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>
29 #include <trace/events/ipi.h>
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/csd.h>
32 #undef CREATE_TRACE_POINTS
35 #include "sched/smp.h"
37 #define CSD_TYPE(_csd) ((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
39 struct call_function_data {
40 call_single_data_t __percpu *csd;
41 cpumask_var_t cpumask;
42 cpumask_var_t cpumask_ipi;
45 static DEFINE_PER_CPU_ALIGNED(struct call_function_data, cfd_data);
47 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
49 static DEFINE_PER_CPU(atomic_t, trigger_backtrace) = ATOMIC_INIT(1);
51 static void __flush_smp_call_function_queue(bool warn_cpu_offline);
53 int smpcfd_prepare_cpu(unsigned int cpu)
55 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
57 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
60 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
62 free_cpumask_var(cfd->cpumask);
65 cfd->csd = alloc_percpu(call_single_data_t);
67 free_cpumask_var(cfd->cpumask);
68 free_cpumask_var(cfd->cpumask_ipi);
75 int smpcfd_dead_cpu(unsigned int cpu)
77 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
79 free_cpumask_var(cfd->cpumask);
80 free_cpumask_var(cfd->cpumask_ipi);
81 free_percpu(cfd->csd);
85 int smpcfd_dying_cpu(unsigned int cpu)
88 * The IPIs for the smp-call-function callbacks queued by other
89 * CPUs might arrive late, either due to hardware latencies or
90 * because this CPU disabled interrupts (inside stop-machine)
91 * before the IPIs were sent. So flush out any pending callbacks
92 * explicitly (without waiting for the IPIs to arrive), to
93 * ensure that the outgoing CPU doesn't go offline with work
96 __flush_smp_call_function_queue(false);
101 void __init call_function_init(void)
105 for_each_possible_cpu(i)
106 init_llist_head(&per_cpu(call_single_queue, i));
108 smpcfd_prepare_cpu(smp_processor_id());
111 static __always_inline void
112 send_call_function_single_ipi(int cpu)
114 if (call_function_single_prep_ipi(cpu)) {
115 trace_ipi_send_cpu(cpu, _RET_IP_,
116 generic_smp_call_function_single_interrupt);
117 arch_send_call_function_single_ipi(cpu);
121 static __always_inline void
122 send_call_function_ipi_mask(struct cpumask *mask)
124 trace_ipi_send_cpumask(mask, _RET_IP_,
125 generic_smp_call_function_single_interrupt);
126 arch_send_call_function_ipi_mask(mask);
129 static __always_inline void
130 csd_do_func(smp_call_func_t func, void *info, struct __call_single_data *csd)
132 trace_csd_function_entry(func, csd);
134 trace_csd_function_exit(func, csd);
137 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
139 static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT, csdlock_debug_enabled);
142 * Parse the csdlock_debug= kernel boot parameter.
144 * If you need to restore the old "ext" value that once provided
145 * additional debugging information, reapply the following commits:
147 * de7b09ef658d ("locking/csd_lock: Prepare more CSD lock debugging")
148 * a5aabace5fb8 ("locking/csd_lock: Add more data to CSD lock debugging")
150 static int __init csdlock_debug(char *str)
153 unsigned int val = 0;
155 ret = get_option(&str, &val);
158 static_branch_enable(&csdlock_debug_enabled);
160 static_branch_disable(&csdlock_debug_enabled);
165 __setup("csdlock_debug=", csdlock_debug);
167 static DEFINE_PER_CPU(call_single_data_t *, cur_csd);
168 static DEFINE_PER_CPU(smp_call_func_t, cur_csd_func);
169 static DEFINE_PER_CPU(void *, cur_csd_info);
171 static ulong csd_lock_timeout = 5000; /* CSD lock timeout in milliseconds. */
172 module_param(csd_lock_timeout, ulong, 0444);
174 static atomic_t csd_bug_count = ATOMIC_INIT(0);
176 /* Record current CSD work for current CPU, NULL to erase. */
177 static void __csd_lock_record(struct __call_single_data *csd)
180 smp_mb(); /* NULL cur_csd after unlock. */
181 __this_cpu_write(cur_csd, NULL);
184 __this_cpu_write(cur_csd_func, csd->func);
185 __this_cpu_write(cur_csd_info, csd->info);
186 smp_wmb(); /* func and info before csd. */
187 __this_cpu_write(cur_csd, csd);
188 smp_mb(); /* Update cur_csd before function call. */
189 /* Or before unlock, as the case may be. */
192 static __always_inline void csd_lock_record(struct __call_single_data *csd)
194 if (static_branch_unlikely(&csdlock_debug_enabled))
195 __csd_lock_record(csd);
198 static int csd_lock_wait_getcpu(struct __call_single_data *csd)
200 unsigned int csd_type;
202 csd_type = CSD_TYPE(csd);
203 if (csd_type == CSD_TYPE_ASYNC || csd_type == CSD_TYPE_SYNC)
204 return csd->node.dst; /* Other CSD_TYPE_ values might not have ->dst. */
209 * Complain if too much time spent waiting. Note that only
210 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
211 * so waiting on other types gets much less information.
