Loading...
1/*
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>
14#include <linux/slab.h>
15#include <linux/list.h>
16#include <linux/spinlock.h>
17#include <linux/anon_inodes.h>
18#include <linux/syscalls.h>
19#include <linux/export.h>
20#include <linux/kref.h>
21#include <linux/eventfd.h>
22#include <linux/proc_fs.h>
23#include <linux/seq_file.h>
24
25struct eventfd_ctx {
26 struct kref kref;
27 wait_queue_head_t wqh;
28 /*
29 * Every time that a write(2) is performed on an eventfd, the
30 * value of the __u64 being written is added to "count" and a
31 * wakeup is performed on "wqh". A read(2) will return the "count"
32 * value to userspace, and will reset "count" to zero. The kernel
33 * side eventfd_signal() also, adds to the "count" counter and
34 * issue a wakeup.
35 */
36 __u64 count;
37 unsigned int flags;
38};
39
40/**
41 * eventfd_signal - Adds @n to the eventfd counter.
42 * @ctx: [in] Pointer to the eventfd context.
43 * @n: [in] Value of the counter to be added to the eventfd internal counter.
44 * The value cannot be negative.
45 *
46 * This function is supposed to be called by the kernel in paths that do not
47 * allow sleeping. In this function we allow the counter to reach the ULLONG_MAX
48 * value, and we signal this as overflow condition by returning a POLLERR
49 * to poll(2).
50 *
51 * Returns the amount by which the counter was incremented. This will be less
52 * than @n if the counter has overflowed.
53 */
54__u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n)
55{
56 unsigned long flags;
57
58 spin_lock_irqsave(&ctx->wqh.lock, flags);
59 if (ULLONG_MAX - ctx->count < n)
60 n = ULLONG_MAX - ctx->count;
61 ctx->count += n;
62 if (waitqueue_active(&ctx->wqh))
63 wake_up_locked_poll(&ctx->wqh, POLLIN);
64 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
65
66 return n;
67}
68EXPORT_SYMBOL_GPL(eventfd_signal);
69
70static void eventfd_free_ctx(struct eventfd_ctx *ctx)
71{
72 kfree(ctx);
73}
74
75static void eventfd_free(struct kref *kref)
76{
77 struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref);
78
79 eventfd_free_ctx(ctx);
80}
81
82/**
83 * eventfd_ctx_get - Acquires a reference to the internal eventfd context.
84 * @ctx: [in] Pointer to the eventfd context.
85 *
86 * Returns: In case of success, returns a pointer to the eventfd context.
87 */
88struct eventfd_ctx *eventfd_ctx_get(struct eventfd_ctx *ctx)
89{
90 kref_get(&ctx->kref);
91 return ctx;
92}
93EXPORT_SYMBOL_GPL(eventfd_ctx_get);
94
95/**
96 * eventfd_ctx_put - Releases a reference to the internal eventfd context.
97 * @ctx: [in] Pointer to eventfd context.
98 *
99 * The eventfd context reference must have been previously acquired either
100 * with eventfd_ctx_get() or eventfd_ctx_fdget().
101 */
102void eventfd_ctx_put(struct eventfd_ctx *ctx)
103{
104 kref_put(&ctx->kref, eventfd_free);
105}
106EXPORT_SYMBOL_GPL(eventfd_ctx_put);
107
108static int eventfd_release(struct inode *inode, struct file *file)
109{
110 struct eventfd_ctx *ctx = file->private_data;
111
112 wake_up_poll(&ctx->wqh, POLLHUP);
113 eventfd_ctx_put(ctx);
114 return 0;
115}
116
117static unsigned int eventfd_poll(struct file *file, poll_table *wait)
118{
119 struct eventfd_ctx *ctx = file->private_data;
120 unsigned int events = 0;
121 u64 count;
122
123 poll_wait(file, &ctx->wqh, wait);
124
125 /*
126 * All writes to ctx->count occur within ctx->wqh.lock. This read
127 * can be done outside ctx->wqh.lock because we know that poll_wait
128 * takes that lock (through add_wait_queue) if our caller will sleep.
