smp_call_function_single(): Run a function on a specific CPU
396 * @cpu: The CPU to run on.
397 * @data: Pre-allocated and setup data structure
398 * @wait: If true, wait until function has completed on specified CPU.
400 * Like smp_call_function_single(), but allow caller to pass in a
401 * pre-allocated data structure. Useful for embedding @data inside
402 * other structures, for instance.
404 void __smp_call_function_single(int cpu
, struct call_single_data
*data
,
407 unsigned int this_cpu
;
410 this_cpu
= get_cpu();
412 * Can deadlock when called with interrupts disabled.
413 * We allow cpu's that are not yet online though, as no one else can
414 * send smp call function interrupt to this cpu and as such deadlocks
417 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait
&& irqs_disabled()
418 && !oops_in_progress
);
420 if (cpu
== this_cpu
) {
421 local_irq_save(flags
);
422 data
->func(data
->info
);
423 local_irq_restore(flags
);
426 generic_exec_single(cpu
, data
, wait
);
432 * smp_call_function_many(): Run a function on a set of other CPUs.
433 * @mask: The set of cpus to run on (only runs on online subset).
434 * @func: The function to run. This must be fast and non-blocking.
435 * @info: An arbitrary pointer to pass to the function.
436 * @wait: If true, wait (atomically) until function has completed
439 * If @wait is true, then returns once @func has returned.
441 * You must not call this function with disabled interrupts or from a
442 * hardware interrupt handler or from a bottom half handler. Preemption
443 * must be disabled when calling this function.
445 void smp_call_function_many(const struct cpumask
*mask
,
446 smp_call_func_t func
, void *info
, bool wait
)
448 struct call_function_data
*data
;
450 int refs
, cpu
, next_cpu
, this_cpu
= smp_processor_id();
453 * Can deadlock when called with interrupts disabled.
454 * We allow cpu's that are not yet online though, as no one else can
455 * send smp call function interrupt to this cpu and as such deadlocks
458 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
459 && !oops_in_progress
&& !early_boot_irqs_disabled
);
461 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
462 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
464 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
466 /* No online cpus? We're done. */
467 if (cpu
>= nr_cpu_ids
)
470 /* Do we have another CPU which isn't us? */
471 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
472 if (next_cpu
== this_cpu
)
473 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
475 /* Fastpath: do that cpu by itself. */
476 if (next_cpu
>= nr_cpu_ids
) {
477 smp_call_function_single(cpu
, func
, info
, wait
);
481 data
= &__get_cpu_var(cfd_data
);
482 csd_lock(&data
->csd
);
484 /* This BUG_ON verifies our reuse assertions and can be removed */
485 BUG_ON(atomic_read(&data
->refs
) || !cpumask_empty(data
->cpumask
));
488 * The global call function queue list add and delete are protected
489 * by a lock, but the list is traversed without any lock, relying
490 * on the rcu list add and delete to allow safe concurrent traversal.
491 * We reuse the call function data without waiting for any grace
492 * period after some other cpu removes it from the global queue.
493 * This means a cpu might find our data block as it is being
496 * We hold off the interrupt handler on the other cpu by
497 * ordering our writes to the cpu mask vs our setting of the
498 * refs counter. We assert only the cpu owning the data block
499 * will set a bit in cpumask, and each bit will only be cleared
500 * by the subject cpu. Each cpu must first find its bit is
501 * set and then check that refs is set indicating the element is
502 * ready to be processed, otherwise it must skip the entry.
504 * On the previous iteration refs was set to 0 by another cpu.
505 * To avoid the use of transitivity, set the counter to 0 here
506 * so the wmb will pair with the rmb in the interrupt handler.