213 static bool csd_lock_wait_toolong(struct __call_single_data *csd, u64 ts0, u64 *ts1, int *bug_id)
219 call_single_data_t *cpu_cur_csd;
220 unsigned int flags = READ_ONCE(csd->node.u_flags);
221 unsigned long long csd_lock_timeout_ns = csd_lock_timeout * NSEC_PER_MSEC;
223 if (!(flags & CSD_FLAG_LOCK)) {
224 if (!unlikely(*bug_id))
226 cpu = csd_lock_wait_getcpu(csd);
227 pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
228 *bug_id, raw_smp_processor_id(), cpu);
233 ts_delta = ts2 - *ts1;
234 if (likely(ts_delta <= csd_lock_timeout_ns || csd_lock_timeout_ns == 0))
237 firsttime = !*bug_id;
239 *bug_id = atomic_inc_return(&csd_bug_count);
240 cpu = csd_lock_wait_getcpu(csd);
241 if (WARN_ONCE(cpu < 0 || cpu >= nr_cpu_ids, "%s: cpu = %d\n", __func__, cpu))
245 cpu_cur_csd = smp_load_acquire(&per_cpu(cur_csd, cpux)); /* Before func and info. */
246 pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
247 firsttime ? "Detected" : "Continued", *bug_id, raw_smp_processor_id(), ts2 - ts0,
248 cpu, csd->func, csd->info);
249 if (cpu_cur_csd && csd != cpu_cur_csd) {
250 pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
251 *bug_id, READ_ONCE(per_cpu(cur_csd_func, cpux)),
252 READ_ONCE(per_cpu(cur_csd_info, cpux)));
254 pr_alert("\tcsd: CSD lock (#%d) %s.\n",
255 *bug_id, !cpu_cur_csd ? "unresponsive" : "handling this request");
258 if (atomic_cmpxchg_acquire(&per_cpu(trigger_backtrace, cpu), 1, 0))
261 pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id, raw_smp_processor_id(), cpu);
262 arch_send_call_function_single_ipi(cpu);
273 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
275 * For non-synchronous ipi calls the csd can still be in use by the
276 * previous function call. For multi-cpu calls its even more interesting
277 * as we'll have to ensure no other cpu is observing our csd.