129 *
130 * The read _can_ therefore seep into add_wait_queue's critical
131 * section, but cannot move above it! add_wait_queue's spin_lock acts
132 * as an acquire barrier and ensures that the read be ordered properly
133 * against the writes. The following CAN happen and is safe:
134 *
135 * poll write
136 * ----------------- ------------
137 * lock ctx->wqh.lock (in poll_wait)
138 * count = ctx->count
139 * __add_wait_queue
140 * unlock ctx->wqh.lock
141 * lock ctx->qwh.lock
142 * ctx->count += n
143 * if (waitqueue_active)
144 * wake_up_locked_poll
145 * unlock ctx->qwh.lock
146 * eventfd_poll returns 0
147 *
148 * but the following, which would miss a wakeup, cannot happen:
149 *
150 * poll write
151 * ----------------- ------------
152 * count = ctx->count (INVALID!)
153 * lock ctx->qwh.lock
154 * ctx->count += n
155 * **waitqueue_active is false**
156 * **no wake_up_locked_poll!**
157 * unlock ctx->qwh.lock
158 * lock ctx->wqh.lock (in poll_wait)
159 * __add_wait_queue
160 * unlock ctx->wqh.lock
161 * eventfd_poll returns 0
162 */
163 count = READ_ONCE(ctx->count);
164
165 if (count > 0)
166 events |= POLLIN;
167 if (count == ULLONG_MAX)
168 events |= POLLERR;
169 if (ULLONG_MAX - 1 > count)
170 events |= POLLOUT;
171
172 return events;
173}
174
175static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
176{
177 *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
178 ctx->count -= *cnt;
179}
180
181/**
182 * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue.
183 * @ctx: [in] Pointer to eventfd context.
184 * @wait: [in] Wait queue to be removed.
185 * @cnt: [out] Pointer to the 64-bit counter value.
186 *
187 * Returns %0 if successful, or the following error codes:
188 *
189 * -EAGAIN : The operation would have blocked.
190 *
191 * This is used to atomically remove a wait queue entry from the eventfd wait
192 * queue head, and read/reset the counter value.
193 */
194int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_t *wait,
195 __u64 *cnt)
196{
197 unsigned long flags;
198
199 spin_lock_irqsave(&ctx->wqh.lock, flags);
200 eventfd_ctx_do_read(ctx, cnt);
201 __remove_wait_queue(&ctx->wqh, wait);
202 if (*cnt != 0 && waitqueue_active(&ctx->wqh))
203 wake_up_locked_poll(&ctx->wqh, POLLOUT);
204 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
205
206 return *cnt != 0 ? 0 : -EAGAIN;
207}
208EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue);
209
210/**
211 * eventfd_ctx_read - Reads the eventfd counter or wait if it is zero.
212 * @ctx: [in] Pointer to eventfd context.
213 * @no_wait: [in] Different from zero if the operation should not block.
214 * @cnt: [out] Pointer to the 64-bit counter value.
215 *
216 * Returns %0 if successful, or the following error codes:
217 *
218 * -EAGAIN : The operation would have blocked but @no_wait was non-zero.
219 * -ERESTARTSYS : A signal interrupted the wait operation.
220 *
221 * If @no_wait is zero, the function might sleep until the eventfd internal
222 * counter becomes greater than zero.
