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/interrupt.h>
18 #include <linux/gfp.h>
19 #include <linux/smp.h>
20 #include <linux/cpu.h>
21 #include <linux/sched.h>
22 #include <linux/sched/idle.h>
23 #include <linux/hypervisor.h>
24 #include <linux/sched/clock.h>
25 #include <linux/nmi.h>
26 #include <linux/sched/debug.h>
27 #include <linux/jump_label.h>
30 #include "sched/smp.h"
32 #define CSD_TYPE(_csd) ((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
35 call_single_data_t csd;
38 struct call_function_data {
39 struct cfd_percpu __percpu *pcpu;
40 cpumask_var_t cpumask;
41 cpumask_var_t cpumask_ipi;
44 static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
46 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
48 static void __flush_smp_call_function_queue(bool warn_cpu_offline);
50 int smpcfd_prepare_cpu(unsigned int cpu)
52 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
54 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
57 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
59 free_cpumask_var(cfd->cpumask);
62 cfd->pcpu = alloc_percpu(struct cfd_percpu);
64 free_cpumask_var(cfd->cpumask);
65 free_cpumask_var(cfd->cpumask_ipi);
72 int smpcfd_dead_cpu(unsigned int cpu)
74 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
76 free_cpumask_var(cfd->cpumask);
77 free_cpumask_var(cfd->cpumask_ipi);
78 free_percpu(cfd->pcpu);
82 int smpcfd_dying_cpu(unsigned int cpu)
85 * The IPIs for the smp-call-function callbacks queued by other
86 * CPUs might arrive late, either due to hardware latencies or
87 * because this CPU disabled interrupts (inside stop-machine)
88 * before the IPIs were sent. So flush out any pending callbacks
89 * explicitly (without waiting for the IPIs to arrive), to
90 * ensure that the outgoing CPU doesn't go offline with work
93 __flush_smp_call_function_queue(false);
98 void __init call_function_init(void)
102 for_each_possible_cpu(i)
103 init_llist_head(&per_cpu(call_single_queue, i));
105 smpcfd_prepare_cpu(smp_processor_id());
108 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
110 static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT, csdlock_debug_enabled);
113 * Parse the csdlock_debug= kernel boot parameter.
115 * If you need to restore the old "ext" value that once provided
116 * additional debugging information, reapply the following commits:
118 * de7b09ef658d ("locking/csd_lock: Prepare more CSD lock debugging")
119 * a5aabace5fb8 ("locking/csd_lock: Add more data to CSD lock debugging")
121 static int __init csdlock_debug(char *str)
123 unsigned int val = 0;
125 get_option(&str, &val);
127 static_branch_enable(&csdlock_debug_enabled);
131 __setup("csdlock_debug=", csdlock_debug);
133 static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
134 static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
135 static DEFINE_PER_CPU(void *, cur_csd_info);
137 static ulong csd_lock_timeout = 5000; /* CSD lock timeout in milliseconds. */
138 module_param(csd_lock_timeout, ulong, 0444);
140 static atomic_t csd_bug_count = ATOMIC_INIT(0);
142 /* Record current CSD work for current CPU, NULL to erase. */
143 static void __csd_lock_record(struct __call_single_data *csd)
146 smp_mb(); /* NULL cur_csd after unlock. */
147 __this_cpu_write(cur_csd, NULL);
150 __this_cpu_write(cur_csd_func, csd->func);
151 __this_cpu_write(cur_csd_info, csd->info);
152 smp_wmb(); /* func and info before csd. */
153 __this_cpu_write(cur_csd, csd);
154 smp_mb(); /* Update cur_csd before function call. */
155 /* Or before unlock, as the case may be. */
158 static __always_inline void csd_lock_record(struct __call_single_data *csd)
160 if (static_branch_unlikely(&csdlock_debug_enabled))
161 __csd_lock_record(csd);
164 static int csd_lock_wait_getcpu(struct __call_single_data *csd)
166 unsigned int csd_type;
168 csd_type = CSD_TYPE(csd);
169 if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
170 return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */
175 * Complain if too much time spent waiting. Note that only
176 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
177 * so waiting on other types gets much less information.
