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/module.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>
17 struct list_head queue;
19 } call_function __cacheline_aligned_in_smp =
21 .queue = LIST_HEAD_INIT(call_function.queue),
22 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
29 struct call_function_data {
30 struct call_single_data csd;
32 cpumask_var_t cpumask;
35 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_function_data, cfd_data);
37 struct call_single_queue {
38 struct list_head list;
42 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_queue, call_single_queue);
45 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
47 long cpu = (long)hcpu;
48 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
52 case CPU_UP_PREPARE_FROZEN:
53 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
58 #ifdef CONFIG_HOTPLUG_CPU
60 case CPU_UP_CANCELED_FROZEN:
64 free_cpumask_var(cfd->cpumask);
72 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
73 .notifier_call = hotplug_cfd,
76 static int __cpuinit init_call_single_data(void)
78 void *cpu = (void *)(long)smp_processor_id();
81 for_each_possible_cpu(i) {
82 struct call_single_queue *q = &per_cpu(call_single_queue, i);
84 raw_spin_lock_init(&q->lock);
85 INIT_LIST_HEAD(&q->list);
88 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
89 register_cpu_notifier(&hotplug_cfd_notifier);
93 early_initcall(init_call_single_data);
96 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
98 * For non-synchronous ipi calls the csd can still be in use by the
99 * previous function call. For multi-cpu calls its even more interesting
100 * as we'll have to ensure no other cpu is observing our csd.
102 static void csd_lock_wait(struct call_single_data *data)
104 while (data->flags & CSD_FLAG_LOCK)
108 static void csd_lock(struct call_single_data *data)
111 data->flags = CSD_FLAG_LOCK;
114 * prevent CPU from reordering the above assignment
115 * to ->flags with any subsequent assignments to other
116 * fields of the specified call_single_data structure:
121 static void csd_unlock(struct call_single_data *data)
123 WARN_ON(!(data->flags & CSD_FLAG_LOCK));
126 * ensure we're all done before releasing data:
130 data->flags &= ~CSD_FLAG_LOCK;
134 * Insert a previously allocated call_single_data element
135 * for execution on the given CPU. data must already have
136 * ->func, ->info, and ->flags set.
139 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
141 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
145 raw_spin_lock_irqsave(&dst->lock, flags);
146 ipi = list_empty(&dst->list);
147 list_add_tail(&data->list, &dst->list);
148 raw_spin_unlock_irqrestore(&dst->lock, flags);
151 * The list addition should be visible before sending the IPI
152 * handler locks the list to pull the entry off it because of
153 * normal cache coherency rules implied by spinlocks.
155 * If IPIs can go out of order to the cache coherency protocol
156 * in an architecture, sufficient synchronisation should be added
157 * to arch code to make it appear to obey cache coherency WRT
158 * locking and barrier primitives. Generic code isn't really
159 * equipped to do the right thing...
162 arch_send_call_function_single_ipi(cpu);
169 * Invoked by arch to handle an IPI for call function. Must be called with
170 * interrupts disabled.
172 void generic_smp_call_function_interrupt(void)
174 struct call_function_data *data;
175 int cpu = smp_processor_id();
178 * Shouldn't receive this interrupt on a cpu that is not yet online.
180 WARN_ON_ONCE(!cpu_online(cpu));
183 * Ensure entry is visible on call_function_queue after we have
184 * entered the IPI. See comment in smp_call_function_many.
185 * If we don't have this, then we may miss an entry on the list
186 * and never get another IPI to process it.
191 * It's ok to use list_for_each_rcu() here even though we may
192 * delete 'pos', since list_del_rcu() doesn't clear ->next
194 list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
197 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask))
200 data->csd.func(data->csd.info);
202 refs = atomic_dec_return(&data->refs);
205 raw_spin_lock(&call_function.lock);
206 list_del_rcu(&data->csd.list);
207 raw_spin_unlock(&call_function.lock);
213 csd_unlock(&data->csd);
219 * Invoked by arch to handle an IPI for call function single. Must be
220 * called from the arch with interrupts disabled.
222 void generic_smp_call_function_single_interrupt(void)
224 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
225 unsigned int data_flags;
229 * Shouldn't receive this interrupt on a cpu that is not yet online.
231 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
233 raw_spin_lock(&q->lock);
234 list_replace_init(&q->list, &list);
235 raw_spin_unlock(&q->lock);
237 while (!list_empty(&list)) {
238 struct call_single_data *data;
240 data = list_entry(list.next, struct call_single_data, list);
241 list_del(&data->list);
244 * 'data' can be invalid after this call if flags == 0
245 * (when called through generic_exec_single()),
246 * so save them away before making the call:
248 data_flags = data->flags;
250 data->func(data->info);
253 * Unlocked CSDs are valid through generic_exec_single():
255 if (data_flags & CSD_FLAG_LOCK)
260 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data, csd_data);
263 * smp_call_function_single - Run a function on a specific CPU
264 * @func: The function to run. This must be fast and non-blocking.
265 * @info: An arbitrary pointer to pass to the function.
266 * @wait: If true, wait until function has completed on other CPUs.
268 * Returns 0 on success, else a negative status code.
270 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
273 struct call_single_data d = {
281 * prevent preemption and reschedule on another processor,
282 * as well as CPU removal
284 this_cpu = get_cpu();
287 * Can deadlock when called with interrupts disabled.
288 * We allow cpu's that are not yet online though, as no one else can
289 * send smp call function interrupt to this cpu and as such deadlocks
292 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
293 && !oops_in_progress);
295 if (cpu == this_cpu) {
296 local_irq_save(flags);
298 local_irq_restore(flags);
300 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
301 struct call_single_data *data = &d;
304 data = &__get_cpu_var(csd_data);
310 generic_exec_single(cpu, data, wait);
312 err = -ENXIO; /* CPU not online */
320 EXPORT_SYMBOL(smp_call_function_single);
323 * smp_call_function_any - Run a function on any of the given cpus
324 * @mask: The mask of cpus it can run on.
