fs/eventfd.c

   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
  23 struct 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  */
  54 int 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 }
  70 EXPORT_SYMBOL_GPL(eventfd_signal);
  71
  72 static void eventfd_free_ctx(struct eventfd_ctx *ctx)
  73 {
  74         kfree(ctx);
  75 }
  76
  77 static 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  */
  90 struct eventfd_ctx *eventfd_ctx_get(struct eventfd_ctx *ctx)
  91 {
  92         kref_get(&ctx->kref);
  93         return ctx;
  94 }
  95 EXPORT_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  */
 104 void eventfd_ctx_put(struct eventfd_ctx *ctx)
 105 {
 106         kref_put(&ctx->kref, eventfd_free);
 107 }
 108 EXPORT_SYMBOL_GPL(eventfd_ctx_put);
 109
 110 static 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
 119 static 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
 139 static 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  */
 158 int 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 }
 172 EXPORT_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  */
 188 ssize_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 }
 226 EXPORT_SYMBOL_GPL(eventfd_ctx_read);
 227
 228 static 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
 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, POLLIN);
 285         }
 286         spin_unlock_irq(&ctx->wqh.lock);
 287
 288         return res;
 289 }
 290
 291 static 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  */
 309 struct 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 }
 323 EXPORT_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  */
 334 struct 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 }
 347 EXPORT_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  */
 358 struct 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 }
 365 EXPORT_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  */
 381 struct 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
 410 SYSCALL_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
 429 err_put_unused_fd:
 430         put_unused_fd(fd);
 431
 432         return error;
 433 }
 434
 435 SYSCALL_DEFINE1(eventfd, unsigned int, count)
 436 {
 437         return sys_eventfd2(count, 0);
 438 }
 439