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>
17 #include <linux/hypervisor.h>
23 CSD_FLAG_SYNCHRONOUS = 0x02,
26 struct call_function_data {
27 struct call_single_data __percpu *csd;
28 cpumask_var_t cpumask;
31 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
33 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
35 static void flush_smp_call_function_queue(bool warn_cpu_offline);
37 int smpcfd_prepare_cpu(unsigned int cpu)
39 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
41 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
44 cfd->csd = alloc_percpu(struct call_single_data);
46 free_cpumask_var(cfd->cpumask);
53 int smpcfd_dead_cpu(unsigned int cpu)
55 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
57 free_cpumask_var(cfd->cpumask);
58 free_percpu(cfd->csd);
62 int smpcfd_dying_cpu(unsigned int cpu)
65 * The IPIs for the smp-call-function callbacks queued by other
66 * CPUs might arrive late, either due to hardware latencies or
67 * because this CPU disabled interrupts (inside stop-machine)
68 * before the IPIs were sent. So flush out any pending callbacks
69 * explicitly (without waiting for the IPIs to arrive), to
70 * ensure that the outgoing CPU doesn't go offline with work
73 flush_smp_call_function_queue(false);
77 void __init call_function_init(void)
81 for_each_possible_cpu(i)
82 init_llist_head(&per_cpu(call_single_queue, i));
84 smpcfd_prepare_cpu(smp_processor_id());
88 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
90 * For non-synchronous ipi calls the csd can still be in use by the
91 * previous function call. For multi-cpu calls its even more interesting
92 * as we'll have to ensure no other cpu is observing our csd.
94 static __always_inline void csd_lock_wait(struct call_single_data *csd)
96 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
99 static __always_inline void csd_lock(struct call_single_data *csd)
102 csd->flags |= CSD_FLAG_LOCK;
105 * prevent CPU from reordering the above assignment
106 * to ->flags with any subsequent assignments to other
107 * fields of the specified call_single_data structure:
112 static __always_inline void csd_unlock(struct call_single_data *csd)
114 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
117 * ensure we're all done before releasing data:
119 smp_store_release(&csd->flags, 0);
122 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
125 * Insert a previously allocated call_single_data element
126 * for execution on the given CPU. data must already have
127 * ->func, ->info, and ->flags set.
129 static int generic_exec_single(int cpu, struct call_single_data *csd,
130 smp_call_func_t func, void *info)
132 if (cpu == smp_processor_id()) {
136 * We can unlock early even for the synchronous on-stack case,
137 * since we're doing this from the same CPU..
140 local_irq_save(flags);
142 local_irq_restore(flags);
147 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
156 * The list addition should be visible before sending the IPI
157 * handler locks the list to pull the entry off it because of
158 * normal cache coherency rules implied by spinlocks.
160 * If IPIs can go out of order to the cache coherency protocol
161 * in an architecture, sufficient synchronisation should be added
162 * to arch code to make it appear to obey cache coherency WRT
163 * locking and barrier primitives. Generic code isn't really
164 * equipped to do the right thing...
166 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
167 arch_send_call_function_single_ipi(cpu);
173 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
175 * Invoked by arch to handle an IPI for call function single.
176 * Must be called with interrupts disabled.
178 void generic_smp_call_function_single_interrupt(void)
180 flush_smp_call_function_queue(true);
184 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
186 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
187 * offline CPU. Skip this check if set to 'false'.
189 * Flush any pending smp-call-function callbacks queued on this CPU. This is
190 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
191 * to ensure that all pending IPI callbacks are run before it goes completely
194 * Loop through the call_single_queue and run all the queued callbacks.
195 * Must be called with interrupts disabled.
197 static void flush_smp_call_function_queue(bool warn_cpu_offline)
199 struct llist_head *head;
200 struct llist_node *entry;
201 struct call_single_data *csd, *csd_next;
204 WARN_ON(!irqs_disabled());
206 head = this_cpu_ptr(&call_single_queue);
207 entry = llist_del_all(head);
208 entry = llist_reverse_order(entry);
210 /* There shouldn't be any pending callbacks on an offline CPU. */
211 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
212 !warned && !llist_empty(head))) {
214 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
217 * We don't have to use the _safe() variant here
218 * because we are not invoking the IPI handlers yet.
