1 /*
2 * Generic helpers for smp ipi calls
3 *
4 * (C) Jens Axboe <jens.axboe@oracle.com> 2008
5 *
6 */
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
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);
17
18 enum {
19 CSD_FLAG_WAIT = 0x01,
20 CSD_FLAG_ALLOC = 0x02,
21 CSD_FLAG_LOCK = 0x04,
22 };
23
24 struct call_function_data {
25 struct call_single_data csd;
26 spinlock_t lock;
27 unsigned int refs;
28 struct rcu_head rcu_head;
29 unsigned long cpumask_bits[];
30 };
31
32 struct call_single_queue {
33 struct list_head list;
34 spinlock_t lock;
35 };
36
init_call_single_data(void)37 static int __cpuinit init_call_single_data(void)
38 {
39 int i;
40
41 for_each_possible_cpu(i) {
42 struct call_single_queue *q = &per_cpu(call_single_queue, i);
43
44 spin_lock_init(&q->lock);
45 INIT_LIST_HEAD(&q->list);
46 }
47 return 0;
48 }
49 early_initcall(init_call_single_data);
50
csd_flag_wait(struct call_single_data * data)51 static void csd_flag_wait(struct call_single_data *data)
52 {
53 /* Wait for response */
54 do {
55 if (!(data->flags & CSD_FLAG_WAIT))
56 break;
57 cpu_relax();
58 } while (1);
59 }
60
61 /*
62 * Insert a previously allocated call_single_data element for execution
63 * on the given CPU. data must already have ->func, ->info, and ->flags set.
64 */
generic_exec_single(int cpu,struct call_single_data * data)65 static void generic_exec_single(int cpu, struct call_single_data *data)
66 {
67 struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
68 int wait = data->flags & CSD_FLAG_WAIT, ipi;
69 unsigned long flags;
70
71 spin_lock_irqsave(&dst->lock, flags);
72 ipi = list_empty(&dst->list);
73 list_add_tail(&data->list, &dst->list);
74 spin_unlock_irqrestore(&dst->lock, flags);
75
76 /*
77 * Make the list addition visible before sending the ipi.
78 */
79 smp_mb();
80
81 if (ipi)
82 arch_send_call_function_single_ipi(cpu);
83
84 if (wait)
85 csd_flag_wait(data);
86 }
87
rcu_free_call_data(struct rcu_head * head)88 static void rcu_free_call_data(struct rcu_head *head)
89 {
90 struct call_function_data *data;
91
92 data = container_of(head, struct call_function_data, rcu_head);
93
94 kfree(data);
95 }
96
97 /*
98 * Invoked by arch to handle an IPI for call function. Must be called with
99 * interrupts disabled.
100 */
generic_smp_call_function_interrupt(void)101 void generic_smp_call_function_interrupt(void)
102 {
103 struct call_function_data *data;
104 int cpu = get_cpu();
105
106 /*
107 * It's ok to use list_for_each_rcu() here even though we may delete
108 * 'pos', since list_del_rcu() doesn't clear ->next
109 */
110 rcu_read_lock();
111 list_for_each_entry_rcu(data, &call_function_queue, csd.list) {
112 int refs;
113
114 if (!cpumask_test_cpu(cpu, to_cpumask(data->cpumask_bits)))
115 continue;
116
117 data->csd.func(data->csd.info);
118
119 spin_lock(&data->lock);
120 cpumask_clear_cpu(cpu, to_cpumask(data->cpumask_bits));
121 WARN_ON(data->refs == 0);
122 data->refs--;
123 refs = data->refs;
124 spin_unlock(&data->lock);
125
126 if (refs)
127 continue;
128
129 spin_lock(&call_function_lock);
130 list_del_rcu(&data->csd.list);
131 spin_unlock(&call_function_lock);
132
133 if (data->csd.flags & CSD_FLAG_WAIT) {
134 /*
135 * serialize stores to data with the flag clear
136 * and wakeup
137 */
138 smp_wmb();
139 data->csd.flags &= ~CSD_FLAG_WAIT;
140 }
141 if (data->csd.flags & CSD_FLAG_ALLOC)
142 call_rcu(&data->rcu_head, rcu_free_call_data);
143 }
144 rcu_read_unlock();
145
146 put_cpu();
147 }
148
149 /*
150 * Invoked by arch to handle an IPI for call function single. Must be called
151 * from the arch with interrupts disabled.
152 */
generic_smp_call_function_single_interrupt(void)153 void generic_smp_call_function_single_interrupt(void)
154 {
155 struct call_single_queue *q = &__get_cpu_var(call_single_queue);
156 LIST_HEAD(list);
157
158 /*
159 * Need to see other stores to list head for checking whether
160 * list is empty without holding q->lock
161 */
162 smp_read_barrier_depends();
163 while (!list_empty(&q->list)) {
164 unsigned int data_flags;
165
166 spin_lock(&q->lock);
167 list_replace_init(&q->list, &list);
168 spin_unlock(&q->lock);
169
170 while (!list_empty(&list)) {
171 struct call_single_data *data;
172
173 data = list_entry(list.next, struct call_single_data,
174 list);
175 list_del(&data->list);
176
177 /*
178 * 'data' can be invalid after this call if
179 * flags == 0 (when called through
180 * generic_exec_single(), so save them away before
181 * making the call.
