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;
35 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
37 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
39 static void flush_smp_call_function_queue(bool warn_cpu_offline);
41 int smpcfd_prepare_cpu(unsigned int cpu)
43 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
45 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
48 cfd->csd = alloc_percpu(struct call_single_data);
50 free_cpumask_var(cfd->cpumask);
57 int smpcfd_dead_cpu(unsigned int cpu)
59 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
61 free_cpumask_var(cfd->cpumask);
62 free_percpu(cfd->csd);
66 int smpcfd_dying_cpu(unsigned int cpu)
69 * The IPIs for the smp-call-function callbacks queued by other
70 * CPUs might arrive late, either due to hardware latencies or
71 * because this CPU disabled interrupts (inside stop-machine)
72 * before the IPIs were sent. So flush out any pending callbacks
73 * explicitly (without waiting for the IPIs to arrive), to
74 * ensure that the outgoing CPU doesn't go offline with work
77 flush_smp_call_function_queue(false);
81 void __init call_function_init(void)
85 for_each_possible_cpu(i)
86 init_llist_head(&per_cpu(call_single_queue, i));
88 smpcfd_prepare_cpu(smp_processor_id());
92 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
94 * For non-synchronous ipi calls the csd can still be in use by the
95 * previous function call. For multi-cpu calls its even more interesting
96 * as we'll have to ensure no other cpu is observing our csd.
98 static __always_inline void csd_lock_wait(struct call_single_data *csd)
100 smp_cond_load_acquire(&csd->flags, !(VAL & CSD_FLAG_LOCK));
103 static __always_inline void csd_lock(struct call_single_data *csd)
106 csd->flags |= CSD_FLAG_LOCK;
109 * prevent CPU from reordering the above assignment
110 * to ->flags with any subsequent assignments to other
111 * fields of the specified call_single_data structure:
116 static __always_inline void csd_unlock(struct call_single_data *csd)
118 WARN_ON(!(csd->flags & CSD_FLAG_LOCK));
121 * ensure we're all done before releasing data:
123 smp_store_release(&csd->flags, 0);
126 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
129 * Insert a previously allocated call_single_data element
130 * for execution on the given CPU. data must already have
131 * ->func, ->info, and ->flags set.
133 static int generic_exec_single(int cpu, struct call_single_data *csd,
134 smp_call_func_t func, void *info)
136 if (cpu == smp_processor_id()) {
140 * We can unlock early even for the synchronous on-stack case,
141 * since we're doing this from the same CPU..
144 local_irq_save(flags);
146 local_irq_restore(flags);
151 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu)) {
160 * The list addition should be visible before sending the IPI
161 * handler locks the list to pull the entry off it because of
162 * normal cache coherency rules implied by spinlocks.
164 * If IPIs can go out of order to the cache coherency protocol
165 * in an architecture, sufficient synchronisation should be added
166 * to arch code to make it appear to obey cache coherency WRT
167 * locking and barrier primitives. Generic code isn't really
168 * equipped to do the right thing...
170 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
171 arch_send_call_function_single_ipi(cpu);
177 * generic_smp_call_function_single_interrupt - Execute SMP IPI callbacks
179 * Invoked by arch to handle an IPI for call function single.
180 * Must be called with interrupts disabled.
182 void generic_smp_call_function_single_interrupt(void)
184 flush_smp_call_function_queue(true);
188 * flush_smp_call_function_queue - Flush pending smp-call-function callbacks
190 * @warn_cpu_offline: If set to 'true', warn if callbacks were queued on an
191 * offline CPU. Skip this check if set to 'false'.
193 * Flush any pending smp-call-function callbacks queued on this CPU. This is
194 * invoked by the generic IPI handler, as well as by a CPU about to go offline,
195 * to ensure that all pending IPI callbacks are run before it goes completely
198 * Loop through the call_single_queue and run all the queued callbacks.
199 * Must be called with interrupts disabled.
201 static void flush_smp_call_function_queue(bool warn_cpu_offline)
203 struct llist_head *head;
204 struct llist_node *entry;
205 struct call_single_data *csd, *csd_next;
208 WARN_ON(!irqs_disabled());
210 head = this_cpu_ptr(&call_single_queue);
211 entry = llist_del_all(head);
212 entry = llist_reverse_order(entry);
214 /* There shouldn't be any pending callbacks on an offline CPU. */
215 if (unlikely(warn_cpu_offline && !cpu_online(smp_processor_id()) &&
216 !warned && !llist_empty(head))) {
218 WARN(1, "IPI on offline CPU %d\n", smp_processor_id());
221 * We don't have to use the _safe() variant here
222 * because we are not invoking the IPI handlers yet.
