| 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/signal.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 | |
| 25 | struct 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 EPOLLERR |
| 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, EPOLLIN); |
| 64 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); |
| 65 | |
| 66 | return n; |
| 67 | } |
| 68 | EXPORT_SYMBOL_GPL(eventfd_signal); |
| 69 | |
| 70 | static void eventfd_free_ctx(struct eventfd_ctx *ctx) |
| 71 | { |
| 72 | kfree(ctx); |
| 73 | } |
| 74 | |
| 75 | static 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_put - Releases a reference to the internal eventfd context. |
| 84 | * @ctx: [in] Pointer to eventfd context. |
| 85 | * |
| 86 | * The eventfd context reference must have been previously acquired either |
| 87 | * with eventfd_ctx_fdget() or eventfd_ctx_fileget(). |
| 88 | */ |
| 89 | void eventfd_ctx_put(struct eventfd_ctx *ctx) |
| 90 | { |
| 91 | kref_put(&ctx->kref, eventfd_free); |
| 92 | } |
| 93 | EXPORT_SYMBOL_GPL(eventfd_ctx_put); |
| 94 | |
| 95 | static int eventfd_release(struct inode *inode, struct file *file) |
| 96 | { |
| 97 | struct eventfd_ctx *ctx = file->private_data; |
| 98 | |
| 99 | wake_up_poll(&ctx->wqh, EPOLLHUP); |
| 100 | eventfd_ctx_put(ctx); |
| 101 | return 0; |
| 102 | } |
| 103 | |
| 104 | static __poll_t eventfd_poll(struct file *file, poll_table *wait) |
| 105 | { |
| 106 | struct eventfd_ctx *ctx = file->private_data; |
| 107 | __poll_t events = 0; |
| 108 | u64 count; |
| 109 | |
| 110 | poll_wait(file, &ctx->wqh, wait); |
| 111 | |
| 112 | /* |
| 113 | * All writes to ctx->count occur within ctx->wqh.lock. This read |
| 114 | * can be done outside ctx->wqh.lock because we know that poll_wait |
| 115 | * takes that lock (through add_wait_queue) if our caller will sleep. |
| 116 | * |
| 117 | * The read _can_ therefore seep into add_wait_queue's critical |
| 118 | * section, but cannot move above it! add_wait_queue's spin_lock acts |
| 119 | * as an acquire barrier and ensures that the read be ordered properly |
| 120 | * against the writes. The following CAN happen and is safe: |
| 121 | * |
| 122 | * poll write |
| 123 | * ----------------- ------------ |
| 124 | * lock ctx->wqh.lock (in poll_wait) |
| 125 | * count = ctx->count |
| 126 | * __add_wait_queue |
| 127 | * unlock ctx->wqh.lock |
| 128 | * lock ctx->qwh.lock |
| 129 | * ctx->count += n |
| 130 | * if (waitqueue_active) |
| 131 | * wake_up_locked_poll |
| 132 | * unlock ctx->qwh.lock |
| 133 | * eventfd_poll returns 0 |
| 134 | * |
| 135 | * but the following, which would miss a wakeup, cannot happen: |
| 136 | * |
| 137 | * poll write |
| 138 | * ----------------- ------------ |
| 139 | * count = ctx->count (INVALID!) |
| 140 | * lock ctx->qwh.lock |
| 141 | * ctx->count += n |
| 142 | * **waitqueue_active is false** |
| 143 | * **no wake_up_locked_poll!** |
| 144 | * unlock ctx->qwh.lock |
| 145 | * lock ctx->wqh.lock (in poll_wait) |
| 146 | * __add_wait_queue |
| 147 | * unlock ctx->wqh.lock |
| 148 | * eventfd_poll returns 0 |
| 149 | */ |
| 150 | count = READ_ONCE(ctx->count); |
| 151 | |
| 152 | if (count > 0) |
| 153 | events |= EPOLLIN; |
| 154 | if (count == ULLONG_MAX) |
| 155 | events |= EPOLLERR; |
| 156 | if (ULLONG_MAX - 1 > count) |
| 157 | events |= EPOLLOUT; |
| 158 | |
| 159 | return events; |
| 160 | } |
| 161 | |
| 162 | static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt) |
| 163 | { |
| 164 | *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count; |
