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>
14 static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
15 static LIST_HEAD(call_function_queue);
16 __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock);
20 CSD_FLAG_ALLOC = 0x02,
23 struct call_function_data {
24 struct call_single_data csd;
28 struct rcu_head rcu_head;
31 struct call_single_queue {
32 struct list_head list;
36 static int __cpuinit init_call_single_data(void)
40 for_each_possible_cpu(i) {
41 struct call_single_queue *q = &per_cpu(call_single_queue, i);
43 spin_lock_init(&q->lock);
44 INIT_LIST_HEAD(&q->list);
48 early_initcall(init_call_single_data);
50 static void csd_flag_wait(struct call_single_data *data)
52 /* Wait for response */
55 * We need to see the flags store in the IPI handler
58 if (!(data->flags & CSD_FLAG_WAIT))
65 * Insert a previously allocated call_single_data element for execution
66 * on the given CPU. data must already have ->func, ->info, and ->flags set.
68 static void generic_exec_single(int cpu, struct call_single_data *data)
70 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
71 int wait = data->flags & CSD_FLAG_WAIT, ipi;
74 spin_lock_irqsave(&dst->lock, flags);
75 ipi = list_empty(&dst->list);
76 list_add_tail(&data->list, &dst->list);
77 spin_unlock_irqrestore(&dst->lock, flags);
80 arch_send_call_function_single_ipi(cpu);
86 static void rcu_free_call_data(struct rcu_head *head)
88 struct call_function_data *data;
90 data = container_of(head, struct call_function_data, rcu_head);
96 * Invoked by arch to handle an IPI for call function. Must be called with
97 * interrupts disabled.
99 void generic_smp_call_function_interrupt(void)
101 struct call_function_data *data;
105 * It's ok to use list_for_each_rcu() here even though we may delete
106 * 'pos', since list_del_rcu() doesn't clear ->next
109 list_for_each_entry_rcu(data, &call_function_queue, csd.list) {
112 if (!cpu_isset(cpu, data->cpumask))
115 data->csd.func(data->csd.info);
117 spin_lock(&data->lock);
118 cpu_clear(cpu, data->cpumask);
119 WARN_ON(data->refs == 0);
122 spin_unlock(&data->lock);
127 spin_lock(&call_function_lock);
128 list_del_rcu(&data->csd.list);
129 spin_unlock(&call_function_lock);
131 if (data->csd.flags & CSD_FLAG_WAIT) {
133 * serialize stores to data with the flag clear
137 data->csd.flags &= ~CSD_FLAG_WAIT;
139 call_rcu(&data->rcu_head, rcu_free_call_data);
147 * Invoked by arch to handle an IPI for call function single. Must be called
148 * from the arch with interrupts disabled.
150 void generic_smp_call_function_single_interrupt(void)
152 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
156 * Need to see other stores to list head for checking whether
157 * list is empty without holding q->lock
160 while (!list_empty(&q->list)) {
161 unsigned int data_flags;
164 list_replace_init(&q->list, &list);
165 spin_unlock(&q->lock);
167 while (!list_empty(&list)) {
168 struct call_single_data *data;
170 data = list_entry(list.next, struct call_single_data,
172 list_del(&data->list);
175 * 'data' can be invalid after this call if
176 * flags == 0 (when called through
177 * generic_exec_single(), so save them away before
180 data_flags = data->flags;
182 data->func(data->info);
184 if (data_flags & CSD_FLAG_WAIT) {
186 data->flags &= ~CSD_FLAG_WAIT;
187 } else if (data_flags & CSD_FLAG_ALLOC)
191 * See comment on outer loop
198 * smp_call_function_single - Run a function on a specific CPU
199 * @func: The function to run. This must be fast and non-blocking.
200 * @info: An arbitrary pointer to pass to the function.
201 * @wait: If true, wait until function has completed on other CPUs.
203 * Returns 0 on success, else a negative status code. Note that @wait
204 * will be implicitly turned on in case of allocation failures, since
205 * we fall back to on-stack allocation.
207 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
210 struct call_single_data d;
212 /* prevent preemption and reschedule on another processor */
215 /* Can deadlock when called with interrupts disabled */
216 WARN_ON(irqs_disabled());
219 local_irq_save(flags);
221 local_irq_restore(flags);
223 struct call_single_data *data = NULL;
226 data = kmalloc(sizeof(*data), GFP_ATOMIC);
228 data->flags = CSD_FLAG_ALLOC;
232 data->flags = CSD_FLAG_WAIT;
237 generic_exec_single(cpu, data);
243 EXPORT_SYMBOL(smp_call_function_single);
246 * __smp_call_function_single(): Run a function on another CPU
247 * @cpu: The CPU to run on.
