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>
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 (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_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
132 * ensure we're all done before releasing data:
136 csd->flags &= ~CSD_FLAG_LOCK;
139 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
142 * Insert a previously allocated call_single_data element
143 * for execution on the given CPU. data must already have
144 * ->func, ->info, and ->flags set.
146 static int generic_exec_single(int cpu, struct call_single_data *csd,
147 smp_call_func_t func, void *info, int wait)
149 struct call_single_data csd_stack = { .flags = 0 };
153 if (cpu == smp_processor_id()) {
154 local_irq_save(flags);
156 local_irq_restore(flags);
161 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu))
168 csd = this_cpu_ptr(&csd_data);
177 csd->flags |= CSD_FLAG_WAIT;
180 * The list addition should be visible before sending the IPI
181 * handler locks the list to pull the entry off it because of
182 * normal cache coherency rules implied by spinlocks.
184 * If IPIs can go out of order to the cache coherency protocol
185 * in an architecture, sufficient synchronisation should be added
186 * to arch code to make it appear to obey cache coherency WRT
187 * locking and barrier primitives. Generic code isn't really
188 * equipped to do the right thing...
190 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
191 arch_send_call_function_single_ipi(cpu);
200 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
202 * Invoked by arch to handle an IPI for call function single.
203 * Must be called with interrupts disabled.
205 void generic_smp_call_function_single_interrupt(void)
207 flush_smp_call_function_queue(true);
211 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
213 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
214 * offline CPU. Skip this check if set to 'false'.
216 * Flush any pending smp-call-function callbacks queued on this CPU. This is
217 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
218 * to ensure that all pending IPI callbacks are run before it goes completely
221 * Loop through the call_single_queue and run all the queued callbacks.
222 * Must be called with interrupts disabled.
224 static void flush_smp_call_function_queue(bool warn_cpu_offline)
226 struct llist_head *head;
227 struct llist_node *entry;
228 struct call_single_data *csd, *csd_next;
231 WARN_ON(!irqs_disabled());
233 head = this_cpu_ptr(&call_single_queue);
234 entry = llist_del_all(head);
235 entry = llist_reverse_order(entry);
237 /* There shouldn't be any pending callbacks on an offline CPU. */
238 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
239 !warned && !llist_empty(head))) {
241 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
244 * We don't have to use the _safe() variant here
245 * because we are not invoking the IPI handlers yet.
247 llist_for_each_entry(csd, entry, llist)
248 pr_warn("IPI callback %pS sent to offline CPU\n",
252 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
253 csd->func(csd->info);
258 * Handle irq works queued remotely by irq_work_queue_on().
259 * Smp functions above are typically synchronous so they
260 * better run first since some other CPUs may be busy waiting
267 * smp_call_function_single - Run a function on a specific CPU
268 * @func: The function to run. This must be fast and non-blocking.
269 * @info: An arbitrary pointer to pass to the function.
270 * @wait: If true, wait until function has completed on other CPUs.
272 * Returns 0 on success, else a negative status code.
274 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
281 * prevent preemption and reschedule on another processor,
282 * as well as CPU removal
284 this_cpu = get_cpu();
287 * Can deadlock when called with interrupts disabled.
288 * We allow cpu's that are not yet online though, as no one else can
289 * send smp call function interrupt to this cpu and as such deadlocks
292 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
293 && !oops_in_progress);
295 err = generic_exec_single(cpu, NULL, func, info, wait);
301 EXPORT_SYMBOL(smp_call_function_single);
304 * smp_call_function_single_async(): Run an asynchronous function on a
306 * @cpu: The CPU to run on.
307 * @csd: Pre-allocated and setup data structure
309 * Like smp_call_function_single(), but the call is asynchonous and
310 * can thus be done from contexts with disabled interrupts.
312 * The caller passes his own pre-allocated data structure
313 * (ie: embedded in an object) and is responsible for synchronizing it
314 * such that the IPIs performed on the @csd are strictly serialized.
316 * NOTE: Be careful, there is unfortunately no current debugging facility to
317 * validate the correctness of this serialization.
319 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
324 err = generic_exec_single(cpu, csd, csd->func, csd->info, 0);
329 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
332 * smp_call_function_any - Run a function on any of the given cpus
333 * @mask: The mask of cpus it can run on.
334 * @func: The function to run. This must be fast and non-blocking.
335 * @info: An arbitrary pointer to pass to the function.
336 * @wait: If true, wait until function has completed.
338 * Returns 0 on success, else a negative status code (if no cpus were online).