223 */
224ssize_t eventfd_ctx_read(struct eventfd_ctx *ctx, int no_wait, __u64 *cnt)
225{
226 ssize_t res;
227 DECLARE_WAITQUEUE(wait, current);
228
229 spin_lock_irq(&ctx->wqh.lock);
230 *cnt = 0;
231 res = -EAGAIN;
232 if (ctx->count > 0)
233 res = 0;
234 else if (!no_wait) {
235 __add_wait_queue(&ctx->wqh, &wait);
236 for (;;) {
237 set_current_state(TASK_INTERRUPTIBLE);
238 if (ctx->count > 0) {
239 res = 0;
240 break;
241 }
242 if (signal_pending(current)) {
243 res = -ERESTARTSYS;
244 break;
245 }
246 spin_unlock_irq(&ctx->wqh.lock);
247 schedule();
248 spin_lock_irq(&ctx->wqh.lock);
249 }
250 __remove_wait_queue(&ctx->wqh, &wait);
251 __set_current_state(TASK_RUNNING);
252 }
253 if (likely(res == 0)) {
254 eventfd_ctx_do_read(ctx, cnt);
255 if (waitqueue_active(&ctx->wqh))
256 wake_up_locked_poll(&ctx->wqh, POLLOUT);
257 }
258 spin_unlock_irq(&ctx->wqh.lock);
259
260 return res;
261}
262EXPORT_SYMBOL_GPL(eventfd_ctx_read);
263
264static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count,
265 loff_t *ppos)
266{
267 struct eventfd_ctx *ctx = file->private_data;
268 ssize_t res;
269 __u64 cnt;
270
271 if (count < sizeof(cnt))
272 return -EINVAL;
273 res = eventfd_ctx_read(ctx, file->f_flags & O_NONBLOCK, &cnt);
274 if (res < 0)
275 return res;
276
277 return put_user(cnt, (__u64 __user *) buf) ? -EFAULT : sizeof(cnt);
278}
279
280static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
281 loff_t *ppos)
282{
283 struct eventfd_ctx *ctx = file->private_data;
284 ssize_t res;
285 __u64 ucnt;
286 DECLARE_WAITQUEUE(wait, current);
287
288 if (count < sizeof(ucnt))
289 return -EINVAL;
290 if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
291 return -EFAULT;
292 if (ucnt == ULLONG_MAX)
293 return -EINVAL;
294 spin_lock_irq(&ctx->wqh.lock);
295 res = -EAGAIN;
296 if (ULLONG_MAX - ctx->count > ucnt)
297 res = sizeof(ucnt);
298 else if (!(file->f_flags & O_NONBLOCK)) {
299 __add_wait_queue(&ctx->wqh, &wait);
300 for (res = 0;;) {
301 set_current_state(TASK_INTERRUPTIBLE);
302 if (ULLONG_MAX - ctx->count > ucnt) {
303 res = sizeof(ucnt);
304 break;
305 }
306 if (signal_pending(current)) {
307 res = -ERESTARTSYS;
308 break;
309 }
310 spin_unlock_irq(&ctx->wqh.lock);
311 schedule();
312 spin_lock_irq(&ctx->wqh.lock);
313 }
314 __remove_wait_queue(&ctx->wqh, &wait);
315 __set_current_state(TASK_RUNNING);
316 }
317 if (likely(res > 0)) {
318 ctx->count += ucnt;
319 if (waitqueue_active(&ctx->wqh))
320 wake_up_locked_poll(&ctx->wqh, POLLIN);
321 }
322 spin_unlock_irq(&ctx->wqh.lock);
323
324 return res;
325}
326
327#ifdef CONFIG_PROC_FS
328static void eventfd_show_fdinfo(struct seq_file *m, struct file *f)
329{
330 struct eventfd_ctx *ctx = f->private_data;
331
332 spin_lock_irq(&ctx->wqh.lock);
333 seq_printf(m, "eventfd-count: %16llx\n",
334 (unsigned long long)ctx->count);
335 spin_unlock_irq(&ctx->wqh.lock);
336}
337#endif
338
339static const struct file_operations eventfd_fops = {
340#ifdef CONFIG_PROC_FS
341 .show_fdinfo = eventfd_show_fdinfo,
342#endif
343 .release = eventfd_release,
344 .poll = eventfd_poll,
345 .read = eventfd_read,
346 .write = eventfd_write,
347 .llseek = noop_llseek,
348};
349
350/**
351 * eventfd_fget - Acquire a reference of an eventfd file descriptor.
352 * @fd: [in] Eventfd file descriptor.
