blob: 6bd3f76fdf881ffc4973bc2d7a0d8346a3464bba [file] [log] [blame]
Davide Libenzie1ad7462007-05-10 22:23:19 -07001/*
2 * fs/eventfd.c
3 *
4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org>
5 *
6 */
7
8#include <linux/file.h>
9#include <linux/poll.h>
10#include <linux/init.h>
11#include <linux/fs.h>
12#include <linux/sched.h>
13#include <linux/kernel.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090014#include <linux/slab.h>
Davide Libenzie1ad7462007-05-10 22:23:19 -070015#include <linux/list.h>
16#include <linux/spinlock.h>
17#include <linux/anon_inodes.h>
Adrian Bunk7747cdb2008-02-06 01:36:49 -080018#include <linux/syscalls.h>
Rusty Russell57186072009-06-12 22:27:09 -060019#include <linux/module.h>
Davide Libenzi13389012009-06-30 11:41:11 -070020#include <linux/kref.h>
21#include <linux/eventfd.h>
Davide Libenzie1ad7462007-05-10 22:23:19 -070022
23struct eventfd_ctx {
Davide Libenzi13389012009-06-30 11:41:11 -070024 struct kref kref;
Davide Libenzie1ad7462007-05-10 22:23:19 -070025 wait_queue_head_t wqh;
26 /*
27 * Every time that a write(2) is performed on an eventfd, the
28 * value of the __u64 being written is added to "count" and a
29 * wakeup is performed on "wqh". A read(2) will return the "count"
30 * value to userspace, and will reset "count" to zero. The kernel
Davide Libenzi13389012009-06-30 11:41:11 -070031 * side eventfd_signal() also, adds to the "count" counter and
Davide Libenzie1ad7462007-05-10 22:23:19 -070032 * issue a wakeup.
33 */
34 __u64 count;
Davide Libenzibcd0b232009-03-31 15:24:18 -070035 unsigned int flags;
Davide Libenzie1ad7462007-05-10 22:23:19 -070036};
37
Davide Libenzi13389012009-06-30 11:41:11 -070038/**
39 * eventfd_signal - Adds @n to the eventfd counter.
40 * @ctx: [in] Pointer to the eventfd context.
41 * @n: [in] Value of the counter to be added to the eventfd internal counter.
42 * The value cannot be negative.
43 *
44 * This function is supposed to be called by the kernel in paths that do not
45 * allow sleeping. In this function we allow the counter to reach the ULLONG_MAX
46 * value, and we signal this as overflow condition by returining a POLLERR
47 * to poll(2).
48 *
49 * Returns @n in case of success, a non-negative number lower than @n in case
50 * of overflow, or the following error codes:
51 *
52 * -EINVAL : The value of @n is negative.
Davide Libenzie1ad7462007-05-10 22:23:19 -070053 */
Davide Libenzi13389012009-06-30 11:41:11 -070054int eventfd_signal(struct eventfd_ctx *ctx, int n)
Davide Libenzie1ad7462007-05-10 22:23:19 -070055{
Davide Libenzie1ad7462007-05-10 22:23:19 -070056 unsigned long flags;
57
58 if (n < 0)
59 return -EINVAL;
Davide Libenzid48eb232007-05-18 12:02:33 -070060 spin_lock_irqsave(&ctx->wqh.lock, flags);
Davide Libenzie1ad7462007-05-10 22:23:19 -070061 if (ULLONG_MAX - ctx->count < n)
62 n = (int) (ULLONG_MAX - ctx->count);
63 ctx->count += n;
64 if (waitqueue_active(&ctx->wqh))
Davide Libenzi39510882009-03-31 15:24:23 -070065 wake_up_locked_poll(&ctx->wqh, POLLIN);
Davide Libenzid48eb232007-05-18 12:02:33 -070066 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
Davide Libenzie1ad7462007-05-10 22:23:19 -070067
68 return n;
69}
Rusty Russell57186072009-06-12 22:27:09 -060070EXPORT_SYMBOL_GPL(eventfd_signal);
Davide Libenzie1ad7462007-05-10 22:23:19 -070071
Davide Libenzi562787a2009-09-22 16:43:57 -070072static void eventfd_free_ctx(struct eventfd_ctx *ctx)
73{
74 kfree(ctx);
75}
76
Davide Libenzi13389012009-06-30 11:41:11 -070077static void eventfd_free(struct kref *kref)
78{
79 struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref);
80
Davide Libenzi562787a2009-09-22 16:43:57 -070081 eventfd_free_ctx(ctx);
Davide Libenzi13389012009-06-30 11:41:11 -070082}
83
84/**
85 * eventfd_ctx_get - Acquires a reference to the internal eventfd context.
