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
20 struct list_head queue;
22 } call_function __cacheline_aligned_in_smp =
24 .queue = LIST_HEAD_INIT(call_function.queue),
25 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
32 struct call_function_data {
33 struct call_single_data csd;
35 cpumask_var_t cpumask;
38 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
40 struct call_single_queue {
41 struct list_head list;
45 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
48 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
50 long cpu = (long)hcpu;
51 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
55 case CPU_UP_PREPARE_FROZEN:
56 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
58 return notifier_from_errno(-ENOMEM);
61 #ifdef CONFIG_HOTPLUG_CPU
63 case CPU_UP_CANCELED_FROZEN:
67 free_cpumask_var(cfd->cpumask);
75 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
76 .notifier_call = hotplug_cfd,
79 void __init call_function_init(void)
81 void *cpu = (void *)(long)smp_processor_id();
84 for_each_possible_cpu(i) {
85 struct call_single_queue *q = &per_cpu(call_single_queue, i);
87 raw_spin_lock_init(&q->lock);
88 INIT_LIST_HEAD(&q->list);
91 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
92 register_cpu_notifier(&hotplug_cfd_notifier);
96 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
98 * For non-synchronous ipi calls the csd can still be in use by the
99 * previous function call. For multi-cpu calls its even more interesting
100 * as we'll have to ensure no other cpu is observing our csd.
102 static void csd_lock_wait(struct call_single_data *data)
104 while (data->flags & CSD_FLAG_LOCK)
108 static void csd_lock(struct call_single_data *data)
111 data->flags = CSD_FLAG_LOCK;
114 * prevent CPU from reordering the above assignment
115 * to ->flags with any subsequent assignments to other
116 * fields of the specified call_single_data structure:
121 static void csd_unlock(struct call_single_data *data)
123 WARN_ON(!(data->flags & CSD_FLAG_LOCK));
126 * ensure we're all done before releasing data:
130 data->flags &= ~CSD_FLAG_LOCK;
134 * Insert a previously allocated call_single_data element
135 * for execution on the given CPU. data must already have
136 * ->func, ->info, and ->flags set.
139 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
141 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
145 raw_spin_lock_irqsave(&dst->lock, flags);
146 ipi = list_empty(&dst->list);
147 list_add_tail(&data->list, &dst->list);
148 raw_spin_unlock_irqrestore(&dst->lock, flags);
151 * The list addition should be visible before sending the IPI
152 * handler locks the list to pull the entry off it because of
153 * normal cache coherency rules implied by spinlocks.
155 * If IPIs can go out of order to the cache coherency protocol
156 * in an architecture, sufficient synchronisation should be added
157 * to arch code to make it appear to obey cache coherency WRT
158 * locking and barrier primitives. Generic code isn't really
159 * equipped to do the right thing...
162 arch_send_call_function_single_ipi(cpu);
169 * Invoked by arch to handle an IPI for call function. Must be called with
170 * interrupts disabled.
172 void generic_smp_call_function_interrupt(void)
174 struct call_function_data *data;
175 int cpu = smp_processor_id();
178 * Shouldn't receive this interrupt on a cpu that is not yet online.
180 WARN_ON_ONCE(!cpu_online(cpu));
183 * Ensure entry is visible on call_function_queue after we have
184 * entered the IPI. See comment in smp_call_function_many.
185 * If we don't have this, then we may miss an entry on the list
186 * and never get another IPI to process it.
191 * It's ok to use list_for_each_rcu() here even though we may
192 * delete 'pos', since list_del_rcu() doesn't clear ->next
194 list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
196 smp_call_func_t func;
199 * Since we walk the list without any locks, we might
200 * see an entry that was completed, removed from the
201 * list and is in the process of being reused.
203 * We must check that the cpu is in the cpumask before
204 * checking the refs, and both must be set before
205 * executing the callback on this cpu.
208 if (!cpumask_test_cpu(cpu, data->cpumask))
213 if (atomic_read(&data->refs) == 0)
216 func = data->csd.func; /* save for later warn */
217 func(data->csd.info);
220 * If the cpu mask is not still set then func enabled
221 * interrupts (BUG), and this cpu took another smp call
222 * function interrupt and executed func(info) twice
223 * on this cpu. That nested execution decremented refs.
225 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask)) {
226 WARN(1, "%pf enabled interrupts and double executed\n", func);
230 refs = atomic_dec_return(&data->refs);
236 WARN_ON(!cpumask_empty(data->cpumask));
238 raw_spin_lock(&call_function.lock);
239 list_del_rcu(&data->csd.list);
240 raw_spin_unlock(&call_function.lock);
242 csd_unlock(&data->csd);
248 * Invoked by arch to handle an IPI for call function single. Must be
249 * called from the arch with interrupts disabled.
