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 if (data->csd.flags & CSD_FLAG_ALLOC)
140 call_rcu(&data->rcu_head, rcu_free_call_data);
148 * Invoked by arch to handle an IPI for call function single. Must be called
149 * from the arch with interrupts disabled.
151 void generic_smp_call_function_single_interrupt(void)
153 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
157 * Need to see other stores to list head for checking whether
158 * list is empty without holding q->lock
161 while (!list_empty(&q->list)) {
162 unsigned int data_flags;
165 list_replace_init(&q->list, &list);
166 spin_unlock(&q->lock);
168 while (!list_empty(&list)) {
169 struct call_single_data *data;
171 data = list_entry(list.next, struct call_single_data,
173 list_del(&data->list);
176 * 'data' can be invalid after this call if
177 * flags == 0 (when called through
178 * generic_exec_single(), so save them away before
181 data_flags = data->flags;
183 data->func(data->info);
185 if (data_flags & CSD_FLAG_WAIT) {
187 data->flags &= ~CSD_FLAG_WAIT;
188 } else if (data_flags & CSD_FLAG_ALLOC)
192 * See comment on outer loop
199 * smp_call_function_single - Run a function on a specific CPU
200 * @func: The function to run. This must be fast and non-blocking.
201 * @info: An arbitrary pointer to pass to the function.
202 * @wait: If true, wait until function has completed on other CPUs.
204 * Returns 0 on success, else a negative status code. Note that @wait
205 * will be implicitly turned on in case of allocation failures, since
206 * we fall back to on-stack allocation.
208 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
211 struct call_single_data d;
213 /* prevent preemption and reschedule on another processor,
214 as well as CPU removal */
218 /* Can deadlock when called with interrupts disabled */
219 WARN_ON(irqs_disabled());
222 local_irq_save(flags);
224 local_irq_restore(flags);
225 } else if ((unsigned)cpu < NR_CPUS && cpu_online(cpu)) {
226 struct call_single_data *data = NULL;
229 data = kmalloc(sizeof(*data), GFP_ATOMIC);
231 data->flags = CSD_FLAG_ALLOC;
235 data->flags = CSD_FLAG_WAIT;
240 generic_exec_single(cpu, data);
242 err = -ENXIO; /* CPU not online */
248 EXPORT_SYMBOL(smp_call_function_single);
251 * __smp_call_function_single(): Run a function on another CPU
252 * @cpu: The CPU to run on.
253 * @data: Pre-allocated and setup data structure
255 * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
256 * data structure. Useful for embedding @data inside other structures, for
260 void __smp_call_function_single(int cpu, struct call_single_data *data)
262 /* Can deadlock when called with interrupts disabled */
263 WARN_ON((data->flags & CSD_FLAG_WAIT) && irqs_disabled());
265 generic_exec_single(cpu, data);
269 static void quiesce_dummy(void *unused)
274 * Ensure stack based data used in call function mask is safe to free.
276 * This is needed by smp_call_function_mask when using on-stack data, because
277 * a single call function queue is shared by all CPUs, and any CPU may pick up
278 * the data item on the queue at any time before it is deleted. So we need to
279 * ensure that all CPUs have transitioned through a quiescent state after
282 * This is a very slow function, implemented by sending synchronous IPIs to
283 * all possible CPUs. For this reason, we have to alloc data rather than use
284 * stack based data even in the case of synchronous calls. The stack based
285 * data is then just used for deadlock/oom fallback which will be very rare.
287 * If a faster scheme can be made, we could go back to preferring stack based
288 * data -- the data allocation/free is non-zero cost.
290 static void smp_call_function_mask_quiesce_stack(cpumask_t mask)
292 struct call_single_data data;
295 data.func = quiesce_dummy;
298 for_each_cpu_mask(cpu, mask) {
299 data.flags = CSD_FLAG_WAIT;
300 generic_exec_single(cpu, &data);
305 * smp_call_function_mask(): Run a function on a set of other CPUs.
306 * @mask: The set of cpus to run on.
307 * @func: The function to run. This must be fast and non-blocking.
308 * @info: An arbitrary pointer to pass to the function.
309 * @wait: If true, wait (atomically) until function has completed on other CPUs.
311 * Returns 0 on success, else a negative status code.
313 * If @wait is true, then returns once @func has returned. Note that @wait
314 * will be implicitly turned on in case of allocation failures, since
315 * we fall back to on-stack allocation.
317 * You must not call this function with disabled interrupts or from a
318 * hardware interrupt handler or from a bottom half handler. Preemption
319 * must be disabled when calling this function.
321 int smp_call_function_mask(cpumask_t mask, void (*func)(void *), void *info,
324 struct call_function_data d;
325 struct call_function_data *data = NULL;
326 cpumask_t allbutself;
331 /* Can deadlock when called with interrupts disabled */
332 WARN_ON(irqs_disabled());
334 cpu = smp_processor_id();
335 allbutself = cpu_online_map;
336 cpu_clear(cpu, allbutself);
337 cpus_and(mask, mask, allbutself);
338 num_cpus = cpus_weight(mask);
341 * If zero CPUs, return. If just a single CPU, turn this request
342 * into a targetted single call instead since it's faster.
346 else if (num_cpus == 1) {
347 cpu = first_cpu(mask);
348 return smp_call_function_single(cpu, func, info, wait);
351 data = kmalloc(sizeof(*data), GFP_ATOMIC);
353 data->csd.flags = CSD_FLAG_ALLOC;
355 data->csd.flags |= CSD_FLAG_WAIT;
358 data->csd.flags = CSD_FLAG_WAIT;
363 spin_lock_init(&data->lock);
364 data->csd.func = func;
365 data->csd.info = info;
366 data->refs = num_cpus;
367 data->cpumask = mask;
369 spin_lock_irqsave(&call_function_lock, flags);
370 list_add_tail_rcu(&data->csd.list, &call_function_queue);
371 spin_unlock_irqrestore(&call_function_lock, flags);
373 /* Send a message to all CPUs in the map */
374 arch_send_call_function_ipi(mask);
376 /* optionally wait for the CPUs to complete */
378 csd_flag_wait(&data->csd);
379 if (unlikely(slowpath))
380 smp_call_function_mask_quiesce_stack(mask);
385 EXPORT_SYMBOL(smp_call_function_mask);
388 * smp_call_function(): Run a function on all other CPUs.
389 * @func: The function to run. This must be fast and non-blocking.
390 * @info: An arbitrary pointer to pass to the function.
391 * @wait: If true, wait (atomically) until function has completed on other CPUs.
393 * Returns 0 on success, else a negative status code.
395 * If @wait is true, then returns once @func has returned; otherwise
396 * it returns just before the target cpu calls @func. In case of allocation
397 * failure, @wait will be implicitly turned on.
399 * You must not call this function with disabled interrupts or from a
400 * hardware interrupt handler or from a bottom half handler.
402 int smp_call_function(void (*func)(void *), void *info, int wait)
407 ret = smp_call_function_mask(cpu_online_map, func, info, wait);
411 EXPORT_SYMBOL(smp_call_function);
413 void ipi_call_lock(void)
415 spin_lock(&call_function_lock);
418 void ipi_call_unlock(void)
420 spin_unlock(&call_function_lock);
423 void ipi_call_lock_irq(void)
425 spin_lock_irq(&call_function_lock);
428 void ipi_call_unlock_irq(void)
430 spin_unlock_irq(&call_function_lock);