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))
169 * The list addition should be visible before sending the IPI
170 * handler locks the list to pull the entry off it because of
171 * normal cache coherency rules implied by spinlocks.
173 * If IPIs can go out of order to the cache coherency protocol
174 * in an architecture, sufficient synchronisation should be added
175 * to arch code to make it appear to obey cache coherency WRT
176 * locking and barrier primitives. Generic code isn't really
177 * equipped to do the right thing...
179 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
180 arch_send_call_function_single_ipi(cpu);
186 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
188 * Invoked by arch to handle an IPI for call function single.
189 * Must be called with interrupts disabled.
191 void generic_smp_call_function_single_interrupt(void)
193 flush_smp_call_function_queue(true);
197 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
199 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
200 * offline CPU. Skip this check if set to 'false'.
202 * Flush any pending smp-call-function callbacks queued on this CPU. This is
203 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
204 * to ensure that all pending IPI callbacks are run before it goes completely
207 * Loop through the call_single_queue and run all the queued callbacks.
208 * Must be called with interrupts disabled.
210 static void flush_smp_call_function_queue(bool warn_cpu_offline)
212 struct llist_head *head;
213 struct llist_node *entry;
214 struct call_single_data *csd, *csd_next;
217 WARN_ON(!irqs_disabled());
219 head = this_cpu_ptr(&call_single_queue);
220 entry = llist_del_all(head);
221 entry = llist_reverse_order(entry);
223 /* There shouldn't be any pending callbacks on an offline CPU. */
224 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
225 !warned && !llist_empty(head))) {
227 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
230 * We don't have to use the _safe() variant here
231 * because we are not invoking the IPI handlers yet.
233 llist_for_each_entry(csd, entry, llist)
234 pr_warn("IPI callback %pS sent to offline CPU\n",
238 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
239 smp_call_func_t func = csd->func;
240 void *info = csd->info;
242 /* Do we wait until *after* callback? */
243 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
253 * Handle irq works queued remotely by irq_work_queue_on().
254 * Smp functions above are typically synchronous so they
255 * better run first since some other CPUs may be busy waiting
262 * smp_call_function_single - Run a function on a specific CPU
263 * @func: The function to run. This must be fast and non-blocking.
264 * @info: An arbitrary pointer to pass to the function.
265 * @wait: If true, wait until function has completed on other CPUs.
267 * Returns 0 on success, else a negative status code.
269 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
272 struct call_single_data *csd;
273 struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS };
278 * prevent preemption and reschedule on another processor,
279 * as well as CPU removal
281 this_cpu = get_cpu();
284 * Can deadlock when called with interrupts disabled.
285 * We allow cpu's that are not yet online though, as no one else can
286 * send smp call function interrupt to this cpu and as such deadlocks
289 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
290 && !oops_in_progress);
294 csd = this_cpu_ptr(&csd_data);
298 err = generic_exec_single(cpu, csd, func, info);
307 EXPORT_SYMBOL(smp_call_function_single);
310 * smp_call_function_single_async(): Run an asynchronous function on a
312 * @cpu: The CPU to run on.
313 * @csd: Pre-allocated and setup data structure
315 * Like smp_call_function_single(), but the call is asynchonous and
316 * can thus be done from contexts with disabled interrupts.
318 * The caller passes his own pre-allocated data structure
319 * (ie: embedded in an object) and is responsible for synchronizing it
320 * such that the IPIs performed on the @csd are strictly serialized.
322 * NOTE: Be careful, there is unfortunately no current debugging facility to
323 * validate the correctness of this serialization.
325 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
331 /* We could deadlock if we have to wait here with interrupts disabled! */
332 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
335 csd->flags = CSD_FLAG_LOCK;
338 err = generic_exec_single(cpu, csd, csd->func, csd->info);
343 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
346 * smp_call_function_any - Run a function on any of the given cpus
347 * @mask: The mask of cpus it can run on.
348 * @func: The function to run. This must be fast and non-blocking.
349 * @info: An arbitrary pointer to pass to the function.
350 * @wait: If true, wait until function has completed.
352 * Returns 0 on success, else a negative status code (if no cpus were online).
