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>
28 #include <linux/string_choices.h>
30 #include <trace/events/ipi.h>
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/csd.h>
33 #undef CREATE_TRACE_POINTS
36 #include "sched/smp.h"
38 #define CSD_TYPE(_csd) ((_csd)->node.u_flags & CSD_FLAG_TYPE_MASK)
40 struct call_function_data
{
41 call_single_data_t __percpu
*csd
;
42 cpumask_var_t cpumask
;
43 cpumask_var_t cpumask_ipi
;
46 static DEFINE_PER_CPU_ALIGNED(struct call_function_data
, cfd_data
);
48 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head
, call_single_queue
);
50 static DEFINE_PER_CPU(atomic_t
, trigger_backtrace
) = ATOMIC_INIT(1);
52 static void __flush_smp_call_function_queue(bool warn_cpu_offline
);
54 int smpcfd_prepare_cpu(unsigned int cpu
)
56 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
58 if (!zalloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
61 if (!zalloc_cpumask_var_node(&cfd
->cpumask_ipi
, GFP_KERNEL
,
63 free_cpumask_var(cfd
->cpumask
);
66 cfd
->csd
= alloc_percpu(call_single_data_t
);
68 free_cpumask_var(cfd
->cpumask
);
69 free_cpumask_var(cfd
->cpumask_ipi
);
76 int smpcfd_dead_cpu(unsigned int cpu
)
78 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
80 free_cpumask_var(cfd
->cpumask
);
81 free_cpumask_var(cfd
->cpumask_ipi
);
82 free_percpu(cfd
->csd
);
86 int smpcfd_dying_cpu(unsigned int cpu
)
89 * The IPIs for the smp-call-function callbacks queued by other
90 * CPUs might arrive late, either due to hardware latencies or
91 * because this CPU disabled interrupts (inside stop-machine)
92 * before the IPIs were sent. So flush out any pending callbacks
93 * explicitly (without waiting for the IPIs to arrive), to
94 * ensure that the outgoing CPU doesn't go offline with work
97 __flush_smp_call_function_queue(false);
102 void __init
call_function_init(void)
106 for_each_possible_cpu(i
)
107 init_llist_head(&per_cpu(call_single_queue
, i
));
109 smpcfd_prepare_cpu(smp_processor_id());
112 static __always_inline
void
113 send_call_function_single_ipi(int cpu
)
115 if (call_function_single_prep_ipi(cpu
)) {
116 trace_ipi_send_cpu(cpu
, _RET_IP_
,
117 generic_smp_call_function_single_interrupt
);
118 arch_send_call_function_single_ipi(cpu
);
122 static __always_inline
void
123 send_call_function_ipi_mask(struct cpumask
*mask
)
125 trace_ipi_send_cpumask(mask
, _RET_IP_
,
126 generic_smp_call_function_single_interrupt
);
127 arch_send_call_function_ipi_mask(mask
);
130 static __always_inline
void
131 csd_do_func(smp_call_func_t func
, void *info
, call_single_data_t
*csd
)
133 trace_csd_function_entry(func
, csd
);
135 trace_csd_function_exit(func
, csd
);
138 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
140 static DEFINE_STATIC_KEY_MAYBE(CONFIG_CSD_LOCK_WAIT_DEBUG_DEFAULT
, csdlock_debug_enabled
);
143 * Parse the csdlock_debug= kernel boot parameter.
