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>
18 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
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_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.
143 void generic_exec_single(int cpu, struct call_single_data *csd, int wait)
145 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
149 raw_spin_lock_irqsave(&dst->lock, flags);
150 ipi = list_empty(&dst->list);
151 list_add_tail(&csd->list, &dst->list);
152 raw_spin_unlock_irqrestore(&dst->lock, flags);
155 * The list addition should be visible before sending the IPI
156 * handler locks the list to pull the entry off it because of
157 * normal cache coherency rules implied by spinlocks.
159 * If IPIs can go out of order to the cache coherency protocol
160 * in an architecture, sufficient synchronisation should be added
161 * to arch code to make it appear to obey cache coherency WRT
162 * locking and barrier primitives. Generic code isn't really
163 * equipped to do the right thing...
166 arch_send_call_function_single_ipi(cpu);
173 * Invoked by arch to handle an IPI for call function single. Must be
174 * called from the arch with interrupts disabled.
176 void generic_smp_call_function_single_interrupt(void)
178 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
182 * Shouldn't receive this interrupt on a cpu that is not yet online.
184 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
186 raw_spin_lock(&q->lock);
187 list_replace_init(&q->list, &list);
188 raw_spin_unlock(&q->lock);
190 while (!list_empty(&list)) {
191 struct call_single_data *csd;
193 csd = list_entry(list.next, struct call_single_data, list);
194 list_del(&csd->list);
196 csd->func(csd->info);
202 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
205 * smp_call_function_single - Run a function on a specific CPU
206 * @func: The function to run. This must be fast and non-blocking.
207 * @info: An arbitrary pointer to pass to the function.
208 * @wait: If true, wait until function has completed on other CPUs.
210 * Returns 0 on success, else a negative status code.
212 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
215 struct call_single_data d = {
223 * prevent preemption and reschedule on another processor,
224 * as well as CPU removal
226 this_cpu = get_cpu();
229 * Can deadlock when called with interrupts disabled.
230 * We allow cpu's that are not yet online though, as no one else can
231 * send smp call function interrupt to this cpu and as such deadlocks
234 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
235 && !oops_in_progress);
237 if (cpu == this_cpu) {
238 local_irq_save(flags);
240 local_irq_restore(flags);
242 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
243 struct call_single_data *csd = &d;
246 csd = &__get_cpu_var(csd_data);
252 generic_exec_single(cpu, csd, wait);
254 err = -ENXIO; /* CPU not online */
262 EXPORT_SYMBOL(smp_call_function_single);
265 * smp_call_function_any - Run a function on any of the given cpus
266 * @mask: The mask of cpus it can run on.
267 * @func: The function to run. This must be fast and non-blocking.
268 * @info: An arbitrary pointer to pass to the function.
269 * @wait: If true, wait until function has completed.
271 * Returns 0 on success, else a negative status code (if no cpus were online).
273 * Selection preference:
274 * 1) current cpu if in @mask
275 * 2) any cpu of current node if in @mask
276 * 3) any other online cpu in @mask
278 int smp_call_function_any(const struct cpumask *mask,
279 smp_call_func_t func, void *info, int wait)
282 const struct cpumask *nodemask;
285 /* Try for same CPU (cheapest) */
287 if (cpumask_test_cpu(cpu, mask))
290 /* Try for same node. */
291 nodemask = cpumask_of_node(cpu_to_node(cpu));
292 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
293 cpu = cpumask_next_and(cpu, nodemask, mask)) {
298 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
299 cpu = cpumask_any_and(mask, cpu_online_mask);
301 ret = smp_call_function_single(cpu, func, info, wait);
305 EXPORT_SYMBOL_GPL(smp_call_function_any);
308 * __smp_call_function_single(): Run a function on a specific CPU
309 * @cpu: The CPU to run on.
310 * @data: Pre-allocated and setup data structure
311 * @wait: If true, wait until function has completed on specified CPU.
313 * Like smp_call_function_single(), but allow caller to pass in a
314 * pre-allocated data structure. Useful for embedding @data inside
315 * other structures, for instance.
317 void __smp_call_function_single(int cpu, struct call_single_data *csd,
320 unsigned int this_cpu;
323 this_cpu = get_cpu();
325 * Can deadlock when called with interrupts disabled.