353 *
354 * Returns a pointer to the eventfd file structure in case of success, or the
355 * following error pointer:
356 *
357 * -EBADF : Invalid @fd file descriptor.
358 * -EINVAL : The @fd file descriptor is not an eventfd file.
359 */
360struct file *eventfd_fget(int fd)
361{
362 struct file *file;
363
364 file = fget(fd);
365 if (!file)
366 return ERR_PTR(-EBADF);
367 if (file->f_op != &eventfd_fops) {
368 fput(file);
369 return ERR_PTR(-EINVAL);
370 }
371
372 return file;
373}
374EXPORT_SYMBOL_GPL(eventfd_fget);
375
376/**
377 * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
378 * @fd: [in] Eventfd file descriptor.
379 *
380 * Returns a pointer to the internal eventfd context, otherwise the error
381 * pointers returned by the following functions:
382 *
383 * eventfd_fget
384 */
385struct eventfd_ctx *eventfd_ctx_fdget(int fd)
386{
387 struct eventfd_ctx *ctx;
388 struct fd f = fdget(fd);
389 if (!f.file)
390 return ERR_PTR(-EBADF);
391 ctx = eventfd_ctx_fileget(f.file);
392 fdput(f);
393 return ctx;
394}
395EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
396
397/**
398 * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
399 * @file: [in] Eventfd file pointer.
400 *
401 * Returns a pointer to the internal eventfd context, otherwise the error
402 * pointer:
403 *
404 * -EINVAL : The @fd file descriptor is not an eventfd file.
405 */
406struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
407{
408 if (file->f_op != &eventfd_fops)
409 return ERR_PTR(-EINVAL);
410
411 return eventfd_ctx_get(file->private_data);
412}
413EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
414
415/**
416 * eventfd_file_create - Creates an eventfd file pointer.
417 * @count: Initial eventfd counter value.
418 * @flags: Flags for the eventfd file.
419 *
420 * This function creates an eventfd file pointer, w/out installing it into
421 * the fd table. This is useful when the eventfd file is used during the
422 * initialization of data structures that require extra setup after the eventfd
423 * creation. So the eventfd creation is split into the file pointer creation
424 * phase, and the file descriptor installation phase.
425 * In this way races with userspace closing the newly installed file descriptor
426 * can be avoided.
427 * Returns an eventfd file pointer, or a proper error pointer.
428 */
429struct file *eventfd_file_create(unsigned int count, int flags)
430{
431 struct file *file;
432 struct eventfd_ctx *ctx;
433
434 /* Check the EFD_* constants for consistency. */
435 BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
436 BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
437
438 if (flags & ~EFD_FLAGS_SET)
439 return ERR_PTR(-EINVAL);
440
441 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
442 if (!ctx)
443 return ERR_PTR(-ENOMEM);
444
445 kref_init(&ctx->kref);
446 init_waitqueue_head(&ctx->wqh);
447 ctx->count = count;
448 ctx->flags = flags;
449
450 file = anon_inode_getfile("[eventfd]", &eventfd_fops, ctx,
451 O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS));
452 if (IS_ERR(file))
453 eventfd_free_ctx(ctx);
454
455 return file;
456}
457
458SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
459{
460 int fd, error;
461 struct file *file;
462
463 error = get_unused_fd_flags(flags & EFD_SHARED_FCNTL_FLAGS);
464 if (error < 0)
465 return error;
466 fd = error;
467
468 file = eventfd_file_create(count, flags);
469 if (IS_ERR(file)) {
470 error = PTR_ERR(file);
471 goto err_put_unused_fd;
472 }
473 fd_install(fd, file);
474
475 return fd;
476
477err_put_unused_fd:
478 put_unused_fd(fd);
479
480 return error;
481}
482
483SYSCALL_DEFINE1(eventfd, unsigned int, count)
484{
485 return sys_eventfd2(count, 0);
486}
487
1/*
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>
14#include <linux/slab.h>
15#include <linux/list.h>
16#include <linux/spinlock.h>
17#include <linux/anon_inodes.h>
18#include <linux/syscalls.h>
19#include <linux/module.h>
20#include <linux/kref.h>
21#include <linux/eventfd.h>
22
23struct eventfd_ctx {
24 struct kref kref;
25 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
31 * side eventfd_signal() also, adds to the "count" counter and
32 * issue a wakeup.