248 * @data: Pre-allocated and setup data structure
250 * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
251 * data structure. Useful for embedding @data inside other structures, for
255 void __smp_call_function_single(int cpu, struct call_single_data *data)
257 /* Can deadlock when called with interrupts disabled */
258 WARN_ON((data->flags & CSD_FLAG_WAIT) && irqs_disabled());
260 generic_exec_single(cpu, data);
264 static void quiesce_dummy(void *unused)
269 * Ensure stack based data used in call function mask is safe to free.
271 * This is needed by smp_call_function_mask when using on-stack data, because
272 * a single call function queue is shared by all CPUs, and any CPU may pick up
273 * the data item on the queue at any time before it is deleted. So we need to
274 * ensure that all CPUs have transitioned through a quiescent state after
277 * This is a very slow function, implemented by sending synchronous IPIs to
278 * all possible CPUs. For this reason, we have to alloc data rather than use
279 * stack based data even in the case of synchronous calls. The stack based
280 * data is then just used for deadlock/oom fallback which will be very rare.
282 * If a faster scheme can be made, we could go back to preferring stack based
283 * data -- the data allocation/free is non-zero cost.
285 static void smp_call_function_mask_quiesce_stack(cpumask_t mask)
287 struct call_single_data data;
290 data.func = quiesce_dummy;
292 data.flags = CSD_FLAG_WAIT;
294 for_each_cpu_mask(cpu, mask)
295 generic_exec_single(cpu, &data);
299 * smp_call_function_mask(): Run a function on a set of other CPUs.
300 * @mask: The set of cpus to run on.
301 * @func: The function to run. This must be fast and non-blocking.
302 * @info: An arbitrary pointer to pass to the function.
303 * @wait: If true, wait (atomically) until function has completed on other CPUs.
305 * Returns 0 on success, else a negative status code.
307 * If @wait is true, then returns once @func has returned. Note that @wait
308 * will be implicitly turned on in case of allocation failures, since
309 * we fall back to on-stack allocation.
311 * You must not call this function with disabled interrupts or from a
312 * hardware interrupt handler or from a bottom half handler. Preemption
313 * must be disabled when calling this function.
315 int smp_call_function_mask(cpumask_t mask, void (*func)(void *), void *info,
318 struct call_function_data d;
319 struct call_function_data *data = NULL;
320 cpumask_t allbutself;
325 /* Can deadlock when called with interrupts disabled */
326 WARN_ON(irqs_disabled());
328 cpu = smp_processor_id();
329 allbutself = cpu_online_map;
330 cpu_clear(cpu, allbutself);
331 cpus_and(mask, mask, allbutself);
332 num_cpus = cpus_weight(mask);
335 * If zero CPUs, return. If just a single CPU, turn this request
336 * into a targetted single call instead since it's faster.
340 else if (num_cpus == 1) {
341 cpu = first_cpu(mask);
342 return smp_call_function_single(cpu, func, info, wait);
345 data = kmalloc(sizeof(*data), GFP_ATOMIC);
347 data->csd.flags = CSD_FLAG_ALLOC;
349 data->csd.flags |= CSD_FLAG_WAIT;
352 data->csd.flags = CSD_FLAG_WAIT;
357 spin_lock_init(&data->lock);
358 data->csd.func = func;
359 data->csd.info = info;
360 data->refs = num_cpus;
361 data->cpumask = mask;
363 spin_lock_irqsave(&call_function_lock, flags);
364 list_add_tail_rcu(&data->csd.list, &call_function_queue);
365 spin_unlock_irqrestore(&call_function_lock, flags);
367 /* Send a message to all CPUs in the map */
368 arch_send_call_function_ipi(mask);
370 /* optionally wait for the CPUs to complete */
372 csd_flag_wait(&data->csd);
373 if (unlikely(slowpath))
374 smp_call_function_mask_quiesce_stack(allbutself);
379 EXPORT_SYMBOL(smp_call_function_mask);
382 * smp_call_function(): Run a function on all other CPUs.
383 * @func: The function to run. This must be fast and non-blocking.
384 * @info: An arbitrary pointer to pass to the function.
385 * @wait: If true, wait (atomically) until function has completed on other CPUs.
387 * Returns 0 on success, else a negative status code.
389 * If @wait is true, then returns once @func has returned; otherwise
390 * it returns just before the target cpu calls @func. In case of allocation
391 * failure, @wait will be implicitly turned on.
393 * You must not call this function with disabled interrupts or from a
394 * hardware interrupt handler or from a bottom half handler.
396 int smp_call_function(void (*func)(void *), void *info, int wait)
401 ret = smp_call_function_mask(cpu_online_map, func, info, wait);
405 EXPORT_SYMBOL(smp_call_function);
407 void ipi_call_lock(void)
409 spin_lock(&call_function_lock);
412 void ipi_call_unlock(void)
414 spin_unlock(&call_function_lock);
417 void ipi_call_lock_irq(void)
419 spin_lock_irq(&call_function_lock);
422 void ipi_call_unlock_irq(void)
424 spin_unlock_irq(&call_function_lock);