251 void generic_smp_call_function_single_interrupt(void)
253 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
254 unsigned int data_flags;
258 * Shouldn't receive this interrupt on a cpu that is not yet online.
260 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
262 raw_spin_lock(&q->lock);
263 list_replace_init(&q->list, &list);
264 raw_spin_unlock(&q->lock);
266 while (!list_empty(&list)) {
267 struct call_single_data *data;
269 data = list_entry(list.next, struct call_single_data, list);
270 list_del(&data->list);
273 * 'data' can be invalid after this call if flags == 0
274 * (when called through generic_exec_single()),
275 * so save them away before making the call:
277 data_flags = data->flags;
279 data->func(data->info);
282 * Unlocked CSDs are valid through generic_exec_single():
284 if (data_flags & CSD_FLAG_LOCK)
289 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
292 * smp_call_function_single - Run a function on a specific CPU
293 * @func: The function to run. This must be fast and non-blocking.
294 * @info: An arbitrary pointer to pass to the function.
295 * @wait: If true, wait until function has completed on other CPUs.
297 * Returns 0 on success, else a negative status code.
299 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
302 struct call_single_data d = {
310 * prevent preemption and reschedule on another processor,
311 * as well as CPU removal
313 this_cpu = get_cpu();
316 * Can deadlock when called with interrupts disabled.
317 * We allow cpu's that are not yet online though, as no one else can
318 * send smp call function interrupt to this cpu and as such deadlocks
321 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
322 && !oops_in_progress);
324 if (cpu == this_cpu) {
325 local_irq_save(flags);
327 local_irq_restore(flags);
329 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
330 struct call_single_data *data = &d;
333 data = &__get_cpu_var(csd_data);
339 generic_exec_single(cpu, data, wait);
341 err = -ENXIO; /* CPU not online */
349 EXPORT_SYMBOL(smp_call_function_single);
352 * smp_call_function_any - Run a function on any of the given cpus
353 * @mask: The mask of cpus it can run on.
354 * @func: The function to run. This must be fast and non-blocking.
355 * @info: An arbitrary pointer to pass to the function.
356 * @wait: If true, wait until function has completed.
358 * Returns 0 on success, else a negative status code (if no cpus were online).
359 * Note that @wait will be implicitly turned on in case of allocation failures,
360 * since we fall back to on-stack allocation.
362 * Selection preference:
363 * 1) current cpu if in @mask
364 * 2) any cpu of current node if in @mask
365 * 3) any other online cpu in @mask
367 int smp_call_function_any(const struct cpumask *mask,
368 smp_call_func_t func, void *info, int wait)
371 const struct cpumask *nodemask;
374 /* Try for same CPU (cheapest) */
376 if (cpumask_test_cpu(cpu, mask))
379 /* Try for same node. */
380 nodemask = cpumask_of_node(cpu_to_node(cpu));
381 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
382 cpu = cpumask_next_and(cpu, nodemask, mask)) {
387 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
388 cpu = cpumask_any_and(mask, cpu_online_mask);
390 ret = smp_call_function_single(cpu, func, info, wait);
394 EXPORT_SYMBOL_GPL(smp_call_function_any);
397 * __smp_call_function_single(): Run a function on a specific CPU
398 * @cpu: The CPU to run on.
399 * @data: Pre-allocated and setup data structure
400 * @wait: If true, wait until function has completed on specified CPU.
402 * Like smp_call_function_single(), but allow caller to pass in a
403 * pre-allocated data structure. Useful for embedding @data inside
404 * other structures, for instance.
406 void __smp_call_function_single(int cpu, struct call_single_data *data,
409 unsigned int this_cpu;
412 this_cpu = get_cpu();
414 * Can deadlock when called with interrupts disabled.
415 * We allow cpu's that are not yet online though, as no one else can
416 * send smp call function interrupt to this cpu and as such deadlocks
419 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
420 && !oops_in_progress);
422 if (cpu == this_cpu) {
423 local_irq_save(flags);
424 data->func(data->info);
425 local_irq_restore(flags);
428 generic_exec_single(cpu, data, wait);
434 * smp_call_function_many(): Run a function on a set of other CPUs.
435 * @mask: The set of cpus to run on (only runs on online subset).
436 * @func: The function to run. This must be fast and non-blocking.
437 * @info: An arbitrary pointer to pass to the function.
438 * @wait: If true, wait (atomically) until function has completed
441 * If @wait is true, then returns once @func has returned.
443 * You must not call this function with disabled interrupts or from a
444 * hardware interrupt handler or from a bottom half handler. Preemption
445 * must be disabled when calling this function.
447 void smp_call_function_many(const struct cpumask *mask,
448 smp_call_func_t func, void *info, bool wait)
450 struct call_function_data *data;
452 int refs, cpu, next_cpu, this_cpu = smp_processor_id();
455 * Can deadlock when called with interrupts disabled.