182 */
183 data_flags = data->flags;
184
185 data->func(data->info);
186
187 if (data_flags & CSD_FLAG_WAIT) {
188 smp_wmb();
189 data->flags &= ~CSD_FLAG_WAIT;
190 } else if (data_flags & CSD_FLAG_LOCK) {
191 smp_wmb();
192 data->flags &= ~CSD_FLAG_LOCK;
193 } else if (data_flags & CSD_FLAG_ALLOC)
194 kfree(data);
195 }
196 /*
197 * See comment on outer loop
198 */
199 smp_read_barrier_depends();
200 }
201 }
202
203 static DEFINE_PER_CPU(struct call_single_data, csd_data);
204
205 /*
206 * smp_call_function_single - Run a function on a specific CPU
207 * @func: The function to run. This must be fast and non-blocking.
208 * @info: An arbitrary pointer to pass to the function.
209 * @wait: If true, wait until function has completed on other CPUs.
210 *
211 * Returns 0 on success, else a negative status code. Note that @wait
212 * will be implicitly turned on in case of allocation failures, since
213 * we fall back to on-stack allocation.
214 */
smp_call_function_single(int cpu,void (* func)(void * info),void * info,int wait)215 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
216 int wait)
217 {
218 struct call_single_data d;
219 unsigned long flags;
220 /* prevent preemption and reschedule on another processor,
221 as well as CPU removal */
222 int me = get_cpu();
223 int err = 0;
224
225 /* Can deadlock when called with interrupts disabled */
226 WARN_ON(irqs_disabled());
227
228 if (cpu == me) {
229 local_irq_save(flags);
230 func(info);
231 local_irq_restore(flags);
232 } else if ((unsigned)cpu < nr_cpu_ids && cpu_online(cpu)) {
233 struct call_single_data *data;
234
235 if (!wait) {
236 /*
237 * We are calling a function on a single CPU
238 * and we are not going to wait for it to finish.
239 * We first try to allocate the data, but if we
240 * fail, we fall back to use a per cpu data to pass
241 * the information to that CPU. Since all callers
242 * of this code will use the same data, we must
243 * synchronize the callers to prevent a new caller
244 * from corrupting the data before the callee
245 * can access it.
246 *
247 * The CSD_FLAG_LOCK is used to let us know when
248 * the IPI handler is done with the data.
249 * The first caller will set it, and the callee
250 * will clear it. The next caller must wait for
251 * it to clear before we set it again. This
252 * will make sure the callee is done with the
253 * data before a new caller will use it.
254 */
255 data = kmalloc(sizeof(*data), GFP_ATOMIC);
256 if (data)
257 data->flags = CSD_FLAG_ALLOC;
258 else {
259 data = &per_cpu(csd_data, me);
260 while (data->flags & CSD_FLAG_LOCK)
261 cpu_relax();
262 data->flags = CSD_FLAG_LOCK;
263 }
264 } else {
265 data = &d;
266 data->flags = CSD_FLAG_WAIT;
267 }
268
269 data->func = func;
270 data->info = info;
271 generic_exec_single(cpu, data);
272 } else {
273 err = -ENXIO; /* CPU not online */
274 }
275
276 put_cpu();
277 return err;
278 }
279 EXPORT_SYMBOL(smp_call_function_single);
280
281 /**
282 * __smp_call_function_single(): Run a function on another CPU
283 * @cpu: The CPU to run on.
284 * @data: Pre-allocated and setup data structure
285 *
286 * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
287 * data structure. Useful for embedding @data inside other structures, for
288 * instance.
289 *
290 */
__smp_call_function_single(int cpu,struct call_single_data * data)291 void __smp_call_function_single(int cpu, struct call_single_data *data)
292 {
293 /* Can deadlock when called with interrupts disabled */
294 WARN_ON((data->flags & CSD_FLAG_WAIT) && irqs_disabled());
295
296 generic_exec_single(cpu, data);
297 }
298
299 /* FIXME: Shim for archs using old arch_send_call_function_ipi API. */
300 #ifndef arch_send_call_function_ipi_mask
301 #define arch_send_call_function_ipi_mask(maskp) \
302 arch_send_call_function_ipi(*(maskp))
303 #endif
304
305 /**
306 * smp_call_function_many(): Run a function on a set of other CPUs.
307 * @mask: The set of cpus to run on (only runs on online subset).
308 * @func: The function to run. This must be fast and non-blocking.
