2 * Generic helpers for smp ipi calls
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9 #include <linux/irq_work.h>
10 #include <linux/rcupdate.h>
11 #include <linux/rculist.h>
12 #include <linux/kernel.h>
13 #include <linux/export.h>
14 #include <linux/percpu.h>
15 #include <linux/init.h>
16 #include <linux/gfp.h>
17 #include <linux/smp.h>
18 #include <linux/cpu.h>
19 #include <linux/sched.h>
20 #include <linux/sched/idle.h>
21 #include <linux/hypervisor.h>
27 CSD_FLAG_SYNCHRONOUS = 0x02,
30 struct call_function_data {
31 struct call_single_data __percpu *csd;
32 cpumask_var_t cpumask;
33 cpumask_var_t cpumask_ipi;
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
38 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
40 static void flush_smp_call_function_queue(bool warn_cpu_offline);
42 int smpcfd_prepare_cpu(unsigned int cpu)
44 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
46 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
49 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
51 free_cpumask_var(cfd->cpumask);
54 cfd->csd = alloc_percpu(struct call_single_data);
56 free_cpumask_var(cfd->cpumask);
57 free_cpumask_var(cfd->cpumask_ipi);
64 int smpcfd_dead_cpu(unsigned int cpu)
66 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
68 free_cpumask_var(cfd->cpumask);
69 free_cpumask_var(cfd->cpumask_ipi);
70 free_percpu(cfd->csd);
74 int smpcfd_dying_cpu(unsigned int cpu)
77 * The IPIs for the smp-call-function callbacks queued by other
78 * CPUs might arrive late, either due to hardware latencies or
79 * because this CPU disabled interrupts (inside stop-machine)
80 * before the IPIs were sent. So flush out any pending callbacks
81 * explicitly (without waiting for the IPIs to arrive), to
82 * ensure that the outgoing CPU doesn't go offline with work
85 flush_smp_call_function_queue(false);
89 void __init call_function_init(void)
93 for_each_possible_cpu(i)
94 init_llist_head(&per_cpu(call_single_queue, i));
96 smpcfd_prepare_cpu(smp_processor_id());
100 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
102 * For non-synchronous ipi calls the csd can still be in use by the
103 * previous function call. For multi-cpu calls its even more interesting
104 * as we'll have to ensure no other cpu is observing our csd.
106 static __always_inline void csd_lock_wait(struct call_single_data *csd)
108 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
111 static __always_inline void csd_lock(struct call_single_data *csd)
114 csd->flags |= CSD_FLAG_LOCK;
117 * prevent CPU from reordering the above assignment
118 * to ->flags with any subsequent assignments to other
119 * fields of the specified call_single_data structure:
124 static __always_inline void csd_unlock(struct call_single_data *csd)
126 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
129 * ensure we're all done before releasing data:
131 smp_store_release(&csd->flags, 0);
134 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
137 * Insert a previously allocated call_single_data element
138 * for execution on the given CPU. data must already have
139 * ->func, ->info, and ->flags set.
141 static int generic_exec_single(int cpu, struct call_single_data *csd,
142 smp_call_func_t func, void *info)
144 if (cpu == smp_processor_id()) {
148 * We can unlock early even for the synchronous on-stack case,
149 * since we're doing this from the same CPU..
152 local_irq_save(flags);
154 local_irq_restore(flags);
159 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
168 * The list addition should be visible before sending the IPI
169 * handler locks the list to pull the entry off it because of
170 * normal cache coherency rules implied by spinlocks.
172 * If IPIs can go out of order to the cache coherency protocol
173 * in an architecture, sufficient synchronisation should be added
174 * to arch code to make it appear to obey cache coherency WRT
175 * locking and barrier primitives. Generic code isn't really
176 * equipped to do the right thing...
178 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
179 arch_send_call_function_single_ipi(cpu);
185 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
187 * Invoked by arch to handle an IPI for call function single.
188 * Must be called with interrupts disabled.
190 void generic_smp_call_function_single_interrupt(void)
192 flush_smp_call_function_queue(true);
196 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
198 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
199 * offline CPU. Skip this check if set to 'false'.
201 * Flush any pending smp-call-function callbacks queued on this CPU. This is
202 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
203 * to ensure that all pending IPI callbacks are run before it goes completely
206 * Loop through the call_single_queue and run all the queued callbacks.
207 * Must be called with interrupts disabled.
