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;
28 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
30 static DEFINE_PER_CPU_SHARED_ALIGNED(struct llist_head, call_single_queue);
33 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
35 long cpu = (long)hcpu;
36 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
40 case CPU_UP_PREPARE_FROZEN:
41 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
43 return notifier_from_errno(-ENOMEM);
44 cfd->csd = alloc_percpu(struct call_single_data);
46 free_cpumask_var(cfd->cpumask);
47 return notifier_from_errno(-ENOMEM);
51 #ifdef CONFIG_HOTPLUG_CPU
53 case CPU_UP_CANCELED_FROZEN:
57 free_cpumask_var(cfd->cpumask);
58 free_percpu(cfd->csd);
66 static struct notifier_block hotplug_cfd_notifier = {
67 .notifier_call = hotplug_cfd,
70 void __init call_function_init(void)
72 void *cpu = (void *)(long)smp_processor_id();
75 for_each_possible_cpu(i)
76 init_llist_head(&per_cpu(call_single_queue, i));
78 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
79 register_cpu_notifier(&hotplug_cfd_notifier);
83 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
85 * For non-synchronous ipi calls the csd can still be in use by the
86 * previous function call. For multi-cpu calls its even more interesting
87 * as we'll have to ensure no other cpu is observing our csd.
89 static void csd_lock_wait(struct call_single_data *csd)
91 while (csd->flags & CSD_FLAG_LOCK)
95 static void csd_lock(struct call_single_data *csd)
98 csd->flags |= CSD_FLAG_LOCK;
101 * prevent CPU from reordering the above assignment
102 * to ->flags with any subsequent assignments to other
103 * fields of the specified call_single_data structure:
108 static void csd_unlock(struct call_single_data *csd)
110 WARN_ON((csd->flags & CSD_FLAG_WAIT) && !(csd->flags & CSD_FLAG_LOCK));
113 * ensure we're all done before releasing data:
117 csd->flags &= ~CSD_FLAG_LOCK;
121 * Insert a previously allocated call_single_data element
122 * for execution on the given CPU. data must already have
123 * ->func, ->info, and ->flags set.
125 static void generic_exec_single(int cpu, struct call_single_data *csd, int wait)
128 csd->flags |= CSD_FLAG_WAIT;
131 * The list addition should be visible before sending the IPI
132 * handler locks the list to pull the entry off it because of
133 * normal cache coherency rules implied by spinlocks.
135 * If IPIs can go out of order to the cache coherency protocol
136 * in an architecture, sufficient synchronisation should be added
137 * to arch code to make it appear to obey cache coherency WRT
138 * locking and barrier primitives. Generic code isn't really
139 * equipped to do the right thing...
141 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
142 arch_send_call_function_single_ipi(cpu);
149 * Invoked by arch to handle an IPI for call function single. Must be
150 * called from the arch with interrupts disabled.
152 void generic_smp_call_function_single_interrupt(void)
154 struct llist_node *entry, *next;
157 * Shouldn't receive this interrupt on a cpu that is not yet online.
159 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
161 entry = llist_del_all(&__get_cpu_var(call_single_queue));
162 entry = llist_reverse_order(entry);
165 struct call_single_data *csd;
169 csd = llist_entry(entry, struct call_single_data, llist);
170 csd->func(csd->info);
177 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
180 * smp_call_function_single - Run a function on a specific CPU
181 * @func: The function to run. This must be fast and non-blocking.
182 * @info: An arbitrary pointer to pass to the function.
183 * @wait: If true, wait until function has completed on other CPUs.
185 * Returns 0 on success, else a negative status code.
187 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
190 struct call_single_data d = {
198 * prevent preemption and reschedule on another processor,
199 * as well as CPU removal
201 this_cpu = get_cpu();
204 * Can deadlock when called with interrupts disabled.
205 * We allow cpu's that are not yet online though, as no one else can
206 * send smp call function interrupt to this cpu and as such deadlocks
209 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
210 && !oops_in_progress);
212 if (cpu == this_cpu) {
213 local_irq_save(flags);
215 local_irq_restore(flags);
217 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
218 struct call_single_data *csd = &d;
221 csd = &__get_cpu_var(csd_data);
227 generic_exec_single(cpu, csd, wait);
229 err = -ENXIO; /* CPU not online */
237 EXPORT_SYMBOL(smp_call_function_single);
240 * smp_call_function_any - Run a function on any of the given cpus
241 * @mask: The mask of cpus it can run on.
242 * @func: The function to run. This must be fast and non-blocking.
243 * @info: An arbitrary pointer to pass to the function.
244 * @wait: If true, wait until function has completed.
246 * Returns 0 on success, else a negative status code (if no cpus were online).