279 static void __csd_lock_wait(struct __call_single_data *csd)
284 ts1 = ts0 = sched_clock();
286 if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
290 smp_acquire__after_ctrl_dep();
293 static __always_inline void csd_lock_wait(struct __call_single_data *csd)
295 if (static_branch_unlikely(&csdlock_debug_enabled)) {
296 __csd_lock_wait(csd);
300 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
303 static void csd_lock_record(struct __call_single_data *csd)
307 static __always_inline void csd_lock_wait(struct __call_single_data *csd)
309 smp_cond_load_acquire(&csd->node.u_flags, !(VAL & CSD_FLAG_LOCK));
313 static __always_inline void csd_lock(struct __call_single_data *csd)
316 csd->node.u_flags |= CSD_FLAG_LOCK;
319 * prevent CPU from reordering the above assignment
320 * to ->flags with any subsequent assignments to other
321 * fields of the specified call_single_data_t structure:
326 static __always_inline void csd_unlock(struct __call_single_data *csd)
328 WARN_ON(!(csd->node.u_flags & CSD_FLAG_LOCK));
331 * ensure we're all done before releasing data:
333 smp_store_release(&csd->node.u_flags, 0);
336 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t, csd_data);
338 void __smp_call_single_queue(int cpu, struct llist_node *node)
341 * We have to check the type of the CSD before queueing it, because
342 * once queued it can have its flags cleared by
343 * flush_smp_call_function_queue()
344 * even if we haven't sent the smp_call IPI yet (e.g. the stopper
345 * executes migration_cpu_stop() on the remote CPU).
347 if (trace_csd_queue_cpu_enabled()) {
348 call_single_data_t *csd;
349 smp_call_func_t func;
351 csd = container_of(node, call_single_data_t, node.llist);
352 func = CSD_TYPE(csd) == CSD_TYPE_TTWU ?
353 sched_ttwu_pending : csd->func;
355 trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
359 * The list addition should be visible to the target CPU when it pops
360 * the head of the list to pull the entry off it in the IPI handler
361 * because of normal cache coherency rules implied by the underlying
364 * If IPIs can go out of order to the cache coherency protocol
365 * in an architecture, sufficient synchronisation should be added
366 * to arch code to make it appear to obey cache coherency WRT
367 * locking and barrier primitives. Generic code isn't really
368 * equipped to do the right thing...
370 if (llist_add(node, &per_cpu(call_single_queue, cpu)))
371 send_call_function_single_ipi(cpu);
375 * Insert a previously allocated call_single_data_t element
376 * for execution on the given CPU. data must already have
377 * ->func, ->info, and ->flags set.
379 static int generic_exec_single(int cpu, struct __call_single_data *csd)
381 if (cpu == smp_processor_id()) {
382 smp_call_func_t func = csd->func;
383 void *info = csd->info;
387 * We can unlock early even for the synchronous on-stack case,
388 * since we're doing this from the same CPU..
390 csd_lock_record(csd);
392 local_irq_save(flags);
393 csd_do_func(func, info, NULL);
394 csd_lock_record(NULL);
395 local_irq_restore(flags);
399 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
404 __smp_call_single_queue(cpu, &csd->node.llist);
410 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
412 * Invoked by arch to handle an IPI for call function single.
413 * Must be called with interrupts disabled.
415 void generic_smp_call_function_single_interrupt(void)
417 __flush_smp_call_function_queue(true);
421 * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks
423 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
424 * offline CPU. Skip this check if set to 'false'.
426 * Flush any pending smp-call-function callbacks queued on this CPU. This is
427 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
428 * to ensure that all pending IPI callbacks are run before it goes completely
431 * Loop through the call_single_queue and run all the queued callbacks.
432 * Must be called with interrupts disabled.
434 static void __flush_smp_call_function_queue(bool warn_cpu_offline)
436 call_single_data_t *csd, *csd_next;
437 struct llist_node *entry, *prev;
438 struct llist_head *head;
442 lockdep_assert_irqs_disabled();
444 /* Allow waiters to send backtrace NMI from here onwards */
445 tbt = this_cpu_ptr(&trigger_backtrace);
446 atomic_set_release(tbt, 1);
448 head = this_cpu_ptr(&call_single_queue);
449 entry = llist_del_all(head);
450 entry = llist_reverse_order(entry);
452 /* There shouldn't be any pending callbacks on an offline CPU. */
453 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
454 !warned && entry != NULL)) {
456 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
459 * We don't have to use the _safe() variant here
460 * because we are not invoking the IPI handlers yet.
462 llist_for_each_entry(csd, entry, node.llist) {
463 switch (CSD_TYPE(csd)) {
466 case CSD_TYPE_IRQ_WORK:
467 pr_warn("IPI callback %pS sent to offline CPU\n",
472 pr_warn("IPI task-wakeup sent to offline CPU\n");
476 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
484 * First; run all SYNC callbacks, people are waiting for us.
