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
24 struct call_function_data {
25 struct call_single_data __percpu *csd;
26 cpumask_var_t cpumask;
27 cpumask_var_t cpumask_ipi;
30 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
32 struct call_single_queue {
33 struct list_head list;
37 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
40 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
42 long cpu = (long)hcpu;
43 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
47 case CPU_UP_PREPARE_FROZEN:
48 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
50 return notifier_from_errno(-ENOMEM);
51 if (!zalloc_cpumask_var_node(&cfd->cpumask_ipi, GFP_KERNEL,
53 free_cpumask_var(cfd->cpumask);
54 return notifier_from_errno(-ENOMEM);
56 cfd->csd = alloc_percpu(struct call_single_data);
58 free_cpumask_var(cfd->cpumask_ipi);
59 free_cpumask_var(cfd->cpumask);
60 return notifier_from_errno(-ENOMEM);
64 #ifdef CONFIG_HOTPLUG_CPU
66 case CPU_UP_CANCELED_FROZEN:
70 free_cpumask_var(cfd->cpumask);
71 free_cpumask_var(cfd->cpumask_ipi);
72 free_percpu(cfd->csd);
80 static struct notifier_block hotplug_cfd_notifier = {
81 .notifier_call = hotplug_cfd,
84 void __init call_function_init(void)
86 void *cpu = (void *)(long)smp_processor_id();
89 for_each_possible_cpu(i) {
90 struct call_single_queue *q = &per_cpu(call_single_queue, i);
92 raw_spin_lock_init(&q->lock);
93 INIT_LIST_HEAD(&q->list);
96 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
97 register_cpu_notifier(&hotplug_cfd_notifier);
101 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
103 * For non-synchronous ipi calls the csd can still be in use by the
104 * previous function call. For multi-cpu calls its even more interesting
105 * as we'll have to ensure no other cpu is observing our csd.
107 static void csd_lock_wait(struct call_single_data *csd)
109 while (csd->flags & CSD_FLAG_LOCK)
113 static void csd_lock(struct call_single_data *csd)
116 csd->flags |= CSD_FLAG_LOCK;
119 * prevent CPU from reordering the above assignment
120 * to ->flags with any subsequent assignments to other
121 * fields of the specified call_single_data structure:
126 static void csd_unlock(struct call_single_data *csd)
128 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
131 * ensure we're all done before releasing data:
135 csd->flags &= ~CSD_FLAG_LOCK;
139 * Insert a previously allocated call_single_data element
140 * for execution on the given CPU. data must already have
141 * ->func, ->info, and ->flags set.
144 void generic_exec_single(int cpu, struct call_single_data *csd, int wait)
146 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
151 csd->flags |= CSD_FLAG_WAIT;
153 raw_spin_lock_irqsave(&dst->lock, flags);
154 ipi = list_empty(&dst->list);
155 list_add_tail(&csd->list, &dst->list);
156 raw_spin_unlock_irqrestore(&dst->lock, flags);
159 * The list addition should be visible before sending the IPI
160 * handler locks the list to pull the entry off it because of
161 * normal cache coherency rules implied by spinlocks.
163 * If IPIs can go out of order to the cache coherency protocol
164 * in an architecture, sufficient synchronisation should be added
165 * to arch code to make it appear to obey cache coherency WRT
166 * locking and barrier primitives. Generic code isn't really
167 * equipped to do the right thing...
170 arch_send_call_function_single_ipi(cpu);
177 * Invoked by arch to handle an IPI for call function single. Must be
178 * called from the arch with interrupts disabled.
180 void generic_smp_call_function_single_interrupt(void)
182 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
186 * Shouldn't receive this interrupt on a cpu that is not yet online.
188 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
190 raw_spin_lock(&q->lock);
191 list_replace_init(&q->list, &list);
192 raw_spin_unlock(&q->lock);
194 while (!list_empty(&list)) {
195 struct call_single_data *csd;
197 csd = list_entry(list.next, struct call_single_data, list);
198 list_del(&csd->list);
200 csd->func(csd->info);
206 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
209 * smp_call_function_single - Run a function on a specific CPU
210 * @func: The function to run. This must be fast and non-blocking.
