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;
120 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
123 * Insert a previously allocated call_single_data element
124 * for execution on the given CPU. data must already have
125 * ->func, ->info, and ->flags set.
127 static int generic_exec_single(int cpu, struct call_single_data *csd,
128 smp_call_func_t func, void *info, int wait)
130 struct call_single_data csd_stack = { .flags = 0 };
134 if (cpu == smp_processor_id()) {
135 local_irq_save(flags);
137 local_irq_restore(flags);
142 if ((unsigned)cpu >= nr_cpu_ids || !cpu_online(cpu))
149 csd = &__get_cpu_var(csd_data);
158 csd->flags |= CSD_FLAG_WAIT;
161 * The list addition should be visible before sending the IPI
162 * handler locks the list to pull the entry off it because of
163 * normal cache coherency rules implied by spinlocks.
165 * If IPIs can go out of order to the cache coherency protocol
166 * in an architecture, sufficient synchronisation should be added
167 * to arch code to make it appear to obey cache coherency WRT
168 * locking and barrier primitives. Generic code isn't really
169 * equipped to do the right thing...
171 if (llist_add(&csd->llist, &per_cpu(call_single_queue, cpu)))
172 arch_send_call_function_single_ipi(cpu);
181 * Invoked by arch to handle an IPI for call function single. Must be
182 * called from the arch with interrupts disabled.
184 void generic_smp_call_function_single_interrupt(void)
186 struct llist_node *entry;
187 struct call_single_data *csd, *csd_next;
190 * Shouldn't receive this interrupt on a cpu that is not yet online.
192 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
194 entry = llist_del_all(&__get_cpu_var(call_single_queue));
195 entry = llist_reverse_order(entry);
197 llist_for_each_entry_safe(csd, csd_next, entry, llist) {
198 csd->func(csd->info);
204 * smp_call_function_single - Run a function on a specific CPU
205 * @func: The function to run. This must be fast and non-blocking.
206 * @info: An arbitrary pointer to pass to the function.
207 * @wait: If true, wait until function has completed on other CPUs.
209 * Returns 0 on success, else a negative status code.
211 int smp_call_function_single(int cpu, smp_call_func_t func, void *info,
218 * prevent preemption and reschedule on another processor,
219 * as well as CPU removal
221 this_cpu = get_cpu();
224 * Can deadlock when called with interrupts disabled.
225 * We allow cpu's that are not yet online though, as no one else can
226 * send smp call function interrupt to this cpu and as such deadlocks
229 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
230 && !oops_in_progress);
232 err = generic_exec_single(cpu, NULL, func, info, wait);
238 EXPORT_SYMBOL(smp_call_function_single);
241 * smp_call_function_single_async(): Run an asynchronous function on a
243 * @cpu: The CPU to run on.
244 * @csd: Pre-allocated and setup data structure
246 * Like smp_call_function_single(), but the call is asynchonous and
247 * can thus be done from contexts with disabled interrupts.
249 * The caller passes his own pre-allocated data structure
250 * (ie: embedded in an object) and is responsible for synchronizing it
251 * such that the IPIs performed on the @csd are strictly serialized.
253 * NOTE: Be careful, there is unfortunately no current debugging facility to
254 * validate the correctness of this serialization.
256 int smp_call_function_single_async(int cpu, struct call_single_data *csd)
261 err = generic_exec_single(cpu, csd, csd->func, csd->info, 0);
266 EXPORT_SYMBOL_GPL(smp_call_function_single_async);
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_many(): Run a function on a set of other CPUs.
313 * @mask: The set of cpus to run on (only runs on online subset).
314 * @func: The function to run. This must be fast and non-blocking.
315 * @info: An arbitrary pointer to pass to the function.
316 * @wait: If true, wait (atomically) until function has completed
319 * If @wait is true, then returns once @func has returned.
321 * You must not call this function with disabled interrupts or from a
322 * hardware interrupt handler or from a bottom half handler. Preemption
323 * must be disabled when calling this function.