248 * Selection preference:
249 * 1) current cpu if in @mask
250 * 2) any cpu of current node if in @mask
251 * 3) any other online cpu in @mask
253 int smp_call_function_any(const struct cpumask *mask,
254 smp_call_func_t func, void *info, int wait)
257 const struct cpumask *nodemask;
260 /* Try for same CPU (cheapest) */
262 if (cpumask_test_cpu(cpu, mask))
265 /* Try for same node. */
266 nodemask = cpumask_of_node(cpu_to_node(cpu));
267 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
268 cpu = cpumask_next_and(cpu, nodemask, mask)) {
273 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
274 cpu = cpumask_any_and(mask, cpu_online_mask);
276 ret = smp_call_function_single(cpu, func, info, wait);
280 EXPORT_SYMBOL_GPL(smp_call_function_any);
283 * __smp_call_function_single(): Run a function on a specific CPU
284 * @cpu: The CPU to run on.
285 * @data: Pre-allocated and setup data structure
286 * @wait: If true, wait until function has completed on specified CPU.
288 * Like smp_call_function_single(), but allow caller to pass in a
289 * pre-allocated data structure. Useful for embedding @data inside
290 * other structures, for instance.
292 void __smp_call_function_single(int cpu, struct call_single_data *csd,
295 unsigned int this_cpu;
298 this_cpu = get_cpu();
300 * Can deadlock when called with interrupts disabled.
301 * We allow cpu's that are not yet online though, as no one else can
302 * send smp call function interrupt to this cpu and as such deadlocks
305 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
306 && !oops_in_progress);
308 if (cpu == this_cpu) {
309 local_irq_save(flags);
310 csd->func(csd->info);
311 local_irq_restore(flags);
314 generic_exec_single(cpu, csd, wait);
318 EXPORT_SYMBOL_GPL(__smp_call_function_single);
321 * smp_call_function_many(): Run a function on a set of other CPUs.
322 * @mask: The set of cpus to run on (only runs on online subset).
323 * @func: The function to run. This must be fast and non-blocking.
324 * @info: An arbitrary pointer to pass to the function.
325 * @wait: If true, wait (atomically) until function has completed
328 * If @wait is true, then returns once @func has returned.
330 * You must not call this function with disabled interrupts or from a
331 * hardware interrupt handler or from a bottom half handler. Preemption
332 * must be disabled when calling this function.
334 void smp_call_function_many(const struct cpumask *mask,
335 smp_call_func_t func, void *info, bool wait)
337 struct call_function_data *cfd;
338 int cpu, next_cpu, this_cpu = smp_processor_id();
341 * Can deadlock when called with interrupts disabled.
342 * We allow cpu's that are not yet online though, as no one else can
343 * send smp call function interrupt to this cpu and as such deadlocks
346 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
347 && !oops_in_progress && !early_boot_irqs_disabled);
349 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
350 cpu = cpumask_first_and(mask, cpu_online_mask);
352 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
354 /* No online cpus? We're done. */
355 if (cpu >= nr_cpu_ids)
358 /* Do we have another CPU which isn't us? */
359 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
360 if (next_cpu == this_cpu)
361 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
363 /* Fastpath: do that cpu by itself. */
364 if (next_cpu >= nr_cpu_ids) {
365 smp_call_function_single(cpu, func, info, wait);
369 cfd = &__get_cpu_var(cfd_data);
371 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
372 cpumask_clear_cpu(this_cpu, cfd->cpumask);
374 /* Some callers race with other cpus changing the passed mask */
375 if (unlikely(!cpumask_weight(cfd->cpumask)))
378 for_each_cpu(cpu, cfd->cpumask) {
379 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
384 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
387 /* Send a message to all CPUs in the map */
388 arch_send_call_function_ipi_mask(cfd->cpumask);
391 for_each_cpu(cpu, cfd->cpumask) {
392 struct call_single_data *csd;
394 csd = per_cpu_ptr(cfd->csd, cpu);
399 EXPORT_SYMBOL(smp_call_function_many);
402 * smp_call_function(): Run a function on all other CPUs.
403 * @func: The function to run. This must be fast and non-blocking.
404 * @info: An arbitrary pointer to pass to the function.
405 * @wait: If true, wait (atomically) until function has completed
410 * If @wait is true, then returns once @func has returned; otherwise
411 * it returns just before the target cpu calls @func.
413 * You must not call this function with disabled interrupts or from a
414 * hardware interrupt handler or from a bottom half handler.
416 int smp_call_function(smp_call_func_t func, void *info, int wait)
419 smp_call_function_many(cpu_online_mask, func, info, wait);
424 EXPORT_SYMBOL(smp_call_function);
426 /* Setup configured maximum number of CPUs to activate */
427 unsigned int setup_max_cpus = NR_CPUS;
428 EXPORT_SYMBOL(setup_max_cpus);
432 * Setup routine for controlling SMP activation
434 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
435 * activation entirely (the MPS table probe still happens, though).