86 * @ctx: [in] Pointer to the eventfd context.
87 *
88 * Returns: In case of success, returns a pointer to the eventfd context.
89 */
90struct eventfd_ctx *eventfd_ctx_get(struct eventfd_ctx *ctx)
91{
92 kref_get(&ctx->kref);
93 return ctx;
94}
95EXPORT_SYMBOL_GPL(eventfd_ctx_get);
96
97/**
98 * eventfd_ctx_put - Releases a reference to the internal eventfd context.
99 * @ctx: [in] Pointer to eventfd context.
100 *
101 * The eventfd context reference must have been previously acquired either
102 * with eventfd_ctx_get() or eventfd_ctx_fdget()).
103 */
104void eventfd_ctx_put(struct eventfd_ctx *ctx)
105{
106 kref_put(&ctx->kref, eventfd_free);
107}
108EXPORT_SYMBOL_GPL(eventfd_ctx_put);
109
Davide Libenzie1ad7462007-05-10 22:23:19 -0700110static int eventfd_release(struct inode *inode, struct file *file)
111{
Davide Libenzi13389012009-06-30 11:41:11 -0700112 struct eventfd_ctx *ctx = file->private_data;
113
114 wake_up_poll(&ctx->wqh, POLLHUP);
115 eventfd_ctx_put(ctx);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700116 return 0;
117}
118
119static unsigned int eventfd_poll(struct file *file, poll_table *wait)
120{
121 struct eventfd_ctx *ctx = file->private_data;
122 unsigned int events = 0;
123 unsigned long flags;
124
125 poll_wait(file, &ctx->wqh, wait);
126
Davide Libenzid48eb232007-05-18 12:02:33 -0700127 spin_lock_irqsave(&ctx->wqh.lock, flags);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700128 if (ctx->count > 0)
129 events |= POLLIN;
130 if (ctx->count == ULLONG_MAX)
131 events |= POLLERR;
132 if (ULLONG_MAX - 1 > ctx->count)
133 events |= POLLOUT;
Davide Libenzid48eb232007-05-18 12:02:33 -0700134 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700135
136 return events;
137}
138
Davide Libenzicb289d62010-01-13 09:34:36 -0800139static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
Davide Libenzie1ad7462007-05-10 22:23:19 -0700140{
Davide Libenzicb289d62010-01-13 09:34:36 -0800141 *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
142 ctx->count -= *cnt;
143}
144
145/**
146 * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue.
147 * @ctx: [in] Pointer to eventfd context.
148 * @wait: [in] Wait queue to be removed.
149 * @cnt: [out] Pointer to the 64bit conter value.
150 *
151 * Returns zero if successful, or the following error codes:
152 *
153 * -EAGAIN : The operation would have blocked.
154 *
155 * This is used to atomically remove a wait queue entry from the eventfd wait
156 * queue head, and read/reset the counter value.