340 * Selection preference:
341 * 1) current cpu if in @mask
342 * 2) any cpu of current node if in @mask
343 * 3) any other online cpu in @mask
345 int smp_call_function_any(const struct cpumask *mask,
346 smp_call_func_t func, void *info, int wait)
349 const struct cpumask *nodemask;
352 /* Try for same CPU (cheapest) */
354 if (cpumask_test_cpu(cpu, mask))
357 /* Try for same node. */
358 nodemask = cpumask_of_node(cpu_to_node(cpu));
359 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
360 cpu = cpumask_next_and(cpu, nodemask, mask)) {
365 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
366 cpu = cpumask_any_and(mask, cpu_online_mask);
368 ret = smp_call_function_single(cpu, func, info, wait);
372 EXPORT_SYMBOL_GPL(smp_call_function_any);
375 * smp_call_function_many(): Run a function on a set of other CPUs.
376 * @mask: The set of cpus to run on (only runs on online subset).
377 * @func: The function to run. This must be fast and non-blocking.
378 * @info: An arbitrary pointer to pass to the function.
379 * @wait: If true, wait (atomically) until function has completed
382 * If @wait is true, then returns once @func has returned.
384 * You must not call this function with disabled interrupts or from a
385 * hardware interrupt handler or from a bottom half handler. Preemption
386 * must be disabled when calling this function.
388 void smp_call_function_many(const struct cpumask *mask,
389 smp_call_func_t func, void *info, bool wait)
391 struct call_function_data *cfd;
392 int cpu, next_cpu, this_cpu = smp_processor_id();
395 * Can deadlock when called with interrupts disabled.
396 * We allow cpu's that are not yet online though, as no one else can
397 * send smp call function interrupt to this cpu and as such deadlocks
400 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
401 && !oops_in_progress && !early_boot_irqs_disabled);
403 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
404 cpu = cpumask_first_and(mask, cpu_online_mask);
406 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
408 /* No online cpus? We're done. */
409 if (cpu >= nr_cpu_ids)
412 /* Do we have another CPU which isn't us? */
413 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
414 if (next_cpu == this_cpu)
415 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
417 /* Fastpath: do that cpu by itself. */
418 if (next_cpu >= nr_cpu_ids) {
419 smp_call_function_single(cpu, func, info, wait);
423 cfd = this_cpu_ptr(&cfd_data);
425 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
426 cpumask_clear_cpu(this_cpu, cfd->cpumask);
428 /* Some callers race with other cpus changing the passed mask */
429 if (unlikely(!cpumask_weight(cfd->cpumask)))
432 for_each_cpu(cpu, cfd->cpumask) {
433 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
438 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
441 /* Send a message to all CPUs in the map */
442 arch_send_call_function_ipi_mask(cfd->cpumask);
445 for_each_cpu(cpu, cfd->cpumask) {
446 struct call_single_data *csd;
448 csd = per_cpu_ptr(cfd->csd, cpu);
453 EXPORT_SYMBOL(smp_call_function_many);
456 * smp_call_function(): Run a function on all other CPUs.
457 * @func: The function to run. This must be fast and non-blocking.
458 * @info: An arbitrary pointer to pass to the function.
459 * @wait: If true, wait (atomically) until function has completed
464 * If @wait is true, then returns once @func has returned; otherwise
465 * it returns just before the target cpu calls @func.
467 * You must not call this function with disabled interrupts or from a
468 * hardware interrupt handler or from a bottom half handler.
470 int smp_call_function(smp_call_func_t func, void *info, int wait)
473 smp_call_function_many(cpu_online_mask, func, info, wait);
478 EXPORT_SYMBOL(smp_call_function);
480 /* Setup configured maximum number of CPUs to activate */
481 unsigned int setup_max_cpus = NR_CPUS;
482 EXPORT_SYMBOL(setup_max_cpus);
486 * Setup routine for controlling SMP activation
488 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
489 * activation entirely (the MPS table probe still happens, though).
