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
,
55 return notifier_from_errno(-ENOMEM
);
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
) {
198 * Since we walk the list without any locks, we might
199 * see an entry that was completed, removed from the
200 * list and is in the process of being reused.
202 * We must check that the cpu is in the cpumask before
203 * checking the refs, and both must be set before
204 * executing the callback on this cpu.
207 if (!cpumask_test_cpu(cpu
, data
->cpumask
))
212 if (atomic_read(&data
->refs
) == 0)
215 if (!cpumask_test_and_clear_cpu(cpu
, data
->cpumask
))
218 data
->csd
.func(data
->csd
.info
);
220 refs
= atomic_dec_return(&data
->refs
);
223 raw_spin_lock(&call_function
.lock
);
224 list_del_rcu(&data
->csd
.list
);
225 raw_spin_unlock(&call_function
.lock
);
231 csd_unlock(&data
->csd
);
237 * Invoked by arch to handle an IPI for call function single. Must be
238 * called from the arch with interrupts disabled.
240 void generic_smp_call_function_single_interrupt(void)
242 struct call_single_queue
*q
= &__get_cpu_var(call_single_queue
);
243 unsigned int data_flags
;
247 * Shouldn't receive this interrupt on a cpu that is not yet online.
249 WARN_ON_ONCE(!cpu_online(smp_processor_id()));
251 raw_spin_lock(&q
->lock
);
252 list_replace_init(&q
->list
, &list
);
253 raw_spin_unlock(&q
->lock
);
255 while (!list_empty(&list
)) {
256 struct call_single_data
*data
;
258 data
= list_entry(list
.next
, struct call_single_data
, list
);
259 list_del(&data
->list
);
262 * 'data' can be invalid after this call if flags == 0
263 * (when called through generic_exec_single()),
264 * so save them away before making the call:
266 data_flags
= data
->flags
;
268 data
->func(data
->info
);
271 * Unlocked CSDs are valid through generic_exec_single():
273 if (data_flags
& CSD_FLAG_LOCK
)
278 static DEFINE_PER_CPU_SHARED_ALIGNED(struct call_single_data
, csd_data
);
281 * smp_call_function_single - Run a function on a specific CPU
282 * @func: The function to run. This must be fast and non-blocking.
283 * @info: An arbitrary pointer to pass to the function.
284 * @wait: If true, wait until function has completed on other CPUs.
286 * Returns 0 on success, else a negative status code.
288 int smp_call_function_single(int cpu
, void (*func
) (void *info
), void *info
,
291 struct call_single_data d
= {
299 * prevent preemption and reschedule on another processor,
300 * as well as CPU removal
302 this_cpu
= get_cpu();
305 * Can deadlock when called with interrupts disabled.
306 * We allow cpu's that are not yet online though, as no one else can
307 * send smp call function interrupt to this cpu and as such deadlocks
310 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
311 && !oops_in_progress
);
313 if (cpu
== this_cpu
) {
314 local_irq_save(flags
);
316 local_irq_restore(flags
);
318 if ((unsigned)cpu
< nr_cpu_ids
&& cpu_online(cpu
)) {
319 struct call_single_data
*data
= &d
;
322 data
= &__get_cpu_var(csd_data
);
328 generic_exec_single(cpu
, data
, wait
);
330 err
= -ENXIO
; /* CPU not online */
338 EXPORT_SYMBOL(smp_call_function_single
);
341 * smp_call_function_any - Run a function on any of the given cpus
342 * @mask: The mask of cpus it can run on.
343 * @func: The function to run. This must be fast and non-blocking.
344 * @info: An arbitrary pointer to pass to the function.
345 * @wait: If true, wait until function has completed.
347 * Returns 0 on success, else a negative status code (if no cpus were online).
348 * Note that @wait will be implicitly turned on in case of allocation failures,
349 * since we fall back to on-stack allocation.
351 * Selection preference:
352 * 1) current cpu if in @mask
353 * 2) any cpu of current node if in @mask
354 * 3) any other online cpu in @mask
356 int smp_call_function_any(const struct cpumask
*mask
,
357 void (*func
)(void *info
), void *info
, int wait
)
360 const struct cpumask
*nodemask
;
363 /* Try for same CPU (cheapest) */
365 if (cpumask_test_cpu(cpu
, mask
))
368 /* Try for same node. */
369 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
370 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
371 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
376 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
377 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
379 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
383 EXPORT_SYMBOL_GPL(smp_call_function_any
);
386 * __smp_call_function_single(): Run a function on another CPU
387 * @cpu: The CPU to run on.
388 * @data: Pre-allocated and setup data structure
390 * Like smp_call_function_single(), but allow caller to pass in a
391 * pre-allocated data structure. Useful for embedding @data inside
392 * other structures, for instance.
394 void __smp_call_function_single(int cpu
, struct call_single_data
*data
,
400 * Can deadlock when called with interrupts disabled.
401 * We allow cpu's that are not yet online though, as no one else can
402 * send smp call function interrupt to this cpu and as such deadlocks
405 WARN_ON_ONCE(cpu_online(smp_processor_id()) && wait
&& irqs_disabled()
406 && !oops_in_progress
);
408 generic_exec_single(cpu
, data
, wait
);
412 * smp_call_function_many(): Run a function on a set of other CPUs.