487 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
488 /* Do we wait until *after* callback? */
489 if (CSD_TYPE(csd) == CSD_TYPE_SYNC) {
490 smp_call_func_t func = csd->func;
491 void *info = csd->info;
494 prev->next = &csd_next->node.llist;
496 entry = &csd_next->node.llist;
499 csd_lock_record(csd);
500 csd_do_func(func, info, csd);
502 csd_lock_record(NULL);
504 prev = &csd->node.llist;
512 * Second; run all !SYNC callbacks.
515 llist_for_each_entry_safe(csd, csd_next, entry, node.llist) {
516 int type = CSD_TYPE(csd);
518 if (type != CSD_TYPE_TTWU) {
520 prev->next = &csd_next->node.llist;
522 entry = &csd_next->node.llist;
525 if (type == CSD_TYPE_ASYNC) {
526 smp_call_func_t func = csd->func;
527 void *info = csd->info;
529 csd_lock_record(csd);
531 csd_do_func(func, info, csd);
532 csd_lock_record(NULL);
533 } else if (type == CSD_TYPE_IRQ_WORK) {
534 irq_work_single(csd);
538 prev = &csd->node.llist;
543 * Third; only CSD_TYPE_TTWU is left, issue those.
546 csd = llist_entry(entry, typeof(*csd), node.llist);
547 csd_do_func(sched_ttwu_pending, entry, csd);
553 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
554 * from task context (idle, migration thread)
556 * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it
557 * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by
558 * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to
559 * handle queued SMP function calls before scheduling.
561 * The migration thread has to ensure that an eventually pending wakeup has
562 * been handled before it migrates a task.
564 void flush_smp_call_function_queue(void)
566 unsigned int was_pending;
569 if (llist_empty(this_cpu_ptr(&call_single_queue)))
572 local_irq_save(flags);
573 /* Get the already pending soft interrupts for RT enabled kernels */
574 was_pending = local_softirq_pending();
575 __flush_smp_call_function_queue(true);
576 if (local_softirq_pending())
577 do_softirq_post_smp_call_flush(was_pending);
579 local_irq_restore(flags);
583 * smp_call_function_single - Run a function on a specific CPU
584 * @func: The function to run. This must be fast and non-blocking.
585 * @info: An arbitrary pointer to pass to the function.
586 * @wait: If true, wait until function has completed on other CPUs.
588 * Returns 0 on success, else a negative status code.
590 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
593 call_single_data_t *csd;
594 call_single_data_t csd_stack = {
595 .node = { .u_flags = CSD_FLAG_LOCK | CSD_TYPE_SYNC, },
601 * prevent preemption and reschedule on another processor,
602 * as well as CPU removal
604 this_cpu = get_cpu();
607 * Can deadlock when called with interrupts disabled.
608 * We allow cpu's that are not yet online though, as no one else can
609 * send smp call function interrupt to this cpu and as such deadlocks
612 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
613 && !oops_in_progress);
616 * When @wait we can deadlock when we interrupt between llist_add() and
617 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
618 * csd_lock() on because the interrupt context uses the same csd
621 WARN_ON_ONCE(!in_task());
625 csd = this_cpu_ptr(&csd_data);
631 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
632 csd->node.src = smp_processor_id();
636 err = generic_exec_single(cpu, csd);
645 EXPORT_SYMBOL(smp_call_function_single);
648 * smp_call_function_single_async() - Run an asynchronous function on a
650 * @cpu: The CPU to run on.
651 * @csd: Pre-allocated and setup data structure
653 * Like smp_call_function_single(), but the call is asynchonous and
654 * can thus be done from contexts with disabled interrupts.
656 * The caller passes his own pre-allocated data structure
657 * (ie: embedded in an object) and is responsible for synchronizing it
658 * such that the IPIs performed on the @csd are strictly serialized.
660 * If the function is called with one csd which has not yet been
661 * processed by previous call to smp_call_function_single_async(), the
662 * function will return immediately with -EBUSY showing that the csd
663 * object is still in progress.