491 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
492 * greater than 0, limits the maximum number of CPUs activated in
496 void __weak arch_disable_smp_support(void) { }
498 static int __init nosmp(char *str)
501 arch_disable_smp_support();
506 early_param("nosmp", nosmp);
508 /* this is hard limit */
509 static int __init nrcpus(char *str)
513 get_option(&str, &nr_cpus);
514 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
515 nr_cpu_ids = nr_cpus;
520 early_param("nr_cpus", nrcpus);
522 static int __init maxcpus(char *str)
524 get_option(&str, &setup_max_cpus);
525 if (setup_max_cpus == 0)
526 arch_disable_smp_support();
531 early_param("maxcpus", maxcpus);
533 /* Setup number of possible processor ids */
534 int nr_cpu_ids __read_mostly = NR_CPUS;
535 EXPORT_SYMBOL(nr_cpu_ids);
537 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
538 void __init setup_nr_cpu_ids(void)
540 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
543 void __weak smp_announce(void)
545 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
548 /* Called by boot processor to activate the rest. */
549 void __init smp_init(void)
555 /* FIXME: This should be done in userspace --RR */
556 for_each_present_cpu(cpu) {
557 if (num_online_cpus() >= setup_max_cpus)
559 if (!cpu_online(cpu))
563 /* Any cleanup work */
565 smp_cpus_done(setup_max_cpus);
569 * Call a function on all processors. May be used during early boot while
570 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
571 * of local_irq_disable/enable().
573 int on_each_cpu(void (*func) (void *info), void *info, int wait)
579 ret = smp_call_function(func, info, wait);
580 local_irq_save(flags);
582 local_irq_restore(flags);
586 EXPORT_SYMBOL(on_each_cpu);
589 * on_each_cpu_mask(): Run a function on processors specified by
590 * cpumask, which may include the local processor.
591 * @mask: The set of cpus to run on (only runs on online subset).
592 * @func: The function to run. This must be fast and non-blocking.
593 * @info: An arbitrary pointer to pass to the function.
594 * @wait: If true, wait (atomically) until function has completed
597 * If @wait is true, then returns once @func has returned.
599 * You must not call this function with disabled interrupts or from a
600 * hardware interrupt handler or from a bottom half handler. The
601 * exception is that it may be used during early boot while
602 * early_boot_irqs_disabled is set.
604 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
605 void *info, bool wait)
609 smp_call_function_many(mask, func, info, wait);
610 if (cpumask_test_cpu(cpu, mask)) {
612 local_irq_save(flags);
614 local_irq_restore(flags);
618 EXPORT_SYMBOL(on_each_cpu_mask);
621 * on_each_cpu_cond(): Call a function on each processor for which
622 * the supplied function cond_func returns true, optionally waiting
623 * for all the required CPUs to finish. This may include the local
625 * @cond_func: A callback function that is passed a cpu id and
626 * the the info parameter. The function is called
627 * with preemption disabled. The function should
628 * return a blooean value indicating whether to IPI
630 * @func: The function to run on all applicable CPUs.
631 * This must be fast and non-blocking.
632 * @info: An arbitrary pointer to pass to both functions.
633 * @wait: If true, wait (atomically) until function has
634 * completed on other CPUs.
635 * @gfp_flags: GFP flags to use when allocating the cpumask
636 * used internally by the function.
638 * The function might sleep if the GFP flags indicates a non
639 * atomic allocation is allowed.
641 * Preemption is disabled to protect against CPUs going offline but not online.
642 * CPUs going online during the call will not be seen or sent an IPI.
644 * You must not call this function with disabled interrupts or
645 * from a hardware interrupt handler or from a bottom half handler.
647 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
648 smp_call_func_t func, void *info, bool wait,
654 might_sleep_if(gfp_flags & __GFP_WAIT);
656 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
658 for_each_online_cpu(cpu)
659 if (cond_func(cpu, info))
660 cpumask_set_cpu(cpu, cpus);
661 on_each_cpu_mask(cpus, func, info, wait);
663 free_cpumask_var(cpus);
666 * No free cpumask, bother. No matter, we'll
667 * just have to IPI them one by one.
670 for_each_online_cpu(cpu)
671 if (cond_func(cpu, info)) {
672 ret = smp_call_function_single(cpu, func,
679 EXPORT_SYMBOL(on_each_cpu_cond);
681 static void do_nothing(void *unused)
686 * kick_all_cpus_sync - Force all cpus out of idle
688 * Used to synchronize the update of pm_idle function pointer. It's
689 * called after the pointer is updated and returns after the dummy
690 * callback function has been executed on all cpus. The execution of
691 * the function can only happen on the remote cpus after they have
692 * left the idle function which had been called via pm_idle function
693 * pointer. So it's guaranteed that nothing uses the previous pointer
696 void kick_all_cpus_sync(void)
698 /* Make sure the change is visible before we kick the cpus */
700 smp_call_function(do_nothing, NULL, 1);
702 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
705 * wake_up_all_idle_cpus - break all cpus out of idle
706 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
707 * including idle polling cpus, for non-idle cpus, we will do nothing
710 void wake_up_all_idle_cpus(void)
715 for_each_online_cpu(cpu) {
716 if (cpu == smp_processor_id())
719 wake_up_if_idle(cpu);
723 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);