309 * @info: An arbitrary pointer to pass to the function.
310 * @wait: If true, wait (atomically) until function has completed on other CPUs.
311 *
312 * If @wait is true, then returns once @func has returned. Note that @wait
313 * will be implicitly turned on in case of allocation failures, since
314 * we fall back to on-stack allocation.
315 *
316 * You must not call this function with disabled interrupts or from a
317 * hardware interrupt handler or from a bottom half handler. Preemption
318 * must be disabled when calling this function.
319 */
smp_call_function_many(const struct cpumask * mask,void (* func)(void *),void * info,bool wait)320 void smp_call_function_many(const struct cpumask *mask,
321 void (*func)(void *), void *info,
322 bool wait)
323 {
324 struct call_function_data *data;
325 unsigned long flags;
326 int cpu, next_cpu;
327
328 /* Can deadlock when called with interrupts disabled */
329 WARN_ON(irqs_disabled());
330
331 /* So, what's a CPU they want? Ignoring this one. */
332 cpu = cpumask_first_and(mask, cpu_online_mask);
333 if (cpu == smp_processor_id())
334 cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
335 /* No online cpus? We're done. */
336 if (cpu >= nr_cpu_ids)
337 return;
338
339 /* Do we have another CPU which isn't us? */
340 next_cpu = cpumask_next_and(cpu, mask, cpu_online_mask);
341 if (next_cpu == smp_processor_id())
342 next_cpu = cpumask_next_and(next_cpu, mask, cpu_online_mask);
343
344 /* Fastpath: do that cpu by itself. */
345 if (next_cpu >= nr_cpu_ids) {
346 smp_call_function_single(cpu, func, info, wait);
347 return;
348 }
349
350 data = kmalloc(sizeof(*data) + cpumask_size(), GFP_ATOMIC);
351 if (unlikely(!data)) {
352 /* Slow path. */
353 for_each_online_cpu(cpu) {
354 if (cpu == smp_processor_id())
355 continue;
356 if (cpumask_test_cpu(cpu, mask))
357 smp_call_function_single(cpu, func, info, wait);
358 }
359 return;
360 }
361
362 spin_lock_init(&data->lock);
363 data->csd.flags = CSD_FLAG_ALLOC;
364 if (wait)
365 data->csd.flags |= CSD_FLAG_WAIT;
366 data->csd.func = func;
367 data->csd.info = info;
368 cpumask_and(to_cpumask(data->cpumask_bits), mask, cpu_online_mask);
369 cpumask_clear_cpu(smp_processor_id(), to_cpumask(data->cpumask_bits));
370 data->refs = cpumask_weight(to_cpumask(data->cpumask_bits));
371
372 spin_lock_irqsave(&call_function_lock, flags);
373 list_add_tail_rcu(&data->csd.list, &call_function_queue);
374 spin_unlock_irqrestore(&call_function_lock, flags);
375
376 /*
377 * Make the list addition visible before sending the ipi.
378 */
379 smp_mb();
380
381 /* Send a message to all CPUs in the map */
382 arch_send_call_function_ipi_mask(to_cpumask(data->cpumask_bits));
383
384 /* optionally wait for the CPUs to complete */
385 if (wait)
386 csd_flag_wait(&data->csd);
387 }
388 EXPORT_SYMBOL(smp_call_function_many);
389
390 /**
391 * smp_call_function(): Run a function on all other CPUs.
392 * @func: The function to run. This must be fast and non-blocking.
393 * @info: An arbitrary pointer to pass to the function.
394 * @wait: If true, wait (atomically) until function has completed on other CPUs.
395 *
396 * Returns 0.
397 *
398 * If @wait is true, then returns once @func has returned; otherwise
399 * it returns just before the target cpu calls @func. In case of allocation
400 * failure, @wait will be implicitly turned on.
401 *
402 * You must not call this function with disabled interrupts or from a
403 * hardware interrupt handler or from a bottom half handler.
404 */
smp_call_function(void (* func)(void *),void * info,int wait)405 int smp_call_function(void (*func)(void *), void *info, int wait)
406 {
407 preempt_disable();
408 smp_call_function_many(cpu_online_mask, func, info, wait);
409 preempt_enable();
410 return 0;
411 }
412 EXPORT_SYMBOL(smp_call_function);
413
ipi_call_lock(void)414 void ipi_call_lock(void)
415 {
416 spin_lock(&call_function_lock);
417 }
418
ipi_call_unlock(void)419 void ipi_call_unlock(void)
420 {
421 spin_unlock(&call_function_lock);
422 }
423
ipi_call_lock_irq(void)424 void ipi_call_lock_irq(void)
425 {
426 spin_lock_irq(&call_function_lock);
427 }
428
ipi_call_unlock_irq(void)429 void ipi_call_unlock_irq(void)
430 {
431 spin_unlock_irq(&call_function_lock);
432 }
433