2 * Generic helpers for smp ipi calls
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
6 #include <linux/rcupdate.h>
7 #include <linux/rculist.h>
8 #include <linux/kernel.h>
9 #include <linux/export.h>
10 #include <linux/percpu.h>
11 #include <linux/init.h>
12 #include <linux/gfp.h>
13 #include <linux/smp.h>
14 #include <linux/cpu.h>
23 struct call_function_data {
24 struct call_single_data __percpu *csd;
25 cpumask_var_t cpumask;
26 cpumask_var_t cpumask_ipi;
29 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
31 struct call_single_queue {
32 struct list_head list;
36 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
39 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
41 long cpu = (long)hcpu;
42 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
46 case CPU_UP_PREPARE_FROZEN:
47 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
49 return notifier_from_errno(-ENOMEM);
50 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
52 free_cpumask_var(cfd->cpumask);
53 return notifier_from_errno(-ENOMEM);
55 cfd->csd = alloc_percpu(struct call_single_data);
57 free_cpumask_var(cfd->cpumask_ipi);
58 free_cpumask_var(cfd->cpumask);
59 return notifier_from_errno(-ENOMEM);
63 #ifdef CONFIG_HOTPLUG_CPU
65 case CPU_UP_CANCELED_FROZEN:
69 free_cpumask_var(cfd->cpumask);
70 free_cpumask_var(cfd->cpumask_ipi);
71 free_percpu(cfd->csd);
79 static struct notifier_block hotplug_cfd_notifier = {
80 .notifier_call = hotplug_cfd,
83 void __init call_function_init(void)
85 void *cpu = (void *)(long)smp_processor_id();
88 for_each_possible_cpu(i) {
89 struct call_single_queue *q = &per_cpu(call_single_queue, i);
91 raw_spin_lock_init(&q->lock);
92 INIT_LIST_HEAD(&q->list);
95 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
96 register_cpu_notifier(&hotplug_cfd_notifier);
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 void csd_lock_wait(struct call_single_data *csd)
108 while (csd->flags & CSD_FLAG_LOCK)
112 static void csd_lock(struct call_single_data *csd)
115 csd->flags |= CSD_FLAG_LOCK;
118 * prevent CPU from reordering the above assignment
119 * to ->flags with any subsequent assignments to other
120 * fields of the specified call_single_data structure:
125 static void csd_unlock(struct call_single_data *csd)
127 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
130 * ensure we're all done before releasing data:
134 csd->flags &= ~CSD_FLAG_LOCK;
138 * Insert a previously allocated call_single_data element
139 * for execution on the given CPU. data must already have
140 * ->func, ->info, and ->flags set.
142 static void generic_exec_single(int cpu, struct call_single_data *csd, int wait)
144 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
149 csd->flags |= CSD_FLAG_WAIT;
151 raw_spin_lock_irqsave(&dst->lock, flags);
152 ipi = list_empty(&dst->list);
153 list_add_tail(&csd->list, &dst->list);
154 raw_spin_unlock_irqrestore(&dst->lock, flags);
157 * The list addition should be visible before sending the IPI
158 * handler locks the list to pull the entry off it because of
159 * normal cache coherency rules implied by spinlocks.
161 * If IPIs can go out of order to the cache coherency protocol
162 * in an architecture, sufficient synchronisation should be added
163 * to arch code to make it appear to obey cache coherency WRT
164 * locking and barrier primitives. Generic code isn't really
165 * equipped to do the right thing...
168 arch_send_call_function_single_ipi(cpu);
175 * Invoked by arch to handle an IPI for call function single. Must be
176 * called from the arch with interrupts disabled.
178 void generic_smp_call_function_single_interrupt(void)
180 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
184 * Shouldn't receive this interrupt on a cpu that is not yet online.