220 llist_for_each_entry(csd, entry, llist)
221 pr_warn("IPI callback %pS sent to offline CPU\n",
225 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
226 smp_call_func_t func = csd->func;
227 void *info = csd->info;
229 /* Do we wait until *after* callback? */
230 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
240 * Handle irq works queued remotely by irq_work_queue_on().
241 * Smp functions above are typically synchronous so they
242 * better run first since some other CPUs may be busy waiting
249 * smp_call_function_single - Run a function on a specific CPU
250 * @func: The function to run. This must be fast and non-blocking.
251 * @info: An arbitrary pointer to pass to the function.
252 * @wait: If true, wait until function has completed on other CPUs.
254 * Returns 0 on success, else a negative status code.
256 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
259 struct call_single_data *csd;
260 struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS };
265 * prevent preemption and reschedule on another processor,
266 * as well as CPU removal
268 this_cpu = get_cpu();
271 * Can deadlock when called with interrupts disabled.
272 * We allow cpu's that are not yet online though, as no one else can
273 * send smp call function interrupt to this cpu and as such deadlocks
276 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
277 && !oops_in_progress);
281 csd = this_cpu_ptr(&csd_data);
285 err = generic_exec_single(cpu, csd, func, info);
294 EXPORT_SYMBOL(smp_call_function_single);
297 * smp_call_function_single_async(): Run an asynchronous function on a
299 * @cpu: The CPU to run on.
300 * @csd: Pre-allocated and setup data structure
302 * Like smp_call_function_single(), but the call is asynchonous and
303 * can thus be done from contexts with disabled interrupts.
305 * The caller passes his own pre-allocated data structure
306 * (ie: embedded in an object) and is responsible for synchronizing it
307 * such that the IPIs performed on the @csd are strictly serialized.
309 * NOTE: Be careful, there is unfortunately no current debugging facility to
310 * validate the correctness of this serialization.
312 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
318 /* We could deadlock if we have to wait here with interrupts disabled! */
319 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
322 csd->flags = CSD_FLAG_LOCK;
325 err = generic_exec_single(cpu, csd, csd->func, csd->info);
330 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
333 * smp_call_function_any - Run a function on any of the given cpus
334 * @mask: The mask of cpus it can run on.
335 * @func: The function to run. This must be fast and non-blocking.
336 * @info: An arbitrary pointer to pass to the function.
337 * @wait: If true, wait until function has completed.
339 * Returns 0 on success, else a negative status code (if no cpus were online).
341 * Selection preference:
342 * 1) current cpu if in @mask
343 * 2) any cpu of current node if in @mask
344 * 3) any other online cpu in @mask
346 int smp_call_function_any(const struct cpumask *mask,
347 smp_call_func_t func, void *info, int wait)
350 const struct cpumask *nodemask;
353 /* Try for same CPU (cheapest) */
355 if (cpumask_test_cpu(cpu, mask))
358 /* Try for same node. */
359 nodemask = cpumask_of_node(cpu_to_node(cpu));
360 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
361 cpu = cpumask_next_and(cpu, nodemask, mask)) {
366 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
367 cpu = cpumask_any_and(mask, cpu_online_mask);
369 ret = smp_call_function_single(cpu, func, info, wait);
373 EXPORT_SYMBOL_GPL(smp_call_function_any);
376 * smp_call_function_many(): Run a function on a set of other CPUs.
377 * @mask: The set of cpus to run on (only runs on online subset).
378 * @func: The function to run. This must be fast and non-blocking.
379 * @info: An arbitrary pointer to pass to the function.
380 * @wait: If true, wait (atomically) until function has completed
383 * If @wait is true, then returns once @func has returned.
385 * You must not call this function with disabled interrupts or from a
386 * hardware interrupt handler or from a bottom half handler. Preemption
387 * must be disabled when calling this function.
389 void smp_call_function_many(const struct cpumask *mask,
390 smp_call_func_t func, void *info, bool wait)
392 struct call_function_data *cfd;
393 int cpu, next_cpu, this_cpu = smp_processor_id();
396 * Can deadlock when called with interrupts disabled.