209 static void flush_smp_call_function_queue(bool warn_cpu_offline)
211 struct llist_head *head;
212 struct llist_node *entry;
213 struct call_single_data *csd, *csd_next;
216 WARN_ON(!irqs_disabled());
218 head = this_cpu_ptr(&call_single_queue);
219 entry = llist_del_all(head);
220 entry = llist_reverse_order(entry);
222 /* There shouldn't be any pending callbacks on an offline CPU. */
223 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
224 !warned && !llist_empty(head))) {
226 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
229 * We don't have to use the _safe() variant here
230 * because we are not invoking the IPI handlers yet.
232 llist_for_each_entry(csd, entry, llist)
233 pr_warn("IPI callback %pS sent to offline CPU\n",
237 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
238 smp_call_func_t func = csd->func;
239 void *info = csd->info;
241 /* Do we wait until *after* callback? */
242 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
252 * Handle irq works queued remotely by irq_work_queue_on().
253 * Smp functions above are typically synchronous so they
254 * better run first since some other CPUs may be busy waiting
261 * smp_call_function_single - Run a function on a specific CPU
262 * @func: The function to run. This must be fast and non-blocking.
263 * @info: An arbitrary pointer to pass to the function.
264 * @wait: If true, wait until function has completed on other CPUs.
266 * Returns 0 on success, else a negative status code.
268 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
271 struct call_single_data *csd;
272 struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS };
277 * prevent preemption and reschedule on another processor,
278 * as well as CPU removal
280 this_cpu = get_cpu();
283 * Can deadlock when called with interrupts disabled.
284 * We allow cpu's that are not yet online though, as no one else can
285 * send smp call function interrupt to this cpu and as such deadlocks
288 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
289 && !oops_in_progress);
293 csd = this_cpu_ptr(&csd_data);
297 err = generic_exec_single(cpu, csd, func, info);
306 EXPORT_SYMBOL(smp_call_function_single);
309 * smp_call_function_single_async(): Run an asynchronous function on a
311 * @cpu: The CPU to run on.
312 * @csd: Pre-allocated and setup data structure
314 * Like smp_call_function_single(), but the call is asynchonous and
315 * can thus be done from contexts with disabled interrupts.
317 * The caller passes his own pre-allocated data structure
318 * (ie: embedded in an object) and is responsible for synchronizing it
319 * such that the IPIs performed on the @csd are strictly serialized.
321 * NOTE: Be careful, there is unfortunately no current debugging facility to
322 * validate the correctness of this serialization.
324 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
330 /* We could deadlock if we have to wait here with interrupts disabled! */
331 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
334 csd->flags = CSD_FLAG_LOCK;
337 err = generic_exec_single(cpu, csd, csd->func, csd->info);
342 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
345 * smp_call_function_any - Run a function on any of the given cpus
346 * @mask: The mask of cpus it can run on.
347 * @func: The function to run. This must be fast and non-blocking.
348 * @info: An arbitrary pointer to pass to the function.
349 * @wait: If true, wait until function has completed.
351 * Returns 0 on success, else a negative status code (if no cpus were online).
353 * Selection preference:
354 * 1) current cpu if in @mask
355 * 2) any cpu of current node if in @mask
356 * 3) any other online cpu in @mask
358 int smp_call_function_any(const struct cpumask *mask,
359 smp_call_func_t func, void *info, int wait)
362 const struct cpumask *nodemask;
365 /* Try for same CPU (cheapest) */
367 if (cpumask_test_cpu(cpu, mask))
370 /* Try for same node. */
371 nodemask = cpumask_of_node(cpu_to_node(cpu));
372 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
373 cpu = cpumask_next_and(cpu, nodemask, mask)) {
378 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
379 cpu = cpumask_any_and(mask, cpu_online_mask);
381 ret = smp_call_function_single(cpu, func, info, wait);
385 EXPORT_SYMBOL_GPL(smp_call_function_any);
388 * smp_call_function_many(): Run a function on a set of other CPUs.
389 * @mask: The set of cpus to run on (only runs on online subset).
390 * @func: The function to run. This must be fast and non-blocking.
391 * @info: An arbitrary pointer to pass to the function.
392 * @wait: If true, wait (atomically) until function has completed
395 * If @wait is true, then returns once @func has returned.
397 * You must not call this function with disabled interrupts or from a
398 * hardware interrupt handler or from a bottom half handler. Preemption
399 * must be disabled when calling this function.
401 void smp_call_function_many(const struct cpumask *mask,
402 smp_call_func_t func, void *info, bool wait)
404 struct call_function_data *cfd;
405 int cpu, next_cpu, this_cpu = smp_processor_id();
408 * Can deadlock when called with interrupts disabled.
