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/smp.h>
13 #include <linux/cpu.h>
15 static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
18 struct list_head queue;
20 } call_function __cacheline_aligned_in_smp =
22 .queue = LIST_HEAD_INIT(call_function.queue),
23 .lock = __RAW_SPIN_LOCK_UNLOCKED(call_function.lock),
30 struct call_function_data {
31 struct call_single_data csd;
33 cpumask_var_t cpumask;
36 struct call_single_queue {
37 struct list_head list;
41 static DEFINE_PER_CPU(struct call_function_data, cfd_data);
44 hotplug_cfd(struct notifier_block *nfb, unsigned long action, void *hcpu)
46 long cpu = (long)hcpu;
47 struct call_function_data *cfd = &per_cpu(cfd_data, cpu);
51 case CPU_UP_PREPARE_FROZEN:
52 if (!zalloc_cpumask_var_node(&cfd->cpumask, GFP_KERNEL,
57 #ifdef CONFIG_HOTPLUG_CPU
59 case CPU_UP_CANCELED_FROZEN:
63 free_cpumask_var(cfd->cpumask);
71 static struct notifier_block __cpuinitdata hotplug_cfd_notifier = {
72 .notifier_call = hotplug_cfd,
75 static int __cpuinit init_call_single_data(void)
77 void *cpu = (void *)(long)smp_processor_id();
80 for_each_possible_cpu(i) {
81 struct call_single_queue *q = &per_cpu(call_single_queue, i);
83 raw_spin_lock_init(&q->lock);
84 INIT_LIST_HEAD(&q->list);
87 hotplug_cfd(&hotplug_cfd_notifier, CPU_UP_PREPARE, cpu);
88 register_cpu_notifier(&hotplug_cfd_notifier);
92 early_initcall(init_call_single_data);
95 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
97 * For non-synchronous ipi calls the csd can still be in use by the
98 * previous function call. For multi-cpu calls its even more interesting
99 * as we'll have to ensure no other cpu is observing our csd.
101 static void csd_lock_wait(struct call_single_data *data)
103 while (data->flags & CSD_FLAG_LOCK)
107 static void csd_lock(struct call_single_data *data)
110 data->flags = CSD_FLAG_LOCK;
113 * prevent CPU from reordering the above assignment
114 * to ->flags with any subsequent assignments to other
115 * fields of the specified call_single_data structure:
120 static void csd_unlock(struct call_single_data *data)
122 WARN_ON(!(data->flags & CSD_FLAG_LOCK));
125 * ensure we're all done before releasing data:
129 data->flags &= ~CSD_FLAG_LOCK;
133 * Insert a previously allocated call_single_data element
134 * for execution on the given CPU. data must already have
135 * ->func, ->info, and ->flags set.
138 void generic_exec_single(int cpu, struct call_single_data *data, int wait)
140 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
144 raw_spin_lock_irqsave(&dst->lock, flags);
145 ipi = list_empty(&dst->list);
146 list_add_tail(&data->list, &dst->list);
147 raw_spin_unlock_irqrestore(&dst->lock, flags);
150 * The list addition should be visible before sending the IPI
151 * handler locks the list to pull the entry off it because of
152 * normal cache coherency rules implied by spinlocks.
154 * If IPIs can go out of order to the cache coherency protocol
155 * in an architecture, sufficient synchronisation should be added
156 * to arch code to make it appear to obey cache coherency WRT
157 * locking and barrier primitives. Generic code isn't really
158 * equipped to do the right thing...
161 arch_send_call_function_single_ipi(cpu);
168 * Invoked by arch to handle an IPI for call function. Must be called with
169 * interrupts disabled.
171 void generic_smp_call_function_interrupt(void)
173 struct call_function_data *data;
174 int cpu = smp_processor_id();
177 * Shouldn't receive this interrupt on a cpu that is not yet online.
179 WARN_ON_ONCE(!cpu_online(cpu));
182 * Ensure entry is visible on call_function_queue after we have
183 * entered the IPI. See comment in smp_call_function_many.
184 * If we don't have this, then we may miss an entry on the list
185 * and never get another IPI to process it.