186 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
188 raw_spin_lock(&q->lock);
189 list_replace_init(&q->list, &list);
190 raw_spin_unlock(&q->lock);
192 while (!list_empty(&list)) {
193 struct call_single_data *csd;
195 csd = list_entry(list.next, struct call_single_data, list);
196 list_del(&csd->list);
198 csd->func(csd->info);
204 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
207 * smp_call_function_single - Run a function on a specific CPU
208 * @func: The function to run. This must be fast and non-blocking.
209 * @info: An arbitrary pointer to pass to the function.
210 * @wait: If true, wait until function has completed on other CPUs.
212 * Returns 0 on success, else a negative status code.
214 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
217 struct call_single_data d = {
225 * prevent preemption and reschedule on another processor,
226 * as well as CPU removal
228 this_cpu = get_cpu();
231 * Can deadlock when called with interrupts disabled.
232 * We allow cpu's that are not yet online though, as no one else can
233 * send smp call function interrupt to this cpu and as such deadlocks
236 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
237 && !oops_in_progress);
239 if (cpu == this_cpu) {
240 local_irq_save(flags);
242 local_irq_restore(flags);
244 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
245 struct call_single_data *csd = &d;
248 csd = &__get_cpu_var(csd_data);
254 generic_exec_single(cpu, csd, wait);
256 err = -ENXIO; /* CPU not online */
264 EXPORT_SYMBOL(smp_call_function_single);
267 * smp_call_function_any - Run a function on any of the given cpus
268 * @mask: The mask of cpus it can run on.
269 * @func: The function to run. This must be fast and non-blocking.
270 * @info: An arbitrary pointer to pass to the function.
271 * @wait: If true, wait until function has completed.
273 * Returns 0 on success, else a negative status code (if no cpus were online).
275 * Selection preference:
276 * 1) current cpu if in @mask
277 * 2) any cpu of current node if in @mask
278 * 3) any other online cpu in @mask
280 int smp_call_function_any(const struct cpumask *mask,
281 smp_call_func_t func, void *info, int wait)
284 const struct cpumask *nodemask;
287 /* Try for same CPU (cheapest) */
289 if (cpumask_test_cpu(cpu, mask))
292 /* Try for same node. */
293 nodemask = cpumask_of_node(cpu_to_node(cpu));
294 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
295 cpu = cpumask_next_and(cpu, nodemask, mask)) {
300 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
301 cpu = cpumask_any_and(mask, cpu_online_mask);
303 ret = smp_call_function_single(cpu, func, info, wait);
307 EXPORT_SYMBOL_GPL(smp_call_function_any);
310 * __smp_call_function_single(): Run a function on a specific CPU
311 * @cpu: The CPU to run on.
312 * @data: Pre-allocated and setup data structure
313 * @wait: If true, wait until function has completed on specified CPU.
315 * Like smp_call_function_single(), but allow caller to pass in a
316 * pre-allocated data structure. Useful for embedding @data inside
317 * other structures, for instance.
319 void __smp_call_function_single(int cpu, struct call_single_data *csd,
322 unsigned int this_cpu;
325 this_cpu = get_cpu();
327 * Can deadlock when called with interrupts disabled.
328 * We allow cpu's that are not yet online though, as no one else can
329 * send smp call function interrupt to this cpu and as such deadlocks
332 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
333 && !oops_in_progress);
335 if (cpu == this_cpu) {
336 local_irq_save(flags);
337 csd->func(csd->info);
338 local_irq_restore(flags);
341 generic_exec_single(cpu, csd, wait);
345 EXPORT_SYMBOL_GPL(__smp_call_function_single);
348 * smp_call_function_many(): Run a function on a set of other CPUs.
349 * @mask: The set of cpus to run on (only runs on online subset).
350 * @func: The function to run. This must be fast and non-blocking.
351 * @info: An arbitrary pointer to pass to the function.
352 * @wait: If true, wait (atomically) until function has completed
355 * If @wait is true, then returns once @func has returned.
357 * You must not call this function with disabled interrupts or from a
358 * hardware interrupt handler or from a bottom half handler. Preemption
359 * must be disabled when calling this function.