665 * NOTE: Be careful, there is unfortunately no current debugging facility to
666 * validate the correctness of this serialization.
668 * Return: %0 on success or negative errno value on error
670 int smp_call_function_single_async(int cpu, struct __call_single_data *csd)
676 if (csd->node.u_flags & CSD_FLAG_LOCK) {
681 csd->node.u_flags = CSD_FLAG_LOCK;
684 err = generic_exec_single(cpu, csd);
691 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
694 * smp_call_function_any - Run a function on any of the given cpus
695 * @mask: The mask of cpus it can run on.
696 * @func: The function to run. This must be fast and non-blocking.
697 * @info: An arbitrary pointer to pass to the function.
698 * @wait: If true, wait until function has completed.
700 * Returns 0 on success, else a negative status code (if no cpus were online).
702 * Selection preference:
703 * 1) current cpu if in @mask
704 * 2) any cpu of current node if in @mask
705 * 3) any other online cpu in @mask
707 int smp_call_function_any(const struct cpumask *mask,
708 smp_call_func_t func, void *info, int wait)
711 const struct cpumask *nodemask;
714 /* Try for same CPU (cheapest) */
716 if (cpumask_test_cpu(cpu, mask))
719 /* Try for same node. */
720 nodemask = cpumask_of_node(cpu_to_node(cpu));
721 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
722 cpu = cpumask_next_and(cpu, nodemask, mask)) {
727 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
728 cpu = cpumask_any_and(mask, cpu_online_mask);
730 ret = smp_call_function_single(cpu, func, info, wait);
734 EXPORT_SYMBOL_GPL(smp_call_function_any);
737 * Flags to be used as scf_flags argument of smp_call_function_many_cond().
739 * %SCF_WAIT: Wait until function execution is completed
740 * %SCF_RUN_LOCAL: Run also locally if local cpu is set in cpumask
742 #define SCF_WAIT (1U << 0)
743 #define SCF_RUN_LOCAL (1U << 1)
745 static void smp_call_function_many_cond(const struct cpumask *mask,
746 smp_call_func_t func, void *info,
747 unsigned int scf_flags,
748 smp_cond_func_t cond_func)
750 int cpu, last_cpu, this_cpu = smp_processor_id();
751 struct call_function_data *cfd;
752 bool wait = scf_flags & SCF_WAIT;
754 bool run_remote = false;
755 bool run_local = false;
757 lockdep_assert_preemption_disabled();
760 * Can deadlock when called with interrupts disabled.
761 * We allow cpu's that are not yet online though, as no one else can
762 * send smp call function interrupt to this cpu and as such deadlocks
765 if (cpu_online(this_cpu) && !oops_in_progress &&
766 !early_boot_irqs_disabled)
767 lockdep_assert_irqs_enabled();
770 * When @wait we can deadlock when we interrupt between llist_add() and
771 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
772 * csd_lock() on because the interrupt context uses the same csd
775 WARN_ON_ONCE(!in_task());
777 /* Check if we need local execution. */
778 if ((scf_flags & SCF_RUN_LOCAL) && cpumask_test_cpu(this_cpu, mask))
781 /* Check if we need remote execution, i.e., any CPU excluding this one. */
782 cpu = cpumask_first_and(mask, cpu_online_mask);
784 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
785 if (cpu < nr_cpu_ids)
789 cfd = this_cpu_ptr(&cfd_data);
790 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
791 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
793 cpumask_clear(cfd->cpumask_ipi);
794 for_each_cpu(cpu, cfd->cpumask) {
795 call_single_data_t *csd = per_cpu_ptr(cfd->csd, cpu);
797 if (cond_func && !cond_func(cpu, info)) {
798 __cpumask_clear_cpu(cpu, cfd->cpumask);
804 csd->node.u_flags |= CSD_TYPE_SYNC;
807 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
808 csd->node.src = smp_processor_id();
811 trace_csd_queue_cpu(cpu, _RET_IP_, func, csd);
813 if (llist_add(&csd->node.llist, &per_cpu(call_single_queue, cpu))) {
814 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
821 * Choose the most efficient way to send an IPI. Note that the
822 * number of CPUs might be zero due to concurrent changes to the
826 send_call_function_single_ipi(last_cpu);
827 else if (likely(nr_cpus > 1))
828 send_call_function_ipi_mask(cfd->cpumask_ipi);
831 if (run_local && (!cond_func || cond_func(this_cpu, info))) {
834 local_irq_save(flags);
835 csd_do_func(func, info, NULL);
836 local_irq_restore(flags);
839 if (run_remote && wait) {
840 for_each_cpu(cpu, cfd->cpumask) {
841 call_single_data_t *csd;
843 csd = per_cpu_ptr(cfd->csd, cpu);
850 * smp_call_function_many(): Run a function on a set of CPUs.
