2 * Generic helpers for smp ipi calls
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/irq_work.h>
10 #include <linux/rcupdate.h>
11 #include <linux/rculist.h>
12 #include <linux/kernel.h>
13 #include <linux/export.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
16 #include <linux/gfp.h>
17 #include <linux/smp.h>
18 #include <linux/cpu.h>
19 #include <linux/sched.h>
20 #include <linux/sched/idle.h>
21 #include <linux/hypervisor.h>
27 CSD_FLAG_SYNCHRONOUS = 0x02,
30 struct call_function_data {
31 call_single_data_t __percpu *csd;
32 cpumask_var_t cpumask;
33 cpumask_var_t cpumask_ipi;
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
38 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
40 static void flush_smp_call_function_queue(bool warn_cpu_offline);
42 int smpcfd_prepare_cpu(unsigned int cpu)
44 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
46 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
49 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
51 free_cpumask_var(cfd->cpumask);
54 cfd->csd = alloc_percpu(call_single_data_t);
56 free_cpumask_var(cfd->cpumask);
57 free_cpumask_var(cfd->cpumask_ipi);
64 int smpcfd_dead_cpu(unsigned int cpu)
66 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
68 free_cpumask_var(cfd->cpumask);
69 free_cpumask_var(cfd->cpumask_ipi);
70 free_percpu(cfd->csd);
74 int smpcfd_dying_cpu(unsigned int cpu)
77 * The IPIs for the smp-call-function callbacks queued by other
78 * CPUs might arrive late, either due to hardware latencies or
79 * because this CPU disabled interrupts (inside stop-machine)
80 * before the IPIs were sent. So flush out any pending callbacks
81 * explicitly (without waiting for the IPIs to arrive), to
82 * ensure that the outgoing CPU doesn't go offline with work
85 flush_smp_call_function_queue(false);
89 void __init call_function_init(void)
93 for_each_possible_cpu(i)
94 init_llist_head(&per_cpu(call_single_queue, i));
96 smpcfd_prepare_cpu(smp_processor_id());
100 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
102 * For non-synchronous ipi calls the csd can still be in use by the
103 * previous function call. For multi-cpu calls its even more interesting
104 * as we'll have to ensure no other cpu is observing our csd.
106 static __always_inline void csd_lock_wait(call_single_data_t *csd)
108 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
111 static __always_inline void csd_lock(call_single_data_t *csd)
114 csd->flags |= CSD_FLAG_LOCK;
117 * prevent CPU from reordering the above assignment
118 * to ->flags with any subsequent assignments to other
119 * fields of the specified call_single_data_t structure:
124 static __always_inline void csd_unlock(call_single_data_t *csd)
126 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
129 * ensure we're all done before releasing data:
131 smp_store_release(&csd->flags, 0);
134 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
137 * Insert a previously allocated call_single_data_t element
138 * for execution on the given CPU. data must already have
139 * ->func, ->info, and ->flags set.
141 static int generic_exec_single(int cpu, call_single_data_t *csd,
142 smp_call_func_t func, void *info)
144 if (cpu == smp_processor_id()) {
148 * We can unlock early even for the synchronous on-stack case,
149 * since we're doing this from the same CPU..
152 local_irq_save(flags);
154 local_irq_restore(flags);
159 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
168 * The list addition should be visible before sending the IPI
169 * handler locks the list to pull the entry off it because of
170 * normal cache coherency rules implied by spinlocks.
172 * If IPIs can go out of order to the cache coherency protocol
173 * in an architecture, sufficient synchronisation should be added
174 * to arch code to make it appear to obey cache coherency WRT
175 * locking and barrier primitives. Generic code isn't really
176 * equipped to do the right thing...
178 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
179 arch_send_call_function_single_ipi(cpu);
185 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
187 * Invoked by arch to handle an IPI for call function single.
188 * Must be called with interrupts disabled.
190 void generic_smp_call_function_single_interrupt(void)
192 flush_smp_call_function_queue(true);
196 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
198 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
199 * offline CPU. Skip this check if set to 'false'.
201 * Flush any pending smp-call-function callbacks queued on this CPU. This is
202 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
203 * to ensure that all pending IPI callbacks are run before it goes completely
206 * Loop through the call_single_queue and run all the queued callbacks.
207 * Must be called with interrupts disabled.