145 * If you need to restore the old "ext" value that once provided
146 * additional debugging information, reapply the following commits:
148 * de7b09ef658d ("locking/csd_lock: Prepare more CSD lock debugging")
149 * a5aabace5fb8 ("locking/csd_lock: Add more data to CSD lock debugging")
151 static int __init
csdlock_debug(char *str
)
154 unsigned int val
= 0;
156 ret
= get_option(&str
, &val
);
159 static_branch_enable(&csdlock_debug_enabled
);
161 static_branch_disable(&csdlock_debug_enabled
);
166 __setup("csdlock_debug=", csdlock_debug
);
168 static DEFINE_PER_CPU(call_single_data_t
*, cur_csd
);
169 static DEFINE_PER_CPU(smp_call_func_t
, cur_csd_func
);
170 static DEFINE_PER_CPU(void *, cur_csd_info
);
172 static ulong csd_lock_timeout
= 5000; /* CSD lock timeout in milliseconds. */
173 module_param(csd_lock_timeout
, ulong
, 0444);
174 static int panic_on_ipistall
; /* CSD panic timeout in milliseconds, 300000 for five minutes. */
175 module_param(panic_on_ipistall
, int, 0444);
177 static atomic_t csd_bug_count
= ATOMIC_INIT(0);
179 /* Record current CSD work for current CPU, NULL to erase. */
180 static void __csd_lock_record(call_single_data_t
*csd
)
183 smp_mb(); /* NULL cur_csd after unlock. */
184 __this_cpu_write(cur_csd
, NULL
);
187 __this_cpu_write(cur_csd_func
, csd
->func
);
188 __this_cpu_write(cur_csd_info
, csd
->info
);
189 smp_wmb(); /* func and info before csd. */
190 __this_cpu_write(cur_csd
, csd
);
191 smp_mb(); /* Update cur_csd before function call. */
192 /* Or before unlock, as the case may be. */
195 static __always_inline
void csd_lock_record(call_single_data_t
*csd
)
197 if (static_branch_unlikely(&csdlock_debug_enabled
))
198 __csd_lock_record(csd
);
201 static int csd_lock_wait_getcpu(call_single_data_t
*csd
)
203 unsigned int csd_type
;
205 csd_type
= CSD_TYPE(csd
);
206 if (csd_type
== CSD_TYPE_ASYNC
|| csd_type
== CSD_TYPE_SYNC
)
207 return csd
->node
.dst
; /* Other CSD_TYPE_ values might not have ->dst. */
212 * Complain if too much time spent waiting. Note that only
213 * the CSD_TYPE_SYNC/ASYNC types provide the destination CPU,
214 * so waiting on other types gets much less information.
216 static bool csd_lock_wait_toolong(call_single_data_t
*csd
, u64 ts0
, u64
*ts1
, int *bug_id
)
222 call_single_data_t
*cpu_cur_csd
;
223 unsigned int flags
= READ_ONCE(csd
->node
.u_flags
);
224 unsigned long long csd_lock_timeout_ns
= csd_lock_timeout
* NSEC_PER_MSEC
;
226 if (!(flags
& CSD_FLAG_LOCK
)) {
227 if (!unlikely(*bug_id
))
229 cpu
= csd_lock_wait_getcpu(csd
);
230 pr_alert("csd: CSD lock (#%d) got unstuck on CPU#%02d, CPU#%02d released the lock.\n",
231 *bug_id
, raw_smp_processor_id(), cpu
);
236 /* How long since we last checked for a stuck CSD lock.*/
237 ts_delta
= ts2
- *ts1
;
238 if (likely(ts_delta
<= csd_lock_timeout_ns
|| csd_lock_timeout_ns
== 0))
241 firsttime
= !*bug_id
;
243 *bug_id
= atomic_inc_return(&csd_bug_count
);
244 cpu
= csd_lock_wait_getcpu(csd
);
245 if (WARN_ONCE(cpu
< 0 || cpu
>= nr_cpu_ids
, "%s: cpu = %d\n", __func__
, cpu
))
249 cpu_cur_csd
= smp_load_acquire(&per_cpu(cur_csd
, cpux
)); /* Before func and info. */
250 /* How long since this CSD lock was stuck. */
251 ts_delta
= ts2
- ts0
;
252 pr_alert("csd: %s non-responsive CSD lock (#%d) on CPU#%d, waiting %llu ns for CPU#%02d %pS(%ps).\n",
253 firsttime
? "Detected" : "Continued", *bug_id
, raw_smp_processor_id(), ts_delta
,
254 cpu
, csd
->func
, csd
->info
);
256 * If the CSD lock is still stuck after 5 minutes, it is unlikely
257 * to become unstuck. Use a signed comparison to avoid triggering
258 * on underflows when the TSC is out of sync between sockets.