409 * We allow cpu's that are not yet online though, as no one else can
410 * send smp call function interrupt to this cpu and as such deadlocks
413 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
414 && !oops_in_progress && !early_boot_irqs_disabled);
416 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
417 cpu = cpumask_first_and(mask, cpu_online_mask);
419 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
421 /* No online cpus? We're done. */
422 if (cpu >= nr_cpu_ids)
425 /* Do we have another CPU which isn't us? */
426 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
427 if (next_cpu == this_cpu)
428 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
430 /* Fastpath: do that cpu by itself. */
431 if (next_cpu >= nr_cpu_ids) {
432 smp_call_function_single(cpu, func, info, wait);
436 cfd = this_cpu_ptr(&cfd_data);
438 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
439 __cpumask_clear_cpu(this_cpu, cfd->cpumask);
441 /* Some callers race with other cpus changing the passed mask */
442 if (unlikely(!cpumask_weight(cfd->cpumask)))
445 cpumask_clear(cfd->cpumask_ipi);
446 for_each_cpu(cpu, cfd->cpumask) {
447 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
451 csd->flags |= CSD_FLAG_SYNCHRONOUS;
454 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
455 __cpumask_set_cpu(cpu, cfd->cpumask_ipi);
458 /* Send a message to all CPUs in the map */
459 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
462 for_each_cpu(cpu, cfd->cpumask) {
463 struct call_single_data *csd;
465 csd = per_cpu_ptr(cfd->csd, cpu);
470 EXPORT_SYMBOL(smp_call_function_many);
473 * smp_call_function(): Run a function on all other CPUs.
474 * @func: The function to run. This must be fast and non-blocking.
475 * @info: An arbitrary pointer to pass to the function.
476 * @wait: If true, wait (atomically) until function has completed
481 * If @wait is true, then returns once @func has returned; otherwise
482 * it returns just before the target cpu calls @func.
484 * You must not call this function with disabled interrupts or from a
485 * hardware interrupt handler or from a bottom half handler.
487 int smp_call_function(smp_call_func_t func, void *info, int wait)
490 smp_call_function_many(cpu_online_mask, func, info, wait);
495 EXPORT_SYMBOL(smp_call_function);
497 /* Setup configured maximum number of CPUs to activate */
498 unsigned int setup_max_cpus = NR_CPUS;
499 EXPORT_SYMBOL(setup_max_cpus);
503 * Setup routine for controlling SMP activation
505 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
506 * activation entirely (the MPS table probe still happens, though).
508 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
509 * greater than 0, limits the maximum number of CPUs activated in
513 void __weak arch_disable_smp_support(void) { }
515 static int __init nosmp(char *str)
518 arch_disable_smp_support();
523 early_param("nosmp", nosmp);
525 /* this is hard limit */
526 static int __init nrcpus(char *str)
530 get_option(&str, &nr_cpus);
531 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
532 nr_cpu_ids = nr_cpus;
537 early_param("nr_cpus", nrcpus);
539 static int __init maxcpus(char *str)
541 get_option(&str, &setup_max_cpus);
542 if (setup_max_cpus == 0)
543 arch_disable_smp_support();
548 early_param("maxcpus", maxcpus);
550 /* Setup number of possible processor ids */
551 int nr_cpu_ids __read_mostly = NR_CPUS;
552 EXPORT_SYMBOL(nr_cpu_ids);
554 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
555 void __init setup_nr_cpu_ids(void)
557 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
560 /* Called by boot processor to activate the rest. */
561 void __init smp_init(void)
563 int num_nodes, num_cpus;
567 cpuhp_threads_init();
569 pr_info("Bringing up secondary CPUs ...\n");
571 /* FIXME: This should be done in userspace --RR */
572 for_each_present_cpu(cpu) {
573 if (num_online_cpus() >= setup_max_cpus)
575 if (!cpu_online(cpu))
579 num_nodes = num_online_nodes();
580 num_cpus = num_online_cpus();
581 pr_info("Brought up %d node%s, %d CPU%s\n",
582 num_nodes, (num_nodes > 1 ? "s" : ""),
583 num_cpus, (num_cpus > 1 ? "s" : ""));
585 /* Any cleanup work */
586 smp_cpus_done(setup_max_cpus);
590 * Call a function on all processors. May be used during early boot while
591 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
592 * of local_irq_disable/enable().
594 int on_each_cpu(void (*func) (void *info), void *info, int wait)
600 ret = smp_call_function(func, info, wait);
601 local_irq_save(flags);
603 local_irq_restore(flags);
607 EXPORT_SYMBOL(on_each_cpu);
610 * on_each_cpu_mask(): Run a function on processors specified by
611 * cpumask, which may include the local processor.