| 165 | ctx->count -= *cnt; |
| 166 | } |
| 167 | |
| 168 | /** |
| 169 | * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue. |
| 170 | * @ctx: [in] Pointer to eventfd context. |
| 171 | * @wait: [in] Wait queue to be removed. |
| 172 | * @cnt: [out] Pointer to the 64-bit counter value. |
| 173 | * |
| 174 | * Returns %0 if successful, or the following error codes: |
| 175 | * |
| 176 | * -EAGAIN : The operation would have blocked. |
| 177 | * |
| 178 | * This is used to atomically remove a wait queue entry from the eventfd wait |
| 179 | * queue head, and read/reset the counter value. |
| 180 | */ |
| 181 | int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *wait, |
| 182 | __u64 *cnt) |
| 183 | { |
| 184 | unsigned long flags; |
| 185 | |
| 186 | spin_lock_irqsave(&ctx->wqh.lock, flags); |
| 187 | eventfd_ctx_do_read(ctx, cnt); |
| 188 | __remove_wait_queue(&ctx->wqh, wait); |
| 189 | if (*cnt != 0 && waitqueue_active(&ctx->wqh)) |
| 190 | wake_up_locked_poll(&ctx->wqh, EPOLLOUT); |
| 191 | spin_unlock_irqrestore(&ctx->wqh.lock, flags); |
| 192 | |
| 193 | return *cnt != 0 ? 0 : -EAGAIN; |
| 194 | } |
| 195 | EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue); |
| 196 | |
| 197 | static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count, |
| 198 | loff_t *ppos) |
| 199 | { |
| 200 | struct eventfd_ctx *ctx = file->private_data; |
| 201 | ssize_t res; |
| 202 | __u64 ucnt = 0; |
| 203 | DECLARE_WAITQUEUE(wait, current); |
| 204 | |
| 205 | if (count < sizeof(ucnt)) |
| 206 | return -EINVAL; |
| 207 | |
| 208 | spin_lock_irq(&ctx->wqh.lock); |
| 209 | res = -EAGAIN; |
| 210 | if (ctx->count > 0) |
| 211 | res = sizeof(ucnt); |
| 212 | else if (!(file->f_flags & O_NONBLOCK)) { |
| 213 | __add_wait_queue(&ctx->wqh, &wait); |
| 214 | for (;;) { |
| 215 | set_current_state(TASK_INTERRUPTIBLE); |
| 216 | if (ctx->count > 0) { |
| 217 | res = sizeof(ucnt); |
| 218 | break; |
| 219 | } |
| 220 | if (signal_pending(current)) { |
| 221 | res = -ERESTARTSYS; |
| 222 | break; |
| 223 | } |
| 224 | spin_unlock_irq(&ctx->wqh.lock); |
| 225 | schedule(); |
| 226 | spin_lock_irq(&ctx->wqh.lock); |
| 227 | } |
| 228 | __remove_wait_queue(&ctx->wqh, &wait); |
| 229 | __set_current_state(TASK_RUNNING); |
| 230 | } |
| 231 | if (likely(res > 0)) { |
| 232 | eventfd_ctx_do_read(ctx, &ucnt); |
| 233 | if (waitqueue_active(&ctx->wqh)) |
| 234 | wake_up_locked_poll(&ctx->wqh, EPOLLOUT); |
| 235 | } |
| 236 | spin_unlock_irq(&ctx->wqh.lock); |
| 237 | |
| 238 | if (res > 0 && put_user(ucnt, (__u64 __user *)buf)) |
| 239 | return -EFAULT; |
| 240 | |
| 241 | return res; |
| 242 | } |
| 243 | |
| 244 | static 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, EPOLLIN); |
| 285 | } |
| 286 | spin_unlock_irq(&ctx->wqh.lock); |
| 287 | |
| 288 | return res; |
| 289 | } |
| 290 | |
| 291 | #ifdef CONFIG_PROC_FS |
| 292 | static void eventfd_show_fdinfo(struct seq_file *m, struct file *f) |
| 293 | { |
| 294 | struct eventfd_ctx *ctx = f->private_data; |
| 295 | |
| 296 | spin_lock_irq(&ctx->wqh.lock); |
| 297 | seq_printf(m, "eventfd-count: %16llx\n", |
| 298 | (unsigned long long)ctx->count); |
| 299 | spin_unlock_irq(&ctx->wqh.lock); |
| 300 | } |
| 301 | #endif |
| 302 | |
| 303 | static const struct file_operations eventfd_fops = { |
| 304 | #ifdef CONFIG_PROC_FS |
| 305 | .show_fdinfo = eventfd_show_fdinfo, |
| 306 | #endif |
| 307 | .release = eventfd_release, |
| 308 | .poll = eventfd_poll, |
| 309 | .read = eventfd_read, |