325 void smp_call_function_many(const struct cpumask *mask,
326 smp_call_func_t func, void *info, bool wait)
328 struct call_function_data *cfd;
329 int cpu, next_cpu, this_cpu = smp_processor_id();
332 * Can deadlock when called with interrupts disabled.
333 * We allow cpu's that are not yet online though, as no one else can
334 * send smp call function interrupt to this cpu and as such deadlocks
337 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
338 && !oops_in_progress && !early_boot_irqs_disabled);
340 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
341 cpu = cpumask_first_and(mask, cpu_online_mask);
343 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
345 /* No online cpus? We're done. */
346 if (cpu >= nr_cpu_ids)
349 /* Do we have another CPU which isn't us? */
350 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
351 if (next_cpu == this_cpu)
352 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
354 /* Fastpath: do that cpu by itself. */
355 if (next_cpu >= nr_cpu_ids) {
356 smp_call_function_single(cpu, func, info, wait);
360 cfd = &__get_cpu_var(cfd_data);
362 cpumask_and(cfd->cpumask, mask, cpu_online_mask);
363 cpumask_clear_cpu(this_cpu, cfd->cpumask);
365 /* Some callers race with other cpus changing the passed mask */
366 if (unlikely(!cpumask_weight(cfd->cpumask)))
369 for_each_cpu(cpu, cfd->cpumask) {
370 struct call_single_data *csd = per_cpu_ptr(cfd->csd, cpu);
375 llist_add(&csd->llist, &per_cpu(call_single_queue, cpu));
378 /* Send a message to all CPUs in the map */
379 arch_send_call_function_ipi_mask(cfd->cpumask);
382 for_each_cpu(cpu, cfd->cpumask) {
383 struct call_single_data *csd;
385 csd = per_cpu_ptr(cfd->csd, cpu);
390 EXPORT_SYMBOL(smp_call_function_many);
393 * smp_call_function(): Run a function on all other CPUs.
394 * @func: The function to run. This must be fast and non-blocking.
395 * @info: An arbitrary pointer to pass to the function.
396 * @wait: If true, wait (atomically) until function has completed
401 * If @wait is true, then returns once @func has returned; otherwise
402 * it returns just before the target cpu calls @func.
404 * You must not call this function with disabled interrupts or from a
405 * hardware interrupt handler or from a bottom half handler.
407 int smp_call_function(smp_call_func_t func, void *info, int wait)
410 smp_call_function_many(cpu_online_mask, func, info, wait);
415 EXPORT_SYMBOL(smp_call_function);
417 /* Setup configured maximum number of CPUs to activate */
418 unsigned int setup_max_cpus = NR_CPUS;
419 EXPORT_SYMBOL(setup_max_cpus);
423 * Setup routine for controlling SMP activation
425 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
426 * activation entirely (the MPS table probe still happens, though).
428 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
429 * greater than 0, limits the maximum number of CPUs activated in
433 void __weak arch_disable_smp_support(void) { }
435 static int __init nosmp(char *str)
438 arch_disable_smp_support();
443 early_param("nosmp", nosmp);
445 /* this is hard limit */
446 static int __init nrcpus(char *str)
450 get_option(&str, &nr_cpus);
451 if (nr_cpus > 0 && nr_cpus < nr_cpu_ids)
452 nr_cpu_ids = nr_cpus;
457 early_param("nr_cpus", nrcpus);
459 static int __init maxcpus(char *str)
461 get_option(&str, &setup_max_cpus);
462 if (setup_max_cpus == 0)
463 arch_disable_smp_support();
468 early_param("maxcpus", maxcpus);
470 /* Setup number of possible processor ids */
471 int nr_cpu_ids __read_mostly = NR_CPUS;
472 EXPORT_SYMBOL(nr_cpu_ids);
474 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
475 void __init setup_nr_cpu_ids(void)
477 nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
480 void __weak smp_announce(void)
482 printk(KERN_INFO "Brought up %d CPUs\n", num_online_cpus());
485 /* Called by boot processor to activate the rest. */
486 void __init smp_init(void)
492 /* FIXME: This should be done in userspace --RR */
493 for_each_present_cpu(cpu) {
494 if (num_online_cpus() >= setup_max_cpus)
496 if (!cpu_online(cpu))
500 /* Any cleanup work */
502 smp_cpus_done(setup_max_cpus);
506 * Call a function on all processors. May be used during early boot while
507 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
508 * of local_irq_disable/enable().
