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(): Run a function on a specific CPU
242 * @cpu: The CPU to run on.
243 * @csd: Pre-allocated and setup data structure
244 * @wait: If true, wait until function has completed on specified CPU.
246 * Like smp_call_function_single(), but allow caller to pass in a
247 * pre-allocated data structure. Useful for embedding @data inside
248 * other structures, for instance.
250 int __smp_call_function_single(int cpu
, struct call_single_data
*csd
, int wait
)
255 this_cpu
= get_cpu();
257 * Can deadlock when called with interrupts disabled.
258 * We allow cpu's that are not yet online though, as no one else can
259 * send smp call function interrupt to this cpu and as such deadlocks
262 WARN_ON_ONCE(cpu_online(this_cpu
) && wait
&& irqs_disabled()
263 && !oops_in_progress
);
265 err
= generic_exec_single(cpu
, csd
, csd
->func
, csd
->info
, wait
);
270 EXPORT_SYMBOL_GPL(__smp_call_function_single
);
273 * smp_call_function_any - Run a function on any of the given cpus
274 * @mask: The mask of cpus it can run on.
275 * @func: The function to run. This must be fast and non-blocking.
276 * @info: An arbitrary pointer to pass to the function.
277 * @wait: If true, wait until function has completed.
279 * Returns 0 on success, else a negative status code (if no cpus were online).
281 * Selection preference:
282 * 1) current cpu if in @mask
283 * 2) any cpu of current node if in @mask
284 * 3) any other online cpu in @mask
286 int smp_call_function_any(const struct cpumask
*mask
,
287 smp_call_func_t func
, void *info
, int wait
)
290 const struct cpumask
*nodemask
;
293 /* Try for same CPU (cheapest) */
295 if (cpumask_test_cpu(cpu
, mask
))
298 /* Try for same node. */
299 nodemask
= cpumask_of_node(cpu_to_node(cpu
));
300 for (cpu
= cpumask_first_and(nodemask
, mask
); cpu
< nr_cpu_ids
;
301 cpu
= cpumask_next_and(cpu
, nodemask
, mask
)) {
306 /* Any online will do: smp_call_function_single handles nr_cpu_ids. */
307 cpu
= cpumask_any_and(mask
, cpu_online_mask
);
309 ret
= smp_call_function_single(cpu
, func
, info
, wait
);
313 EXPORT_SYMBOL_GPL(smp_call_function_any
);
316 * smp_call_function_many(): Run a function on a set of other CPUs.
317 * @mask: The set of cpus to run on (only runs on online subset).
318 * @func: The function to run. This must be fast and non-blocking.
319 * @info: An arbitrary pointer to pass to the function.
320 * @wait: If true, wait (atomically) until function has completed
323 * If @wait is true, then returns once @func has returned.
325 * You must not call this function with disabled interrupts or from a
326 * hardware interrupt handler or from a bottom half handler. Preemption
327 * must be disabled when calling this function.
329 void smp_call_function_many(const struct cpumask
*mask
,
330 smp_call_func_t func
, void *info
, bool wait
)
332 struct call_function_data
*cfd
;
333 int cpu
, next_cpu
, this_cpu
= smp_processor_id();
336 * Can deadlock when called with interrupts disabled.