224 llist_for_each_entry(csd, entry, llist)
225 pr_warn("IPI callback %pS sent to offline CPU\n",
229 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
230 smp_call_func_t func = csd->func;
231 void *info = csd->info;
233 /* Do we wait until *after* callback? */
234 if (csd->flags & CSD_FLAG_SYNCHRONOUS) {
244 * Handle irq works queued remotely by irq_work_queue_on().
245 * Smp functions above are typically synchronous so they
246 * better run first since some other CPUs may be busy waiting
253 * smp_call_function_single - Run a function on a specific CPU
254 * @func: The function to run. This must be fast and non-blocking.
255 * @info: An arbitrary pointer to pass to the function.
256 * @wait: If true, wait until function has completed on other CPUs.
258 * Returns 0 on success, else a negative status code.
260 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
263 struct call_single_data *csd;
264 struct call_single_data csd_stack = { .flags = CSD_FLAG_LOCK | CSD_FLAG_SYNCHRONOUS };
269 * prevent preemption and reschedule on another processor,
270 * as well as CPU removal
272 this_cpu = get_cpu();
275 * Can deadlock when called with interrupts disabled.
276 * We allow cpu's that are not yet online though, as no one else can
277 * send smp call function interrupt to this cpu and as such deadlocks
280 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
281 && !oops_in_progress);
285 csd = this_cpu_ptr(&csd_data);
289 err = generic_exec_single(cpu, csd, func, info);
298 EXPORT_SYMBOL(smp_call_function_single);
301 * smp_call_function_single_async(): Run an asynchronous function on a
303 * @cpu: The CPU to run on.
304 * @csd: Pre-allocated and setup data structure
306 * Like smp_call_function_single(), but the call is asynchonous and
307 * can thus be done from contexts with disabled interrupts.
309 * The caller passes his own pre-allocated data structure
310 * (ie: embedded in an object) and is responsible for synchronizing it
311 * such that the IPIs performed on the @csd are strictly serialized.
313 * NOTE: Be careful, there is unfortunately no current debugging facility to
314 * validate the correctness of this serialization.
316 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
322 /* We could deadlock if we have to wait here with interrupts disabled! */
323 if (WARN_ON_ONCE(csd->flags & CSD_FLAG_LOCK))
326 csd->flags = CSD_FLAG_LOCK;
329 err = generic_exec_single(cpu, csd, csd->func, csd->info);
334 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
337 * smp_call_function_any - Run a function on any of the given cpus
338 * @mask: The mask of cpus it can run on.
339 * @func: The function to run. This must be fast and non-blocking.
340 * @info: An arbitrary pointer to pass to the function.
341 * @wait: If true, wait until function has completed.
343 * Returns 0 on success, else a negative status code (if no cpus were online).
345 * Selection preference:
346 * 1) current cpu if in @mask
347 * 2) any cpu of current node if in @mask
348 * 3) any other online cpu in @mask
350 int smp_call_function_any(const struct cpumask *mask,
351 smp_call_func_t func, void *info, int wait)
354 const struct cpumask *nodemask;
357 /* Try for same CPU (cheapest) */
359 if (cpumask_test_cpu(cpu, mask))
362 /* Try for same node. */
363 nodemask = cpumask_of_node(cpu_to_node(cpu));
364 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
365 cpu = cpumask_next_and(cpu, nodemask, mask)) {
370 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
371 cpu = cpumask_any_and(mask, cpu_online_mask);
373 ret = smp_call_function_single(cpu, func, info, wait);
377 EXPORT_SYMBOL_GPL(smp_call_function_any);
380 * smp_call_function_many(): Run a function on a set of other CPUs.
381 * @mask: The set of cpus to run on (only runs on online subset).
382 * @func: The function to run. This must be fast and non-blocking.
383 * @info: An arbitrary pointer to pass to the function.
384 * @wait: If true, wait (atomically) until function has completed
387 * If @wait is true, then returns once @func has returned.
389 * You must not call this function with disabled interrupts or from a
390 * hardware interrupt handler or from a bottom half handler. Preemption
391 * must be disabled when calling this function.
393 void smp_call_function_many(const struct cpumask *mask,
394 smp_call_func_t func, void *info, bool wait)
396 struct call_function_data *cfd;
397 int cpu, next_cpu, this_cpu = smp_processor_id();
400 * Can deadlock when called with interrupts disabled.