397 * We allow cpu's that are not yet online though, as no one else can
398 * send smp call function interrupt to this cpu and as such deadlocks
401 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
402 && !oops_in_progress && !early_boot_irqs_disabled);
404 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
405 cpu = cpumask_first_and(mask, cpu_online_mask);
407 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
409 /* No online cpus? We're done. */
410 if (cpu >= nr_cpu_ids)
413 /* Do we have another CPU which isn't us? */
414 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
415 if (next_cpu == this_cpu)
416 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
418 /* Fastpath: do that cpu by itself. */
419 if (next_cpu >= nr_cpu_ids) {
420 smp_call_function_single(cpu, func, info, wait);
424 cfd = this_cpu_ptr(&cfd_data);
426 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
427 cpumask_clear_cpu(this_cpu, cfd->cpumask);
429 /* Some callers race with other cpus changing the passed mask */
430 if (unlikely(!cpumask_weight(cfd->cpumask)))
433 for_each_cpu(cpu, cfd->cpumask) {
434 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
438 csd->flags |= CSD_FLAG_SYNCHRONOUS;
441 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
444 /* Send a message to all CPUs in the map */
445 arch_send_call_function_ipi_mask(cfd->cpumask);
448 for_each_cpu(cpu, cfd->cpumask) {
449 struct call_single_data *csd;
451 csd = per_cpu_ptr(cfd->csd, cpu);
456 EXPORT_SYMBOL(smp_call_function_many);
459 * smp_call_function(): Run a function on all other CPUs.
460 * @func: The function to run. This must be fast and non-blocking.
461 * @info: An arbitrary pointer to pass to the function.
462 * @wait: If true, wait (atomically) until function has completed
467 * If @wait is true, then returns once @func has returned; otherwise
468 * it returns just before the target cpu calls @func.
470 * You must not call this function with disabled interrupts or from a
471 * hardware interrupt handler or from a bottom half handler.
473 int smp_call_function(smp_call_func_t func, void *info, int wait)
476 smp_call_function_many(cpu_online_mask, func, info, wait);
481 EXPORT_SYMBOL(smp_call_function);
483 /* Setup configured maximum number of CPUs to activate */
484 unsigned int setup_max_cpus = NR_CPUS;
485 EXPORT_SYMBOL(setup_max_cpus);
489 * Setup routine for controlling SMP activation
491 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
492 * activation entirely (the MPS table probe still happens, though).
494 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
495 * greater than 0, limits the maximum number of CPUs activated in
499 void __weak arch_disable_smp_support(void) { }
501 static int __init nosmp(char *str)
504 arch_disable_smp_support();
509 early_param("nosmp", nosmp);
511 /* this is hard limit */
512 static int __init nrcpus(char *str)
516 get_option(&str, &nr_cpus);
517 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
518 nr_cpu_ids = nr_cpus;
523 early_param("nr_cpus", nrcpus);
525 static int __init maxcpus(char *str)
527 get_option(&str, &setup_max_cpus);
528 if (setup_max_cpus == 0)
529 arch_disable_smp_support();
534 early_param("maxcpus", maxcpus);
536 /* Setup number of possible processor ids */
537 int nr_cpu_ids __read_mostly = NR_CPUS;
538 EXPORT_SYMBOL(nr_cpu_ids);
540 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
541 void __init setup_nr_cpu_ids(void)
543 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
546 void __weak smp_announce(void)
548 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
551 /* Called by boot processor to activate the rest. */
552 void __init smp_init(void)
557 cpuhp_threads_init();
559 /* FIXME: This should be done in userspace --RR */
560 for_each_present_cpu(cpu) {
561 if (num_online_cpus() >= setup_max_cpus)
563 if (!cpu_online(cpu))
567 /* Any cleanup work */
569 smp_cpus_done(setup_max_cpus);
573 * Call a function on all processors. May be used during early boot while
574 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
575 * of local_irq_disable/enable().
577 int on_each_cpu(void (*func) (void *info), void *info, int wait)
583 ret = smp_call_function(func, info, wait);
584 local_irq_save(flags);
586 local_irq_restore(flags);
590 EXPORT_SYMBOL(on_each_cpu);
593 * on_each_cpu_mask(): Run a function on processors specified by
594 * cpumask, which may include the local processor.
595 * @mask: The set of cpus to run on (only runs on online subset).
596 * @func: The function to run. This must be fast and non-blocking.
597 * @info: An arbitrary pointer to pass to the function.