361 void smp_call_function_many(const struct cpumask *mask,
362 smp_call_func_t func, void *info, bool wait)
364 struct call_function_data *cfd;
365 int cpu, next_cpu, this_cpu = smp_processor_id();
368 * Can deadlock when called with interrupts disabled.
369 * We allow cpu's that are not yet online though, as no one else can
370 * send smp call function interrupt to this cpu and as such deadlocks
373 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
374 && !oops_in_progress && !early_boot_irqs_disabled);
376 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
377 cpu = cpumask_first_and(mask, cpu_online_mask);
379 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
381 /* No online cpus? We're done. */
382 if (cpu >= nr_cpu_ids)
385 /* Do we have another CPU which isn't us? */
386 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
387 if (next_cpu == this_cpu)
388 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
390 /* Fastpath: do that cpu by itself. */
391 if (next_cpu >= nr_cpu_ids) {
392 smp_call_function_single(cpu, func, info, wait);
396 cfd = &__get_cpu_var(cfd_data);
398 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
399 cpumask_clear_cpu(this_cpu, cfd->cpumask);
401 /* Some callers race with other cpus changing the passed mask */
402 if (unlikely(!cpumask_weight(cfd->cpumask)))
406 * After we put an entry into the list, cfd->cpumask may be cleared
407 * again when another CPU sends another IPI for a SMP function call, so
408 * cfd->cpumask will be zero.
410 cpumask_copy(cfd->cpumask_ipi, cfd->cpumask);
412 for_each_cpu(cpu, cfd->cpumask) {
413 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
414 struct call_single_queue *dst =
415 &per_cpu(call_single_queue, cpu);
422 raw_spin_lock_irqsave(&dst->lock, flags);
423 list_add_tail(&csd->list, &dst->list);
424 raw_spin_unlock_irqrestore(&dst->lock, flags);
427 /* Send a message to all CPUs in the map */
428 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
431 for_each_cpu(cpu, cfd->cpumask) {
432 struct call_single_data *csd;
434 csd = per_cpu_ptr(cfd->csd, cpu);
439 EXPORT_SYMBOL(smp_call_function_many);
442 * smp_call_function(): Run a function on all other CPUs.
443 * @func: The function to run. This must be fast and non-blocking.
444 * @info: An arbitrary pointer to pass to the function.
445 * @wait: If true, wait (atomically) until function has completed
450 * If @wait is true, then returns once @func has returned; otherwise
451 * it returns just before the target cpu calls @func.
453 * You must not call this function with disabled interrupts or from a
454 * hardware interrupt handler or from a bottom half handler.
456 int smp_call_function(smp_call_func_t func, void *info, int wait)
459 smp_call_function_many(cpu_online_mask, func, info, wait);
464 EXPORT_SYMBOL(smp_call_function);
466 /* Setup configured maximum number of CPUs to activate */
467 unsigned int setup_max_cpus = NR_CPUS;
468 EXPORT_SYMBOL(setup_max_cpus);
472 * Setup routine for controlling SMP activation
474 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
475 * activation entirely (the MPS table probe still happens, though).
477 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
478 * greater than 0, limits the maximum number of CPUs activated in
482 void __weak arch_disable_smp_support(void) { }
484 static int __init nosmp(char *str)
487 arch_disable_smp_support();
492 early_param("nosmp", nosmp);
494 /* this is hard limit */
495 static int __init nrcpus(char *str)
499 get_option(&str, &nr_cpus);
500 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
501 nr_cpu_ids = nr_cpus;
506 early_param("nr_cpus", nrcpus);
508 static int __init maxcpus(char *str)
510 get_option(&str, &setup_max_cpus);
511 if (setup_max_cpus == 0)
512 arch_disable_smp_support();
517 early_param("maxcpus", maxcpus);
519 /* Setup number of possible processor ids */
520 int nr_cpu_ids __read_mostly = NR_CPUS;
521 EXPORT_SYMBOL(nr_cpu_ids);
523 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
524 void __init setup_nr_cpu_ids(void)
526 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
529 void __weak smp_announce(void)
531 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
534 /* Called by boot processor to activate the rest. */
535 void __init smp_init(void)
541 /* FIXME: This should be done in userspace --RR */
542 for_each_present_cpu(cpu) {
543 if (num_online_cpus() >= setup_max_cpus)
545 if (!cpu_online(cpu))
549 /* Any cleanup work */
551 smp_cpus_done(setup_max_cpus);
555 * Call a function on all processors. May be used during early boot while
556 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
557 * of local_irq_disable/enable().