179 static bool csd_lock_wait_toolong(struct __call_single_data *csd, u64 ts0, u64 *ts1, int *bug_id)
185 call_single_data_t *cpu_cur_csd;
186 unsigned int flags = READ_ONCE(csd->node.u_flags);
187 unsigned long long csd_lock_timeout_ns = csd_lock_timeout * NSEC_PER_MSEC;
189 if (!(flags & CSD_FLAG_LOCK)) {
190 if (!unlikely(*bug_id))
192 cpu = csd_lock_wait_getcpu(csd);
193 pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
194 *bug_id, raw_smp_processor_id(), cpu);
199 ts_delta = ts2 - *ts1;
200 if (likely(ts_delta <= csd_lock_timeout_ns || csd_lock_timeout_ns == 0))
203 firsttime = !*bug_id;
205 *bug_id = atomic_inc_return(&csd_bug_count);
206 cpu = csd_lock_wait_getcpu(csd);
207 if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
211 cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
212 pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
213 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0,
214 cpu, csd->func, csd->info);
215 if (cpu_cur_csd && csd != cpu_cur_csd) {
216 pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
217 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
218 READ_ONCE(per_cpu(cur_csd_info, cpux)));
220 pr_alert("\tcsd: CSD lock (#%d) %s.\n",
221 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
226 pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
227 arch_send_call_function_single_ipi(cpu);
237 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
239 * For non-synchronous ipi calls the csd can still be in use by the
240 * previous function call. For multi-cpu calls its even more interesting
241 * as we'll have to ensure no other cpu is observing our csd.
243 static void __csd_lock_wait(struct __call_single_data *csd)
248 ts1 = ts0 = sched_clock();
250 if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
254 smp_acquire__after_ctrl_dep();
257 static __always_inline void csd_lock_wait(struct __call_single_data *csd)
259 if (static_branch_unlikely(&csdlock_debug_enabled)) {
260 __csd_lock_wait(csd);
264 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
267 static void csd_lock_record(struct __call_single_data *csd)
271 static __always_inline void csd_lock_wait(struct __call_single_data *csd)
273 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
277 static __always_inline void csd_lock(struct __call_single_data *csd)
280 csd->node.u_flags |= CSD_FLAG_LOCK;
283 * prevent CPU from reordering the above assignment
284 * to ->flags with any subsequent assignments to other
285 * fields of the specified call_single_data_t structure:
290 static __always_inline void csd_unlock(struct __call_single_data *csd)
292 WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK));
295 * ensure we're all done before releasing data:
297 smp_store_release(&csd->node.u_flags, 0);
300 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
302 void __smp_call_single_queue(int cpu, struct llist_node *node)
305 * The list addition should be visible before sending the IPI
306 * handler locks the list to pull the entry off it because of
307 * normal cache coherency rules implied by spinlocks.
309 * If IPIs can go out of order to the cache coherency protocol
310 * in an architecture, sufficient synchronisation should be added
311 * to arch code to make it appear to obey cache coherency WRT
312 * locking and barrier primitives. Generic code isn't really
313 * equipped to do the right thing...
315 if (llist_add(node, &per_cpu(call_single_queue, cpu)))
316 send_call_function_single_ipi(cpu);
320 * Insert a previously allocated call_single_data_t element
321 * for execution on the given CPU. data must already have
322 * ->func, ->info, and ->flags set.
324 static int generic_exec_single(int cpu, struct __call_single_data *csd)
326 if (cpu == smp_processor_id()) {
327 smp_call_func_t func = csd->func;
328 void *info = csd->info;
332 * We can unlock early even for the synchronous on-stack case,
333 * since we're doing this from the same CPU..
335 csd_lock_record(csd);
337 local_irq_save(flags);
339 csd_lock_record(NULL);
340 local_irq_restore(flags);
344 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
349 __smp_call_single_queue(cpu, &csd->node.llist);
355 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
357 * Invoked by arch to handle an IPI for call function single.
358 * Must be called with interrupts disabled.