326 * We allow cpu's that are not yet online though, as no one else can
327 * send smp call function interrupt to this cpu and as such deadlocks
330 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
331 && !oops_in_progress);
333 if (cpu == this_cpu) {
334 local_irq_save(flags);
335 csd->func(csd->info);
336 local_irq_restore(flags);
339 generic_exec_single(cpu, csd, wait);
345 * smp_call_function_many(): Run a function on a set of other CPUs.
346 * @mask: The set of cpus to run on (only runs on online subset).
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 (atomically) until function has completed
352 * If @wait is true, then returns once @func has returned.
354 * You must not call this function with disabled interrupts or from a
355 * hardware interrupt handler or from a bottom half handler. Preemption
356 * must be disabled when calling this function.
358 void smp_call_function_many(const struct cpumask *mask,
359 smp_call_func_t func, void *info, bool wait)
361 struct call_function_data *cfd;
362 int cpu, next_cpu, this_cpu = smp_processor_id();
365 * Can deadlock when called with interrupts disabled.
366 * We allow cpu's that are not yet online though, as no one else can
367 * send smp call function interrupt to this cpu and as such deadlocks
370 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
371 && !oops_in_progress && !early_boot_irqs_disabled);
373 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
374 cpu = cpumask_first_and(mask, cpu_online_mask);
376 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
378 /* No online cpus? We're done. */
379 if (cpu >= nr_cpu_ids)
382 /* Do we have another CPU which isn't us? */
383 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
384 if (next_cpu == this_cpu)
385 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
387 /* Fastpath: do that cpu by itself. */
388 if (next_cpu >= nr_cpu_ids) {
389 smp_call_function_single(cpu, func, info, wait);
393 cfd = &__get_cpu_var(cfd_data);
395 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
396 cpumask_clear_cpu(this_cpu, cfd->cpumask);
398 /* Some callers race with other cpus changing the passed mask */
399 if (unlikely(!cpumask_weight(cfd->cpumask)))
403 * After we put an entry into the list, cfd->cpumask may be cleared
404 * again when another CPU sends another IPI for a SMP function call, so
405 * cfd->cpumask will be zero.
407 cpumask_copy(cfd->cpumask_ipi, cfd->cpumask);
409 for_each_cpu(cpu, cfd->cpumask) {
410 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
411 struct call_single_queue *dst =
412 &per_cpu(call_single_queue, cpu);
419 raw_spin_lock_irqsave(&dst->lock, flags);
420 list_add_tail(&csd->list, &dst->list);
421 raw_spin_unlock_irqrestore(&dst->lock, flags);
424 /* Send a message to all CPUs in the map */
425 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
428 for_each_cpu(cpu, cfd->cpumask) {
429 struct call_single_data *csd;
431 csd = per_cpu_ptr(cfd->csd, cpu);
436 EXPORT_SYMBOL(smp_call_function_many);
439 * smp_call_function(): Run a function on all other CPUs.
440 * @func: The function to run. This must be fast and non-blocking.
441 * @info: An arbitrary pointer to pass to the function.
442 * @wait: If true, wait (atomically) until function has completed
447 * If @wait is true, then returns once @func has returned; otherwise
448 * it returns just before the target cpu calls @func.
450 * You must not call this function with disabled interrupts or from a
451 * hardware interrupt handler or from a bottom half handler.
453 int smp_call_function(smp_call_func_t func, void *info, int wait)
456 smp_call_function_many(cpu_online_mask, func, info, wait);
461 EXPORT_SYMBOL(smp_call_function);
462 #endif /* USE_GENERIC_SMP_HELPERS */
464 /* Setup configured maximum number of CPUs to activate */
465 unsigned int setup_max_cpus = NR_CPUS;
466 EXPORT_SYMBOL(setup_max_cpus);
470 * Setup routine for controlling SMP activation
472 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
473 * activation entirely (the MPS table probe still happens, though).