337 * We allow cpu's that are not yet online though, as no one else can
338 * send smp call function interrupt to this cpu and as such deadlocks
341 WARN_ON_ONCE(cpu_online(this_cpu
) && irqs_disabled()
342 && !oops_in_progress
&& !early_boot_irqs_disabled
);
344 /* Try to fastpath. So, what's a CPU they want? Ignoring this one. */
345 cpu
= cpumask_first_and(mask
, cpu_online_mask
);
347 cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
349 /* No online cpus? We're done. */
350 if (cpu
>= nr_cpu_ids
)
353 /* Do we have another CPU which isn't us? */
354 next_cpu
= cpumask_next_and(cpu
, mask
, cpu_online_mask
);
355 if (next_cpu
== this_cpu
)
356 next_cpu
= cpumask_next_and(next_cpu
, mask
, cpu_online_mask
);
358 /* Fastpath: do that cpu by itself. */
359 if (next_cpu
>= nr_cpu_ids
) {
360 smp_call_function_single(cpu
, func
, info
, wait
);
364 cfd
= &__get_cpu_var(cfd_data
);
366 cpumask_and(cfd
->cpumask
, mask
, cpu_online_mask
);
367 cpumask_clear_cpu(this_cpu
, cfd
->cpumask
);
369 /* Some callers race with other cpus changing the passed mask */
370 if (unlikely(!cpumask_weight(cfd
->cpumask
)))
373 for_each_cpu(cpu
, cfd
->cpumask
) {
374 struct call_single_data
*csd
= per_cpu_ptr(cfd
->csd
, cpu
);
379 llist_add(&csd
->llist
, &per_cpu(call_single_queue
, cpu
));
382 /* Send a message to all CPUs in the map */
383 arch_send_call_function_ipi_mask(cfd
->cpumask
);
386 for_each_cpu(cpu
, cfd
->cpumask
) {
387 struct call_single_data
*csd
;
389 csd
= per_cpu_ptr(cfd
->csd
, cpu
);
394 EXPORT_SYMBOL(smp_call_function_many
);
397 * smp_call_function(): Run a function on all other CPUs.
398 * @func: The function to run. This must be fast and non-blocking.
399 * @info: An arbitrary pointer to pass to the function.
400 * @wait: If true, wait (atomically) until function has completed
405 * If @wait is true, then returns once @func has returned; otherwise
406 * it returns just before the target cpu calls @func.
408 * You must not call this function with disabled interrupts or from a
409 * hardware interrupt handler or from a bottom half handler.
411 int smp_call_function(smp_call_func_t func
, void *info
, int wait
)
414 smp_call_function_many(cpu_online_mask
, func
, info
, wait
);
419 EXPORT_SYMBOL(smp_call_function
);
421 /* Setup configured maximum number of CPUs to activate */
422 unsigned int setup_max_cpus
= NR_CPUS
;
423 EXPORT_SYMBOL(setup_max_cpus
);
427 * Setup routine for controlling SMP activation
429 * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
430 * activation entirely (the MPS table probe still happens, though).
432 * Command-line option of "maxcpus=<NUM>", where <NUM> is an integer
433 * greater than 0, limits the maximum number of CPUs activated in
437 void __weak
arch_disable_smp_support(void) { }
439 static int __init
nosmp(char *str
)
442 arch_disable_smp_support();
447 early_param("nosmp", nosmp
);
449 /* this is hard limit */
450 static int __init
nrcpus(char *str
)
454 get_option(&str
, &nr_cpus
);
455 if (nr_cpus
> 0 && nr_cpus
< nr_cpu_ids
)
456 nr_cpu_ids
= nr_cpus
;
461 early_param("nr_cpus", nrcpus
);
463 static int __init
maxcpus(char *str
)
465 get_option(&str
, &setup_max_cpus
);
466 if (setup_max_cpus
== 0)
467 arch_disable_smp_support();
472 early_param("maxcpus", maxcpus
);
474 /* Setup number of possible processor ids */
475 int nr_cpu_ids __read_mostly
= NR_CPUS
;
476 EXPORT_SYMBOL(nr_cpu_ids
);
478 /* An arch may set nr_cpu_ids earlier if needed, so this would be redundant */
479 void __init
setup_nr_cpu_ids(void)
481 nr_cpu_ids
= find_last_bit(cpumask_bits(cpu_possible_mask
),NR_CPUS
) + 1;
484 void __weak
smp_announce(void)
486 printk(KERN_INFO
"Brought up %d CPUs\n", num_online_cpus());
489 /* Called by boot processor to activate the rest. */
490 void __init
smp_init(void)
496 /* FIXME: This should be done in userspace --RR */
497 for_each_present_cpu(cpu
) {
498 if (num_online_cpus() >= setup_max_cpus
)
500 if (!cpu_online(cpu
))
504 /* Any cleanup work */
506 smp_cpus_done(setup_max_cpus
);
510 * Call a function on all processors. May be used during early boot while
511 * early_boot_irqs_disabled is set. Use local_irq_save/restore() instead
512 * of local_irq_disable/enable().