851 * @mask: The set of cpus to run on (only runs on online subset).
852 * @func: The function to run. This must be fast and non-blocking.
853 * @info: An arbitrary pointer to pass to the function.
854 * @wait: Bitmask that controls the operation. If %SCF_WAIT is set, wait
855 * (atomically) until function has completed on other CPUs. If
856 * %SCF_RUN_LOCAL is set, the function will also be run locally
857 * if the local CPU is set in the @cpumask.
859 * If @wait is true, then returns once @func has returned.
861 * You must not call this function with disabled interrupts or from a
862 * hardware interrupt handler or from a bottom half handler. Preemption
863 * must be disabled when calling this function.
865 void smp_call_function_many(const struct cpumask *mask,
866 smp_call_func_t func, void *info, bool wait)
868 smp_call_function_many_cond(mask, func, info, wait * SCF_WAIT, NULL);
870 EXPORT_SYMBOL(smp_call_function_many);
873 * smp_call_function(): Run a function on all other CPUs.
874 * @func: The function to run. This must be fast and non-blocking.
875 * @info: An arbitrary pointer to pass to the function.
876 * @wait: If true, wait (atomically) until function has completed
881 * If @wait is true, then returns once @func has returned; otherwise
882 * it returns just before the target cpu calls @func.
884 * You must not call this function with disabled interrupts or from a
885 * hardware interrupt handler or from a bottom half handler.
887 void smp_call_function(smp_call_func_t func, void *info, int wait)
890 smp_call_function_many(cpu_online_mask, func, info, wait);
893 EXPORT_SYMBOL(smp_call_function);
895 /* Setup configured maximum number of CPUs to activate */
896 unsigned int setup_max_cpus = NR_CPUS;
897 EXPORT_SYMBOL(setup_max_cpus);
901 * Setup routine for controlling SMP activation
903 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
904 * activation entirely (the MPS table probe still happens, though).
906 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
907 * greater than 0, limits the maximum number of CPUs activated in
911 void __weak __init arch_disable_smp_support(void) { }
913 static int __init nosmp(char *str)
916 arch_disable_smp_support();
921 early_param("nosmp", nosmp);
923 /* this is hard limit */
924 static int __init nrcpus(char *str)
928 if (get_option(&str, &nr_cpus) && nr_cpus > 0 && nr_cpus < nr_cpu_ids)
929 set_nr_cpu_ids(nr_cpus);
934 early_param("nr_cpus", nrcpus);
936 static int __init maxcpus(char *str)
938 get_option(&str, &setup_max_cpus);
939 if (setup_max_cpus == 0)
940 arch_disable_smp_support();
945 early_param("maxcpus", maxcpus);
947 #if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS)
948 /* Setup number of possible processor ids */
949 unsigned int nr_cpu_ids __read_mostly = NR_CPUS;
950 EXPORT_SYMBOL(nr_cpu_ids);
953 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
954 void __init setup_nr_cpu_ids(void)
956 set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask), NR_CPUS) + 1);
959 /* Called by boot processor to activate the rest. */
960 void __init smp_init(void)
962 int num_nodes, num_cpus;
965 cpuhp_threads_init();
967 pr_info("Bringing up secondary CPUs ...\n");
969 bringup_nonboot_cpus(setup_max_cpus);
971 num_nodes = num_online_nodes();
972 num_cpus = num_online_cpus();
973 pr_info("Brought up %d node%s, %d CPU%s\n",
974 num_nodes, (num_nodes > 1 ? "s" : ""),
975 num_cpus, (num_cpus > 1 ? "s" : ""));
977 /* Any cleanup work */
978 smp_cpus_done(setup_max_cpus);
982 * on_each_cpu_cond(): Call a function on each processor for which
983 * the supplied function cond_func returns true, optionally waiting
984 * for all the required CPUs to finish. This may include the local
986 * @cond_func: A callback function that is passed a cpu id and
987 * the info parameter. The function is called
988 * with preemption disabled. The function should
989 * return a blooean value indicating whether to IPI
991 * @func: The function to run on all applicable CPUs.