209 static void flush_smp_call_function_queue(bool warn_cpu_offline)
211 struct llist_head *head;
212 struct llist_node *entry;
213 call_single_data_t *csd, *csd_next;
216 lockdep_assert_irqs_disabled();
218 head = this_cpu_ptr(&call_single_queue);
219 entry = llist_del_all(head);
220 entry = llist_reverse_order(entry);
222 /* There shouldn't be any pending callbacks on an offline CPU. */
223 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
224 !warned && !llist_empty(head))) {
226 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
229 * We don't have to use the _safe() variant here
230 * because we are not invoking the IPI handlers yet.
232 llist_for_each_entry(csd, entry, llist)
233 pr_warn("IPI callback %pS sent to offline CPU\n",
237 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
238 smp_call_func_t func = csd->func;
239 void *info = csd->info;
241 /* Do we wait until *after* callback? */
242 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
252 * Handle irq works queued remotely by irq_work_queue_on().
253 * Smp functions above are typically synchronous so they
254 * better run first since some other CPUs may be busy waiting
261 * smp_call_function_single - Run a function on a specific CPU
262 * @func: The function to run. This must be fast and non-blocking.
263 * @info: An arbitrary pointer to pass to the function.
264 * @wait: If true, wait until function has completed on other CPUs.
266 * Returns 0 on success, else a negative status code.
268 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
271 call_single_data_t *csd;
272 call_single_data_t csd_stack = {
273 .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS,
279 * prevent preemption and reschedule on another processor,
280 * as well as CPU removal
282 this_cpu = get_cpu();
285 * Can deadlock when called with interrupts disabled.
286 * We allow cpu's that are not yet online though, as no one else can
287 * send smp call function interrupt to this cpu and as such deadlocks
290 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
291 && !oops_in_progress);
295 csd = this_cpu_ptr(&csd_data);
299 err = generic_exec_single(cpu, csd, func, info);
308 EXPORT_SYMBOL(smp_call_function_single);
311 * smp_call_function_single_async(): Run an asynchronous function on a
313 * @cpu: The CPU to run on.
314 * @csd: Pre-allocated and setup data structure
316 * Like smp_call_function_single(), but the call is asynchonous and
317 * can thus be done from contexts with disabled interrupts.
319 * The caller passes his own pre-allocated data structure
320 * (ie: embedded in an object) and is responsible for synchronizing it
321 * such that the IPIs performed on the @csd are strictly serialized.
323 * NOTE: Be careful, there is unfortunately no current debugging facility to
324 * validate the correctness of this serialization.
326 int smp_call_function_single_async(int cpu, call_single_data_t *csd)
332 /* We could deadlock if we have to wait here with interrupts disabled! */
333 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
336 csd->flags = CSD_FLAG_LOCK;
339 err = generic_exec_single(cpu, csd, csd->func, csd->info);
344 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
347 * smp_call_function_any - Run a function on any of the given cpus
348 * @mask: The mask of cpus it can run on.
349 * @func: The function to run. This must be fast and non-blocking.
350 * @info: An arbitrary pointer to pass to the function.
351 * @wait: If true, wait until function has completed.
353 * Returns 0 on success, else a negative status code (if no cpus were online).
355 * Selection preference:
356 * 1) current cpu if in @mask
357 * 2) any cpu of current node if in @mask
358 * 3) any other online cpu in @mask
360 int smp_call_function_any(const struct cpumask *mask,
361 smp_call_func_t func, void *info, int wait)
364 const struct cpumask *nodemask;
367 /* Try for same CPU (cheapest) */
369 if (cpumask_test_cpu(cpu, mask))
372 /* Try for same node. */
373 nodemask = cpumask_of_node(cpu_to_node(cpu));
374 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
375 cpu = cpumask_next_and(cpu, nodemask, mask)) {
380 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
381 cpu = cpumask_any_and(mask, cpu_online_mask);
383 ret = smp_call_function_single(cpu, func, info, wait);
387 EXPORT_SYMBOL_GPL(smp_call_function_any);
390 * smp_call_function_many(): Run a function on a set of other CPUs.
391 * @mask: The set of cpus to run on (only runs on online subset).
392 * @func: The function to run. This must be fast and non-blocking.
393 * @info: An arbitrary pointer to pass to the function.
394 * @wait: If true, wait (atomically) until function has completed
397 * If @wait is true, then returns once @func has returned.
399 * You must not call this function with disabled interrupts or from a
400 * hardware interrupt handler or from a bottom half handler. Preemption
401 * must be disabled when calling this function.
403 void smp_call_function_many(const struct cpumask *mask,
404 smp_call_func_t func, void *info, bool wait)
406 struct call_function_data *cfd;
407 int cpu, next_cpu, this_cpu = smp_processor_id();
410 * Can deadlock when called with interrupts disabled.
411 * We allow cpu's that are not yet online though, as no one else can
412 * send smp call function interrupt to this cpu and as such deadlocks
415 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
416 && !oops_in_progress && !early_boot_irqs_disabled);
418 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
419 cpu = cpumask_first_and(mask, cpu_online_mask);
421 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
423 /* No online cpus? We're done. */
424 if (cpu >= nr_cpu_ids)
427 /* Do we have another CPU which isn't us? */
428 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
429 if (next_cpu == this_cpu)
430 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
432 /* Fastpath: do that cpu by itself. */
433 if (next_cpu >= nr_cpu_ids) {
434 smp_call_function_single(cpu, func, info, wait);
438 cfd = this_cpu_ptr(&cfd_data);
440 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
441 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
443 /* Some callers race with other cpus changing the passed mask */
444 if (unlikely(!cpumask_weight(cfd->cpumask)))
447 cpumask_clear(cfd->cpumask_ipi);
448 for_each_cpu(cpu, cfd->cpumask) {
449 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
453 csd->flags |= CSD_FLAG_SYNCHRONOUS;
456 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
457 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
460 /* Send a message to all CPUs in the map */
461 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
464 for_each_cpu(cpu, cfd->cpumask) {
465 call_single_data_t *csd;
467 csd = per_cpu_ptr(cfd->csd, cpu);
472 EXPORT_SYMBOL(smp_call_function_many);
475 * smp_call_function(): Run a function on all other CPUs.
476 * @func: The function to run. This must be fast and non-blocking.
477 * @info: An arbitrary pointer to pass to the function.
478 * @wait: If true, wait (atomically) until function has completed
483 * If @wait is true, then returns once @func has returned; otherwise
484 * it returns just before the target cpu calls @func.
486 * You must not call this function with disabled interrupts or from a
487 * hardware interrupt handler or from a bottom half handler.
489 int smp_call_function(smp_call_func_t func, void *info, int wait)
492 smp_call_function_many(cpu_online_mask, func, info, wait);
497 EXPORT_SYMBOL(smp_call_function);
499 /* Setup configured maximum number of CPUs to activate */
500 unsigned int setup_max_cpus = NR_CPUS;
501 EXPORT_SYMBOL(setup_max_cpus);
505 * Setup routine for controlling SMP activation
507 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
508 * activation entirely (the MPS table probe still happens, though).
510 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
511 * greater than 0, limits the maximum number of CPUs activated in
515 void __weak arch_disable_smp_support(void) { }
517 static int __init nosmp(char *str)
520 arch_disable_smp_support();
525 early_param("nosmp", nosmp);
527 /* this is hard limit */
528 static int __init nrcpus(char *str)
532 get_option(&str, &nr_cpus);
533 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
534 nr_cpu_ids = nr_cpus;
539 early_param("nr_cpus", nrcpus);
541 static int __init maxcpus(char *str)
543 get_option(&str, &setup_max_cpus);
544 if (setup_max_cpus == 0)
545 arch_disable_smp_support();
550 early_param("maxcpus", maxcpus);
552 /* Setup number of possible processor ids */
553 unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
554 EXPORT_SYMBOL(nr_cpu_ids);
556 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
557 void __init setup_nr_cpu_ids(void)
559 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
562 /* Called by boot processor to activate the rest. */
563 void __init smp_init(void)
565 int num_nodes, num_cpus;
569 cpuhp_threads_init();
571 pr_info("Bringing up secondary CPUs ...\n");
573 /* FIXME: This should be done in userspace --RR */
574 for_each_present_cpu(cpu) {
575 if (num_online_cpus() >= setup_max_cpus)
577 if (!cpu_online(cpu))
581 num_nodes = num_online_nodes();
582 num_cpus = num_online_cpus();
583 pr_info("Brought up %d node%s, %d CPU%s\n",
584 num_nodes, (num_nodes > 1 ? "s" : ""),
585 num_cpus, (num_cpus > 1 ? "s" : ""));
587 /* Any cleanup work */
588 smp_cpus_done(setup_max_cpus);
592 * Call a function on all processors. May be used during early boot while
593 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
594 * of local_irq_disable/enable().
596 int on_each_cpu(void (*func) (void *info), void *info, int wait)
602 ret = smp_call_function(func, info, wait);
603 local_irq_save(flags);
605 local_irq_restore(flags);
609 EXPORT_SYMBOL(on_each_cpu);
612 * on_each_cpu_mask(): Run a function on processors specified by
613 * cpumask, which may include the local processor.
614 * @mask: The set of cpus to run on (only runs on online subset).
615 * @func: The function to run. This must be fast and non-blocking.
616 * @info: An arbitrary pointer to pass to the function.
617 * @wait: If true, wait (atomically) until function has completed
620 * If @wait is true, then returns once @func has returned.
622 * You must not call this function with disabled interrupts or from a
623 * hardware interrupt handler or from a bottom half handler. The
624 * exception is that it may be used during early boot while
625 * early_boot_irqs_disabled is set.
627 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
628 void *info, bool wait)
632 smp_call_function_many(mask, func, info, wait);
633 if (cpumask_test_cpu(cpu, mask)) {
635 local_irq_save(flags);
637 local_irq_restore(flags);
641 EXPORT_SYMBOL(on_each_cpu_mask);
644 * on_each_cpu_cond(): Call a function on each processor for which
645 * the supplied function cond_func returns true, optionally waiting
646 * for all the required CPUs to finish. This may include the local
648 * @cond_func: A callback function that is passed a cpu id and
649 * the the info parameter. The function is called
650 * with preemption disabled. The function should
651 * return a blooean value indicating whether to IPI
653 * @func: The function to run on all applicable CPUs.
654 * This must be fast and non-blocking.
655 * @info: An arbitrary pointer to pass to both functions.
656 * @wait: If true, wait (atomically) until function has
657 * completed on other CPUs.
658 * @gfp_flags: GFP flags to use when allocating the cpumask
659 * used internally by the function.
661 * The function might sleep if the GFP flags indicates a non
662 * atomic allocation is allowed.
664 * Preemption is disabled to protect against CPUs going offline but not online.
665 * CPUs going online during the call will not be seen or sent an IPI.
667 * You must not call this function with disabled interrupts or
668 * from a hardware interrupt handler or from a bottom half handler.
670 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
671 smp_call_func_t func, void *info, bool wait,
677 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
679 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
681 for_each_online_cpu(cpu)
682 if (cond_func(cpu, info))
683 cpumask_set_cpu(cpu, cpus);
684 on_each_cpu_mask(cpus, func, info, wait);
686 free_cpumask_var(cpus);
689 * No free cpumask, bother. No matter, we'll
690 * just have to IPI them one by one.
693 for_each_online_cpu(cpu)
694 if (cond_func(cpu, info)) {
695 ret = smp_call_function_single(cpu, func,
702 EXPORT_SYMBOL(on_each_cpu_cond);
704 static void do_nothing(void *unused)
709 * kick_all_cpus_sync - Force all cpus out of idle
711 * Used to synchronize the update of pm_idle function pointer. It's
712 * called after the pointer is updated and returns after the dummy
713 * callback function has been executed on all cpus. The execution of
714 * the function can only happen on the remote cpus after they have
715 * left the idle function which had been called via pm_idle function
716 * pointer. So it's guaranteed that nothing uses the previous pointer
719 void kick_all_cpus_sync(void)
721 /* Make sure the change is visible before we kick the cpus */
723 smp_call_function(do_nothing, NULL, 1);
725 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
728 * wake_up_all_idle_cpus - break all cpus out of idle
729 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
730 * including idle polling cpus, for non-idle cpus, we will do nothing
733 void wake_up_all_idle_cpus(void)
738 for_each_online_cpu(cpu) {
739 if (cpu == smp_processor_id())
742 wake_up_if_idle(cpu);
746 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
749 * smp_call_on_cpu - Call a function on a specific cpu
751 * Used to call a function on a specific cpu and wait for it to return.
752 * Optionally make sure the call is done on a specified physical cpu via vcpu
753 * pinning in order to support virtualized environments.
755 struct smp_call_on_cpu_struct {
756 struct work_struct work;
757 struct completion done;
764 static void smp_call_on_cpu_callback(struct work_struct *work)
766 struct smp_call_on_cpu_struct *sscs;
768 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
770 hypervisor_pin_vcpu(sscs->cpu);
771 sscs->ret = sscs->func(sscs->data);
773 hypervisor_pin_vcpu(-1);
775 complete(&sscs->done);
778 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
780 struct smp_call_on_cpu_struct sscs = {
781 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
784 .cpu = phys ? cpu : -1,
787 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
789 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
792 queue_work_on(cpu, system_wq, &sscs.work);
793 wait_for_completion(&sscs.done);
797 EXPORT_SYMBOL_GPL(smp_call_on_cpu);