360 void generic_smp_call_function_single_interrupt(void)
362 __flush_smp_call_function_queue(true);
366 * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks
368 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
369 * offline CPU. Skip this check if set to 'false'.
371 * Flush any pending smp-call-function callbacks queued on this CPU. This is
372 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
373 * to ensure that all pending IPI callbacks are run before it goes completely
376 * Loop through the call_single_queue and run all the queued callbacks.
377 * Must be called with interrupts disabled.
379 static void __flush_smp_call_function_queue(bool warn_cpu_offline)
381 call_single_data_t *csd, *csd_next;
382 struct llist_node *entry, *prev;
383 struct llist_head *head;
386 lockdep_assert_irqs_disabled();
388 head = this_cpu_ptr(&call_single_queue);
389 entry = llist_del_all(head);
390 entry = llist_reverse_order(entry);
392 /* There shouldn't be any pending callbacks on an offline CPU. */
393 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
394 !warned && entry != NULL)) {
396 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
399 * We don't have to use the _safe() variant here
400 * because we are not invoking the IPI handlers yet.
402 llist_for_each_entry(csd, entry, node.llist) {
403 switch (CSD_TYPE(csd)) {
406 case CSD_TYPE_IRQ_WORK:
407 pr_warn("IPI callback %pS sent to offline CPU\n",
412 pr_warn("IPI task-wakeup sent to offline CPU\n");
416 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
424 * First; run all SYNC callbacks, people are waiting for us.
427 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
428 /* Do we wait until *after* callback? */
429 if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
430 smp_call_func_t func = csd->func;
431 void *info = csd->info;
434 prev->next = &csd_next->node.llist;
436 entry = &csd_next->node.llist;
439 csd_lock_record(csd);
442 csd_lock_record(NULL);
444 prev = &csd->node.llist;
452 * Second; run all !SYNC callbacks.
455 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
456 int type = CSD_TYPE(csd);
458 if (type != CSD_TYPE_TTWU) {
460 prev->next = &csd_next->node.llist;
462 entry = &csd_next->node.llist;
465 if (type == CSD_TYPE_ASYNC) {
466 smp_call_func_t func = csd->func;
467 void *info = csd->info;
469 csd_lock_record(csd);
472 csd_lock_record(NULL);
473 } else if (type == CSD_TYPE_IRQ_WORK) {
474 irq_work_single(csd);
478 prev = &csd->node.llist;
483 * Third; only CSD_TYPE_TTWU is left, issue those.
486 sched_ttwu_pending(entry);
491 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
492 * from task context (idle, migration thread)
494 * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it
495 * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by
496 * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to
497 * handle queued SMP function calls before scheduling.
499 * The migration thread has to ensure that an eventually pending wakeup has
500 * been handled before it migrates a task.
502 void flush_smp_call_function_queue(void)
504 unsigned int was_pending;
507 if (llist_empty(this_cpu_ptr(&call_single_queue)))
510 local_irq_save(flags);
511 /* Get the already pending soft interrupts for RT enabled kernels */
512 was_pending = local_softirq_pending();
513 __flush_smp_call_function_queue(true);
514 if (local_softirq_pending())
515 do_softirq_post_smp_call_flush(was_pending);
517 local_irq_restore(flags);
521 * smp_call_function_single - Run a function on a specific CPU
522 * @func: The function to run. This must be fast and non-blocking.
523 * @info: An arbitrary pointer to pass to the function.
524 * @wait: If true, wait until function has completed on other CPUs.
526 * Returns 0 on success, else a negative status code.
528 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
531 call_single_data_t *csd;
532 call_single_data_t csd_stack = {
533 .node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, },
539 * prevent preemption and reschedule on another processor,
540 * as well as CPU removal
542 this_cpu = get_cpu();
545 * Can deadlock when called with interrupts disabled.