992 * This must be fast and non-blocking.
993 * @info: An arbitrary pointer to pass to both functions.
994 * @wait: If true, wait (atomically) until function has
995 * completed on other CPUs.
997 * Preemption is disabled to protect against CPUs going offline but not online.
998 * CPUs going online during the call will not be seen or sent an IPI.
1000 * You must not call this function with disabled interrupts or
1001 * from a hardware interrupt handler or from a bottom half handler.
1003 void on_each_cpu_cond_mask(smp_cond_func_t cond_func, smp_call_func_t func,
1004 void *info, bool wait, const struct cpumask *mask)
1006 unsigned int scf_flags = SCF_RUN_LOCAL;
1009 scf_flags |= SCF_WAIT;
1012 smp_call_function_many_cond(mask, func, info, scf_flags, cond_func);
1015 EXPORT_SYMBOL(on_each_cpu_cond_mask);
1017 static void do_nothing(void *unused)
1022 * kick_all_cpus_sync - Force all cpus out of idle
1024 * Used to synchronize the update of pm_idle function pointer. It's
1025 * called after the pointer is updated and returns after the dummy
1026 * callback function has been executed on all cpus. The execution of
1027 * the function can only happen on the remote cpus after they have
1028 * left the idle function which had been called via pm_idle function
1029 * pointer. So it's guaranteed that nothing uses the previous pointer
1032 void kick_all_cpus_sync(void)
1034 /* Make sure the change is visible before we kick the cpus */
1036 smp_call_function(do_nothing, NULL, 1);
1038 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
1041 * wake_up_all_idle_cpus - break all cpus out of idle
1042 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
1043 * including idle polling cpus, for non-idle cpus, we will do nothing
1046 void wake_up_all_idle_cpus(void)
1050 for_each_possible_cpu(cpu) {
1052 if (cpu != smp_processor_id() && cpu_online(cpu))
1053 wake_up_if_idle(cpu);
1057 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
1060 * struct smp_call_on_cpu_struct - Call a function on a specific CPU
1061 * @work: &work_struct
1062 * @done: &completion to signal
1063 * @func: function to call
1064 * @data: function's data argument
1065 * @ret: return value from @func
1066 * @cpu: target CPU (%-1 for any CPU)
1068 * Used to call a function on a specific cpu and wait for it to return.
1069 * Optionally make sure the call is done on a specified physical cpu via vcpu
1070 * pinning in order to support virtualized environments.
1072 struct smp_call_on_cpu_struct {
1073 struct work_struct work;
1074 struct completion done;
1075 int (*func)(void *);
1081 static void smp_call_on_cpu_callback(struct work_struct *work)
1083 struct smp_call_on_cpu_struct *sscs;
1085 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
1087 hypervisor_pin_vcpu(sscs->cpu);
1088 sscs->ret = sscs->func(sscs->data);
1090 hypervisor_pin_vcpu(-1);
1092 complete(&sscs->done);
1095 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
1097 struct smp_call_on_cpu_struct sscs = {
1098 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
1101 .cpu = phys ? cpu : -1,
1104 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
1106 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1109 queue_work_on(cpu, system_wq, &sscs.work);
1110 wait_for_completion(&sscs.done);
1114 EXPORT_SYMBOL_GPL(smp_call_on_cpu);