598 * @wait: If true, wait (atomically) until function has completed
601 * If @wait is true, then returns once @func has returned.
603 * You must not call this function with disabled interrupts or from a
604 * hardware interrupt handler or from a bottom half handler. The
605 * exception is that it may be used during early boot while
606 * early_boot_irqs_disabled is set.
608 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
609 void *info, bool wait)
613 smp_call_function_many(mask, func, info, wait);
614 if (cpumask_test_cpu(cpu, mask)) {
616 local_irq_save(flags);
618 local_irq_restore(flags);
622 EXPORT_SYMBOL(on_each_cpu_mask);
625 * on_each_cpu_cond(): Call a function on each processor for which
626 * the supplied function cond_func returns true, optionally waiting
627 * for all the required CPUs to finish. This may include the local
629 * @cond_func: A callback function that is passed a cpu id and
630 * the the info parameter. The function is called
631 * with preemption disabled. The function should
632 * return a blooean value indicating whether to IPI
634 * @func: The function to run on all applicable CPUs.
635 * This must be fast and non-blocking.
636 * @info: An arbitrary pointer to pass to both functions.
637 * @wait: If true, wait (atomically) until function has
638 * completed on other CPUs.
639 * @gfp_flags: GFP flags to use when allocating the cpumask
640 * used internally by the function.
642 * The function might sleep if the GFP flags indicates a non
643 * atomic allocation is allowed.
645 * Preemption is disabled to protect against CPUs going offline but not online.
646 * CPUs going online during the call will not be seen or sent an IPI.
648 * You must not call this function with disabled interrupts or
649 * from a hardware interrupt handler or from a bottom half handler.
651 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
652 smp_call_func_t func, void *info, bool wait,
658 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
660 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
662 for_each_online_cpu(cpu)
663 if (cond_func(cpu, info))
664 cpumask_set_cpu(cpu, cpus);
665 on_each_cpu_mask(cpus, func, info, wait);
667 free_cpumask_var(cpus);
670 * No free cpumask, bother. No matter, we'll
671 * just have to IPI them one by one.
674 for_each_online_cpu(cpu)
675 if (cond_func(cpu, info)) {
676 ret = smp_call_function_single(cpu, func,
683 EXPORT_SYMBOL(on_each_cpu_cond);
685 static void do_nothing(void *unused)
690 * kick_all_cpus_sync - Force all cpus out of idle
692 * Used to synchronize the update of pm_idle function pointer. It's
693 * called after the pointer is updated and returns after the dummy
694 * callback function has been executed on all cpus. The execution of
695 * the function can only happen on the remote cpus after they have
696 * left the idle function which had been called via pm_idle function
697 * pointer. So it's guaranteed that nothing uses the previous pointer
700 void kick_all_cpus_sync(void)
702 /* Make sure the change is visible before we kick the cpus */
704 smp_call_function(do_nothing, NULL, 1);
706 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
709 * wake_up_all_idle_cpus - break all cpus out of idle
710 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
711 * including idle polling cpus, for non-idle cpus, we will do nothing
714 void wake_up_all_idle_cpus(void)
719 for_each_online_cpu(cpu) {
720 if (cpu == smp_processor_id())
723 wake_up_if_idle(cpu);
727 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
730 * smp_call_on_cpu - Call a function on a specific cpu
732 * Used to call a function on a specific cpu and wait for it to return.
733 * Optionally make sure the call is done on a specified physical cpu via vcpu
734 * pinning in order to support virtualized environments.
736 struct smp_call_on_cpu_struct {
737 struct work_struct work;
738 struct completion done;
745 static void smp_call_on_cpu_callback(struct work_struct *work)
747 struct smp_call_on_cpu_struct *sscs;
749 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
751 hypervisor_pin_vcpu(sscs->cpu);
752 sscs->ret = sscs->func(sscs->data);
754 hypervisor_pin_vcpu(-1);
756 complete(&sscs->done);
759 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
761 struct smp_call_on_cpu_struct sscs = {
762 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
765 .cpu = phys ? cpu : -1,
768 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
770 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
773 queue_work_on(cpu, system_wq, &sscs.work);
774 wait_for_completion(&sscs.done);
778 EXPORT_SYMBOL_GPL(smp_call_on_cpu);