475 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
476 * greater than 0, limits the maximum number of CPUs activated in
480 void __weak arch_disable_smp_support(void) { }
482 static int __init nosmp(char *str)
485 arch_disable_smp_support();
490 early_param("nosmp", nosmp);
492 /* this is hard limit */
493 static int __init nrcpus(char *str)
497 get_option(&str, &nr_cpus);
498 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
499 nr_cpu_ids = nr_cpus;
504 early_param("nr_cpus", nrcpus);
506 static int __init maxcpus(char *str)
508 get_option(&str, &setup_max_cpus);
509 if (setup_max_cpus == 0)
510 arch_disable_smp_support();
515 early_param("maxcpus", maxcpus);
517 /* Setup number of possible processor ids */
518 int nr_cpu_ids __read_mostly = NR_CPUS;
519 EXPORT_SYMBOL(nr_cpu_ids);
521 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
522 void __init setup_nr_cpu_ids(void)
524 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
527 /* Called by boot processor to activate the rest. */
528 void __init smp_init(void)
534 /* FIXME: This should be done in userspace --RR */
535 for_each_present_cpu(cpu) {
536 if (num_online_cpus() >= setup_max_cpus)
538 if (!cpu_online(cpu))
542 /* Any cleanup work */
543 printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
544 smp_cpus_done(setup_max_cpus);
548 * Call a function on all processors. May be used during early boot while
549 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
550 * of local_irq_disable/enable().
552 int on_each_cpu(void (*func) (void *info), void *info, int wait)
558 ret = smp_call_function(func, info, wait);
559 local_irq_save(flags);
561 local_irq_restore(flags);
565 EXPORT_SYMBOL(on_each_cpu);
568 * on_each_cpu_mask(): Run a function on processors specified by
569 * cpumask, which may include the local processor.
570 * @mask: The set of cpus to run on (only runs on online subset).
571 * @func: The function to run. This must be fast and non-blocking.
572 * @info: An arbitrary pointer to pass to the function.
573 * @wait: If true, wait (atomically) until function has completed
576 * If @wait is true, then returns once @func has returned.
578 * You must not call this function with disabled interrupts or from a
579 * hardware interrupt handler or from a bottom half handler. The
580 * exception is that it may be used during early boot while
581 * early_boot_irqs_disabled is set.
583 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
584 void *info, bool wait)
588 smp_call_function_many(mask, func, info, wait);
589 if (cpumask_test_cpu(cpu, mask)) {
591 local_irq_save(flags);
593 local_irq_restore(flags);
597 EXPORT_SYMBOL(on_each_cpu_mask);
600 * on_each_cpu_cond(): Call a function on each processor for which
601 * the supplied function cond_func returns true, optionally waiting
602 * for all the required CPUs to finish. This may include the local
604 * @cond_func: A callback function that is passed a cpu id and
605 * the the info parameter. The function is called
606 * with preemption disabled. The function should
607 * return a blooean value indicating whether to IPI
609 * @func: The function to run on all applicable CPUs.
610 * This must be fast and non-blocking.
611 * @info: An arbitrary pointer to pass to both functions.
612 * @wait: If true, wait (atomically) until function has
613 * completed on other CPUs.
614 * @gfp_flags: GFP flags to use when allocating the cpumask
615 * used internally by the function.
617 * The function might sleep if the GFP flags indicates a non
618 * atomic allocation is allowed.
620 * Preemption is disabled to protect against CPUs going offline but not online.
621 * CPUs going online during the call will not be seen or sent an IPI.
623 * You must not call this function with disabled interrupts or
624 * from a hardware interrupt handler or from a bottom half handler.
626 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
627 smp_call_func_t func, void *info, bool wait,
633 might_sleep_if(gfp_flags & __GFP_WAIT);
635 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
637 for_each_online_cpu(cpu)
638 if (cond_func(cpu, info))
639 cpumask_set_cpu(cpu, cpus);
640 on_each_cpu_mask(cpus, func, info, wait);
642 free_cpumask_var(cpus);
645 * No free cpumask, bother. No matter, we'll
646 * just have to IPI them one by one.
649 for_each_online_cpu(cpu)
650 if (cond_func(cpu, info)) {
651 ret = smp_call_function_single(cpu, func,
658 EXPORT_SYMBOL(on_each_cpu_cond);
660 static void do_nothing(void *unused)
665 * kick_all_cpus_sync - Force all cpus out of idle
667 * Used to synchronize the update of pm_idle function pointer. It's
668 * called after the pointer is updated and returns after the dummy
669 * callback function has been executed on all cpus. The execution of
670 * the function can only happen on the remote cpus after they have
671 * left the idle function which had been called via pm_idle function
672 * pointer. So it's guaranteed that nothing uses the previous pointer
675 void kick_all_cpus_sync(void)
677 /* Make sure the change is visible before we kick the cpus */
679 smp_call_function(do_nothing, NULL, 1);
681 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);