510 int on_each_cpu(void (*func) (void *info), void *info, int wait)
516 ret = smp_call_function(func, info, wait);
517 local_irq_save(flags);
519 local_irq_restore(flags);
523 EXPORT_SYMBOL(on_each_cpu);
526 * on_each_cpu_mask(): Run a function on processors specified by
527 * cpumask, which may include the local processor.
528 * @mask: The set of cpus to run on (only runs on online subset).
529 * @func: The function to run. This must be fast and non-blocking.
530 * @info: An arbitrary pointer to pass to the function.
531 * @wait: If true, wait (atomically) until function has completed
534 * If @wait is true, then returns once @func has returned.
536 * You must not call this function with disabled interrupts or from a
537 * hardware interrupt handler or from a bottom half handler. The
538 * exception is that it may be used during early boot while
539 * early_boot_irqs_disabled is set.
541 void on_each_cpu_mask(const struct cpumask *mask, smp_call_func_t func,
542 void *info, bool wait)
546 smp_call_function_many(mask, func, info, wait);
547 if (cpumask_test_cpu(cpu, mask)) {
549 local_irq_save(flags);
551 local_irq_restore(flags);
555 EXPORT_SYMBOL(on_each_cpu_mask);
558 * on_each_cpu_cond(): Call a function on each processor for which
559 * the supplied function cond_func returns true, optionally waiting
560 * for all the required CPUs to finish. This may include the local
562 * @cond_func: A callback function that is passed a cpu id and
563 * the the info parameter. The function is called
564 * with preemption disabled. The function should
565 * return a blooean value indicating whether to IPI
567 * @func: The function to run on all applicable CPUs.
568 * This must be fast and non-blocking.
569 * @info: An arbitrary pointer to pass to both functions.
570 * @wait: If true, wait (atomically) until function has
571 * completed on other CPUs.
572 * @gfp_flags: GFP flags to use when allocating the cpumask
573 * used internally by the function.
575 * The function might sleep if the GFP flags indicates a non
576 * atomic allocation is allowed.
578 * Preemption is disabled to protect against CPUs going offline but not online.
579 * CPUs going online during the call will not be seen or sent an IPI.
581 * You must not call this function with disabled interrupts or
582 * from a hardware interrupt handler or from a bottom half handler.
584 void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
585 smp_call_func_t func, void *info, bool wait,
591 might_sleep_if(gfp_flags & __GFP_WAIT);
593 if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
595 for_each_online_cpu(cpu)
596 if (cond_func(cpu, info))
597 cpumask_set_cpu(cpu, cpus);
598 on_each_cpu_mask(cpus, func, info, wait);
600 free_cpumask_var(cpus);
603 * No free cpumask, bother. No matter, we'll
604 * just have to IPI them one by one.
607 for_each_online_cpu(cpu)
608 if (cond_func(cpu, info)) {
609 ret = smp_call_function_single(cpu, func,
616 EXPORT_SYMBOL(on_each_cpu_cond);
618 static void do_nothing(void *unused)
623 * kick_all_cpus_sync - Force all cpus out of idle
625 * Used to synchronize the update of pm_idle function pointer. It's
626 * called after the pointer is updated and returns after the dummy
627 * callback function has been executed on all cpus. The execution of
628 * the function can only happen on the remote cpus after they have
629 * left the idle function which had been called via pm_idle function
630 * pointer. So it's guaranteed that nothing uses the previous pointer
633 void kick_all_cpus_sync(void)
635 /* Make sure the change is visible before we kick the cpus */
637 smp_call_function(do_nothing, NULL, 1);
639 EXPORT_SYMBOL_GPL(kick_all_cpus_sync);