2 * Generic helpers for smp ipi calls
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
7 #include <linux/init.h>
8 #include <linux/module.h>
9 #include <linux/percpu.h>
10 #include <linux/rcupdate.h>
11 #include <linux/rculist.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
= {
21 .queue
= LIST_HEAD_INIT(call_function
.queue
),
22 .lock
= __SPIN_LOCK_UNLOCKED(call_function
.lock
),
29 struct call_function_data
{
30 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
) = {
42 .lock
= __SPIN_LOCK_UNLOCKED(cfd_data
.lock
),
46 hotplug_cfd(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
48 long cpu
= (long)hcpu
;
49 struct call_function_data
*cfd
= &per_cpu(cfd_data
, cpu
);
53 case CPU_UP_PREPARE_FROZEN
:
54 if (!alloc_cpumask_var_node(&cfd
->cpumask
, GFP_KERNEL
,
59 #ifdef CONFIG_CPU_HOTPLUG
61 case CPU_UP_CANCELED_FROZEN
:
65 free_cpumask_var(cfd
->cpumask
);
73 static struct notifier_block __cpuinitdata hotplug_cfd_notifier
= {
74 .notifier_call
= hotplug_cfd
,
77 static int __cpuinit
init_call_single_data(void)
79 void *cpu
= (void *)(long)smp_processor_id();
82 for_each_possible_cpu(i
) {
83 struct call_single_queue
*q
= &per_cpu(call_single_queue
, i
);
85 spin_lock_init(&q
->lock
);
86 INIT_LIST_HEAD(&q
->list
);
89 hotplug_cfd(&hotplug_cfd_notifier
, CPU_UP_PREPARE
, cpu
);
90 register_cpu_notifier(&hotplug_cfd_notifier
);
94 early_initcall(init_call_single_data
);
97 * csd_lock/csd_unlock used to serialize access to per-cpu csd resources
99 * For non-synchronous ipi calls the csd can still be in use by the previous
100 * function call. For multi-cpu calls its even more interesting as we'll have
101 * to ensure no other cpu is observing our csd.
103 static void csd_lock_wait(struct call_single_data
*data
)
105 while (data
->flags
& CSD_FLAG_LOCK
)
109 static void csd_lock(struct call_single_data
*data
)
112 data
->flags
= CSD_FLAG_LOCK
;
115 * prevent CPU from reordering the above assignment to ->flags
116 * with any subsequent assignments to other fields of the
117 * specified call_single_data structure.
123 static void csd_unlock(struct call_single_data
*data
)
125 WARN_ON(!(data
->flags
& CSD_FLAG_LOCK
));
127 * ensure we're all done before releasing data
130 data
->flags
&= ~CSD_FLAG_LOCK
;
134 * Insert a previously allocated call_single_data element for execution
135 * on the given CPU. data must already have ->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 spin_lock_irqsave(&dst
->lock
, flags
);
145 ipi
= list_empty(&dst
->list
);
146 list_add_tail(&data
->list
, &dst
->list
);
147 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 equipped
158 * 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
;
178 * Ensure entry is visible on call_function_queue after we have
179 * entered the IPI. See comment in smp_call_function_many.
180 * If we don't have this, then we may miss an entry on the list
181 * and never get another IPI to process it.
186 * It's ok to use list_for_each_rcu() here even though we may delete
187 * 'pos', since list_del_rcu() doesn't clear ->next
189 list_for_each_entry_rcu(data
, &call_function
.queue
, csd
.list
) {
192 spin_lock(&data
->lock
);
193 if (!cpumask_test_cpu(cpu
, data
->cpumask
)) {
194 spin_unlock(&data
->lock
);
197 cpumask_clear_cpu(cpu
, data
->cpumask
);
198 spin_unlock(&data
->lock
);
200 data
->csd
.func(data
->csd
.info
);
202 spin_lock(&data
->lock
);
203 WARN_ON(data
->refs
== 0);
206 spin_lock(&call_function
.lock
);
207 list_del_rcu(&data
->csd
.list
);
208 spin_unlock(&call_function
.lock
);
210 spin_unlock(&data
->lock
);
215 csd_unlock(&data
->csd
);
222 * Invoked by arch to handle an IPI for call function single. Must be called
223 * from the arch with interrupts disabled.
225 void generic_smp_call_function_single_interrupt(void)
227 struct call_single_queue
*q
= &__get_cpu_var(call_single_queue
);
229 unsigned int data_flags
;
232 list_replace_init(&q
->list
, &list
);
233 spin_unlock(&q
->lock
);
235 while (!list_empty(&list
)) {
236 struct call_single_data
*data
;
238 data
= list_entry(list
.next
, struct call_single_data
,
240 list_del(&data
->list
);
243 * 'data' can be invalid after this call if
244 * flags == 0 (when called through
245 * generic_exec_single(), so save them away before
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(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. Note that @wait
269 * will be implicitly turned on in case of allocation failures, since
270 * we fall back to on-stack allocation.