260 BUG_ON(panic_on_ipistall
> 0 && (s64
)ts_delta
> ((s64
)panic_on_ipistall
* NSEC_PER_MSEC
));
261 if (cpu_cur_csd
&& csd
!= cpu_cur_csd
) {
262 pr_alert("\tcsd: CSD lock (#%d) handling prior %pS(%ps) request.\n",
263 *bug_id
, READ_ONCE(per_cpu(cur_csd_func
, cpux
)),
264 READ_ONCE(per_cpu(cur_csd_info
, cpux
)));
266 pr_alert("\tcsd: CSD lock (#%d) %s.\n",
267 *bug_id
, !cpu_cur_csd
? "unresponsive" : "handling this request");
270 if (atomic_cmpxchg_acquire(&per_cpu(trigger_backtrace
, cpu
), 1, 0))
273 pr_alert("csd: Re-sending CSD lock (#%d) IPI from CPU#%02d to CPU#%02d\n", *bug_id
, raw_smp_processor_id(), cpu
);
274 arch_send_call_function_single_ipi(cpu
);
285 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
287 * For non-synchronous ipi calls the csd can still be in use by the
288 * previous function call. For multi-cpu calls its even more interesting
289 * as we'll have to ensure no other cpu is observing our csd.
291 static void __csd_lock_wait(call_single_data_t
*csd
)
296 ts1
= ts0
= sched_clock();
298 if (csd_lock_wait_toolong(csd
, ts0
, &ts1
, &bug_id
))
302 smp_acquire__after_ctrl_dep();
305 static __always_inline
void csd_lock_wait(call_single_data_t
*csd
)
307 if (static_branch_unlikely(&csdlock_debug_enabled
)) {
308 __csd_lock_wait(csd
);
312 smp_cond_load_acquire(&csd
->node
.u_flags
, !(VAL
& CSD_FLAG_LOCK
));
315 static void csd_lock_record(call_single_data_t
*csd
)
319 static __always_inline
void csd_lock_wait(call_single_data_t
*csd
)
321 smp_cond_load_acquire(&csd
->node
.u_flags
, !(VAL
& CSD_FLAG_LOCK
));
325 static __always_inline
void csd_lock(call_single_data_t
*csd
)
328 csd
->node
.u_flags
|= CSD_FLAG_LOCK
;
331 * prevent CPU from reordering the above assignment
332 * to ->flags with any subsequent assignments to other
333 * fields of the specified call_single_data_t structure:
338 static __always_inline
void csd_unlock(call_single_data_t
*csd
)
340 WARN_ON(!(csd
->node
.u_flags
& CSD_FLAG_LOCK
));
343 * ensure we're all done before releasing data:
345 smp_store_release(&csd
->node
.u_flags
, 0);
348 static DEFINE_PER_CPU_SHARED_ALIGNED(call_single_data_t
, csd_data
);
350 void __smp_call_single_queue(int cpu
, struct llist_node
*node
)
353 * We have to check the type of the CSD before queueing it, because
354 * once queued it can have its flags cleared by
355 * flush_smp_call_function_queue()
356 * even if we haven't sent the smp_call IPI yet (e.g. the stopper
357 * executes migration_cpu_stop() on the remote CPU).
359 if (trace_csd_queue_cpu_enabled()) {
360 call_single_data_t
*csd
;
361 smp_call_func_t func
;
363 csd
= container_of(node
, call_single_data_t
, node
.llist
);
364 func
= CSD_TYPE(csd
) == CSD_TYPE_TTWU
?
365 sched_ttwu_pending
: csd
->func
;
367 trace_csd_queue_cpu(cpu
, _RET_IP_
, func
, csd
);
371 * The list addition should be visible to the target CPU when it pops
372 * the head of the list to pull the entry off it in the IPI handler
373 * because of normal cache coherency rules implied by the underlying
376 * If IPIs can go out of order to the cache coherency protocol
377 * in an architecture, sufficient synchronisation should be added
378 * to arch code to make it appear to obey cache coherency WRT
379 * locking and barrier primitives. Generic code isn't really
380 * equipped to do the right thing...
382 if (llist_add(node
, &per_cpu(call_single_queue
, cpu
)))
383 send_call_function_single_ipi(cpu
);
387 * Insert a previously allocated call_single_data_t element
388 * for execution on the given CPU. data must already have
389 * ->func, ->info, and ->flags set.