211 * @info: An arbitrary pointer to pass to the function.
212 * @wait: If true, wait until function has completed on other CPUs.
214 * Returns 0 on success, else a negative status code.
216 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
219 struct call_single_data d = {
227 * prevent preemption and reschedule on another processor,
228 * as well as CPU removal
230 this_cpu = get_cpu();
233 * Can deadlock when called with interrupts disabled.
234 * We allow cpu's that are not yet online though, as no one else can
235 * send smp call function interrupt to this cpu and as such deadlocks
238 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
239 && !oops_in_progress);
241 if (cpu == this_cpu) {
242 local_irq_save(flags);
244 local_irq_restore(flags);
246 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
247 struct call_single_data *csd = &d;
250 csd = &__get_cpu_var(csd_data);
256 generic_exec_single(cpu, csd, wait);
258 err = -ENXIO; /* CPU not online */
266 EXPORT_SYMBOL(smp_call_function_single);
269 * smp_call_function_any - Run a function on any of the given cpus
270 * @mask: The mask of cpus it can run on.
271 * @func: The function to run. This must be fast and non-blocking.
272 * @info: An arbitrary pointer to pass to the function.
273 * @wait: If true, wait until function has completed.
275 * Returns 0 on success, else a negative status code (if no cpus were online).
277 * Selection preference:
278 * 1) current cpu if in @mask
279 * 2) any cpu of current node if in @mask
280 * 3) any other online cpu in @mask
282 int smp_call_function_any(const struct cpumask *mask,
283 smp_call_func_t func, void *info, int wait)
286 const struct cpumask *nodemask;
289 /* Try for same CPU (cheapest) */
291 if (cpumask_test_cpu(cpu, mask))
294 /* Try for same node. */
295 nodemask = cpumask_of_node(cpu_to_node(cpu));
296 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
297 cpu = cpumask_next_and(cpu, nodemask, mask)) {
302 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
303 cpu = cpumask_any_and(mask, cpu_online_mask);
305 ret = smp_call_function_single(cpu, func, info, wait);
309 EXPORT_SYMBOL_GPL(smp_call_function_any);
312 * __smp_call_function_single(): Run a function on a specific CPU
313 * @cpu: The CPU to run on.
314 * @data: Pre-allocated and setup data structure
315 * @wait: If true, wait until function has completed on specified CPU.
317 * Like smp_call_function_single(), but allow caller to pass in a
318 * pre-allocated data structure. Useful for embedding @data inside
319 * other structures, for instance.
321 void __smp_call_function_single(int cpu, struct call_single_data *csd,
324 unsigned int this_cpu;
327 this_cpu = get_cpu();
329 * Can deadlock when called with interrupts disabled.
330 * We allow cpu's that are not yet online though, as no one else can
331 * send smp call function interrupt to this cpu and as such deadlocks
334 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
335 && !oops_in_progress);
337 if (cpu == this_cpu) {
338 local_irq_save(flags);
339 csd->func(csd->info);
340 local_irq_restore(flags);
343 generic_exec_single(cpu, csd, wait);
347 EXPORT_SYMBOL_GPL(__smp_call_function_single);
350 * smp_call_function_many(): Run a function on a set of other CPUs.
351 * @mask: The set of cpus to run on (only runs on online subset).
352 * @func: The function to run. This must be fast and non-blocking.
353 * @info: An arbitrary pointer to pass to the function.
354 * @wait: If true, wait (atomically) until function has completed
357 * If @wait is true, then returns once @func has returned.
359 * You must not call this function with disabled interrupts or from a
360 * hardware interrupt handler or from a bottom half handler. Preemption
361 * must be disabled when calling this function.
363 void smp_call_function_many(const struct cpumask *mask,
364 smp_call_func_t func, void *info, bool wait)
366 struct call_function_data *cfd;
367 int cpu, next_cpu, this_cpu = smp_processor_id();
370 * Can deadlock when called with interrupts disabled.