325 * @func: The function to run. This must be fast and non-blocking.
326 * @info: An arbitrary pointer to pass to the function.
327 * @wait: If true, wait until function has completed.
329 * Returns 0 on success, else a negative status code (if no cpus were online).
330 * Note that @wait will be implicitly turned on in case of allocation failures,
331 * since we fall back to on-stack allocation.
333 * Selection preference:
334 * 1) current cpu if in @mask
335 * 2) any cpu of current node if in @mask
336 * 3) any other online cpu in @mask
338 int smp_call_function_any(const struct cpumask *mask,
339 void (*func)(void *info), void *info, int wait)
342 const struct cpumask *nodemask;
345 /* Try for same CPU (cheapest) */
347 if (cpumask_test_cpu(cpu, mask))
350 /* Try for same node. */
351 nodemask = cpumask_of_node(cpu_to_node(cpu));
352 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
353 cpu = cpumask_next_and(cpu, nodemask, mask)) {
358 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
359 cpu = cpumask_any_and(mask, cpu_online_mask);
361 ret = smp_call_function_single(cpu, func, info, wait);
365 EXPORT_SYMBOL_GPL(smp_call_function_any);
368 * __smp_call_function_single(): Run a function on another CPU
369 * @cpu: The CPU to run on.
370 * @data: Pre-allocated and setup data structure
372 * Like smp_call_function_single(), but allow caller to pass in a
373 * pre-allocated data structure. Useful for embedding @data inside
374 * other structures, for instance.
376 void __smp_call_function_single(int cpu, struct call_single_data *data,
382 * Can deadlock when called with interrupts disabled.
383 * We allow cpu's that are not yet online though, as no one else can
384 * send smp call function interrupt to this cpu and as such deadlocks
387 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
388 && !oops_in_progress);
390 generic_exec_single(cpu, data, wait);
394 * smp_call_function_many(): Run a function on a set of other CPUs.
395 * @mask: The set of cpus to run on (only runs on online subset).
396 * @func: The function to run. This must be fast and non-blocking.
397 * @info: An arbitrary pointer to pass to the function.
398 * @wait: If true, wait (atomically) until function has completed
401 * If @wait is true, then returns once @func has returned.
403 * You must not call this function with disabled interrupts or from a
404 * hardware interrupt handler or from a bottom half handler. Preemption
405 * must be disabled when calling this function.
407 void smp_call_function_many(const struct cpumask *mask,
408 void (*func)(void *), void *info, bool wait)
410 struct call_function_data *data;
412 int cpu, next_cpu, this_cpu = smp_processor_id();
415 * Can deadlock when called with interrupts disabled.
416 * We allow cpu's that are not yet online though, as no one else can
417 * send smp call function interrupt to this cpu and as such deadlocks
420 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
421 && !oops_in_progress);
423 /* So, what's a CPU they want? Ignoring this one. */
424 cpu = cpumask_first_and(mask, cpu_online_mask);
426 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
428 /* No online cpus? We're done. */
429 if (cpu >= nr_cpu_ids)
432 /* Do we have another CPU which isn't us? */
433 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
434 if (next_cpu == this_cpu)
435 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
437 /* Fastpath: do that cpu by itself. */
438 if (next_cpu >= nr_cpu_ids) {
439 smp_call_function_single(cpu, func, info, wait);
443 data = &__get_cpu_var(cfd_data);
444 csd_lock(&data->csd);
446 data->csd.func = func;
447 data->csd.info = info;
448 cpumask_and(data->cpumask, mask, cpu_online_mask);
449 cpumask_clear_cpu(this_cpu, data->cpumask);
450 atomic_set(&data->refs, cpumask_weight(data->cpumask));
452 raw_spin_lock_irqsave(&call_function.lock, flags);
454 * Place entry at the _HEAD_ of the list, so that any cpu still
455 * observing the entry in generic_smp_call_function_interrupt()
456 * will not miss any other list entries:
458 list_add_rcu(&data->csd.list, &call_function.queue);
459 raw_spin_unlock_irqrestore(&call_function.lock, flags);
462 * Make the list addition visible before sending the ipi.
463 * (IPIs must obey or appear to obey normal Linux cache
464 * coherency rules -- see comment in generic_exec_single).
468 /* Send a message to all CPUs in the map */
469 arch_send_call_function_ipi_mask(data->cpumask);
471 /* Optionally wait for the CPUs to complete */
473 csd_lock_wait(&data->csd);
475 EXPORT_SYMBOL(smp_call_function_many);
478 * smp_call_function(): Run a function on all other CPUs.
479 * @func: The function to run. This must be fast and non-blocking.
480 * @info: An arbitrary pointer to pass to the function.
481 * @wait: If true, wait (atomically) until function has completed
486 * If @wait is true, then returns once @func has returned; otherwise
487 * it returns just before the target cpu calls @func.
489 * You must not call this function with disabled interrupts or from a
490 * hardware interrupt handler or from a bottom half handler.
492 int smp_call_function(void (*func)(void *), void *info, int wait)
495 smp_call_function_many(cpu_online_mask, func, info, wait);
500 EXPORT_SYMBOL(smp_call_function);
502 void ipi_call_lock(void)
504 raw_spin_lock(&call_function.lock);
507 void ipi_call_unlock(void)
509 raw_spin_unlock(&call_function.lock);
512 void ipi_call_lock_irq(void)
514 raw_spin_lock_irq(&call_function.lock);
517 void ipi_call_unlock_irq(void)
519 raw_spin_unlock_irq(&call_function.lock);