391 static int generic_exec_single(int cpu
, call_single_data_t
*csd
)
393 if (cpu
== smp_processor_id()) {
394 smp_call_func_t func
= csd
->func
;
395 void *info
= csd
->info
;
399 * We can unlock early even for the synchronous on-stack case,
400 * since we're doing this from the same CPU..
402 csd_lock_record(csd
);
404 local_irq_save(flags
);
405 csd_do_func(func
, info
, NULL
);
406 csd_lock_record(NULL
);
407 local_irq_restore(flags
);
411 if ((unsigned)cpu
>= nr_cpu_ids
|| !cpu_online(cpu
)) {
416 __smp_call_single_queue(cpu
, &csd
->node
.llist
);
422 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
424 * Invoked by arch to handle an IPI for call function single.
425 * Must be called with interrupts disabled.
427 void generic_smp_call_function_single_interrupt(void)
429 __flush_smp_call_function_queue(true);
433 * __flush_smp_call_function_queue - Flush pending smp-call-function callbacks
435 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
436 * offline CPU. Skip this check if set to 'false'.
438 * Flush any pending smp-call-function callbacks queued on this CPU. This is
439 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
440 * to ensure that all pending IPI callbacks are run before it goes completely
443 * Loop through the call_single_queue and run all the queued callbacks.
444 * Must be called with interrupts disabled.
446 static void __flush_smp_call_function_queue(bool warn_cpu_offline
)
448 call_single_data_t
*csd
, *csd_next
;
449 struct llist_node
*entry
, *prev
;
450 struct llist_head
*head
;
454 lockdep_assert_irqs_disabled();
456 /* Allow waiters to send backtrace NMI from here onwards */
457 tbt
= this_cpu_ptr(&trigger_backtrace
);
458 atomic_set_release(tbt
, 1);
460 head
= this_cpu_ptr(&call_single_queue
);
461 entry
= llist_del_all(head
);
462 entry
= llist_reverse_order(entry
);
464 /* There shouldn't be any pending callbacks on an offline CPU. */
465 if (unlikely(warn_cpu_offline
&& !cpu_online(smp_processor_id()) &&
466 !warned
&& entry
!= NULL
)) {
468 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
471 * We don't have to use the _safe() variant here
472 * because we are not invoking the IPI handlers yet.
474 llist_for_each_entry(csd
, entry
, node
.llist
) {
475 switch (CSD_TYPE(csd
)) {
478 case CSD_TYPE_IRQ_WORK
:
479 pr_warn("IPI callback %pS sent to offline CPU\n",
484 pr_warn("IPI task-wakeup sent to offline CPU\n");
488 pr_warn("IPI callback, unknown type %d, sent to offline CPU\n",
496 * First; run all SYNC callbacks, people are waiting for us.
499 llist_for_each_entry_safe(csd
, csd_next
, entry
, node
.llist
) {
500 /* Do we wait until *after* callback? */
501 if (CSD_TYPE(csd
) == CSD_TYPE_SYNC
) {
502 smp_call_func_t func
= csd
->func
;
503 void *info
= csd
->info
;
506 prev
->next
= &csd_next
->node
.llist
;
508 entry
= &csd_next
->node
.llist
;
511 csd_lock_record(csd
);
512 csd_do_func(func
, info
, csd
);
514 csd_lock_record(NULL
);
516 prev
= &csd
->node
.llist
;
524 * Second; run all !SYNC callbacks.
527 llist_for_each_entry_safe(csd
, csd_next
, entry
, node
.llist
) {
528 int type
= CSD_TYPE(csd
);
530 if (type
!= CSD_TYPE_TTWU
) {
532 prev
->next
= &csd_next
->node
.llist
;
534 entry
= &csd_next
->node
.llist
;
537 if (type
== CSD_TYPE_ASYNC
) {
538 smp_call_func_t func
= csd
->func
;
539 void *info
= csd
->info
;
541 csd_lock_record(csd
);
543 csd_do_func(func
, info
, csd
);
544 csd_lock_record(NULL
);
545 } else if (type
== CSD_TYPE_IRQ_WORK
) {
546 irq_work_single(csd
);
550 prev
= &csd
->node
.llist
;
555 * Third; only CSD_TYPE_TTWU is left, issue those.