546 * We allow cpu's that are not yet online though, as no one else can
547 * send smp call function interrupt to this cpu and as such deadlocks
550 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
551 && !oops_in_progress);
554 * When @wait we can deadlock when we interrupt between llist_add() and
555 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
556 * csd_lock() on because the interrupt context uses the same csd
559 WARN_ON_ONCE(!in_task());
563 csd = this_cpu_ptr(&csd_data);
569 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
570 csd->node.src = smp_processor_id();
574 err = generic_exec_single(cpu, csd);
583 EXPORT_SYMBOL(smp_call_function_single);
586 * smp_call_function_single_async() - Run an asynchronous function on a
588 * @cpu: The CPU to run on.
589 * @csd: Pre-allocated and setup data structure
591 * Like smp_call_function_single(), but the call is asynchonous and
592 * can thus be done from contexts with disabled interrupts.
594 * The caller passes his own pre-allocated data structure
595 * (ie: embedded in an object) and is responsible for synchronizing it
596 * such that the IPIs performed on the @csd are strictly serialized.
598 * If the function is called with one csd which has not yet been
599 * processed by previous call to smp_call_function_single_async(), the
600 * function will return immediately with -EBUSY showing that the csd
601 * object is still in progress.
603 * NOTE: Be careful, there is unfortunately no current debugging facility to
604 * validate the correctness of this serialization.
606 * Return: %0 on success or negative errno value on error
608 int smp_call_function_single_async(int cpu, struct __call_single_data *csd)
614 if (csd->node.u_flags & CSD_FLAG_LOCK) {
619 csd->node.u_flags = CSD_FLAG_LOCK;
622 err = generic_exec_single(cpu, csd);
629 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
632 * smp_call_function_any - Run a function on any of the given cpus
633 * @mask: The mask of cpus it can run on.
634 * @func: The function to run. This must be fast and non-blocking.
635 * @info: An arbitrary pointer to pass to the function.
636 * @wait: If true, wait until function has completed.
638 * Returns 0 on success, else a negative status code (if no cpus were online).
640 * Selection preference:
641 * 1) current cpu if in @mask
642 * 2) any cpu of current node if in @mask
643 * 3) any other online cpu in @mask
645 int smp_call_function_any(const struct cpumask *mask,
646 smp_call_func_t func, void *info, int wait)
649 const struct cpumask *nodemask;
652 /* Try for same CPU (cheapest) */
654 if (cpumask_test_cpu(cpu, mask))
657 /* Try for same node. */
658 nodemask = cpumask_of_node(cpu_to_node(cpu));
659 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
660 cpu = cpumask_next_and(cpu, nodemask, mask)) {
665 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
666 cpu = cpumask_any_and(mask, cpu_online_mask);
668 ret = smp_call_function_single(cpu, func, info, wait);
672 EXPORT_SYMBOL_GPL(smp_call_function_any);
675 * Flags to be used as scf_flags argument of smp_call_function_many_cond().
677 * %SCF_WAIT: Wait until function execution is completed
678 * %SCF_RUN_LOCAL: Run also locally if local cpu is set in cpumask
680 #define SCF_WAIT (1U << 0)
681 #define SCF_RUN_LOCAL (1U << 1)
683 static void smp_call_function_many_cond(const struct cpumask *mask,
684 smp_call_func_t func, void *info,
685 unsigned int scf_flags,
686 smp_cond_func_t cond_func)
688 int cpu, last_cpu, this_cpu = smp_processor_id();
689 struct call_function_data *cfd;
690 bool wait = scf_flags & SCF_WAIT;
691 bool run_remote = false;
692 bool run_local = false;
695 lockdep_assert_preemption_disabled();
698 * Can deadlock when called with interrupts disabled.