508 atomic_set(&data
->refs
, 0); /* convert 3rd to 1st party write */
510 data
->csd
.func
= func
;
511 data
->csd
.info
= info
;
513 /* Ensure 0 refs is visible before mask. Also orders func and info */
516 /* We rely on the "and" being processed before the store */
517 cpumask_and(data
->cpumask
, mask
, cpu_online_mask
);
518 cpumask_clear_cpu(this_cpu
, data
->cpumask
);
519 refs
= cpumask_weight(data
->cpumask
);
521 /* Some callers race with other cpus changing the passed mask */
522 if (unlikely(!refs
)) {
523 csd_unlock(&data
->csd
);
527 raw_spin_lock_irqsave(&call_function
.lock
, flags
);
529 * Place entry at the _HEAD_ of the list, so that any cpu still
530 * observing the entry in generic_smp_call_function_interrupt()
531 * will not miss any other list entries:
533 list_add_rcu(&data
->csd
.list
, &call_function
.queue
);
535 * We rely on the wmb() in list_add_rcu to complete our writes
536 * to the cpumask before this write to refs, which indicates
537 * data is on the list and is ready to be processed.
539 atomic_set(&data
->refs
, refs
);
540 raw_spin_unlock_irqrestore(&call_function
.lock
, flags
);
543 * Make the list addition visible before sending the ipi.
544 * (IPIs must obey or appear to obey normal Linux cache
545 * coherency rules -- see comment in generic_exec_single).
549 /* Send a message to all CPUs in the map */
550 arch_send_call_function_ipi_mask(data
->cpumask
);
552 /* Optionally wait for the CPUs to complete */
554 csd_lock_wait(&data
->csd
);
556 EXPORT_SYMBOL(smp_call_function_many
);
559 * smp_call_function(): Run a function on all other CPUs.
560 * @func: The function to run. This must be fast and non-blocking.
561 * @info: An arbitrary pointer to pass to the function.
562 * @wait: If true, wait (atomically) until function has completed
567 * If @wait is true, then returns once @func has returned; otherwise
568 * it returns just before the target cpu calls @func.
570 * You must not call this function with disabled interrupts or from a
571 * hardware interrupt handler or from a bottom half handler.
573 int smp_call_function(smp_call_func_t func
, void *info
, int wait
)
576 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
581 EXPORT_SYMBOL(smp_call_function
);
583 void ipi_call_lock(void)
585 raw_spin_lock(&call_function
.lock
);
588 void ipi_call_unlock(void)
590 raw_spin_unlock(&call_function
.lock
);
593 void ipi_call_lock_irq(void)
595 raw_spin_lock_irq(&call_function
.lock
);
598 void ipi_call_unlock_irq(void)
600 raw_spin_unlock_irq(&call_function
.lock
);
602 #endif /* USE_GENERIC_SMP_HELPERS */
604 /* Setup configured maximum number of CPUs to activate */
605 unsigned int setup_max_cpus
= NR_CPUS
;
606 EXPORT_SYMBOL(setup_max_cpus
);
610 * Setup routine for controlling SMP activation
612 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
613 * activation entirely (the MPS table probe still happens, though).
615 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
616 * greater than 0, limits the maximum number of CPUs activated in
620 void __weak
arch_disable_smp_support(void) { }
622 static int __init
nosmp(char *str
)
625 arch_disable_smp_support();
630 early_param("nosmp", nosmp
);
632 /* this is hard limit */
633 static int __init
nrcpus(char *str
)
637 get_option(&str
, &nr_cpus
);
638 if (nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
639 nr_cpu_ids
= nr_cpus
;
644 early_param("nr_cpus", nrcpus
);
646 static int __init
maxcpus(char *str
)
648 get_option(&str
, &setup_max_cpus
);
649 if (setup_max_cpus
== 0)
650 arch_disable_smp_support();
655 early_param("maxcpus", maxcpus
);
657 /* Setup number of possible processor ids */
658 int nr_cpu_ids __read_mostly
= NR_CPUS
;
659 EXPORT_SYMBOL(nr_cpu_ids
);
661 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
662 void __init
setup_nr_cpu_ids(void)
664 nr_cpu_ids
= find_last_bit(cpumask_bits(cpu_possible_mask
),NR_CPUS
) + 1;
667 /* Called by boot processor to activate the rest. */
668 void __init
smp_init(void)
672 /* FIXME: This should be done in userspace --RR */
673 for_each_present_cpu(cpu
) {
674 if (num_online_cpus() >= setup_max_cpus
)
676 if (!cpu_online(cpu
))
680 /* Any cleanup work */
681 printk(KERN_INFO
"Brought up %ld CPUs\n", (long)num_online_cpus());
682 smp_cpus_done(setup_max_cpus
);
686 * Call a function on all processors. May be used during early boot while
687 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
688 * of local_irq_disable/enable().