558 csd
= llist_entry(entry
, typeof(*csd
), node
.llist
);
559 csd_do_func(sched_ttwu_pending
, entry
, csd
);
565 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
566 * from task context (idle, migration thread)
568 * When TIF_POLLING_NRFLAG is supported and a CPU is in idle and has it
569 * set, then remote CPUs can avoid sending IPIs and wake the idle CPU by
570 * setting TIF_NEED_RESCHED. The idle task on the woken up CPU has to
571 * handle queued SMP function calls before scheduling.
573 * The migration thread has to ensure that an eventually pending wakeup has
574 * been handled before it migrates a task.
576 void flush_smp_call_function_queue(void)
578 unsigned int was_pending
;
581 if (llist_empty(this_cpu_ptr(&call_single_queue
)))
584 local_irq_save(flags
);
585 /* Get the already pending soft interrupts for RT enabled kernels */
586 was_pending
= local_softirq_pending();
587 __flush_smp_call_function_queue(true);
588 if (local_softirq_pending())
589 do_softirq_post_smp_call_flush(was_pending
);
591 local_irq_restore(flags
);
595 * smp_call_function_single - Run a function on a specific CPU
596 * @func: The function to run. This must be fast and non-blocking.
597 * @info: An arbitrary pointer to pass to the function.
598 * @wait: If true, wait until function has completed on other CPUs.
600 * Returns 0 on success, else a negative status code.
602 int smp_call_function_single(int cpu
, smp_call_func_t func
, void *info
,
605 call_single_data_t
*csd
;
606 call_single_data_t csd_stack
= {
607 .node
= { .u_flags
= CSD_FLAG_LOCK
| CSD_TYPE_SYNC
, },
613 * prevent preemption and reschedule on another processor,
614 * as well as CPU removal
616 this_cpu
= get_cpu();
619 * Can deadlock when called with interrupts disabled.
620 * We allow cpu's that are not yet online though, as no one else can
621 * send smp call function interrupt to this cpu and as such deadlocks
624 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
625 && !oops_in_progress
);
628 * When @wait we can deadlock when we interrupt between llist_add() and
629 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
630 * csd_lock() on because the interrupt context uses the same csd
633 WARN_ON_ONCE(!in_task());
637 csd
= this_cpu_ptr(&csd_data
);
643 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
644 csd
->node
.src
= smp_processor_id();
648 err
= generic_exec_single(cpu
, csd
);
657 EXPORT_SYMBOL(smp_call_function_single
);
660 * smp_call_function_single_async() - Run an asynchronous function on a
662 * @cpu: The CPU to run on.
663 * @csd: Pre-allocated and setup data structure
665 * Like smp_call_function_single(), but the call is asynchonous and
666 * can thus be done from contexts with disabled interrupts.
668 * The caller passes his own pre-allocated data structure
669 * (ie: embedded in an object) and is responsible for synchronizing it
670 * such that the IPIs performed on the @csd are strictly serialized.
672 * If the function is called with one csd which has not yet been
673 * processed by previous call to smp_call_function_single_async(), the
674 * function will return immediately with -EBUSY showing that the csd
675 * object is still in progress.
677 * NOTE: Be careful, there is unfortunately no current debugging facility to
678 * validate the correctness of this serialization.
680 * Return: %0 on success or negative errno value on error
682 int smp_call_function_single_async(int cpu
, call_single_data_t
*csd
)
688 if (csd
->node
.u_flags
& CSD_FLAG_LOCK
) {
693 csd
->node
.u_flags
= CSD_FLAG_LOCK
;
696 err
= generic_exec_single(cpu
, csd
);
703 EXPORT_SYMBOL_GPL(smp_call_function_single_async
);
706 * smp_call_function_any - Run a function on any of the given cpus
707 * @mask: The mask of cpus it can run on.
708 * @func: The function to run. This must be fast and non-blocking.
709 * @info: An arbitrary pointer to pass to the function.
710 * @wait: If true, wait until function has completed.
712 * Returns 0 on success, else a negative status code (if no cpus were online).