514 int on_each_cpu(void (*func
) (void *info
), void *info
, int wait
)
520 ret
= smp_call_function(func
, info
, wait
);
521 local_irq_save(flags
);
523 local_irq_restore(flags
);
527 EXPORT_SYMBOL(on_each_cpu
);
530 * on_each_cpu_mask(): Run a function on processors specified by
531 * cpumask, which may include the local processor.
532 * @mask: The set of cpus to run on (only runs on online subset).
533 * @func: The function to run. This must be fast and non-blocking.
534 * @info: An arbitrary pointer to pass to the function.
535 * @wait: If true, wait (atomically) until function has completed
538 * If @wait is true, then returns once @func has returned.
540 * You must not call this function with disabled interrupts or from a
541 * hardware interrupt handler or from a bottom half handler. The
542 * exception is that it may be used during early boot while
543 * early_boot_irqs_disabled is set.
545 void on_each_cpu_mask(const struct cpumask
*mask
, smp_call_func_t func
,
546 void *info
, bool wait
)
550 smp_call_function_many(mask
, func
, info
, wait
);
551 if (cpumask_test_cpu(cpu
, mask
)) {
553 local_irq_save(flags
);
555 local_irq_restore(flags
);
559 EXPORT_SYMBOL(on_each_cpu_mask
);
562 * on_each_cpu_cond(): Call a function on each processor for which
563 * the supplied function cond_func returns true, optionally waiting
564 * for all the required CPUs to finish. This may include the local
566 * @cond_func: A callback function that is passed a cpu id and
567 * the the info parameter. The function is called
568 * with preemption disabled. The function should
569 * return a blooean value indicating whether to IPI
571 * @func: The function to run on all applicable CPUs.
572 * This must be fast and non-blocking.
573 * @info: An arbitrary pointer to pass to both functions.
574 * @wait: If true, wait (atomically) until function has
575 * completed on other CPUs.
576 * @gfp_flags: GFP flags to use when allocating the cpumask
577 * used internally by the function.
579 * The function might sleep if the GFP flags indicates a non
580 * atomic allocation is allowed.
582 * Preemption is disabled to protect against CPUs going offline but not online.
583 * CPUs going online during the call will not be seen or sent an IPI.
585 * You must not call this function with disabled interrupts or
586 * from a hardware interrupt handler or from a bottom half handler.
588 void on_each_cpu_cond(bool (*cond_func
)(int cpu
, void *info
),
589 smp_call_func_t func
, void *info
, bool wait
,
595 might_sleep_if(gfp_flags
& __GFP_WAIT
);
597 if (likely(zalloc_cpumask_var(&cpus
, (gfp_flags
|__GFP_NOWARN
)))) {
599 for_each_online_cpu(cpu
)
600 if (cond_func(cpu
, info
))
601 cpumask_set_cpu(cpu
, cpus
);
602 on_each_cpu_mask(cpus
, func
, info
, wait
);
604 free_cpumask_var(cpus
);
607 * No free cpumask, bother. No matter, we'll
608 * just have to IPI them one by one.
611 for_each_online_cpu(cpu
)
612 if (cond_func(cpu
, info
)) {
613 ret
= smp_call_function_single(cpu
, func
,
620 EXPORT_SYMBOL(on_each_cpu_cond
);
622 static void do_nothing(void *unused
)
627 * kick_all_cpus_sync - Force all cpus out of idle
629 * Used to synchronize the update of pm_idle function pointer. It's
630 * called after the pointer is updated and returns after the dummy
631 * callback function has been executed on all cpus. The execution of
632 * the function can only happen on the remote cpus after they have
633 * left the idle function which had been called via pm_idle function
634 * pointer. So it's guaranteed that nothing uses the previous pointer
637 void kick_all_cpus_sync(void)
639 /* Make sure the change is visible before we kick the cpus */
641 smp_call_function(do_nothing
, NULL
, 1);
643 EXPORT_SYMBOL_GPL(kick_all_cpus_sync
);