612 * @mask: The set of cpus to run on (only runs on online subset).
613 * @func: The function to run. This must be fast and non-blocking.
614 * @info: An arbitrary pointer to pass to the function.
615 * @wait: If true, wait (atomically) until function has completed
618 * If @wait is true, then returns once @func has returned.
620 * You must not call this function with disabled interrupts or from a
621 * hardware interrupt handler or from a bottom half handler. The
622 * exception is that it may be used during early boot while
623 * early_boot_irqs_disabled is set.
625 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
626 void *info, bool wait)
630 smp_call_function_many(mask, func, info, wait);
631 if (cpumask_test_cpu(cpu, mask)) {
633 local_irq_save(flags);
635 local_irq_restore(flags);
639 EXPORT_SYMBOL(on_each_cpu_mask);
642 * on_each_cpu_cond(): Call a function on each processor for which
643 * the supplied function cond_func returns true, optionally waiting
644 * for all the required CPUs to finish. This may include the local
646 * @cond_func: A callback function that is passed a cpu id and
647 * the the info parameter. The function is called
648 * with preemption disabled. The function should
649 * return a blooean value indicating whether to IPI
651 * @func: The function to run on all applicable CPUs.
652 * This must be fast and non-blocking.
653 * @info: An arbitrary pointer to pass to both functions.
654 * @wait: If true, wait (atomically) until function has
655 * completed on other CPUs.
656 * @gfp_flags: GFP flags to use when allocating the cpumask
657 * used internally by the function.
659 * The function might sleep if the GFP flags indicates a non
660 * atomic allocation is allowed.
662 * Preemption is disabled to protect against CPUs going offline but not online.
663 * CPUs going online during the call will not be seen or sent an IPI.
665 * You must not call this function with disabled interrupts or
666 * from a hardware interrupt handler or from a bottom half handler.
668 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
669 smp_call_func_t func, void *info, bool wait,
675 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
677 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
679 for_each_online_cpu(cpu)
680 if (cond_func(cpu, info))
681 cpumask_set_cpu(cpu, cpus);
682 on_each_cpu_mask(cpus, func, info, wait);
684 free_cpumask_var(cpus);
687 * No free cpumask, bother. No matter, we'll
688 * just have to IPI them one by one.
691 for_each_online_cpu(cpu)
692 if (cond_func(cpu, info)) {
693 ret = smp_call_function_single(cpu, func,
700 EXPORT_SYMBOL(on_each_cpu_cond);
702 static void do_nothing(void *unused)
707 * kick_all_cpus_sync - Force all cpus out of idle
709 * Used to synchronize the update of pm_idle function pointer. It's
710 * called after the pointer is updated and returns after the dummy
711 * callback function has been executed on all cpus. The execution of
712 * the function can only happen on the remote cpus after they have
713 * left the idle function which had been called via pm_idle function
714 * pointer. So it's guaranteed that nothing uses the previous pointer
717 void kick_all_cpus_sync(void)
719 /* Make sure the change is visible before we kick the cpus */
721 smp_call_function(do_nothing, NULL, 1);
723 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
726 * wake_up_all_idle_cpus - break all cpus out of idle
727 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
728 * including idle polling cpus, for non-idle cpus, we will do nothing
731 void wake_up_all_idle_cpus(void)
736 for_each_online_cpu(cpu) {
737 if (cpu == smp_processor_id())
740 wake_up_if_idle(cpu);
744 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
747 * smp_call_on_cpu - Call a function on a specific cpu
749 * Used to call a function on a specific cpu and wait for it to return.
750 * Optionally make sure the call is done on a specified physical cpu via vcpu
751 * pinning in order to support virtualized environments.
753 struct smp_call_on_cpu_struct {
754 struct work_struct work;
755 struct completion done;
762 static void smp_call_on_cpu_callback(struct work_struct *work)
764 struct smp_call_on_cpu_struct *sscs;
766 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
768 hypervisor_pin_vcpu(sscs->cpu);
769 sscs->ret = sscs->func(sscs->data);
771 hypervisor_pin_vcpu(-1);
773 complete(&sscs->done);
776 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
778 struct smp_call_on_cpu_struct sscs = {
779 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
782 .cpu = phys ? cpu : -1,
785 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
787 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
790 queue_work_on(cpu, system_wq, &sscs.work);
791 wait_for_completion(&sscs.done);
795 EXPORT_SYMBOL_GPL(smp_call_on_cpu);