1 | // SPDX-License-Identifier: GPL-2.0-only |
2 | /* |
3 | * fs/eventfd.c |
4 | * |
5 | * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> |
6 | * |
7 | */ |
8 | |
9 | #include <linux/file.h> |
10 | #include <linux/poll.h> |
11 | #include <linux/init.h> |
12 | #include <linux/fs.h> |
13 | #include <linux/sched/signal.h> |
14 | #include <linux/kernel.h> |
15 | #include <linux/slab.h> |
16 | #include <linux/list.h> |
17 | #include <linux/spinlock.h> |
18 | #include <linux/anon_inodes.h> |
19 | #include <linux/syscalls.h> |
20 | #include <linux/export.h> |
21 | #include <linux/kref.h> |
22 | #include <linux/eventfd.h> |
23 | #include <linux/proc_fs.h> |
24 | #include <linux/seq_file.h> |
25 | #include <linux/idr.h> |
26 | #include <linux/uio.h> |
27 | |
28 | static DEFINE_IDA(eventfd_ida); |
29 | |
30 | struct eventfd_ctx { |
31 | struct kref kref; |
32 | wait_queue_head_t wqh; |
33 | /* |
34 | * Every time that a write(2) is performed on an eventfd, the |
35 | * value of the __u64 being written is added to "count" and a |
36 | * wakeup is performed on "wqh". If EFD_SEMAPHORE flag was not |
37 | * specified, a read(2) will return the "count" value to userspace, |
38 | * and will reset "count" to zero. The kernel side eventfd_signal() |
39 | * also, adds to the "count" counter and issue a wakeup. |
40 | */ |
41 | __u64 count; |
42 | unsigned int flags; |
43 | int id; |
44 | }; |
45 | |
46 | __u64 eventfd_signal_mask(struct eventfd_ctx *ctx, __u64 n, __poll_t mask) |
47 | { |
48 | unsigned long flags; |
49 | |
50 | /* |
51 | * Deadlock or stack overflow issues can happen if we recurse here |
52 | * through waitqueue wakeup handlers. If the caller users potentially |
53 | * nested waitqueues with custom wakeup handlers, then it should |
54 | * check eventfd_signal_allowed() before calling this function. If |
55 | * it returns false, the eventfd_signal() call should be deferred to a |
56 | * safe context. |
57 | */ |
58 | if (WARN_ON_ONCE(current->in_eventfd)) |
59 | return 0; |
60 | |
61 | spin_lock_irqsave(&ctx->wqh.lock, flags); |
62 | current->in_eventfd = 1; |
63 | if (ULLONG_MAX - ctx->count < n) |
64 | n = ULLONG_MAX - ctx->count; |
65 | ctx->count += n; |
66 | if (waitqueue_active(wq_head: &ctx->wqh)) |
67 | wake_up_locked_poll(&ctx->wqh, EPOLLIN | mask); |
68 | current->in_eventfd = 0; |
69 | spin_unlock_irqrestore(lock: &ctx->wqh.lock, flags); |
70 | |
71 | return n; |
72 | } |
73 | |
74 | /** |
75 | * eventfd_signal - Adds @n to the eventfd counter. |
76 | * @ctx: [in] Pointer to the eventfd context. |
77 | * @n: [in] Value of the counter to be added to the eventfd internal counter. |
78 | * The value cannot be negative. |
79 | * |
80 | * This function is supposed to be called by the kernel in paths that do not |
81 | * allow sleeping. In this function we allow the counter to reach the ULLONG_MAX |
82 | * value, and we signal this as overflow condition by returning a EPOLLERR |
83 | * to poll(2). |
84 | * |
85 | * Returns the amount by which the counter was incremented. This will be less |
86 | * than @n if the counter has overflowed. |
87 | */ |
88 | __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n) |
89 | { |
90 | return eventfd_signal_mask(ctx, n, mask: 0); |
91 | } |
92 | EXPORT_SYMBOL_GPL(eventfd_signal); |
93 | |
94 | static void eventfd_free_ctx(struct eventfd_ctx *ctx) |
95 | { |
96 | if (ctx->id >= 0) |
97 | ida_simple_remove(&eventfd_ida, ctx->id); |
98 | kfree(objp: ctx); |
99 | } |
100 | |
101 | static void eventfd_free(struct kref *kref) |
102 | { |
103 | struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref); |
104 | |
105 | eventfd_free_ctx(ctx); |