371 * We allow cpu's that are not yet online though, as no one else can
372 * send smp call function interrupt to this cpu and as such deadlocks
375 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
376 && !oops_in_progress && !early_boot_irqs_disabled);
378 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
379 cpu = cpumask_first_and(mask, cpu_online_mask);
381 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
383 /* No online cpus? We're done. */
384 if (cpu >= nr_cpu_ids)
387 /* Do we have another CPU which isn't us? */
388 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
389 if (next_cpu == this_cpu)
390 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
392 /* Fastpath: do that cpu by itself. */
393 if (next_cpu >= nr_cpu_ids) {
394 smp_call_function_single(cpu, func, info, wait);
398 cfd = &__get_cpu_var(cfd_data);
400 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
401 cpumask_clear_cpu(this_cpu, cfd->cpumask);
403 /* Some callers race with other cpus changing the passed mask */
404 if (unlikely(!cpumask_weight(cfd->cpumask)))
408 * After we put an entry into the list, cfd->cpumask may be cleared
409 * again when another CPU sends another IPI for a SMP function call, so
410 * cfd->cpumask will be zero.
412 cpumask_copy(cfd->cpumask_ipi, cfd->cpumask);
414 for_each_cpu(cpu, cfd->cpumask) {
415 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
416 struct call_single_queue *dst =
417 &per_cpu(call_single_queue, cpu);
424 raw_spin_lock_irqsave(&dst->lock, flags);
425 list_add_tail(&csd->list, &dst->list);
426 raw_spin_unlock_irqrestore(&dst->lock, flags);
429 /* Send a message to all CPUs in the map */
430 arch_send_call_function_ipi_mask(cfd->cpumask_ipi);
433 for_each_cpu(cpu, cfd->cpumask) {
434 struct call_single_data *csd;
436 csd = per_cpu_ptr(cfd->csd, cpu);
441 EXPORT_SYMBOL(smp_call_function_many);
444 * smp_call_function(): Run a function on all other CPUs.
445 * @func: The function to run. This must be fast and non-blocking.
446 * @info: An arbitrary pointer to pass to the function.
447 * @wait: If true, wait (atomically) until function has completed
452 * If @wait is true, then returns once @func has returned; otherwise
453 * it returns just before the target cpu calls @func.
455 * You must not call this function with disabled interrupts or from a
456 * hardware interrupt handler or from a bottom half handler.
458 int smp_call_function(smp_call_func_t func, void *info, int wait)
461 smp_call_function_many(cpu_online_mask, func, info, wait);
466 EXPORT_SYMBOL(smp_call_function);
467 #endif /* USE_GENERIC_SMP_HELPERS */
469 /* Setup configured maximum number of CPUs to activate */
470 unsigned int setup_max_cpus = NR_CPUS;
471 EXPORT_SYMBOL(setup_max_cpus);
475 * Setup routine for controlling SMP activation
477 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
478 * activation entirely (the MPS table probe still happens, though).
480 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
481 * greater than 0, limits the maximum number of CPUs activated in
485 void __weak arch_disable_smp_support(void) { }
487 static int __init nosmp(char *str)
490 arch_disable_smp_support();
495 early_param("nosmp", nosmp);
497 /* this is hard limit */
498 static int __init nrcpus(char *str)
502 get_option(&str, &nr_cpus);
503 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
504 nr_cpu_ids = nr_cpus;
509 early_param("nr_cpus", nrcpus);
511 static int __init maxcpus(char *str)
513 get_option(&str, &setup_max_cpus);
514 if (setup_max_cpus == 0)
515 arch_disable_smp_support();
520 early_param("maxcpus", maxcpus);
522 /* Setup number of possible processor ids */
523 int nr_cpu_ids __read_mostly = NR_CPUS;
524 EXPORT_SYMBOL(nr_cpu_ids);
526 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
527 void __init setup_nr_cpu_ids(void)
529 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
532 void __weak smp_announce(void)
534 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
537 /* Called by boot processor to activate the rest. */
538 void __init smp_init(void)
544 /* FIXME: This should be done in userspace --RR */
545 for_each_present_cpu(cpu) {
546 if (num_online_cpus() >= setup_max_cpus)
548 if (!cpu_online(cpu))
552 /* Any cleanup work */
554 smp_cpus_done(setup_max_cpus);
558 * Call a function on all processors. May be used during early boot while
559 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
560 * of local_irq_disable/enable().