690 int on_each_cpu(void (*func
) (void *info
), void *info
, int wait
)
696 ret
= smp_call_function(func
, info
, wait
);
697 local_irq_save(flags
);
699 local_irq_restore(flags
);
703 EXPORT_SYMBOL(on_each_cpu
);
706 * on_each_cpu_mask(): Run a function on processors specified by
707 * cpumask, which may include the local processor.
708 * @mask: The set of cpus to run on (only runs on online subset).
709 * @func: The function to run. This must be fast and non-blocking.
710 * @info: An arbitrary pointer to pass to the function.
711 * @wait: If true, wait (atomically) until function has completed
714 * If @wait is true, then returns once @func has returned.
716 * You must not call this function with disabled interrupts or
717 * from a hardware interrupt handler or from a bottom half handler.
719 void on_each_cpu_mask(const struct cpumask
*mask
, smp_call_func_t func
,
720 void *info
, bool wait
)
724 smp_call_function_many(mask
, func
, info
, wait
);
725 if (cpumask_test_cpu(cpu
, mask
)) {
732 EXPORT_SYMBOL(on_each_cpu_mask
);
735 * on_each_cpu_cond(): Call a function on each processor for which
736 * the supplied function cond_func returns true, optionally waiting
737 * for all the required CPUs to finish. This may include the local
739 * @cond_func: A callback function that is passed a cpu id and
740 * the the info parameter. The function is called
741 * with preemption disabled. The function should
742 * return a blooean value indicating whether to IPI
744 * @func: The function to run on all applicable CPUs.
745 * This must be fast and non-blocking.
746 * @info: An arbitrary pointer to pass to both functions.
747 * @wait: If true, wait (atomically) until function has
748 * completed on other CPUs.
749 * @gfp_flags: GFP flags to use when allocating the cpumask
750 * used internally by the function.
752 * The function might sleep if the GFP flags indicates a non
753 * atomic allocation is allowed.
755 * Preemption is disabled to protect against CPUs going offline but not online.
756 * CPUs going online during the call will not be seen or sent an IPI.
758 * You must not call this function with disabled interrupts or
759 * from a hardware interrupt handler or from a bottom half handler.
761 void on_each_cpu_cond(bool (*cond_func
)(int cpu
, void *info
),
762 smp_call_func_t func
, void *info
, bool wait
,
768 might_sleep_if(gfp_flags
& __GFP_WAIT
);
770 if (likely(zalloc_cpumask_var(&cpus
, (gfp_flags
|__GFP_NOWARN
)))) {
772 for_each_online_cpu(cpu
)
773 if (cond_func(cpu
, info
))
774 cpumask_set_cpu(cpu
, cpus
);
775 on_each_cpu_mask(cpus
, func
, info
, wait
);
777 free_cpumask_var(cpus
);
780 * No free cpumask, bother. No matter, we'll
781 * just have to IPI them one by one.
784 for_each_online_cpu(cpu
)
785 if (cond_func(cpu
, info
)) {
786 ret
= smp_call_function_single(cpu
, func
,
793 EXPORT_SYMBOL(on_each_cpu_cond
);