699 * We allow cpu's that are not yet online though, as no one else can
700 * send smp call function interrupt to this cpu and as such deadlocks
703 if (cpu_online(this_cpu) && !oops_in_progress &&
704 !early_boot_irqs_disabled)
705 lockdep_assert_irqs_enabled();
708 * When @wait we can deadlock when we interrupt between llist_add() and
709 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
710 * csd_lock() on because the interrupt context uses the same csd
713 WARN_ON_ONCE(!in_task());
715 /* Check if we need local execution. */
716 if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask))
719 /* Check if we need remote execution, i.e., any CPU excluding this one. */
720 cpu = cpumask_first_and(mask, cpu_online_mask);
722 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
723 if (cpu < nr_cpu_ids)
727 cfd = this_cpu_ptr(&cfd_data);
728 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
729 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
731 cpumask_clear(cfd->cpumask_ipi);
732 for_each_cpu(cpu, cfd->cpumask) {
733 call_single_data_t *csd = &per_cpu_ptr(cfd->pcpu, cpu)->csd;
735 if (cond_func && !cond_func(cpu, info))
740 csd->node.u_flags |= CSD_TYPE_SYNC;
743 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
744 csd->node.src = smp_processor_id();
747 if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
748 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
755 * Choose the most efficient way to send an IPI. Note that the
756 * number of CPUs might be zero due to concurrent changes to the
760 send_call_function_single_ipi(last_cpu);
761 else if (likely(nr_cpus > 1))
762 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
765 if (run_local && (!cond_func || cond_func(this_cpu, info))) {
768 local_irq_save(flags);
770 local_irq_restore(flags);
773 if (run_remote && wait) {
774 for_each_cpu(cpu, cfd->cpumask) {
775 call_single_data_t *csd;
777 csd = &per_cpu_ptr(cfd->pcpu, cpu)->csd;
784 * smp_call_function_many(): Run a function on a set of CPUs.
785 * @mask: The set of cpus to run on (only runs on online subset).
786 * @func: The function to run. This must be fast and non-blocking.
787 * @info: An arbitrary pointer to pass to the function.
788 * @wait: Bitmask that controls the operation. If %SCF_WAIT is set, wait
789 * (atomically) until function has completed on other CPUs. If
790 * %SCF_RUN_LOCAL is set, the function will also be run locally
791 * if the local CPU is set in the @cpumask.
793 * If @wait is true, then returns once @func has returned.
795 * You must not call this function with disabled interrupts or from a
796 * hardware interrupt handler or from a bottom half handler. Preemption
797 * must be disabled when calling this function.
799 void smp_call_function_many(const struct cpumask *mask,
800 smp_call_func_t func, void *info, bool wait)
802 smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL);
804 EXPORT_SYMBOL(smp_call_function_many);
807 * smp_call_function(): Run a function on all other CPUs.
808 * @func: The function to run. This must be fast and non-blocking.
809 * @info: An arbitrary pointer to pass to the function.
810 * @wait: If true, wait (atomically) until function has completed
815 * If @wait is true, then returns once @func has returned; otherwise
816 * it returns just before the target cpu calls @func.
818 * You must not call this function with disabled interrupts or from a
819 * hardware interrupt handler or from a bottom half handler.
821 void smp_call_function(smp_call_func_t func, void *info, int wait)
824 smp_call_function_many(cpu_online_mask, func, info, wait);
827 EXPORT_SYMBOL(smp_call_function);
829 /* Setup configured maximum number of CPUs to activate */
830 unsigned int setup_max_cpus = NR_CPUS;
831 EXPORT_SYMBOL(setup_max_cpus);
835 * Setup routine for controlling SMP activation
837 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
838 * activation entirely (the MPS table probe still happens, though).
840 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
841 * greater than 0, limits the maximum number of CPUs activated in
845 void __weak arch_disable_smp_support(void) { }
847 static int __init nosmp(char *str)
850 arch_disable_smp_support();
855 early_param("nosmp", nosmp);
857 /* this is hard limit */
858 static int __init nrcpus(char *str)
862 if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
863 set_nr_cpu_ids(nr_cpus);
868 early_param("nr_cpus", nrcpus);
870 static int __init maxcpus(char *str)
872 get_option(&str, &setup_max_cpus);
873 if (setup_max_cpus == 0)
874 arch_disable_smp_support();
879 early_param("maxcpus", maxcpus);
881 #if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS)
882 /* Setup number of possible processor ids */
883 unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
884 EXPORT_SYMBOL(nr_cpu_ids);
887 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
888 void __init setup_nr_cpu_ids(void)
890 set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask), NR_CPUS) + 1);
893 /* Called by boot processor to activate the rest. */
894 void __init smp_init(void)
896 int num_nodes, num_cpus;
899 cpuhp_threads_init();
901 pr_info("Bringing up secondary CPUs ...\n");
903 bringup_nonboot_cpus(setup_max_cpus);
905 num_nodes = num_online_nodes();
906 num_cpus = num_online_cpus();
907 pr_info("Brought up %d node%s, %d CPU%s\n",
908 num_nodes, (num_nodes > 1 ? "s" : ""),
909 num_cpus, (num_cpus > 1 ? "s" : ""));
911 /* Any cleanup work */
912 smp_cpus_done(setup_max_cpus);
916 * on_each_cpu_cond(): Call a function on each processor for which
917 * the supplied function cond_func returns true, optionally waiting
918 * for all the required CPUs to finish. This may include the local
920 * @cond_func: A callback function that is passed a cpu id and
921 * the info parameter. The function is called
922 * with preemption disabled. The function should
923 * return a blooean value indicating whether to IPI
925 * @func: The function to run on all applicable CPUs.