437 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
438 * greater than 0, limits the maximum number of CPUs activated in
442 void __weak arch_disable_smp_support(void) { }
444 static int __init nosmp(char *str)
447 arch_disable_smp_support();
452 early_param("nosmp", nosmp);
454 /* this is hard limit */
455 static int __init nrcpus(char *str)
459 get_option(&str, &nr_cpus);
460 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
461 nr_cpu_ids = nr_cpus;
466 early_param("nr_cpus", nrcpus);
468 static int __init maxcpus(char *str)
470 get_option(&str, &setup_max_cpus);
471 if (setup_max_cpus == 0)
472 arch_disable_smp_support();
477 early_param("maxcpus", maxcpus);
479 /* Setup number of possible processor ids */
480 int nr_cpu_ids __read_mostly = NR_CPUS;
481 EXPORT_SYMBOL(nr_cpu_ids);
483 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
484 void __init setup_nr_cpu_ids(void)
486 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
489 void __weak smp_announce(void)
491 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
494 /* Called by boot processor to activate the rest. */
495 void __init smp_init(void)
501 /* FIXME: This should be done in userspace --RR */
502 for_each_present_cpu(cpu) {
503 if (num_online_cpus() >= setup_max_cpus)
505 if (!cpu_online(cpu))
509 /* Any cleanup work */
511 smp_cpus_done(setup_max_cpus);
515 * Call a function on all processors. May be used during early boot while
516 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
517 * of local_irq_disable/enable().
519 int on_each_cpu(void (*func) (void *info), void *info, int wait)
525 ret = smp_call_function(func, info, wait);
526 local_irq_save(flags);
528 local_irq_restore(flags);
532 EXPORT_SYMBOL(on_each_cpu);
535 * on_each_cpu_mask(): Run a function on processors specified by
536 * cpumask, which may include the local processor.
537 * @mask: The set of cpus to run on (only runs on online subset).
538 * @func: The function to run. This must be fast and non-blocking.
539 * @info: An arbitrary pointer to pass to the function.
540 * @wait: If true, wait (atomically) until function has completed
543 * If @wait is true, then returns once @func has returned.
545 * You must not call this function with disabled interrupts or from a
546 * hardware interrupt handler or from a bottom half handler. The
547 * exception is that it may be used during early boot while
548 * early_boot_irqs_disabled is set.
550 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
551 void *info, bool wait)
555 smp_call_function_many(mask, func, info, wait);
556 if (cpumask_test_cpu(cpu, mask)) {
558 local_irq_save(flags);
560 local_irq_restore(flags);
564 EXPORT_SYMBOL(on_each_cpu_mask);
567 * on_each_cpu_cond(): Call a function on each processor for which
568 * the supplied function cond_func returns true, optionally waiting
569 * for all the required CPUs to finish. This may include the local
571 * @cond_func: A callback function that is passed a cpu id and
572 * the the info parameter. The function is called
573 * with preemption disabled. The function should
574 * return a blooean value indicating whether to IPI
576 * @func: The function to run on all applicable CPUs.
577 * This must be fast and non-blocking.
578 * @info: An arbitrary pointer to pass to both functions.
579 * @wait: If true, wait (atomically) until function has
580 * completed on other CPUs.
581 * @gfp_flags: GFP flags to use when allocating the cpumask
582 * used internally by the function.
584 * The function might sleep if the GFP flags indicates a non
585 * atomic allocation is allowed.
587 * Preemption is disabled to protect against CPUs going offline but not online.
588 * CPUs going online during the call will not be seen or sent an IPI.
590 * You must not call this function with disabled interrupts or
591 * from a hardware interrupt handler or from a bottom half handler.
593 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
594 smp_call_func_t func, void *info, bool wait,
600 might_sleep_if(gfp_flags & __GFP_WAIT);
602 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
604 for_each_online_cpu(cpu)
605 if (cond_func(cpu, info))
606 cpumask_set_cpu(cpu, cpus);
607 on_each_cpu_mask(cpus, func, info, wait);
609 free_cpumask_var(cpus);
612 * No free cpumask, bother. No matter, we'll
613 * just have to IPI them one by one.
616 for_each_online_cpu(cpu)
617 if (cond_func(cpu, info)) {
618 ret = smp_call_function_single(cpu, func,
625 EXPORT_SYMBOL(on_each_cpu_cond);
627 static void do_nothing(void *unused)
632 * kick_all_cpus_sync - Force all cpus out of idle
634 * Used to synchronize the update of pm_idle function pointer. It's
635 * called after the pointer is updated and returns after the dummy
636 * callback function has been executed on all cpus. The execution of
637 * the function can only happen on the remote cpus after they have
638 * left the idle function which had been called via pm_idle function
639 * pointer. So it's guaranteed that nothing uses the previous pointer
642 void kick_all_cpus_sync(void)
644 /* Make sure the change is visible before we kick the cpus */
646 smp_call_function(do_nothing, NULL, 1);
648 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);