401 * We allow cpu's that are not yet online though, as no one else can
402 * send smp call function interrupt to this cpu and as such deadlocks
405 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
406 && !oops_in_progress && !early_boot_irqs_disabled);
408 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
409 cpu = cpumask_first_and(mask, cpu_online_mask);
411 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
413 /* No online cpus? We're done. */
414 if (cpu >= nr_cpu_ids)
417 /* Do we have another CPU which isn't us? */
418 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
419 if (next_cpu == this_cpu)
420 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
422 /* Fastpath: do that cpu by itself. */
423 if (next_cpu >= nr_cpu_ids) {
424 smp_call_function_single(cpu, func, info, wait);
428 cfd = this_cpu_ptr(&cfd_data);
430 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
431 cpumask_clear_cpu(this_cpu, cfd->cpumask);
433 /* Some callers race with other cpus changing the passed mask */
434 if (unlikely(!cpumask_weight(cfd->cpumask)))
437 for_each_cpu(cpu, cfd->cpumask) {
438 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
442 csd->flags |= CSD_FLAG_SYNCHRONOUS;
445 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
448 /* Send a message to all CPUs in the map */
449 arch_send_call_function_ipi_mask(cfd->cpumask);
452 for_each_cpu(cpu, cfd->cpumask) {
453 struct call_single_data *csd;
455 csd = per_cpu_ptr(cfd->csd, cpu);
460 EXPORT_SYMBOL(smp_call_function_many);
463 * smp_call_function(): Run a function on all other CPUs.
464 * @func: The function to run. This must be fast and non-blocking.
465 * @info: An arbitrary pointer to pass to the function.
466 * @wait: If true, wait (atomically) until function has completed
471 * If @wait is true, then returns once @func has returned; otherwise
472 * it returns just before the target cpu calls @func.
474 * You must not call this function with disabled interrupts or from a
475 * hardware interrupt handler or from a bottom half handler.
477 int smp_call_function(smp_call_func_t func, void *info, int wait)
480 smp_call_function_many(cpu_online_mask, func, info, wait);
485 EXPORT_SYMBOL(smp_call_function);
487 /* Setup configured maximum number of CPUs to activate */
488 unsigned int setup_max_cpus = NR_CPUS;
489 EXPORT_SYMBOL(setup_max_cpus);
493 * Setup routine for controlling SMP activation
495 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
496 * activation entirely (the MPS table probe still happens, though).
498 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
499 * greater than 0, limits the maximum number of CPUs activated in
503 void __weak arch_disable_smp_support(void) { }
505 static int __init nosmp(char *str)
508 arch_disable_smp_support();
513 early_param("nosmp", nosmp);
515 /* this is hard limit */
516 static int __init nrcpus(char *str)
520 get_option(&str, &nr_cpus);
521 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
522 nr_cpu_ids = nr_cpus;
527 early_param("nr_cpus", nrcpus);
529 static int __init maxcpus(char *str)
531 get_option(&str, &setup_max_cpus);
532 if (setup_max_cpus == 0)
533 arch_disable_smp_support();
538 early_param("maxcpus", maxcpus);
540 /* Setup number of possible processor ids */
541 int nr_cpu_ids __read_mostly = NR_CPUS;
542 EXPORT_SYMBOL(nr_cpu_ids);
544 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
545 void __init setup_nr_cpu_ids(void)
547 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
550 /* Called by boot processor to activate the rest. */
551 void __init smp_init(void)
553 int num_nodes, num_cpus;
557 cpuhp_threads_init();
559 pr_info("Bringing up secondary CPUs ...\n");
561 /* FIXME: This should be done in userspace --RR */
562 for_each_present_cpu(cpu) {
563 if (num_online_cpus() >= setup_max_cpus)
565 if (!cpu_online(cpu))
569 num_nodes = num_online_nodes();
570 num_cpus = num_online_cpus();
571 pr_info("Brought up %d node%s, %d CPU%s\n",
572 num_nodes, (num_nodes > 1 ? "s" : ""),
573 num_cpus, (num_cpus > 1 ? "s" : ""));
575 /* Any cleanup work */
576 smp_cpus_done(setup_max_cpus);
580 * Call a function on all processors. May be used during early boot while
581 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
582 * of local_irq_disable/enable().
584 int on_each_cpu(void (*func) (void *info), void *info, int wait)
590 ret = smp_call_function(func, info, wait);
591 local_irq_save(flags);
593 local_irq_restore(flags);
597 EXPORT_SYMBOL(on_each_cpu);
600 * on_each_cpu_mask(): Run a function on processors specified by
601 * cpumask, which may include the local processor.
602 * @mask: The set of cpus to run on (only runs on online subset).
603 * @func: The function to run. This must be fast and non-blocking.
604 * @info: An arbitrary pointer to pass to the function.
605 * @wait: If true, wait (atomically) until function has completed
608 * If @wait is true, then returns once @func has returned.