562 int on_each_cpu(void (*func) (void *info), void *info, int wait)
568 ret = smp_call_function(func, info, wait);
569 local_irq_save(flags);
571 local_irq_restore(flags);
575 EXPORT_SYMBOL(on_each_cpu);
578 * on_each_cpu_mask(): Run a function on processors specified by
579 * cpumask, which may include the local processor.
580 * @mask: The set of cpus to run on (only runs on online subset).
581 * @func: The function to run. This must be fast and non-blocking.
582 * @info: An arbitrary pointer to pass to the function.
583 * @wait: If true, wait (atomically) until function has completed
586 * If @wait is true, then returns once @func has returned.
588 * You must not call this function with disabled interrupts or from a
589 * hardware interrupt handler or from a bottom half handler. The
590 * exception is that it may be used during early boot while
591 * early_boot_irqs_disabled is set.
593 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
594 void *info, bool wait)
598 smp_call_function_many(mask, func, info, wait);
599 if (cpumask_test_cpu(cpu, mask)) {
601 local_irq_save(flags);
603 local_irq_restore(flags);
607 EXPORT_SYMBOL(on_each_cpu_mask);
610 * on_each_cpu_cond(): Call a function on each processor for which
611 * the supplied function cond_func returns true, optionally waiting
612 * for all the required CPUs to finish. This may include the local
614 * @cond_func: A callback function that is passed a cpu id and
615 * the the info parameter. The function is called
616 * with preemption disabled. The function should
617 * return a blooean value indicating whether to IPI
619 * @func: The function to run on all applicable CPUs.
620 * This must be fast and non-blocking.
621 * @info: An arbitrary pointer to pass to both functions.
622 * @wait: If true, wait (atomically) until function has
623 * completed on other CPUs.
624 * @gfp_flags: GFP flags to use when allocating the cpumask
625 * used internally by the function.
627 * The function might sleep if the GFP flags indicates a non
628 * atomic allocation is allowed.
630 * Preemption is disabled to protect against CPUs going offline but not online.
631 * CPUs going online during the call will not be seen or sent an IPI.
633 * You must not call this function with disabled interrupts or
634 * from a hardware interrupt handler or from a bottom half handler.
636 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
637 smp_call_func_t func, void *info, bool wait,
643 might_sleep_if(gfp_flags & __GFP_WAIT);
645 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
647 for_each_online_cpu(cpu)
648 if (cond_func(cpu, info))
649 cpumask_set_cpu(cpu, cpus);
650 on_each_cpu_mask(cpus, func, info, wait);
652 free_cpumask_var(cpus);
655 * No free cpumask, bother. No matter, we'll
656 * just have to IPI them one by one.
659 for_each_online_cpu(cpu)
660 if (cond_func(cpu, info)) {
661 ret = smp_call_function_single(cpu, func,
668 EXPORT_SYMBOL(on_each_cpu_cond);
670 static void do_nothing(void *unused)
675 * kick_all_cpus_sync - Force all cpus out of idle
677 * Used to synchronize the update of pm_idle function pointer. It's
678 * called after the pointer is updated and returns after the dummy
679 * callback function has been executed on all cpus. The execution of
680 * the function can only happen on the remote cpus after they have
681 * left the idle function which had been called via pm_idle function
682 * pointer. So it's guaranteed that nothing uses the previous pointer
685 void kick_all_cpus_sync(void)
687 /* Make sure the change is visible before we kick the cpus */
689 smp_call_function(do_nothing, NULL, 1);
691 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);