456 * We allow cpu's that are not yet online though, as no one else can
457 * send smp call function interrupt to this cpu and as such deadlocks
460 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
461 && !oops_in_progress && !early_boot_irqs_disabled);
463 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
464 cpu = cpumask_first_and(mask, cpu_online_mask);
466 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
468 /* No online cpus? We're done. */
469 if (cpu >= nr_cpu_ids)
472 /* Do we have another CPU which isn't us? */
473 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
474 if (next_cpu == this_cpu)
475 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
477 /* Fastpath: do that cpu by itself. */
478 if (next_cpu >= nr_cpu_ids) {
479 smp_call_function_single(cpu, func, info, wait);
483 data = &__get_cpu_var(cfd_data);
484 csd_lock(&data->csd);
486 /* This BUG_ON verifies our reuse assertions and can be removed */
487 BUG_ON(atomic_read(&data->refs) || !cpumask_empty(data->cpumask));
490 * The global call function queue list add and delete are protected
491 * by a lock, but the list is traversed without any lock, relying
492 * on the rcu list add and delete to allow safe concurrent traversal.
493 * We reuse the call function data without waiting for any grace
494 * period after some other cpu removes it from the global queue.
495 * This means a cpu might find our data block as it is being
498 * We hold off the interrupt handler on the other cpu by
499 * ordering our writes to the cpu mask vs our setting of the
500 * refs counter. We assert only the cpu owning the data block
501 * will set a bit in cpumask, and each bit will only be cleared
502 * by the subject cpu. Each cpu must first find its bit is
503 * set and then check that refs is set indicating the element is
504 * ready to be processed, otherwise it must skip the entry.
506 * On the previous iteration refs was set to 0 by another cpu.
507 * To avoid the use of transitivity, set the counter to 0 here
508 * so the wmb will pair with the rmb in the interrupt handler.
510 atomic_set(&data->refs, 0); /* convert 3rd to 1st party write */
512 data->csd.func = func;
513 data->csd.info = info;
515 /* Ensure 0 refs is visible before mask. Also orders func and info */
518 /* We rely on the "and" being processed before the store */
519 cpumask_and(data->cpumask, mask, cpu_online_mask);
520 cpumask_clear_cpu(this_cpu, data->cpumask);
521 refs = cpumask_weight(data->cpumask);
523 /* Some callers race with other cpus changing the passed mask */
524 if (unlikely(!refs)) {
525 csd_unlock(&data->csd);
529 raw_spin_lock_irqsave(&call_function.lock, flags);
531 * Place entry at the _HEAD_ of the list, so that any cpu still
532 * observing the entry in generic_smp_call_function_interrupt()
533 * will not miss any other list entries:
535 list_add_rcu(&data->csd.list, &call_function.queue);
537 * We rely on the wmb() in list_add_rcu to complete our writes
538 * to the cpumask before this write to refs, which indicates
539 * data is on the list and is ready to be processed.
541 atomic_set(&data->refs, refs);
542 raw_spin_unlock_irqrestore(&call_function.lock, flags);
545 * Make the list addition visible before sending the ipi.
546 * (IPIs must obey or appear to obey normal Linux cache
547 * coherency rules -- see comment in generic_exec_single).
551 /* Send a message to all CPUs in the map */
552 arch_send_call_function_ipi_mask(data->cpumask);
554 /* Optionally wait for the CPUs to complete */
556 csd_lock_wait(&data->csd);
558 EXPORT_SYMBOL(smp_call_function_many);
561 * smp_call_function(): Run a function on all other CPUs.
562 * @func: The function to run. This must be fast and non-blocking.
563 * @info: An arbitrary pointer to pass to the function.
564 * @wait: If true, wait (atomically) until function has completed
569 * If @wait is true, then returns once @func has returned; otherwise
570 * it returns just before the target cpu calls @func.
572 * You must not call this function with disabled interrupts or from a
573 * hardware interrupt handler or from a bottom half handler.
575 int smp_call_function(smp_call_func_t func, void *info, int wait)
578 smp_call_function_many(cpu_online_mask, func, info, wait);
583 EXPORT_SYMBOL(smp_call_function);
585 void ipi_call_lock(void)
587 raw_spin_lock(&call_function.lock);
590 void ipi_call_unlock(void)
592 raw_spin_unlock(&call_function.lock);
595 void ipi_call_lock_irq(void)
597 raw_spin_lock_irq(&call_function.lock);
600 void ipi_call_unlock_irq(void)
602 raw_spin_unlock_irq(&call_function.lock);
604 #endif /* USE_GENERIC_SMP_HELPERS */
606 /* Setup configured maximum number of CPUs to activate */
607 unsigned int setup_max_cpus = NR_CPUS;
608 EXPORT_SYMBOL(setup_max_cpus);
612 * Setup routine for controlling SMP activation
614 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
615 * activation entirely (the MPS table probe still happens, though).