157 */
158int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_t *wait,
159 __u64 *cnt)
160{
161 unsigned long flags;
162
163 spin_lock_irqsave(&ctx->wqh.lock, flags);
164 eventfd_ctx_do_read(ctx, cnt);
165 __remove_wait_queue(&ctx->wqh, wait);
166 if (*cnt != 0 && waitqueue_active(&ctx->wqh))
167 wake_up_locked_poll(&ctx->wqh, POLLOUT);
168 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
169
170 return *cnt != 0 ? 0 : -EAGAIN;
171}
172EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue);
173
174/**
175 * eventfd_ctx_read - Reads the eventfd counter or wait if it is zero.
176 * @ctx: [in] Pointer to eventfd context.
177 * @no_wait: [in] Different from zero if the operation should not block.
178 * @cnt: [out] Pointer to the 64bit conter value.
179 *
180 * Returns zero if successful, or the following error codes:
181 *
182 * -EAGAIN : The operation would have blocked but @no_wait was nonzero.
183 * -ERESTARTSYS : A signal interrupted the wait operation.
184 *
185 * If @no_wait is zero, the function might sleep until the eventfd internal
186 * counter becomes greater than zero.
187 */
188ssize_t eventfd_ctx_read(struct eventfd_ctx *ctx, int no_wait, __u64 *cnt)
189{
Davide Libenzie1ad7462007-05-10 22:23:19 -0700190 ssize_t res;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700191 DECLARE_WAITQUEUE(wait, current);
192
Davide Libenzid48eb232007-05-18 12:02:33 -0700193 spin_lock_irq(&ctx->wqh.lock);
Davide Libenzicb289d62010-01-13 09:34:36 -0800194 *cnt = 0;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700195 res = -EAGAIN;
Davide Libenzibcd0b232009-03-31 15:24:18 -0700196 if (ctx->count > 0)
Davide Libenzicb289d62010-01-13 09:34:36 -0800197 res = 0;
198 else if (!no_wait) {
Davide Libenzie1ad7462007-05-10 22:23:19 -0700199 __add_wait_queue(&ctx->wqh, &wait);
Davide Libenzicb289d62010-01-13 09:34:36 -0800200 for (;;) {
Davide Libenzie1ad7462007-05-10 22:23:19 -0700201 set_current_state(TASK_INTERRUPTIBLE);
202 if (ctx->count > 0) {
Davide Libenzicb289d62010-01-13 09:34:36 -0800203 res = 0;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700204 break;
205 }
206 if (signal_pending(current)) {
207 res = -ERESTARTSYS;
208 break;
209 }
Davide Libenzid48eb232007-05-18 12:02:33 -0700210 spin_unlock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700211 schedule();
Davide Libenzid48eb232007-05-18 12:02:33 -0700212 spin_lock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700213 }
214 __remove_wait_queue(&ctx->wqh, &wait);
215 __set_current_state(TASK_RUNNING);
216 }
Davide Libenzicb289d62010-01-13 09:34:36 -0800217 if (likely(res == 0)) {
218 eventfd_ctx_do_read(ctx, cnt);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700219 if (waitqueue_active(&ctx->wqh))
Davide Libenzi39510882009-03-31 15:24:23 -0700220 wake_up_locked_poll(&ctx->wqh, POLLOUT);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700221 }
Davide Libenzid48eb232007-05-18 12:02:33 -0700222 spin_unlock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700223
224 return res;
225}
Davide Libenzicb289d62010-01-13 09:34:36 -0800226EXPORT_SYMBOL_GPL(eventfd_ctx_read);
227
228static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count,
229 loff_t *ppos)
230{
231 struct eventfd_ctx *ctx = file->private_data;
232 ssize_t res;
233 __u64 cnt;
234
235 if (count < sizeof(cnt))
236 return -EINVAL;
237 res = eventfd_ctx_read(ctx, file->f_flags & O_NONBLOCK, &cnt);
238 if (res < 0)
239 return res;
240
241 return put_user(cnt, (__u64 __user *) buf) ? -EFAULT : sizeof(cnt);
242}
Davide Libenzie1ad7462007-05-10 22:23:19 -0700243