926 * This must be fast and non-blocking.
927 * @info: An arbitrary pointer to pass to both functions.
928 * @wait: If true, wait (atomically) until function has
929 * completed on other CPUs.
931 * Preemption is disabled to protect against CPUs going offline but not online.
932 * CPUs going online during the call will not be seen or sent an IPI.
934 * You must not call this function with disabled interrupts or
935 * from a hardware interrupt handler or from a bottom half handler.
937 void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
938 void *info, bool wait, const struct cpumask *mask)
940 unsigned int scf_flags = SCF_RUN_LOCAL;
943 scf_flags |= SCF_WAIT;
946 smp_call_function_many_cond(mask, func, info, scf_flags, cond_func);
949 EXPORT_SYMBOL(on_each_cpu_cond_mask);
951 static void do_nothing(void *unused)
956 * kick_all_cpus_sync - Force all cpus out of idle
958 * Used to synchronize the update of pm_idle function pointer. It's
959 * called after the pointer is updated and returns after the dummy
960 * callback function has been executed on all cpus. The execution of
961 * the function can only happen on the remote cpus after they have
962 * left the idle function which had been called via pm_idle function
963 * pointer. So it's guaranteed that nothing uses the previous pointer
966 void kick_all_cpus_sync(void)
968 /* Make sure the change is visible before we kick the cpus */
970 smp_call_function(do_nothing, NULL, 1);
972 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
975 * wake_up_all_idle_cpus - break all cpus out of idle
976 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
977 * including idle polling cpus, for non-idle cpus, we will do nothing
980 void wake_up_all_idle_cpus(void)
984 for_each_possible_cpu(cpu) {
986 if (cpu != smp_processor_id() && cpu_online(cpu))
987 wake_up_if_idle(cpu);
991 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
994 * struct smp_call_on_cpu_struct - Call a function on a specific CPU
995 * @work: &work_struct
996 * @done: &completion to signal
997 * @func: function to call
998 * @data: function's data argument
999 * @ret: return value from @func
1000 * @cpu: target CPU (%-1 for any CPU)
1002 * Used to call a function on a specific cpu and wait for it to return.
1003 * Optionally make sure the call is done on a specified physical cpu via vcpu
1004 * pinning in order to support virtualized environments.
1006 struct smp_call_on_cpu_struct {
1007 struct work_struct work;
1008 struct completion done;
1009 int (*func)(void *);
1015 static void smp_call_on_cpu_callback(struct work_struct *work)
1017 struct smp_call_on_cpu_struct *sscs;
1019 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
1021 hypervisor_pin_vcpu(sscs->cpu);
1022 sscs->ret = sscs->func(sscs->data);
1024 hypervisor_pin_vcpu(-1);
1026 complete(&sscs->done);
1029 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
1031 struct smp_call_on_cpu_struct sscs = {
1032 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
1035 .cpu = phys ? cpu : -1,
1038 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
1040 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1043 queue_work_on(cpu, system_wq, &sscs.work);
1044 wait_for_completion(&sscs.done);
1048 EXPORT_SYMBOL_GPL(smp_call_on_cpu);