617 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
618 * greater than 0, limits the maximum number of CPUs activated in
622 void __weak arch_disable_smp_support(void) { }
624 static int __init nosmp(char *str)
627 arch_disable_smp_support();
632 early_param("nosmp", nosmp);
634 /* this is hard limit */
635 static int __init nrcpus(char *str)
639 get_option(&str, &nr_cpus);
640 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
641 nr_cpu_ids = nr_cpus;
646 early_param("nr_cpus", nrcpus);
648 static int __init maxcpus(char *str)
650 get_option(&str, &setup_max_cpus);
651 if (setup_max_cpus == 0)
652 arch_disable_smp_support();
657 early_param("maxcpus", maxcpus);
659 /* Setup number of possible processor ids */
660 int nr_cpu_ids __read_mostly = NR_CPUS;
661 EXPORT_SYMBOL(nr_cpu_ids);
663 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
664 void __init setup_nr_cpu_ids(void)
666 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
669 /* Called by boot processor to activate the rest. */
670 void __init smp_init(void)
676 /* FIXME: This should be done in userspace --RR */
677 for_each_present_cpu(cpu) {
678 if (num_online_cpus() >= setup_max_cpus)
680 if (!cpu_online(cpu))
684 /* Any cleanup work */
685 printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
686 smp_cpus_done(setup_max_cpus);
690 * Call a function on all processors. May be used during early boot while
691 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
692 * of local_irq_disable/enable().
694 int on_each_cpu(void (*func) (void *info), void *info, int wait)
700 ret = smp_call_function(func, info, wait);
701 local_irq_save(flags);
703 local_irq_restore(flags);
707 EXPORT_SYMBOL(on_each_cpu);
710 * on_each_cpu_mask(): Run a function on processors specified by
711 * cpumask, which may include the local processor.
712 * @mask: The set of cpus to run on (only runs on online subset).
713 * @func: The function to run. This must be fast and non-blocking.
714 * @info: An arbitrary pointer to pass to the function.
715 * @wait: If true, wait (atomically) until function has completed
718 * If @wait is true, then returns once @func has returned.
720 * You must not call this function with disabled interrupts or
721 * from a hardware interrupt handler or from a bottom half handler.
723 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
724 void *info, bool wait)
728 smp_call_function_many(mask, func, info, wait);
729 if (cpumask_test_cpu(cpu, mask)) {
736 EXPORT_SYMBOL(on_each_cpu_mask);
739 * on_each_cpu_cond(): Call a function on each processor for which
740 * the supplied function cond_func returns true, optionally waiting
741 * for all the required CPUs to finish. This may include the local
743 * @cond_func: A callback function that is passed a cpu id and
744 * the the info parameter. The function is called
745 * with preemption disabled. The function should
746 * return a blooean value indicating whether to IPI
748 * @func: The function to run on all applicable CPUs.
749 * This must be fast and non-blocking.
750 * @info: An arbitrary pointer to pass to both functions.
751 * @wait: If true, wait (atomically) until function has
752 * completed on other CPUs.
753 * @gfp_flags: GFP flags to use when allocating the cpumask
754 * used internally by the function.
756 * The function might sleep if the GFP flags indicates a non
757 * atomic allocation is allowed.
759 * Preemption is disabled to protect against CPUs going offline but not online.
760 * CPUs going online during the call will not be seen or sent an IPI.
762 * You must not call this function with disabled interrupts or
763 * from a hardware interrupt handler or from a bottom half handler.
765 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
766 smp_call_func_t func, void *info, bool wait,
772 might_sleep_if(gfp_flags & __GFP_WAIT);
774 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
776 for_each_online_cpu(cpu)
777 if (cond_func(cpu, info))
778 cpumask_set_cpu(cpu, cpus);
779 on_each_cpu_mask(cpus, func, info, wait);
781 free_cpumask_var(cpus);
784 * No free cpumask, bother. No matter, we'll
785 * just have to IPI them one by one.
788 for_each_online_cpu(cpu)
789 if (cond_func(cpu, info)) {
790 ret = smp_call_function_single(cpu, func,
797 EXPORT_SYMBOL(on_each_cpu_cond);
799 static void do_nothing(void *unused)
804 * kick_all_cpus_sync - Force all cpus out of idle
806 * Used to synchronize the update of pm_idle function pointer. It's
807 * called after the pointer is updated and returns after the dummy
808 * callback function has been executed on all cpus. The execution of
809 * the function can only happen on the remote cpus after they have
810 * left the idle function which had been called via pm_idle function
811 * pointer. So it's guaranteed that nothing uses the previous pointer
814 void kick_all_cpus_sync(void)
816 /* Make sure the change is visible before we kick the cpus */
818 smp_call_function(do_nothing, NULL, 1);
820 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);