106 | } |
107 | |
108 | /** |
109 | * eventfd_ctx_put - Releases a reference to the internal eventfd context. |
110 | * @ctx: [in] Pointer to eventfd context. |
111 | * |
112 | * The eventfd context reference must have been previously acquired either |
113 | * with eventfd_ctx_fdget() or eventfd_ctx_fileget(). |
114 | */ |
115 | void eventfd_ctx_put(struct eventfd_ctx *ctx) |
116 | { |
117 | kref_put(kref: &ctx->kref, release: eventfd_free); |
118 | } |
119 | EXPORT_SYMBOL_GPL(eventfd_ctx_put); |
120 | |
121 | static int eventfd_release(struct inode *inode, struct file *file) |
122 | { |
123 | struct eventfd_ctx *ctx = file->private_data; |
124 | |
125 | wake_up_poll(&ctx->wqh, EPOLLHUP); |
126 | eventfd_ctx_put(ctx); |
127 | return 0; |
128 | } |
129 | |
130 | static __poll_t eventfd_poll(struct file *file, poll_table *wait) |
131 | { |
132 | struct eventfd_ctx *ctx = file->private_data; |
133 | __poll_t events = 0; |
134 | u64 count; |
135 | |
136 | poll_wait(filp: file, wait_address: &ctx->wqh, p: wait); |
137 | |
138 | /* |
139 | * All writes to ctx->count occur within ctx->wqh.lock. This read |
140 | * can be done outside ctx->wqh.lock because we know that poll_wait |
141 | * takes that lock (through add_wait_queue) if our caller will sleep. |
142 | * |
143 | * The read _can_ therefore seep into add_wait_queue's critical |
144 | * section, but cannot move above it! add_wait_queue's spin_lock acts |
145 | * as an acquire barrier and ensures that the read be ordered properly |
146 | * against the writes. The following CAN happen and is safe: |
147 | * |
148 | * poll write |
149 | * ----------------- ------------ |
150 | * lock ctx->wqh.lock (in poll_wait) |
151 | * count = ctx->count |
152 | * __add_wait_queue |
153 | * unlock ctx->wqh.lock |
154 | * lock ctx->qwh.lock |
155 | * ctx->count += n |
156 | * if (waitqueue_active) |
157 | * wake_up_locked_poll |
158 | * unlock ctx->qwh.lock |
159 | * eventfd_poll returns 0 |
160 | * |
161 | * but the following, which would miss a wakeup, cannot happen: |
162 | * |
163 | * poll write |
164 | * ----------------- ------------ |
165 | * count = ctx->count (INVALID!) |
166 | * lock ctx->qwh.lock |
167 | * ctx->count += n |
168 | * **waitqueue_active is false** |
169 | * **no wake_up_locked_poll!** |
170 | * unlock ctx->qwh.lock |
171 | * lock ctx->wqh.lock (in poll_wait) |
172 | * __add_wait_queue |
173 | * unlock ctx->wqh.lock |
174 | * eventfd_poll returns 0 |
175 | */ |
176 | count = READ_ONCE(ctx->count); |
177 | |
178 | if (count > 0) |
179 | events |= EPOLLIN; |
180 | if (count == ULLONG_MAX) |
181 | events |= EPOLLERR; |
182 | if (ULLONG_MAX - 1 > count) |
183 | events |= EPOLLOUT; |
184 | |
185 | return events; |
186 | } |
187 | |
188 | void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt) |
189 | { |
190 | lockdep_assert_held(&ctx->wqh.lock); |
191 | |
192 | *cnt = ((ctx->flags & EFD_SEMAPHORE) && ctx->count) ? 1 : ctx->count; |
193 | ctx->count -= *cnt; |
194 | } |
195 | EXPORT_SYMBOL_GPL(eventfd_ctx_do_read); |
196 | |
197 | /** |
198 | * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue. |
199 | * @ctx: [in] Pointer to eventfd context. |
200 | * @wait: [in] Wait queue to be removed. |
201 | * @cnt: [out] Pointer to the 64-bit counter value. |
202 | * |
203 | * Returns %0 if successful, or the following error codes: |
204 | * |
205 | * -EAGAIN : The operation would have blocked. |
206 | * |
207 | * This is used to atomically remove a wait queue entry from the eventfd wait |
208 | * queue head, and read/reset the counter value. |
209 | */ |
210 | int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *wait, |