354 * Selection preference:
355 * 1) current cpu if in @mask
356 * 2) any cpu of current node if in @mask
357 * 3) any other online cpu in @mask
359 int smp_call_function_any(const struct cpumask *mask,
360 smp_call_func_t func, void *info, int wait)
363 const struct cpumask *nodemask;
366 /* Try for same CPU (cheapest) */
368 if (cpumask_test_cpu(cpu, mask))
371 /* Try for same node. */
372 nodemask = cpumask_of_node(cpu_to_node(cpu));
373 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
374 cpu = cpumask_next_and(cpu, nodemask, mask)) {
379 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
380 cpu = cpumask_any_and(mask, cpu_online_mask);
382 ret = smp_call_function_single(cpu, func, info, wait);
386 EXPORT_SYMBOL_GPL(smp_call_function_any);
389 * smp_call_function_many(): Run a function on a set of other CPUs.
390 * @mask: The set of cpus to run on (only runs on online subset).
391 * @func: The function to run. This must be fast and non-blocking.
392 * @info: An arbitrary pointer to pass to the function.
393 * @wait: If true, wait (atomically) until function has completed
396 * If @wait is true, then returns once @func has returned.
398 * You must not call this function with disabled interrupts or from a
399 * hardware interrupt handler or from a bottom half handler. Preemption
400 * must be disabled when calling this function.
402 void smp_call_function_many(const struct cpumask *mask,
403 smp_call_func_t func, void *info, bool wait)
405 struct call_function_data *cfd;
406 int cpu, next_cpu, this_cpu = smp_processor_id();
409 * Can deadlock when called with interrupts disabled.
410 * We allow cpu's that are not yet online though, as no one else can
411 * send smp call function interrupt to this cpu and as such deadlocks
414 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
415 && !oops_in_progress && !early_boot_irqs_disabled);
417 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
418 cpu = cpumask_first_and(mask, cpu_online_mask);
420 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
422 /* No online cpus? We're done. */
423 if (cpu >= nr_cpu_ids)
426 /* Do we have another CPU which isn't us? */
427 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
428 if (next_cpu == this_cpu)
429 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
431 /* Fastpath: do that cpu by itself. */
432 if (next_cpu >= nr_cpu_ids) {
433 smp_call_function_single(cpu, func, info, wait);
437 cfd = this_cpu_ptr(&cfd_data);
439 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
440 cpumask_clear_cpu(this_cpu, cfd->cpumask);
442 /* Some callers race with other cpus changing the passed mask */
443 if (unlikely(!cpumask_weight(cfd->cpumask)))
446 for_each_cpu(cpu, cfd->cpumask) {
447 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
451 csd->flags |= CSD_FLAG_SYNCHRONOUS;
454 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
457 /* Send a message to all CPUs in the map */
458 arch_send_call_function_ipi_mask(cfd->cpumask);
461 for_each_cpu(cpu, cfd->cpumask) {
462 struct call_single_data *csd;
464 csd = per_cpu_ptr(cfd->csd, cpu);
469 EXPORT_SYMBOL(smp_call_function_many);
472 * smp_call_function(): Run a function on all other CPUs.
473 * @func: The function to run. This must be fast and non-blocking.
474 * @info: An arbitrary pointer to pass to the function.
475 * @wait: If true, wait (atomically) until function has completed
480 * If @wait is true, then returns once @func has returned; otherwise
481 * it returns just before the target cpu calls @func.
483 * You must not call this function with disabled interrupts or from a
484 * hardware interrupt handler or from a bottom half handler.
486 int smp_call_function(smp_call_func_t func, void *info, int wait)
489 smp_call_function_many(cpu_online_mask, func, info, wait);
494 EXPORT_SYMBOL(smp_call_function);
496 /* Setup configured maximum number of CPUs to activate */
497 unsigned int setup_max_cpus = NR_CPUS;
498 EXPORT_SYMBOL(setup_max_cpus);
502 * Setup routine for controlling SMP activation
504 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
505 * activation entirely (the MPS table probe still happens, though).
507 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
508 * greater than 0, limits the maximum number of CPUs activated in
512 void __weak arch_disable_smp_support(void) { }
514 static int __init nosmp(char *str)
517 arch_disable_smp_support();
522 early_param("nosmp", nosmp);
524 /* this is hard limit */
525 static int __init nrcpus(char *str)
529 get_option(&str, &nr_cpus);
530 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
531 nr_cpu_ids = nr_cpus;
536 early_param("nr_cpus", nrcpus);
538 static int __init maxcpus(char *str)
540 get_option(&str, &setup_max_cpus);
541 if (setup_max_cpus == 0)
542 arch_disable_smp_support();
547 early_param("maxcpus", maxcpus);
549 /* Setup number of possible processor ids */
550 int nr_cpu_ids __read_mostly = NR_CPUS;
551 EXPORT_SYMBOL(nr_cpu_ids);
553 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
554 void __init setup_nr_cpu_ids(void)
556 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
559 void __weak smp_announce(void)
561 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
564 /* Called by boot processor to activate the rest. */
565 void __init smp_init(void)
571 /* FIXME: This should be done in userspace --RR */
572 for_each_present_cpu(cpu) {
573 if (num_online_cpus() >= setup_max_cpus)
575 if (!cpu_online(cpu))
579 /* Any cleanup work */
581 smp_cpus_done(setup_max_cpus);
585 * Call a function on all processors. May be used during early boot while
586 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
587 * of local_irq_disable/enable().