272 int smp_call_function_single(int cpu
, void (*func
) (void *info
), void *info
,
275 struct call_single_data d
= {
279 /* prevent preemption and reschedule on another processor,
280 as well as CPU removal */
284 /* Can deadlock when called with interrupts disabled */
285 WARN_ON(irqs_disabled());
288 local_irq_save(flags
);
290 local_irq_restore(flags
);
291 } else if ((unsigned)cpu
< nr_cpu_ids
&& cpu_online(cpu
)) {
292 struct call_single_data
*data
= &d
;
295 data
= &__get_cpu_var(csd_data
);
301 generic_exec_single(cpu
, data
, wait
);
303 err
= -ENXIO
; /* CPU not online */
309 EXPORT_SYMBOL(smp_call_function_single
);
312 * __smp_call_function_single(): Run a function on another CPU
313 * @cpu: The CPU to run on.
314 * @data: Pre-allocated and setup data structure
316 * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
317 * data structure. Useful for embedding @data inside other structures, for
321 void __smp_call_function_single(int cpu
, struct call_single_data
*data
,
326 /* Can deadlock when called with interrupts disabled */
327 WARN_ON(wait
&& irqs_disabled());
329 generic_exec_single(cpu
, data
, wait
);
332 /* FIXME: Shim for archs using old arch_send_call_function_ipi API. */
333 #ifndef arch_send_call_function_ipi_mask
334 #define arch_send_call_function_ipi_mask(maskp) \
335 arch_send_call_function_ipi(*(maskp))
339 * smp_call_function_many(): Run a function on a set of other CPUs.
340 * @mask: The set of cpus to run on (only runs on online subset).
341 * @func: The function to run. This must be fast and non-blocking.
342 * @info: An arbitrary pointer to pass to the function.
343 * @wait: If true, wait (atomically) until function has completed on other CPUs.
345 * If @wait is true, then returns once @func has returned. Note that @wait
346 * will be implicitly turned on in case of allocation failures, since
347 * we fall back to on-stack allocation.
349 * You must not call this function with disabled interrupts or from a
350 * hardware interrupt handler or from a bottom half handler. Preemption
351 * must be disabled when calling this function.
353 void smp_call_function_many(const struct cpumask
*mask
,
354 void (*func
)(void *), void *info
,
357 struct call_function_data
*data
;
359 int cpu
, next_cpu
, me
= smp_processor_id();
361 /* Can deadlock when called with interrupts disabled */
362 WARN_ON(irqs_disabled());
364 /* So, what's a CPU they want? Ignoring this one. */
365 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
367 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
368 /* No online cpus? We're done. */
369 if (cpu
>= nr_cpu_ids
)
372 /* Do we have another CPU which isn't us? */
373 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
375 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
377 /* Fastpath: do that cpu by itself. */
378 if (next_cpu
>= nr_cpu_ids
) {
379 smp_call_function_single(cpu
, func
, info
, wait
);
383 data
= &__get_cpu_var(cfd_data
);
384 csd_lock(&data
->csd
);
386 spin_lock_irqsave(&data
->lock
, flags
);
387 data
->csd
.func
= func
;
388 data
->csd
.info
= info
;
389 cpumask_and(data
->cpumask
, mask
, cpu_online_mask
);
390 cpumask_clear_cpu(me
, data
->cpumask
);
391 data
->refs
= cpumask_weight(data
->cpumask
);
393 spin_lock(&call_function
.lock
);
395 * Place entry at the _HEAD_ of the list, so that any cpu still
396 * observing the entry in generic_smp_call_function_interrupt() will
397 * not miss any other list entries.
399 list_add_rcu(&data
->csd
.list
, &call_function
.queue
);
400 spin_unlock(&call_function
.lock
);
401 spin_unlock_irqrestore(&data
->lock
, flags
);
404 * Make the list addition visible before sending the ipi.
405 * (IPIs must obey or appear to obey normal Linux cache coherency
406 * rules -- see comment in generic_exec_single).
410 /* Send a message to all CPUs in the map */
411 arch_send_call_function_ipi_mask(data
->cpumask
);
413 /* optionally wait for the CPUs to complete */
415 csd_lock_wait(&data
->csd
);
417 EXPORT_SYMBOL(smp_call_function_many
);
420 * smp_call_function(): Run a function on all other CPUs.
421 * @func: The function to run. This must be fast and non-blocking.
422 * @info: An arbitrary pointer to pass to the function.
423 * @wait: If true, wait (atomically) until function has completed on other CPUs.
427 * If @wait is true, then returns once @func has returned; otherwise
428 * it returns just before the target cpu calls @func. In case of allocation
429 * failure, @wait will be implicitly turned on.
431 * You must not call this function with disabled interrupts or from a
432 * hardware interrupt handler or from a bottom half handler.
434 int smp_call_function(void (*func
)(void *), void *info
, int wait
)
437 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
441 EXPORT_SYMBOL(smp_call_function
);
443 void ipi_call_lock(void)
445 spin_lock(&call_function
.lock
);
448 void ipi_call_unlock(void)
450 spin_unlock(&call_function
.lock
);
453 void ipi_call_lock_irq(void)
455 spin_lock_irq(&call_function
.lock
);
458 void ipi_call_unlock_irq(void)
460 spin_unlock_irq(&call_function
.lock
);