190 * It's ok to use list_for_each_rcu() here even though we may
191 * delete 'pos', since list_del_rcu() doesn't clear ->next
193 list_for_each_entry_rcu(data, &call_function.queue, csd.list) {
196 if (!cpumask_test_and_clear_cpu(cpu, data->cpumask))
199 data->csd.func(data->csd.info);
201 refs = atomic_dec_return(&data->refs);
204 raw_spin_lock(&call_function.lock);
205 list_del_rcu(&data->csd.list);
206 raw_spin_unlock(&call_function.lock);
212 csd_unlock(&data->csd);
218 * Invoked by arch to handle an IPI for call function single. Must be
219 * called from the arch with interrupts disabled.
221 void generic_smp_call_function_single_interrupt(void)
223 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
224 unsigned int data_flags;
228 * Shouldn't receive this interrupt on a cpu that is not yet online.
230 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
232 raw_spin_lock(&q->lock);
233 list_replace_init(&q->list, &list);
234 raw_spin_unlock(&q->lock);
236 while (!list_empty(&list)) {
237 struct call_single_data *data;
239 data = list_entry(list.next, struct call_single_data, list);
240 list_del(&data->list);
243 * 'data' can be invalid after this call if flags == 0
244 * (when called through generic_exec_single()),
245 * so save them away before making the call:
247 data_flags = data->flags;
249 data->func(data->info);
252 * Unlocked CSDs are valid through generic_exec_single():
254 if (data_flags & CSD_FLAG_LOCK)
259 static DEFINE_PER_CPU(struct call_single_data, csd_data);
262 * smp_call_function_single - Run a function on a specific CPU
263 * @func: The function to run. This must be fast and non-blocking.
264 * @info: An arbitrary pointer to pass to the function.
265 * @wait: If true, wait until function has completed on other CPUs.
267 * Returns 0 on success, else a negative status code.
269 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
272 struct call_single_data d = {
280 * prevent preemption and reschedule on another processor,
281 * as well as CPU removal
283 this_cpu = get_cpu();
286 * Can deadlock when called with interrupts disabled.
287 * We allow cpu's that are not yet online though, as no one else can
288 * send smp call function interrupt to this cpu and as such deadlocks
291 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
292 && !oops_in_progress);
294 if (cpu == this_cpu) {
295 local_irq_save(flags);
297 local_irq_restore(flags);
299 if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
300 struct call_single_data *data = &d;
303 data = &__get_cpu_var(csd_data);
309 generic_exec_single(cpu, data, wait);
311 err = -ENXIO; /* CPU not online */
319 EXPORT_SYMBOL(smp_call_function_single);
322 * smp_call_function_any - Run a function on any of the given cpus
323 * @mask: The mask of cpus it can run on.
324 * @func: The function to run. This must be fast and non-blocking.
325 * @info: An arbitrary pointer to pass to the function.
326 * @wait: If true, wait until function has completed.
328 * Returns 0 on success, else a negative status code (if no cpus were online).
329 * Note that @wait will be implicitly turned on in case of allocation failures,
330 * since we fall back to on-stack allocation.
332 * Selection preference:
333 * 1) current cpu if in @mask
334 * 2) any cpu of current node if in @mask
335 * 3) any other online cpu in @mask
337 int smp_call_function_any(const struct cpumask *mask,
338 void (*func)(void *info), void *info, int wait)
341 const struct cpumask *nodemask;
344 /* Try for same CPU (cheapest) */
346 if (cpumask_test_cpu(cpu, mask))
349 /* Try for same node. */
350 nodemask = cpumask_of_node(cpu_to_node(cpu));
351 for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
352 cpu = cpumask_next_and(cpu, nodemask, mask)) {
357 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
358 cpu = cpumask_any_and(mask, cpu_online_mask);
360 ret = smp_call_function_single(cpu, func, info, wait);
364 EXPORT_SYMBOL_GPL(smp_call_function_any);
367 * __smp_call_function_single(): Run a function on another CPU
368 * @cpu: The CPU to run on.
369 * @data: Pre-allocated and setup data structure
371 * Like smp_call_function_single(), but allow caller to pass in a
372 * pre-allocated data structure. Useful for embedding @data inside
373 * other structures, for instance.