589 int on_each_cpu(void (*func) (void *info), void *info, int wait)
595 ret = smp_call_function(func, info, wait);
596 local_irq_save(flags);
598 local_irq_restore(flags);
602 EXPORT_SYMBOL(on_each_cpu);
605 * on_each_cpu_mask(): Run a function on processors specified by
606 * cpumask, which may include the local processor.
607 * @mask: The set of cpus to run on (only runs on online subset).
608 * @func: The function to run. This must be fast and non-blocking.
609 * @info: An arbitrary pointer to pass to the function.
610 * @wait: If true, wait (atomically) until function has completed
613 * If @wait is true, then returns once @func has returned.
615 * You must not call this function with disabled interrupts or from a
616 * hardware interrupt handler or from a bottom half handler. The
617 * exception is that it may be used during early boot while
618 * early_boot_irqs_disabled is set.
620 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
621 void *info, bool wait)
625 smp_call_function_many(mask, func, info, wait);
626 if (cpumask_test_cpu(cpu, mask)) {
628 local_irq_save(flags);
630 local_irq_restore(flags);
634 EXPORT_SYMBOL(on_each_cpu_mask);
637 * on_each_cpu_cond(): Call a function on each processor for which
638 * the supplied function cond_func returns true, optionally waiting
639 * for all the required CPUs to finish. This may include the local
641 * @cond_func: A callback function that is passed a cpu id and
642 * the the info parameter. The function is called
643 * with preemption disabled. The function should
644 * return a blooean value indicating whether to IPI
646 * @func: The function to run on all applicable CPUs.
647 * This must be fast and non-blocking.
648 * @info: An arbitrary pointer to pass to both functions.
649 * @wait: If true, wait (atomically) until function has
650 * completed on other CPUs.
651 * @gfp_flags: GFP flags to use when allocating the cpumask
652 * used internally by the function.
654 * The function might sleep if the GFP flags indicates a non
655 * atomic allocation is allowed.
657 * Preemption is disabled to protect against CPUs going offline but not online.
658 * CPUs going online during the call will not be seen or sent an IPI.
660 * You must not call this function with disabled interrupts or
661 * from a hardware interrupt handler or from a bottom half handler.
663 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
664 smp_call_func_t func, void *info, bool wait,
670 might_sleep_if(gfp_flags & __GFP_WAIT);
672 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
674 for_each_online_cpu(cpu)
675 if (cond_func(cpu, info))
676 cpumask_set_cpu(cpu, cpus);
677 on_each_cpu_mask(cpus, func, info, wait);
679 free_cpumask_var(cpus);
682 * No free cpumask, bother. No matter, we'll
683 * just have to IPI them one by one.
686 for_each_online_cpu(cpu)
687 if (cond_func(cpu, info)) {
688 ret = smp_call_function_single(cpu, func,
695 EXPORT_SYMBOL(on_each_cpu_cond);
697 static void do_nothing(void *unused)
702 * kick_all_cpus_sync - Force all cpus out of idle
704 * Used to synchronize the update of pm_idle function pointer. It's
705 * called after the pointer is updated and returns after the dummy
706 * callback function has been executed on all cpus. The execution of
707 * the function can only happen on the remote cpus after they have
708 * left the idle function which had been called via pm_idle function
709 * pointer. So it's guaranteed that nothing uses the previous pointer
712 void kick_all_cpus_sync(void)
714 /* Make sure the change is visible before we kick the cpus */
716 smp_call_function(do_nothing, NULL, 1);
718 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
721 * wake_up_all_idle_cpus - break all cpus out of idle
722 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
723 * including idle polling cpus, for non-idle cpus, we will do nothing
726 void wake_up_all_idle_cpus(void)
731 for_each_online_cpu(cpu) {
732 if (cpu == smp_processor_id())
735 wake_up_if_idle(cpu);
739 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);