211 | __u64 *cnt) |
212 | { |
213 | unsigned long flags; |
214 | |
215 | spin_lock_irqsave(&ctx->wqh.lock, flags); |
216 | eventfd_ctx_do_read(ctx, cnt); |
217 | __remove_wait_queue(wq_head: &ctx->wqh, wq_entry: wait); |
218 | if (*cnt != 0 && waitqueue_active(wq_head: &ctx->wqh)) |
219 | wake_up_locked_poll(&ctx->wqh, EPOLLOUT); |
220 | spin_unlock_irqrestore(lock: &ctx->wqh.lock, flags); |
221 | |
222 | return *cnt != 0 ? 0 : -EAGAIN; |
223 | } |
224 | EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue); |
225 | |
226 | static ssize_t eventfd_read(struct kiocb *iocb, struct iov_iter *to) |
227 | { |
228 | struct file *file = iocb->ki_filp; |
229 | struct eventfd_ctx *ctx = file->private_data; |
230 | __u64 ucnt = 0; |
231 | |
232 | if (iov_iter_count(i: to) < sizeof(ucnt)) |
233 | return -EINVAL; |
234 | spin_lock_irq(lock: &ctx->wqh.lock); |
235 | if (!ctx->count) { |
236 | if ((file->f_flags & O_NONBLOCK) || |
237 | (iocb->ki_flags & IOCB_NOWAIT)) { |
238 | spin_unlock_irq(lock: &ctx->wqh.lock); |
239 | return -EAGAIN; |
240 | } |
241 | |
242 | if (wait_event_interruptible_locked_irq(ctx->wqh, ctx->count)) { |
243 | spin_unlock_irq(lock: &ctx->wqh.lock); |
244 | return -ERESTARTSYS; |
245 | } |
246 | } |
247 | eventfd_ctx_do_read(ctx, &ucnt); |
248 | current->in_eventfd = 1; |
249 | if (waitqueue_active(wq_head: &ctx->wqh)) |
250 | wake_up_locked_poll(&ctx->wqh, EPOLLOUT); |
251 | current->in_eventfd = 0; |
252 | spin_unlock_irq(lock: &ctx->wqh.lock); |
253 | if (unlikely(copy_to_iter(&ucnt, sizeof(ucnt), to) != sizeof(ucnt))) |
254 | return -EFAULT; |
255 | |
256 | return sizeof(ucnt); |
257 | } |
258 | |
259 | static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count, |
260 | loff_t *ppos) |
261 | { |
262 | struct eventfd_ctx *ctx = file->private_data; |
263 | ssize_t res; |
264 | __u64 ucnt; |
265 | |
266 | if (count < sizeof(ucnt)) |
267 | return -EINVAL; |
268 | if (copy_from_user(to: &ucnt, from: buf, n: sizeof(ucnt))) |
269 | return -EFAULT; |
270 | if (ucnt == ULLONG_MAX) |
271 | return -EINVAL; |
272 | spin_lock_irq(lock: &ctx->wqh.lock); |
273 | res = -EAGAIN; |
274 | if (ULLONG_MAX - ctx->count > ucnt) |
275 | res = sizeof(ucnt); |
276 | else if (!(file->f_flags & O_NONBLOCK)) { |
277 | res = wait_event_interruptible_locked_irq(ctx->wqh, |
278 | ULLONG_MAX - ctx->count > ucnt); |
279 | if (!res) |
280 | res = sizeof(ucnt); |
281 | } |
282 | if (likely(res > 0)) { |
283 | ctx->count += ucnt; |
284 | current->in_eventfd = 1; |
285 | if (waitqueue_active(wq_head: &ctx->wqh)) |
286 | wake_up_locked_poll(&ctx->wqh, EPOLLIN); |
287 | current->in_eventfd = 0; |
288 | } |
289 | spin_unlock_irq(lock: &ctx->wqh.lock); |
290 | |
291 | return res; |
292 | } |
293 | |
294 | #ifdef CONFIG_PROC_FS |
295 | static void eventfd_show_fdinfo(struct seq_file *m, struct file *f) |
296 | { |
297 | struct eventfd_ctx *ctx = f->private_data; |
298 | |
299 | spin_lock_irq(lock: &ctx->wqh.lock); |
300 | seq_printf(m, fmt: "eventfd-count: %16llx\n" , |
301 | (unsigned long long)ctx->count); |
302 | spin_unlock_irq(lock: &ctx->wqh.lock); |
303 | seq_printf(m, fmt: "eventfd-id: %d\n" , ctx->id); |
304 | seq_printf(m, fmt: "eventfd-semaphore: %d\n" , |
305 | !!(ctx->flags & EFD_SEMAPHORE)); |
306 | } |
307 | #endif |
308 | |
309 | static const struct file_operations eventfd_fops = { |
310 | #ifdef CONFIG_PROC_FS |
311 | .show_fdinfo = eventfd_show_fdinfo, |
312 | #endif |
313 | .release = eventfd_release, |
314 | .poll = eventfd_poll, |
315 | .read_iter = eventfd_read, |
316 | .write = eventfd_write, |
317 | .llseek = noop_llseek, |
318 | }; |
319 | |
320 | /** |
