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/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 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 }
  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_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  */
  88 struct eventfd_ctx *eventfd_ctx_get(struct eventfd_ctx *ctx)
  89 {
  90         kref_get(&ctx->kref);
  91         return ctx;
  92 }
  93 EXPORT_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  */
 102 void eventfd_ctx_put(struct eventfd_ctx *ctx)
 103 {
 104         kref_put(&ctx->kref, eventfd_free);
 105 }
 106 EXPORT_SYMBOL_GPL(eventfd_ctx_put);
 107
 108 static 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
 117 static 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         smp_rmb();
 125         count = ctx->count;
 126
 127         if (count > 0)
 128                 events |= POLLIN;
 129         if (count == ULLONG_MAX)
 130                 events |= POLLERR;
 131         if (ULLONG_MAX - 1 > count)
 132                 events |= POLLOUT;
 133
 134         return events;
 135 }
 136
 137 static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
 138 {
 139         *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
 140         ctx->count -= *cnt;
 141 }
 142
 143 /**
 144  * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue.
 145  * @ctx: [in] Pointer to eventfd context.
 146  * @wait: [in] Wait queue to be removed.
 147  * @cnt: [out] Pointer to the 64-bit counter value.
 148  *
 149  * