1 // SPDX-License-Identifier: GPL-2.0-only
3 * Generic helpers for smp ipi calls
5 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 #include <linux/irq_work.h>
11 #include <linux/rcupdate.h>
12 #include <linux/rculist.h>
13 #include <linux/kernel.h>
14 #include <linux/export.h>
15 #include <linux/percpu.h>
16 #include <linux/init.h>
17 #include <linux/gfp.h>
18 #include <linux/smp.h>
19 #include <linux/cpu.h>
20 #include <linux/sched.h>
21 #include <linux/sched/idle.h>
22 #include <linux/hypervisor.h>
28 CSD_FLAG_SYNCHRONOUS = 0x02,
31 struct call_function_data {
32 call_single_data_t __percpu *csd;
33 cpumask_var_t cpumask;
34 cpumask_var_t cpumask_ipi;
37 static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
39 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
41 static void flush_smp_call_function_queue(bool warn_cpu_offline);
43 int smpcfd_prepare_cpu(unsigned int cpu)
45 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
47 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
50 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
52 free_cpumask_var(cfd->cpumask);
55 cfd->csd = alloc_percpu(call_single_data_t);
57 free_cpumask_var(cfd->cpumask);
58 free_cpumask_var(cfd->cpumask_ipi);
65 int smpcfd_dead_cpu(unsigned int cpu)
67 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
69 free_cpumask_var(cfd->cpumask);
70 free_cpumask_var(cfd->cpumask_ipi);
71 free_percpu(cfd->csd);
75 int smpcfd_dying_cpu(unsigned int cpu)
78 * The IPIs for the smp-call-function callbacks queued by other
79 * CPUs might arrive late, either due to hardware latencies or
80 * because this CPU disabled interrupts (inside stop-machine)
81 * before the IPIs were sent. So flush out any pending callbacks
82 * explicitly (without waiting for the IPIs to arrive), to
83 * ensure that the outgoing CPU doesn't go offline with work
86 flush_smp_call_function_queue(false);
90 void __init call_function_init(void)
94 for_each_possible_cpu(i)
95 init_llist_head(&per_cpu(call_single_queue, i));
97 smpcfd_prepare_cpu(smp_processor_id());
101 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
103 * For non-synchronous ipi calls the csd can still be in use by the
104 * previous function call. For multi-cpu calls its even more interesting
105 * as we'll have to ensure no other cpu is observing our csd.
107 static __always_inline void csd_lock_wait(call_single_data_t *csd)
109 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
112 static __always_inline void csd_lock(call_single_data_t *csd)
115 csd->flags |= CSD_FLAG_LOCK;
118 * prevent CPU from reordering the above assignment
119 * to ->flags with any subsequent assignments to other
120 * fields of the specified call_single_data_t structure:
125 static __always_inline void csd_unlock(call_single_data_t *csd)
127 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
130 * ensure we're all done before releasing data:
132 smp_store_release(&csd->flags, 0);
135 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
138 * Insert a previously allocated call_single_data_t element
139 * for execution on the given CPU. data must already have
140 * ->func, ->info, and ->flags set.
142 static int generic_exec_single(int cpu, call_single_data_t *csd,
143 smp_call_func_t func, void *info)
145 if (cpu == smp_processor_id()) {
149 * We can unlock early even for the synchronous on-stack case,
150 * since we're doing this from the same CPU..
153 local_irq_save(flags);
155 local_irq_restore(flags);
160 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 call_single_data_t *csd, *csd_next;
217 lockdep_assert_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 call_single_data_t *csd;
273 call_single_data_t csd_stack = {
274 .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS,
280 * prevent preemption and reschedule on another processor,
281 * as well as CPU removal
283 this_cpu = get_cpu();
286 * Can deadlock when called with interrupts disabled.
287 * We allow cpu's that are not yet online though, as no one else can
288 * send smp call function interrupt to this cpu and as such deadlocks
291 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
292 && !oops_in_progress);
295 * When @wait we can deadlock when we interrupt between llist_add() and
296 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
297 * csd_lock() on because the interrupt context uses the same csd
300 WARN_ON_ONCE(!in_task());
304 csd = this_cpu_ptr(&csd_data);
308 err = generic_exec_single(cpu, csd, func, info);
317 EXPORT_SYMBOL(smp_call_function_single);
320 * smp_call_function_single_async(): Run an asynchronous function on a
322 * @cpu: The CPU to run on.
323 * @csd: Pre-allocated and setup data structure
325 * Like smp_call_function_single(), but the call is asynchonous and
326 * can thus be done from contexts with disabled interrupts.
328 * The caller passes his own pre-allocated data structure
329 * (ie: embedded in an object) and is responsible for synchronizing it
330 * such that the IPIs performed on the @csd are strictly serialized.
332 * NOTE: Be careful, there is unfortunately no current debugging facility to
333 * validate the correctness of this serialization.
335 int smp_call_function_single_async(int cpu, call_single_data_t *csd)
341 /* We could deadlock if we have to wait here with interrupts disabled! */
342 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
345 csd->flags = CSD_FLAG_LOCK;
348 err = generic_exec_single(cpu, csd, csd->func, csd->info);
353 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
356 * smp_call_function_any - Run a function on any of the given cpus
357 * @mask: The mask of cpus it can run on.
358 * @func: The function to run. This must be fast and non-blocking.
359 * @info: An arbitrary pointer to pass to the function.
360 * @wait: If true, wait until function has completed.
362 * Returns 0 on success, else a negative status code (if no cpus were online).
364 * Selection preference:
365 * 1) current cpu if in @mask
366 * 2) any cpu of current node if in @mask
367 * 3) any other online cpu in @mask
369 int smp_call_function_any(const struct cpumask *mask,
370 smp_call_func_t func, void *info, int wait)
373 const struct cpumask *nodemask;
376 /* Try for same CPU (cheapest) */
378 if (cpumask_test_cpu(cpu, mask))
381 /* Try for same node. */
382 nodemask = cpumask_of_node(cpu_to_node(cpu));
383 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
384 cpu = cpumask_next_and(cpu, nodemask, mask)) {
389 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
390 cpu = cpumask_any_and(mask, cpu_online_mask);
392 ret = smp_call_function_single(cpu, func, info, wait);
396 EXPORT_SYMBOL_GPL(smp_call_function_any);
399 * smp_call_function_many(): Run a function on a set of other CPUs.
400 * @mask: The set of cpus to run on (only runs on online subset).
401 * @func: The function to run. This must be fast and non-blocking.
402 * @info: An arbitrary pointer to pass to the function.
403 * @wait: If true, wait (atomically) until function has completed
406 * If @wait is true, then returns once @func has returned.
408 * You must not call this function with disabled interrupts or from a
409 * hardware interrupt handler or from a bottom half handler. Preemption
410 * must be disabled when calling this function.
412 void smp_call_function_many(const struct cpumask *mask,
413 smp_call_func_t func, void *info, bool wait)
415 struct call_function_data *cfd;
416 int cpu, next_cpu, this_cpu = smp_processor_id();
419 * Can deadlock when called with interrupts disabled.
420 * We allow cpu's that are not yet online though, as no one else can
421 * send smp call function interrupt to this cpu and as such deadlocks
424 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
425 && !oops_in_progress && !early_boot_irqs_disabled);
428 * When @wait we can deadlock when we interrupt between llist_add() and
429 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
430 * csd_lock() on because the interrupt context uses the same csd
433 WARN_ON_ONCE(!in_task());
435 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
436 cpu = cpumask_first_and(mask, cpu_online_mask);
438 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
440 /* No online cpus? We're done. */
441 if (cpu >= nr_cpu_ids)
444 /* Do we have another CPU which isn't us? */
445 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
446 if (next_cpu == this_cpu)
447 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
449 /* Fastpath: do that cpu by itself. */
450 if (next_cpu >= nr_cpu_ids) {
451 smp_call_function_single(cpu, func, info, wait);
455 cfd = this_cpu_ptr(&cfd_data);
457 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
458 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
460 /* Some callers race with other cpus changing the passed mask */
461 if (unlikely(!cpumask_weight(cfd->cpumask)))
464 cpumask_clear(cfd->cpumask_ipi);
465 for_each_cpu(cpu, cfd->cpumask) {
466 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
470 csd->flags |= CSD_FLAG_SYNCHRONOUS;
473 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
474 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
477 /* Send a message to all CPUs in the map */
478 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
481 for_each_cpu(cpu, cfd->cpumask) {
482 call_single_data_t *csd;
484 csd = per_cpu_ptr(cfd->csd, cpu);
489 EXPORT_SYMBOL(smp_call_function_many);
492 * smp_call_function(): Run a function on all other CPUs.
493 * @func: The function to run. This must be fast and non-blocking.
494 * @info: An arbitrary pointer to pass to the function.
495 * @wait: If true, wait (atomically) until function has completed
500 * If @wait is true, then returns once @func has returned; otherwise
501 * it returns just before the target cpu calls @func.
503 * You must not call this function with disabled interrupts or from a
504 * hardware interrupt handler or from a bottom half handler.
506 void smp_call_function(smp_call_func_t func, void *info, int wait)
509 smp_call_function_many(cpu_online_mask, func, info, wait);
512 EXPORT_SYMBOL(smp_call_function);
514 /* Setup configured maximum number of CPUs to activate */
515 unsigned int setup_max_cpus = NR_CPUS;
516 EXPORT_SYMBOL(setup_max_cpus);
520 * Setup routine for controlling SMP activation
522 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
523 * activation entirely (the MPS table probe still happens, though).
525 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
526 * greater than 0, limits the maximum number of CPUs activated in
530 void __weak arch_disable_smp_support(void) { }
532 static int __init nosmp(char *str)
535 arch_disable_smp_support();
540 early_param("nosmp", nosmp);
542 /* this is hard limit */
543 static int __init nrcpus(char *str)
547 get_option(&str, &nr_cpus);
548 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
549 nr_cpu_ids = nr_cpus;
554 early_param("nr_cpus", nrcpus);
556 static int __init maxcpus(char *str)
558 get_option(&str, &setup_max_cpus);
559 if (setup_max_cpus == 0)
560 arch_disable_smp_support();
565 early_param("maxcpus", maxcpus);
567 /* Setup number of possible processor ids */
568 unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
569 EXPORT_SYMBOL(nr_cpu_ids);
571 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
572 void __init setup_nr_cpu_ids(void)
574 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
577 /* Called by boot processor to activate the rest. */
578 void __init smp_init(void)
580 int num_nodes, num_cpus;
584 cpuhp_threads_init();
586 pr_info("Bringing up secondary CPUs ...\n");
588 /* FIXME: This should be done in userspace --RR */
589 for_each_present_cpu(cpu) {
590 if (num_online_cpus() >= setup_max_cpus)
592 if (!cpu_online(cpu))
596 num_nodes = num_online_nodes();
597 num_cpus = num_online_cpus();
598 pr_info("Brought up %d node%s, %d CPU%s\n",
599 num_nodes, (num_nodes > 1 ? "s" : ""),
600 num_cpus, (num_cpus > 1 ? "s" : ""));
602 /* Any cleanup work */
603 smp_cpus_done(setup_max_cpus);
607 * Call a function on all processors. May be used during early boot while
608 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
609 * of local_irq_disable/enable().
611 void on_each_cpu(void (*func) (void *info), void *info, int wait)
616 smp_call_function(func, info, wait);
617 local_irq_save(flags);
619 local_irq_restore(flags);
622 EXPORT_SYMBOL(on_each_cpu);
625 * on_each_cpu_mask(): Run a function on processors specified by
626 * cpumask, which may include the local processor.
627 * @mask: The set of cpus to run on (only runs on online subset).
628 * @func: The function to run. This must be fast and non-blocking.
629 * @info: An arbitrary pointer to pass to the function.
630 * @wait: If true, wait (atomically) until function has completed
633 * If @wait is true, then returns once @func has returned.
635 * You must not call this function with disabled interrupts or from a
636 * hardware interrupt handler or from a bottom half handler. The
637 * exception is that it may be used during early boot while
638 * early_boot_irqs_disabled is set.
640 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
641 void *info, bool wait)
645 smp_call_function_many(mask, func, info, wait);
646 if (cpumask_test_cpu(cpu, mask)) {
648 local_irq_save(flags);
650 local_irq_restore(flags);
654 EXPORT_SYMBOL(on_each_cpu_mask);
657 * on_each_cpu_cond(): Call a function on each processor for which
658 * the supplied function cond_func returns true, optionally waiting
659 * for all the required CPUs to finish. This may include the local
661 * @cond_func: A callback function that is passed a cpu id and
662 * the the info parameter. The function is called
663 * with preemption disabled. The function should
664 * return a blooean value indicating whether to IPI
666 * @func: The function to run on all applicable CPUs.
667 * This must be fast and non-blocking.
668 * @info: An arbitrary pointer to pass to both functions.
669 * @wait: If true, wait (atomically) until function has
670 * completed on other CPUs.
671 * @gfp_flags: GFP flags to use when allocating the cpumask
672 * used internally by the function.
674 * The function might sleep if the GFP flags indicates a non
675 * atomic allocation is allowed.
677 * Preemption is disabled to protect against CPUs going offline but not online.
678 * CPUs going online during the call will not be seen or sent an IPI.
680 * You must not call this function with disabled interrupts or
681 * from a hardware interrupt handler or from a bottom half handler.
683 void on_each_cpu_cond_mask(bool (*cond_func)(int cpu, void *info),
684 smp_call_func_t func, void *info, bool wait,
685 gfp_t gfp_flags, const struct cpumask *mask)
690 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
692 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
694 for_each_cpu(cpu, mask)
695 if (cond_func(cpu, info))
696 __cpumask_set_cpu(cpu, cpus);
697 on_each_cpu_mask(cpus, func, info, wait);
699 free_cpumask_var(cpus);
702 * No free cpumask, bother. No matter, we'll
703 * just have to IPI them one by one.
706 for_each_cpu(cpu, mask)
707 if (cond_func(cpu, info)) {
708 ret = smp_call_function_single(cpu, func,
715 EXPORT_SYMBOL(on_each_cpu_cond_mask);
717 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
718 smp_call_func_t func, void *info, bool wait,
721 on_each_cpu_cond_mask(cond_func, func, info, wait, gfp_flags,
724 EXPORT_SYMBOL(on_each_cpu_cond);
726 static void do_nothing(void *unused)
731 * kick_all_cpus_sync - Force all cpus out of idle
733 * Used to synchronize the update of pm_idle function pointer. It's
734 * called after the pointer is updated and returns after the dummy
735 * callback function has been executed on all cpus. The execution of
736 * the function can only happen on the remote cpus after they have
737 * left the idle function which had been called via pm_idle function
738 * pointer. So it's guaranteed that nothing uses the previous pointer
741 void kick_all_cpus_sync(void)
743 /* Make sure the change is visible before we kick the cpus */
745 smp_call_function(do_nothing, NULL, 1);
747 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
750 * wake_up_all_idle_cpus - break all cpus out of idle
751 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
752 * including idle polling cpus, for non-idle cpus, we will do nothing
755 void wake_up_all_idle_cpus(void)
760 for_each_online_cpu(cpu) {
761 if (cpu == smp_processor_id())
764 wake_up_if_idle(cpu);
768 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
771 * smp_call_on_cpu - Call a function on a specific cpu
773 * Used to call a function on a specific cpu and wait for it to return.
774 * Optionally make sure the call is done on a specified physical cpu via vcpu
775 * pinning in order to support virtualized environments.
777 struct smp_call_on_cpu_struct {
778 struct work_struct work;
779 struct completion done;
786 static void smp_call_on_cpu_callback(struct work_struct *work)
788 struct smp_call_on_cpu_struct *sscs;
790 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
792 hypervisor_pin_vcpu(sscs->cpu);
793 sscs->ret = sscs->func(sscs->data);
795 hypervisor_pin_vcpu(-1);
797 complete(&sscs->done);
800 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
802 struct smp_call_on_cpu_struct sscs = {
803 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
806 .cpu = phys ? cpu : -1,
809 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
811 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
814 queue_work_on(cpu, system_wq, &sscs.work);
815 wait_for_completion(&sscs.done);
819 EXPORT_SYMBOL_GPL(smp_call_on_cpu);