714 * Selection preference:
715 * 1) current cpu if in @mask
716 * 2) any cpu of current node if in @mask
717 * 3) any other online cpu in @mask
719 int smp_call_function_any(const struct cpumask
*mask
,
720 smp_call_func_t func
, void *info
, int wait
)
723 const struct cpumask
*nodemask
;
726 /* Try for same CPU (cheapest) */
728 if (cpumask_test_cpu(cpu
, mask
))
731 /* Try for same node. */
732 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
733 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
734 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
739 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
740 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
742 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
746 EXPORT_SYMBOL_GPL(smp_call_function_any
);
749 * Flags to be used as scf_flags argument of smp_call_function_many_cond().
751 * %SCF_WAIT: Wait until function execution is completed
752 * %SCF_RUN_LOCAL: Run also locally if local cpu is set in cpumask
754 #define SCF_WAIT (1U << 0)
755 #define SCF_RUN_LOCAL (1U << 1)
757 static void smp_call_function_many_cond(const struct cpumask
*mask
,
758 smp_call_func_t func
, void *info
,
759 unsigned int scf_flags
,
760 smp_cond_func_t cond_func
)
762 int cpu
, last_cpu
, this_cpu
= smp_processor_id();
763 struct call_function_data
*cfd
;
764 bool wait
= scf_flags
& SCF_WAIT
;
766 bool run_remote
= false;
767 bool run_local
= false;
769 lockdep_assert_preemption_disabled();
772 * Can deadlock when called with interrupts disabled.
773 * We allow cpu's that are not yet online though, as no one else can
774 * send smp call function interrupt to this cpu and as such deadlocks
777 if (cpu_online(this_cpu
) && !oops_in_progress
&&
778 !early_boot_irqs_disabled
)
779 lockdep_assert_irqs_enabled();
782 * When @wait we can deadlock when we interrupt between llist_add() and
783 * arch_send_call_function_ipi*(); when !@wait we can deadlock due to
784 * csd_lock() on because the interrupt context uses the same csd
787 WARN_ON_ONCE(!in_task());
789 /* Check if we need local execution. */
790 if ((scf_flags
& SCF_RUN_LOCAL
) && cpumask_test_cpu(this_cpu
, mask
))
793 /* Check if we need remote execution, i.e., any CPU excluding this one. */
794 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
796 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
797 if (cpu
< nr_cpu_ids
)
801 cfd
= this_cpu_ptr(&cfd_data
);
802 cpumask_and(cfd
->cpumask
, mask
, cpu_online_mask
);
803 __cpumask_clear_cpu(this_cpu
, cfd
->cpumask
);
805 cpumask_clear(cfd
->cpumask_ipi
);
806 for_each_cpu(cpu
, cfd
->cpumask
) {
807 call_single_data_t
*csd
= per_cpu_ptr(cfd
->csd
, cpu
);
809 if (cond_func
&& !cond_func(cpu
, info
)) {
810 __cpumask_clear_cpu(cpu
, cfd
->cpumask
);
816 csd
->node
.u_flags
|= CSD_TYPE_SYNC
;
819 #ifdef CONFIG_CSD_LOCK_WAIT_DEBUG
820 csd
->node
.src
= smp_processor_id();
823 trace_csd_queue_cpu(cpu
, _RET_IP_
, func
, csd
);
825 if (llist_add(&csd
->node
.llist
, &per_cpu(call_single_queue
, cpu
))) {
826 __cpumask_set_cpu(cpu
, cfd
->cpumask_ipi
);
833 * Choose the most efficient way to send an IPI. Note that the
834 * number of CPUs might be zero due to concurrent changes to the
838 send_call_function_single_ipi(last_cpu
);
839 else if (likely(nr_cpus
> 1))
840 send_call_function_ipi_mask(cfd
->cpumask_ipi
);
843 if (run_local
&& (!cond_func
|| cond_func(this_cpu
, info
))) {
846 local_irq_save(flags
);
847 csd_do_func(func
, info
, NULL
);
848 local_irq_restore(flags
);
851 if (run_remote
&& wait
) {
852 for_each_cpu(cpu
, cfd
->cpumask
) {
853 call_single_data_t
*csd
;
855 csd
= per_cpu_ptr(cfd
->csd
, cpu
);
862 * smp_call_function_many(): Run a function on a set of CPUs.