375 void __smp_call_function_single(int cpu, struct call_single_data *data,
381 * Can deadlock when called with interrupts disabled.
382 * We allow cpu's that are not yet online though, as no one else can
383 * send smp call function interrupt to this cpu and as such deadlocks
386 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait && irqs_disabled()
387 && !oops_in_progress);
389 generic_exec_single(cpu, data, wait);
393 * smp_call_function_many(): Run a function on a set of other CPUs.
394 * @mask: The set of cpus to run on (only runs on online subset).
395 * @func: The function to run. This must be fast and non-blocking.
396 * @info: An arbitrary pointer to pass to the function.
397 * @wait: If true, wait (atomically) until function has completed
400 * If @wait is true, then returns once @func has returned.
402 * You must not call this function with disabled interrupts or from a
403 * hardware interrupt handler or from a bottom half handler. Preemption
404 * must be disabled when calling this function.
406 void smp_call_function_many(const struct cpumask *mask,
407 void (*func)(void *), void *info, bool wait)
409 struct call_function_data *data;
411 int cpu, next_cpu, this_cpu = smp_processor_id();
414 * Can deadlock when called with interrupts disabled.
415 * We allow cpu's that are not yet online though, as no one else can
416 * send smp call function interrupt to this cpu and as such deadlocks
419 WARN_ON_ONCE(cpu_online(this_cpu) && irqs_disabled()
420 && !oops_in_progress);
422 /* So, what's a CPU they want? Ignoring this one. */
423 cpu = cpumask_first_and(mask, cpu_online_mask);
425 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
427 /* No online cpus? We're done. */
428 if (cpu >= nr_cpu_ids)
431 /* Do we have another CPU which isn't us? */
432 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
433 if (next_cpu == this_cpu)
434 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
436 /* Fastpath: do that cpu by itself. */
437 if (next_cpu >= nr_cpu_ids) {
438 smp_call_function_single(cpu, func, info, wait);
442 data = &__get_cpu_var(cfd_data);
443 csd_lock(&data->csd);
445 data->csd.func = func;
446 data->csd.info = info;
447 cpumask_and(data->cpumask, mask, cpu_online_mask);
448 cpumask_clear_cpu(this_cpu, data->cpumask);
449 atomic_set(&data->refs, cpumask_weight(data->cpumask));
451 raw_spin_lock_irqsave(&call_function.lock, flags);
453 * Place entry at the _HEAD_ of the list, so that any cpu still
454 * observing the entry in generic_smp_call_function_interrupt()
455 * will not miss any other list entries:
457 list_add_rcu(&data->csd.list, &call_function.queue);
458 raw_spin_unlock_irqrestore(&call_function.lock, flags);
461 * Make the list addition visible before sending the ipi.
462 * (IPIs must obey or appear to obey normal Linux cache
463 * coherency rules -- see comment in generic_exec_single).
467 /* Send a message to all CPUs in the map */
468 arch_send_call_function_ipi_mask(data->cpumask);
470 /* Optionally wait for the CPUs to complete */
472 csd_lock_wait(&data->csd);
474 EXPORT_SYMBOL(smp_call_function_many);
477 * smp_call_function(): Run a function on all other CPUs.
478 * @func: The function to run. This must be fast and non-blocking.
479 * @info: An arbitrary pointer to pass to the function.
480 * @wait: If true, wait (atomically) until function has completed
485 * If @wait is true, then returns once @func has returned; otherwise
486 * it returns just before the target cpu calls @func.
488 * You must not call this function with disabled interrupts or from a
489 * hardware interrupt handler or from a bottom half handler.
491 int smp_call_function(void (*func)(void *), void *info, int wait)
494 smp_call_function_many(cpu_online_mask, func, info, wait);
499 EXPORT_SYMBOL(smp_call_function);
501 void ipi_call_lock(void)
503 raw_spin_lock(&call_function.lock);
506 void ipi_call_unlock(void)
508 raw_spin_unlock(&call_function.lock);
511 void ipi_call_lock_irq(void)
513 raw_spin_lock_irq(&call_function.lock);
516 void ipi_call_unlock_irq(void)
518 raw_spin_unlock_irq(&call_function.lock);