244static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
245 loff_t *ppos)
246{
247 struct eventfd_ctx *ctx = file->private_data;
248 ssize_t res;
249 __u64 ucnt;
250 DECLARE_WAITQUEUE(wait, current);
251
252 if (count < sizeof(ucnt))
253 return -EINVAL;
254 if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
255 return -EFAULT;
256 if (ucnt == ULLONG_MAX)
257 return -EINVAL;
Davide Libenzid48eb232007-05-18 12:02:33 -0700258 spin_lock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700259 res = -EAGAIN;
260 if (ULLONG_MAX - ctx->count > ucnt)
261 res = sizeof(ucnt);
262 else if (!(file->f_flags & O_NONBLOCK)) {
263 __add_wait_queue(&ctx->wqh, &wait);
264 for (res = 0;;) {
265 set_current_state(TASK_INTERRUPTIBLE);
266 if (ULLONG_MAX - ctx->count > ucnt) {
267 res = sizeof(ucnt);
268 break;
269 }
270 if (signal_pending(current)) {
271 res = -ERESTARTSYS;
272 break;
273 }
Davide Libenzid48eb232007-05-18 12:02:33 -0700274 spin_unlock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700275 schedule();
Davide Libenzid48eb232007-05-18 12:02:33 -0700276 spin_lock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700277 }
278 __remove_wait_queue(&ctx->wqh, &wait);
279 __set_current_state(TASK_RUNNING);
280 }
Davide Libenzibcd0b232009-03-31 15:24:18 -0700281 if (likely(res > 0)) {
Davide Libenzie1ad7462007-05-10 22:23:19 -0700282 ctx->count += ucnt;
283 if (waitqueue_active(&ctx->wqh))
Davide Libenzi39510882009-03-31 15:24:23 -0700284 wake_up_locked_poll(&ctx->wqh, POLLIN);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700285 }
Davide Libenzid48eb232007-05-18 12:02:33 -0700286 spin_unlock_irq(&ctx->wqh.lock);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700287
288 return res;
289}
290
291static const struct file_operations eventfd_fops = {
292 .release = eventfd_release,
293 .poll = eventfd_poll,
294 .read = eventfd_read,
295 .write = eventfd_write,
296};
297
Davide Libenzi13389012009-06-30 11:41:11 -0700298/**
299 * eventfd_fget - Acquire a reference of an eventfd file descriptor.
300 * @fd: [in] Eventfd file descriptor.
301 *
302 * Returns a pointer to the eventfd file structure in case of success, or the
303 * following error pointer:
304 *
305 * -EBADF : Invalid @fd file descriptor.
306 * -EINVAL : The @fd file descriptor is not an eventfd file.
307 */
Davide Libenzie1ad7462007-05-10 22:23:19 -0700308struct file *eventfd_fget(int fd)
309{
310 struct file *file;
311
312 file = fget(fd);
313 if (!file)
314 return ERR_PTR(-EBADF);
315 if (file->f_op != &eventfd_fops) {
316 fput(file);
317 return ERR_PTR(-EINVAL);
318 }
319
320 return file;
321}
Rusty Russell57186072009-06-12 22:27:09 -0600322EXPORT_SYMBOL_GPL(eventfd_fget);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700323
Davide Libenzi13389012009-06-30 11:41:11 -0700324/**
325 * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
326 * @fd: [in] Eventfd file descriptor.
327 *
328 * Returns a pointer to the internal eventfd context, otherwise the error
329 * pointers returned by the following functions:
330 *
331 * eventfd_fget
332 */
333struct eventfd_ctx *eventfd_ctx_fdget(int fd)
334{
335 struct file *file;
336 struct eventfd_ctx *ctx;
337
338 file = eventfd_fget(fd);
339 if (IS_ERR(file))
340 return (struct eventfd_ctx *) file;
341 ctx = eventfd_ctx_get(file->private_data);
342 fput(file);
343
344 return ctx;
345}
346EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
347
348/**
349 * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
350 * @file: [in] Eventfd file pointer.