33 */
34 __u64 count;
35 unsigned int flags;
36};
37
38/**
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.
53 */
54int eventfd_signal(struct eventfd_ctx *ctx, int n)
55{
56 unsigned long flags;
57
58 if (n < 0)
59 return -EINVAL;
60 spin_lock_irqsave(&ctx->wqh.lock, flags);
61 if (ULLONG_MAX - ctx->count < n)
62 n = (int) (ULLONG_MAX - ctx->count);
63 ctx->count += n;
64 if (waitqueue_active(&ctx->wqh))
65 wake_up_locked_poll(&ctx->wqh, POLLIN);
66 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
67
68 return n;
69}
70EXPORT_SYMBOL_GPL(eventfd_signal);
71
72static void eventfd_free_ctx(struct eventfd_ctx *ctx)
73{
74 kfree(ctx);
75}
76
77static void eventfd_free(struct kref *kref)
78{
79 struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref);
80
81 eventfd_free_ctx(ctx);
82}
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
110static int eventfd_release(struct inode *inode, struct file *file)
111{
112 struct eventfd_ctx *ctx = file->private_data;
113
114 wake_up_poll(&ctx->wqh, POLLHUP);
115 eventfd_ctx_put(ctx);
116 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
127 spin_lock_irqsave(&ctx->wqh.lock, flags);
128 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;
134 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
135
136 return events;
137}
138
139static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
140{
141 *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 64-bit counter value.
150 *
151 * Returns %0 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 64-bit counter value.
179 *
180 * Returns %0 if successful, or the following error codes:
181 *
182 * -EAGAIN : The operation would have blocked but @no_wait was non-zero.
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{
190 ssize_t res;
191 DECLARE_WAITQUEUE(wait, current);
192
193 spin_lock_irq(&ctx->wqh.lock);
194 *cnt = 0;
195 res = -EAGAIN;
196 if (ctx->count > 0)
197 res = 0;
198 else if (!no_wait) {
199 __add_wait_queue(&ctx->wqh, &wait);
200 for (;;) {
201 set_current_state(TASK_INTERRUPTIBLE);
202 if (ctx->count > 0) {
203 res = 0;
204 break;
205 }
206 if (signal_pending(current)) {
207 res = -ERESTARTSYS;
208 break;
209 }
210 spin_unlock_irq(&ctx->wqh.lock);
211 schedule();
212 spin_lock_irq(&ctx->wqh.lock);
213 }
214 __remove_wait_queue(&ctx->wqh, &wait);
215 __set_current_state(TASK_RUNNING);
216 }
217 if (likely(res == 0)) {
218 eventfd_ctx_do_read(ctx, cnt);
219 if (waitqueue_active(&ctx->wqh))
220 wake_up_locked_poll(&ctx->wqh, POLLOUT);
221 }
222 spin_unlock_irq(&ctx->wqh.lock);
223
224 return res;
225}
226EXPORT_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}
243
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;
258 spin_lock_irq(&ctx->wqh.lock);
259 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 }
274 spin_unlock_irq(&ctx->wqh.lock);
275 schedule();
276 spin_lock_irq(&ctx->wqh.lock);
277 }
278 __remove_wait_queue(&ctx->wqh, &wait);
279 __set_current_state(TASK_RUNNING);
280 }
281 if (likely(res > 0)) {
282 ctx->count += ucnt;
283 if (waitqueue_active(&ctx->wqh))
284 wake_up_locked_poll(&ctx->wqh, POLLIN);
285 }
286 spin_unlock_irq(&ctx->wqh.lock);
287
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 .llseek = noop_llseek,
297};
298
299/**
300 * eventfd_fget - Acquire a reference of an eventfd file descriptor.