863 * @mask: The set of cpus to run on (only runs on online subset).
864 * @func: The function to run. This must be fast and non-blocking.
865 * @info: An arbitrary pointer to pass to the function.
866 * @wait: Bitmask that controls the operation. If %SCF_WAIT is set, wait
867 * (atomically) until function has completed on other CPUs. If
868 * %SCF_RUN_LOCAL is set, the function will also be run locally
869 * if the local CPU is set in the @cpumask.
871 * If @wait is true, then returns once @func has returned.
873 * You must not call this function with disabled interrupts or from a
874 * hardware interrupt handler or from a bottom half handler. Preemption
875 * must be disabled when calling this function.
877 void smp_call_function_many(const struct cpumask
*mask
,
878 smp_call_func_t func
, void *info
, bool wait
)
880 smp_call_function_many_cond(mask
, func
, info
, wait
* SCF_WAIT
, NULL
);
882 EXPORT_SYMBOL(smp_call_function_many
);
885 * smp_call_function(): Run a function on all other CPUs.
886 * @func: The function to run. This must be fast and non-blocking.
887 * @info: An arbitrary pointer to pass to the function.
888 * @wait: If true, wait (atomically) until function has completed
893 * If @wait is true, then returns once @func has returned; otherwise
894 * it returns just before the target cpu calls @func.
896 * You must not call this function with disabled interrupts or from a
897 * hardware interrupt handler or from a bottom half handler.
899 void smp_call_function(smp_call_func_t func
, void *info
, int wait
)
902 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
905 EXPORT_SYMBOL(smp_call_function
);
907 /* Setup configured maximum number of CPUs to activate */
908 unsigned int setup_max_cpus
= NR_CPUS
;
909 EXPORT_SYMBOL(setup_max_cpus
);
913 * Setup routine for controlling SMP activation
915 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
916 * activation entirely (the MPS table probe still happens, though).
918 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
919 * greater than 0, limits the maximum number of CPUs activated in
923 void __weak __init
arch_disable_smp_support(void) { }
925 static int __init
nosmp(char *str
)
928 arch_disable_smp_support();
933 early_param("nosmp", nosmp
);
935 /* this is hard limit */
936 static int __init
nrcpus(char *str
)
940 if (get_option(&str
, &nr_cpus
) && nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
941 set_nr_cpu_ids(nr_cpus
);
946 early_param("nr_cpus", nrcpus
);
948 static int __init
maxcpus(char *str
)
950 get_option(&str
, &setup_max_cpus
);
951 if (setup_max_cpus
== 0)
952 arch_disable_smp_support();
957 early_param("maxcpus", maxcpus
);
959 #if (NR_CPUS > 1) && !defined(CONFIG_FORCE_NR_CPUS)
960 /* Setup number of possible processor ids */
961 unsigned int nr_cpu_ids __read_mostly
= NR_CPUS
;
962 EXPORT_SYMBOL(nr_cpu_ids
);
965 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
966 void __init
setup_nr_cpu_ids(void)
968 set_nr_cpu_ids(find_last_bit(cpumask_bits(cpu_possible_mask
), NR_CPUS
) + 1);
971 /* Called by boot processor to activate the rest. */
972 void __init
smp_init(void)
974 int num_nodes
, num_cpus
;
977 cpuhp_threads_init();
979 pr_info("Bringing up secondary CPUs ...\n");
981 bringup_nonboot_cpus(setup_max_cpus
);
983 num_nodes
= num_online_nodes();
984 num_cpus
= num_online_cpus();
985 pr_info("Brought up %d node%s, %d CPU%s\n",
986 num_nodes
, str_plural(num_nodes
), num_cpus
, str_plural(num_cpus
));
988 /* Any cleanup work */
989 smp_cpus_done(setup_max_cpus
);
993 * on_each_cpu_cond(): Call a function on each processor for which
994 * the supplied function cond_func returns true, optionally waiting
995 * for all the required CPUs to finish. This may include the local
997 * @cond_func: A callback function that is passed a cpu id and
998 * the info parameter. The function is called
999 * with preemption disabled. The function should
1000 * return a blooean value indicating whether to IPI
1001 * the specified CPU.