413 * @mask: The set of cpus to run on (only runs on online subset).
414 * @func: The function to run. This must be fast and non-blocking.
415 * @info: An arbitrary pointer to pass to the function.
416 * @wait: If true, wait (atomically) until function has completed
419 * If @wait is true, then returns once @func has returned.
421 * You must not call this function with disabled interrupts or from a
422 * hardware interrupt handler or from a bottom half handler. Preemption
423 * must be disabled when calling this function.
425 void smp_call_function_many(const struct cpumask
*mask
,
426 void (*func
)(void *), void *info
, bool wait
)
428 struct call_function_data
*data
;
430 int refs
, cpu
, next_cpu
, this_cpu
= smp_processor_id();
433 * Can deadlock when called with interrupts disabled.
434 * We allow cpu's that are not yet online though, as no one else can
435 * send smp call function interrupt to this cpu and as such deadlocks
438 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
439 && !oops_in_progress
);
441 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
442 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
444 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
446 /* No online cpus? We're done. */
447 if (cpu
>= nr_cpu_ids
)
450 /* Do we have another CPU which isn't us? */
451 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
452 if (next_cpu
== this_cpu
)
453 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
455 /* Fastpath: do that cpu by itself. */
456 if (next_cpu
>= nr_cpu_ids
) {
457 smp_call_function_single(cpu
, func
, info
, wait
);
461 data
= &__get_cpu_var(cfd_data
);
462 csd_lock(&data
->csd
);
464 /* This BUG_ON verifies our reuse assertions and can be removed */
465 BUG_ON(atomic_read(&data
->refs
) || !cpumask_empty(data
->cpumask
));
468 * The global call function queue list add and delete are protected
469 * by a lock, but the list is traversed without any lock, relying
470 * on the rcu list add and delete to allow safe concurrent traversal.
471 * We reuse the call function data without waiting for any grace
472 * period after some other cpu removes it from the global queue.
473 * This means a cpu might find our data block as it is being
476 * We hold off the interrupt handler on the other cpu by
477 * ordering our writes to the cpu mask vs our setting of the
478 * refs counter. We assert only the cpu owning the data block
479 * will set a bit in cpumask, and each bit will only be cleared
480 * by the subject cpu. Each cpu must first find its bit is
481 * set and then check that refs is set indicating the element is
482 * ready to be processed, otherwise it must skip the entry.
484 * On the previous iteration refs was set to 0 by another cpu.
485 * To avoid the use of transitivity, set the counter to 0 here
486 * so the wmb will pair with the rmb in the interrupt handler.
488 atomic_set(&data
->refs
, 0); /* convert 3rd to 1st party write */
490 data
->csd
.func
= func
;
491 data
->csd
.info
= info
;
493 /* Ensure 0 refs is visible before mask. Also orders func and info */
496 /* We rely on the "and" being processed before the store */
497 cpumask_and(data
->cpumask
, mask
, cpu_online_mask
);
498 cpumask_clear_cpu(this_cpu
, data
->cpumask
);
499 refs
= cpumask_weight(data
->cpumask
);
501 /* Some callers race with other cpus changing the passed mask */
502 if (unlikely(!refs
)) {
503 csd_unlock(&data
->csd
);
507 raw_spin_lock_irqsave(&call_function
.lock
, flags
);
509 * Place entry at the _HEAD_ of the list, so that any cpu still
510 * observing the entry in generic_smp_call_function_interrupt()
511 * will not miss any other list entries:
513 list_add_rcu(&data
->csd
.list
, &call_function
.queue
);
515 * We rely on the wmb() in list_add_rcu to complete our writes
516 * to the cpumask before this write to refs, which indicates
517 * data is on the list and is ready to be processed.
519 atomic_set(&data
->refs
, refs
);
520 raw_spin_unlock_irqrestore(&call_function
.lock
, flags
);
523 * Make the list addition visible before sending the ipi.
524 * (IPIs must obey or appear to obey normal Linux cache
525 * coherency rules -- see comment in generic_exec_single).
529 /* Send a message to all CPUs in the map */
530 arch_send_call_function_ipi_mask(data
->cpumask
);
532 /* Optionally wait for the CPUs to complete */
534 csd_lock_wait(&data
->csd
);
536 EXPORT_SYMBOL(smp_call_function_many
);
539 * smp_call_function(): Run a function on all other CPUs.
540 * @func: The function to run. This must be fast and non-blocking.
541 * @info: An arbitrary pointer to pass to the function.
542 * @wait: If true, wait (atomically) until function has completed
547 * If @wait is true, then returns once @func has returned; otherwise
548 * it returns just before the target cpu calls @func.
550 * You must not call this function with disabled interrupts or from a
551 * hardware interrupt handler or from a bottom half handler.
553 int smp_call_function(void (*func
)(void *), void *info
, int wait
)
556 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
561 EXPORT_SYMBOL(smp_call_function
);
563 void ipi_call_lock(void)
565 raw_spin_lock(&call_function
.lock
);
568 void ipi_call_unlock(void)
570 raw_spin_unlock(&call_function
.lock
);
573 void ipi_call_lock_irq(void)
575 raw_spin_lock_irq(&call_function
.lock
);
578 void ipi_call_unlock_irq(void)
580 raw_spin_unlock_irq(&call_function
.lock
);