301 * @fd: [in] Eventfd file descriptor.
302 *
303 * Returns a pointer to the eventfd file structure in case of success, or the
304 * following error pointer:
305 *
306 * -EBADF : Invalid @fd file descriptor.
307 * -EINVAL : The @fd file descriptor is not an eventfd file.
308 */
309struct file *eventfd_fget(int fd)
310{
311 struct file *file;
312
313 file = fget(fd);
314 if (!file)
315 return ERR_PTR(-EBADF);
316 if (file->f_op != &eventfd_fops) {
317 fput(file);
318 return ERR_PTR(-EINVAL);
319 }
320
321 return file;
322}
323EXPORT_SYMBOL_GPL(eventfd_fget);
324
325/**
326 * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
327 * @fd: [in] Eventfd file descriptor.
328 *
329 * Returns a pointer to the internal eventfd context, otherwise the error
330 * pointers returned by the following functions:
331 *
332 * eventfd_fget
333 */
334struct eventfd_ctx *eventfd_ctx_fdget(int fd)
335{
336 struct file *file;
337 struct eventfd_ctx *ctx;
338
339 file = eventfd_fget(fd);
340 if (IS_ERR(file))
341 return (struct eventfd_ctx *) file;
342 ctx = eventfd_ctx_get(file->private_data);
343 fput(file);
344
345 return ctx;
346}
347EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
348
349/**
350 * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
351 * @file: [in] Eventfd file pointer.
352 *
353 * Returns a pointer to the internal eventfd context, otherwise the error
354 * pointer:
355 *
356 * -EINVAL : The @fd file descriptor is not an eventfd file.
357 */
358struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
359{
360 if (file->f_op != &eventfd_fops)
361 return ERR_PTR(-EINVAL);
362
363 return eventfd_ctx_get(file->private_data);
364}
365EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
366
367/**
368 * eventfd_file_create - Creates an eventfd file pointer.
369 * @count: Initial eventfd counter value.
370 * @flags: Flags for the eventfd file.
371 *
372 * This function creates an eventfd file pointer, w/out installing it into
373 * the fd table. This is useful when the eventfd file is used during the
374 * initialization of data structures that require extra setup after the eventfd
375 * creation. So the eventfd creation is split into the file pointer creation
376 * phase, and the file descriptor installation phase.
377 * In this way races with userspace closing the newly installed file descriptor
378 * can be avoided.
379 * Returns an eventfd file pointer, or a proper error pointer.
380 */
381struct file *eventfd_file_create(unsigned int count, int flags)
382{
383 struct file *file;
384 struct eventfd_ctx *ctx;
385
386 /* Check the EFD_* constants for consistency. */
387 BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
388 BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
389
390 if (flags & ~EFD_FLAGS_SET)
391 return ERR_PTR(-EINVAL);
392
393 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
394 if (!ctx)
395 return ERR_PTR(-ENOMEM);
396
397 kref_init(&ctx->kref);
398 init_waitqueue_head(&ctx->wqh);
399 ctx->count = count;
400 ctx->flags = flags;
401
402 file = anon_inode_getfile("[eventfd]", &eventfd_fops, ctx,
403 O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS));
404 if (IS_ERR(file))
405 eventfd_free_ctx(ctx);
406
407 return file;
408}
409
410SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
411{
412 int fd, error;
413 struct file *file;
414
415 error = get_unused_fd_flags(flags & EFD_SHARED_FCNTL_FLAGS);
416 if (error < 0)
417 return error;
418 fd = error;
419
420 file = eventfd_file_create(count, flags);
421 if (IS_ERR(file)) {
422 error = PTR_ERR(file);
423 goto err_put_unused_fd;
424 }
425 fd_install(fd, file);
426
427 return fd;
428
429err_put_unused_fd:
430 put_unused_fd(fd);
431
432 return error;
433}
434
435SYSCALL_DEFINE1(eventfd, unsigned int, count)
436{
437 return sys_eventfd2(count, 0);
438}
439