1002 * @func: The function to run on all applicable CPUs.
1003 * This must be fast and non-blocking.
1004 * @info: An arbitrary pointer to pass to both functions.
1005 * @wait: If true, wait (atomically) until function has
1006 * completed on other CPUs.
1008 * Preemption is disabled to protect against CPUs going offline but not online.
1009 * CPUs going online during the call will not be seen or sent an IPI.
1011 * You must not call this function with disabled interrupts or
1012 * from a hardware interrupt handler or from a bottom half handler.
1014 void on_each_cpu_cond_mask(smp_cond_func_t cond_func
, smp_call_func_t func
,
1015 void *info
, bool wait
, const struct cpumask
*mask
)
1017 unsigned int scf_flags
= SCF_RUN_LOCAL
;
1020 scf_flags
|= SCF_WAIT
;
1023 smp_call_function_many_cond(mask
, func
, info
, scf_flags
, cond_func
);
1026 EXPORT_SYMBOL(on_each_cpu_cond_mask
);
1028 static void do_nothing(void *unused
)
1033 * kick_all_cpus_sync - Force all cpus out of idle
1035 * Used to synchronize the update of pm_idle function pointer. It's
1036 * called after the pointer is updated and returns after the dummy
1037 * callback function has been executed on all cpus. The execution of
1038 * the function can only happen on the remote cpus after they have
1039 * left the idle function which had been called via pm_idle function
1040 * pointer. So it's guaranteed that nothing uses the previous pointer
1043 void kick_all_cpus_sync(void)
1045 /* Make sure the change is visible before we kick the cpus */
1047 smp_call_function(do_nothing
, NULL
, 1);
1049 EXPORT_SYMBOL_GPL(kick_all_cpus_sync
);
1052 * wake_up_all_idle_cpus - break all cpus out of idle
1053 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
1054 * including idle polling cpus, for non-idle cpus, we will do nothing
1057 void wake_up_all_idle_cpus(void)
1061 for_each_possible_cpu(cpu
) {
1063 if (cpu
!= smp_processor_id() && cpu_online(cpu
))
1064 wake_up_if_idle(cpu
);
1068 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus
);
1071 * struct smp_call_on_cpu_struct - Call a function on a specific CPU
1072 * @work: &work_struct
1073 * @done: &completion to signal
1074 * @func: function to call
1075 * @data: function's data argument
1076 * @ret: return value from @func
1077 * @cpu: target CPU (%-1 for any CPU)
1079 * Used to call a function on a specific cpu and wait for it to return.
1080 * Optionally make sure the call is done on a specified physical cpu via vcpu
1081 * pinning in order to support virtualized environments.
1083 struct smp_call_on_cpu_struct
{
1084 struct work_struct work
;
1085 struct completion done
;
1086 int (*func
)(void *);
1092 static void smp_call_on_cpu_callback(struct work_struct
*work
)
1094 struct smp_call_on_cpu_struct
*sscs
;
1096 sscs
= container_of(work
, struct smp_call_on_cpu_struct
, work
);
1098 hypervisor_pin_vcpu(sscs
->cpu
);
1099 sscs
->ret
= sscs
->func(sscs
->data
);
1101 hypervisor_pin_vcpu(-1);
1103 complete(&sscs
->done
);
1106 int smp_call_on_cpu(unsigned int cpu
, int (*func
)(void *), void *par
, bool phys
)
1108 struct smp_call_on_cpu_struct sscs
= {
1109 .done
= COMPLETION_INITIALIZER_ONSTACK(sscs
.done
),
1112 .cpu
= phys
? cpu
: -1,
1115 INIT_WORK_ONSTACK(&sscs
.work
, smp_call_on_cpu_callback
);
1117 if (cpu
>= nr_cpu_ids
|| !cpu_online(cpu
))
1120 queue_work_on(cpu
, system_wq
, &sscs
.work
);
1121 wait_for_completion(&sscs
.done
);
1122 destroy_work_on_stack(&sscs
.work
);
1126 EXPORT_SYMBOL_GPL(smp_call_on_cpu
);