351 *
352 * Returns a pointer to the internal eventfd context, otherwise the error
353 * pointer:
354 *
355 * -EINVAL : The @fd file descriptor is not an eventfd file.
356 */
357struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
358{
359 if (file->f_op != &eventfd_fops)
360 return ERR_PTR(-EINVAL);
361
362 return eventfd_ctx_get(file->private_data);
363}
364EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
365
Davide Libenzi562787a2009-09-22 16:43:57 -0700366/**
367 * eventfd_file_create - Creates an eventfd file pointer.
368 * @count: Initial eventfd counter value.
369 * @flags: Flags for the eventfd file.
370 *
371 * This function creates an eventfd file pointer, w/out installing it into
372 * the fd table. This is useful when the eventfd file is used during the
373 * initialization of data structures that require extra setup after the eventfd
374 * creation. So the eventfd creation is split into the file pointer creation
375 * phase, and the file descriptor installation phase.
376 * In this way races with userspace closing the newly installed file descriptor
377 * can be avoided.
378 * Returns an eventfd file pointer, or a proper error pointer.
379 */
380struct file *eventfd_file_create(unsigned int count, int flags)
Davide Libenzie1ad7462007-05-10 22:23:19 -0700381{
Davide Libenzi562787a2009-09-22 16:43:57 -0700382 struct file *file;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700383 struct eventfd_ctx *ctx;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700384
Ulrich Dreppere38b36f2008-07-23 21:29:42 -0700385 /* Check the EFD_* constants for consistency. */
386 BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
387 BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
388
Davide Libenzibcd0b232009-03-31 15:24:18 -0700389 if (flags & ~EFD_FLAGS_SET)
Davide Libenzi562787a2009-09-22 16:43:57 -0700390 return ERR_PTR(-EINVAL);
Ulrich Drepperb087498e2008-07-23 21:29:25 -0700391
Davide Libenzie1ad7462007-05-10 22:23:19 -0700392 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
393 if (!ctx)
Davide Libenzi562787a2009-09-22 16:43:57 -0700394 return ERR_PTR(-ENOMEM);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700395
Davide Libenzi13389012009-06-30 11:41:11 -0700396 kref_init(&ctx->kref);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700397 init_waitqueue_head(&ctx->wqh);
Davide Libenzie1ad7462007-05-10 22:23:19 -0700398 ctx->count = count;
Davide Libenzibcd0b232009-03-31 15:24:18 -0700399 ctx->flags = flags;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700400
Davide Libenzi562787a2009-09-22 16:43:57 -0700401 file = anon_inode_getfile("[eventfd]", &eventfd_fops, ctx,
Roland Dreier628ff7c2009-12-18 09:41:24 -0800402 O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS));
Davide Libenzi562787a2009-09-22 16:43:57 -0700403 if (IS_ERR(file))
404 eventfd_free_ctx(ctx);
405
406 return file;
407}
408
409SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
410{
411 int fd, error;
412 struct file *file;
413
414 error = get_unused_fd_flags(flags & EFD_SHARED_FCNTL_FLAGS);
415 if (error < 0)
416 return error;
417 fd = error;
418
419 file = eventfd_file_create(count, flags);
420 if (IS_ERR(file)) {
421 error = PTR_ERR(file);
422 goto err_put_unused_fd;
423 }
424 fd_install(fd, file);
425
Al Viro2030a422008-02-23 06:46:49 -0500426 return fd;
Davide Libenzi562787a2009-09-22 16:43:57 -0700427
428err_put_unused_fd:
429 put_unused_fd(fd);
430
431 return error;
Davide Libenzie1ad7462007-05-10 22:23:19 -0700432}
433
Heiko Carstensd4e82042009-01-14 14:14:34 +0100434SYSCALL_DEFINE1(eventfd, unsigned int, count)
Ulrich Drepperb087498e2008-07-23 21:29:25 -0700435{
436 return sys_eventfd2(count, 0);
437}
Davide Libenzibcd0b232009-03-31 15:24:18 -0700438