321 | * eventfd_fget - Acquire a reference of an eventfd file descriptor. |
322 | * @fd: [in] Eventfd file descriptor. |
323 | * |
324 | * Returns a pointer to the eventfd file structure in case of success, or the |
325 | * following error pointer: |
326 | * |
327 | * -EBADF : Invalid @fd file descriptor. |
328 | * -EINVAL : The @fd file descriptor is not an eventfd file. |
329 | */ |
330 | struct file *eventfd_fget(int fd) |
331 | { |
332 | struct file *file; |
333 | |
334 | file = fget(fd); |
335 | if (!file) |
336 | return ERR_PTR(error: -EBADF); |
337 | if (file->f_op != &eventfd_fops) { |
338 | fput(file); |
339 | return ERR_PTR(error: -EINVAL); |
340 | } |
341 | |
342 | return file; |
343 | } |
344 | EXPORT_SYMBOL_GPL(eventfd_fget); |
345 | |
346 | /** |
347 | * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context. |
348 | * @fd: [in] Eventfd file descriptor. |
349 | * |
350 | * Returns a pointer to the internal eventfd context, otherwise the error |
351 | * pointers returned by the following functions: |
352 | * |
353 | * eventfd_fget |
354 | */ |
355 | struct eventfd_ctx *eventfd_ctx_fdget(int fd) |
356 | { |
357 | struct eventfd_ctx *ctx; |
358 | struct fd f = fdget(fd); |
359 | if (!f.file) |
360 | return ERR_PTR(error: -EBADF); |
361 | ctx = eventfd_ctx_fileget(file: f.file); |
362 | fdput(fd: f); |
363 | return ctx; |
364 | } |
365 | EXPORT_SYMBOL_GPL(eventfd_ctx_fdget); |
366 | |
367 | /** |
368 | * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context. |
369 | * @file: [in] Eventfd file pointer. |
370 | * |
371 | * Returns a pointer to the internal eventfd context, otherwise the error |
372 | * pointer: |
373 | * |
374 | * -EINVAL : The @fd file descriptor is not an eventfd file. |
375 | */ |
376 | struct eventfd_ctx *eventfd_ctx_fileget(struct file *file) |
377 | { |
378 | struct eventfd_ctx *ctx; |
379 | |
380 | if (file->f_op != &eventfd_fops) |
381 | return ERR_PTR(error: -EINVAL); |
382 | |
383 | ctx = file->private_data; |
384 | kref_get(kref: &ctx->kref); |
385 | return ctx; |
386 | } |
387 | EXPORT_SYMBOL_GPL(eventfd_ctx_fileget); |
388 | |
389 | static int do_eventfd(unsigned int count, int flags) |
390 | { |
391 | struct eventfd_ctx *ctx; |
392 | struct file *file; |
393 | int fd; |
394 | |
395 | /* Check the EFD_* constants for consistency. */ |
396 | BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC); |
397 | BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK); |
398 | |
399 | if (flags & ~EFD_FLAGS_SET) |
400 | return -EINVAL; |
401 | |
402 | ctx = kmalloc(size: sizeof(*ctx), GFP_KERNEL); |
403 | if (!ctx) |
404 | return -ENOMEM; |
405 | |
406 | kref_init(kref: &ctx->kref); |
407 | init_waitqueue_head(&ctx->wqh); |
408 | ctx->count = count; |
409 | ctx->flags = flags; |
410 | ctx->id = ida_simple_get(&eventfd_ida, 0, 0, GFP_KERNEL); |
411 | |
412 | flags &= EFD_SHARED_FCNTL_FLAGS; |
413 | flags |= O_RDWR; |
414 | fd = get_unused_fd_flags(flags); |
415 | if (fd < 0) |
416 | goto err; |
417 | |
418 | file = anon_inode_getfile(name: "[eventfd]" , fops: &eventfd_fops, priv: ctx, flags); |
419 | if (IS_ERR(ptr: file)) { |
420 | put_unused_fd(fd); |
421 | fd = PTR_ERR(ptr: file); |
422 | goto err; |
423 | } |
424 | |
425 | file->f_mode |= FMODE_NOWAIT; |
426 | fd_install(fd, file); |
427 | return fd; |
428 | err: |
429 | eventfd_free_ctx(ctx); |
430 | return fd; |
431 | } |
432 | |
433 | SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags) |
434 | { |
435 | return do_eventfd(count, flags); |
436 | } |
437 | |
438 | SYSCALL_DEFINE1(eventfd, unsigned int, count) |
439 | { |
440 | return do_eventfd(count, flags: 0); |
441 | } |
442 | |
443 | |