610 * You must not call this function with disabled interrupts or from a
611 * hardware interrupt handler or from a bottom half handler. The
612 * exception is that it may be used during early boot while
613 * early_boot_irqs_disabled is set.
615 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
616 void *info, bool wait)
620 smp_call_function_many(mask, func, info, wait);
621 if (cpumask_test_cpu(cpu, mask)) {
623 local_irq_save(flags);
625 local_irq_restore(flags);
629 EXPORT_SYMBOL(on_each_cpu_mask);
632 * on_each_cpu_cond(): Call a function on each processor for which
633 * the supplied function cond_func returns true, optionally waiting
634 * for all the required CPUs to finish. This may include the local
636 * @cond_func: A callback function that is passed a cpu id and
637 * the the info parameter. The function is called
638 * with preemption disabled. The function should
639 * return a blooean value indicating whether to IPI
641 * @func: The function to run on all applicable CPUs.
642 * This must be fast and non-blocking.
643 * @info: An arbitrary pointer to pass to both functions.
644 * @wait: If true, wait (atomically) until function has
645 * completed on other CPUs.
646 * @gfp_flags: GFP flags to use when allocating the cpumask
647 * used internally by the function.
649 * The function might sleep if the GFP flags indicates a non
650 * atomic allocation is allowed.
652 * Preemption is disabled to protect against CPUs going offline but not online.
653 * CPUs going online during the call will not be seen or sent an IPI.
655 * You must not call this function with disabled interrupts or
656 * from a hardware interrupt handler or from a bottom half handler.
658 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
659 smp_call_func_t func, void *info, bool wait,
665 might_sleep_if(gfpflags_allow_blocking(gfp_flags));
667 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
669 for_each_online_cpu(cpu)
670 if (cond_func(cpu, info))
671 cpumask_set_cpu(cpu, cpus);
672 on_each_cpu_mask(cpus, func, info, wait);
674 free_cpumask_var(cpus);
677 * No free cpumask, bother. No matter, we'll
678 * just have to IPI them one by one.
681 for_each_online_cpu(cpu)
682 if (cond_func(cpu, info)) {
683 ret = smp_call_function_single(cpu, func,
690 EXPORT_SYMBOL(on_each_cpu_cond);
692 static void do_nothing(void *unused)
697 * kick_all_cpus_sync - Force all cpus out of idle
699 * Used to synchronize the update of pm_idle function pointer. It's
700 * called after the pointer is updated and returns after the dummy
701 * callback function has been executed on all cpus. The execution of
702 * the function can only happen on the remote cpus after they have
703 * left the idle function which had been called via pm_idle function
704 * pointer. So it's guaranteed that nothing uses the previous pointer
707 void kick_all_cpus_sync(void)
709 /* Make sure the change is visible before we kick the cpus */
711 smp_call_function(do_nothing, NULL, 1);
713 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);
716 * wake_up_all_idle_cpus - break all cpus out of idle
717 * wake_up_all_idle_cpus try to break all cpus which is in idle state even
718 * including idle polling cpus, for non-idle cpus, we will do nothing
721 void wake_up_all_idle_cpus(void)
726 for_each_online_cpu(cpu) {
727 if (cpu == smp_processor_id())
730 wake_up_if_idle(cpu);
734 EXPORT_SYMBOL_GPL(wake_up_all_idle_cpus);
737 * smp_call_on_cpu - Call a function on a specific cpu
739 * Used to call a function on a specific cpu and wait for it to return.
740 * Optionally make sure the call is done on a specified physical cpu via vcpu
741 * pinning in order to support virtualized environments.
743 struct smp_call_on_cpu_struct {
744 struct work_struct work;
745 struct completion done;
752 static void smp_call_on_cpu_callback(struct work_struct *work)
754 struct smp_call_on_cpu_struct *sscs;
756 sscs = container_of(work, struct smp_call_on_cpu_struct, work);
758 hypervisor_pin_vcpu(sscs->cpu);
759 sscs->ret = sscs->func(sscs->data);
761 hypervisor_pin_vcpu(-1);
763 complete(&sscs->done);
766 int smp_call_on_cpu(unsigned int cpu, int (*func)(void *), void *par, bool phys)
768 struct smp_call_on_cpu_struct sscs = {
769 .done = COMPLETION_INITIALIZER_ONSTACK(sscs.done),
772 .cpu = phys ? cpu : -1,
775 INIT_WORK_ONSTACK(&sscs.work, smp_call_on_cpu_callback);
777 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
780 queue_work_on(cpu, system_wq, &sscs.work);
781 wait_for_completion(&sscs.done);
785 EXPORT_SYMBOL_GPL(smp_call_on_cpu);