| 310 | .write = eventfd_write, |
| 311 | .llseek = noop_llseek, |
| 312 | }; |
| 313 | |
| 314 | /** |
| 315 | * eventfd_fget - Acquire a reference of an eventfd file descriptor. |
| 316 | * @fd: [in] Eventfd file descriptor. |
| 317 | * |
| 318 | * Returns a pointer to the eventfd file structure in case of success, or the |
| 319 | * following error pointer: |
| 320 | * |
| 321 | * -EBADF : Invalid @fd file descriptor. |
| 322 | * -EINVAL : The @fd file descriptor is not an eventfd file. |
| 323 | */ |
| 324 | struct file *eventfd_fget(int fd) |
| 325 | { |
| 326 | struct file *file; |
| 327 | |
| 328 | file = fget(fd); |
| 329 | if (!file) |
| 330 | return ERR_PTR(-EBADF); |
| 331 | if (file->f_op != &eventfd_fops) { |
| 332 | fput(file); |
| 333 | return ERR_PTR(-EINVAL); |
| 334 | } |
| 335 | |
| 336 | return file; |
| 337 | } |
| 338 | EXPORT_SYMBOL_GPL(eventfd_fget); |
| 339 | |
| 340 | /** |
| 341 | * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context. |
| 342 | * @fd: [in] Eventfd file descriptor. |
| 343 | * |
| 344 | * Returns a pointer to the internal eventfd context, otherwise the error |
| 345 | * pointers returned by the following functions: |
| 346 | * |
| 347 | * eventfd_fget |
| 348 | */ |
| 349 | struct eventfd_ctx *eventfd_ctx_fdget(int fd) |
| 350 | { |
| 351 | struct eventfd_ctx *ctx; |
| 352 | struct fd f = fdget(fd); |
| 353 | if (!f.file) |
| 354 | return ERR_PTR(-EBADF); |
| 355 | ctx = eventfd_ctx_fileget(f.file); |
| 356 | fdput(f); |
| 357 | return ctx; |
| 358 | } |
| 359 | EXPORT_SYMBOL_GPL(eventfd_ctx_fdget); |
| 360 | |
| 361 | /** |
| 362 | * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context. |
| 363 | * @file: [in] Eventfd file pointer. |
| 364 | * |
| 365 | * Returns a pointer to the internal eventfd context, otherwise the error |
| 366 | * pointer: |
| 367 | * |
| 368 | * -EINVAL : The @fd file descriptor is not an eventfd file. |
| 369 | */ |
| 370 | struct eventfd_ctx *eventfd_ctx_fileget(struct file *file) |
| 371 | { |
| 372 | struct eventfd_ctx *ctx; |
| 373 | |
| 374 | if (file->f_op != &eventfd_fops) |
| 375 | return ERR_PTR(-EINVAL); |
| 376 | |
| 377 | ctx = file->private_data; |
| 378 | kref_get(&ctx->kref); |
| 379 | return ctx; |
| 380 | } |
| 381 | EXPORT_SYMBOL_GPL(eventfd_ctx_fileget); |
| 382 | |
| 383 | static int do_eventfd(unsigned int count, int flags) |
| 384 | { |
| 385 | struct eventfd_ctx *ctx; |
| 386 | int fd; |
| 387 | |
| 388 | /* Check the EFD_* constants for consistency. */ |
| 389 | BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC); |
| 390 | BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK); |
| 391 | |
| 392 | if (flags & ~EFD_FLAGS_SET) |
| 393 | return -EINVAL; |
| 394 | |
| 395 | ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); |
| 396 | if (!ctx) |
| 397 | return -ENOMEM; |
| 398 | |
| 399 | kref_init(&ctx->kref); |
| 400 | init_waitqueue_head(&ctx->wqh); |
| 401 | ctx->count = count; |
| 402 | ctx->flags = flags; |
| 403 | |
| 404 | fd = anon_inode_getfd("[eventfd]", &eventfd_fops, ctx, |
| 405 | O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS)); |
| 406 | if (fd < 0) |
| 407 | eventfd_free_ctx(ctx); |
| 408 | |
| 409 | return fd; |
| 410 | } |
| 411 | |
| 412 | SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags) |
| 413 | { |
| 414 | return do_eventfd(count, flags); |
| 415 | } |
| 416 | |
| 417 | SYSCALL_DEFINE1(eventfd, unsigned int, count) |
| 418 | { |
| 419 | return do_eventfd(count, 0); |
| 420 | } |
| 421 | |