559 int on_each_cpu(void (*func) (void *info), void *info, int wait)
565 ret = smp_call_function(func, info, wait);
566 local_irq_save(flags);
568 local_irq_restore(flags);
572 EXPORT_SYMBOL(on_each_cpu);
575 * on_each_cpu_mask(): Run a function on processors specified by
576 * cpumask, which may include the local processor.
577 * @mask: The set of cpus to run on (only runs on online subset).
578 * @func: The function to run. This must be fast and non-blocking.
579 * @info: An arbitrary pointer to pass to the function.
580 * @wait: If true, wait (atomically) until function has completed
583 * If @wait is true, then returns once @func has returned.
585 * You must not call this function with disabled interrupts or from a
586 * hardware interrupt handler or from a bottom half handler. The
587 * exception is that it may be used during early boot while
588 * early_boot_irqs_disabled is set.
590 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
591 void *info, bool wait)
595 smp_call_function_many(mask, func, info, wait);
596 if (cpumask_test_cpu(cpu, mask)) {
598 local_irq_save(flags);
600 local_irq_restore(flags);
604 EXPORT_SYMBOL(on_each_cpu_mask);
607 * on_each_cpu_cond(): Call a function on each processor for which
608 * the supplied function cond_func returns true, optionally waiting
609 * for all the required CPUs to finish. This may include the local
611 * @cond_func: A callback function that is passed a cpu id and
612 * the the info parameter. The function is called
613 * with preemption disabled. The function should
614 * return a blooean value indicating whether to IPI
616 * @func: The function to run on all applicable CPUs.
617 * This must be fast and non-blocking.
618 * @info: An arbitrary pointer to pass to both functions.
619 * @wait: If true, wait (atomically) until function has
620 * completed on other CPUs.
621 * @gfp_flags: GFP flags to use when allocating the cpumask
622 * used internally by the function.
624 * The function might sleep if the GFP flags indicates a non
625 * atomic allocation is allowed.
627 * Preemption is disabled to protect against CPUs going offline but not online.
628 * CPUs going online during the call will not be seen or sent an IPI.
630 * You must not call this function with disabled interrupts or
631 * from a hardware interrupt handler or from a bottom half handler.
633 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
634 smp_call_func_t func, void *info, bool wait,
640 might_sleep_if(gfp_flags & __GFP_WAIT);
642 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
644 for_each_online_cpu(cpu)
645 if (cond_func(cpu, info))
646 cpumask_set_cpu(cpu, cpus);
647 on_each_cpu_mask(cpus, func, info, wait);
649 free_cpumask_var(cpus);
652 * No free cpumask, bother. No matter, we'll
653 * just have to IPI them one by one.
656 for_each_online_cpu(cpu)
657 if (cond_func(cpu, info)) {
658 ret = smp_call_function_single(cpu, func,
665 EXPORT_SYMBOL(on_each_cpu_cond);
667 static void do_nothing(void *unused)
672 * kick_all_cpus_sync - Force all cpus out of idle
674 * Used to synchronize the update of pm_idle function pointer. It's
675 * called after the pointer is updated and returns after the dummy
676 * callback function has been executed on all cpus. The execution of
677 * the function can only happen on the remote cpus after they have
678 * left the idle function which had been called via pm_idle function
679 * pointer. So it's guaranteed that nothing uses the previous pointer
682 void kick_all_cpus_sync(void)
684 /* Make sure the change is visible before we kick the cpus */
686 smp_call_function(do_nothing, NULL, 1);
688 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);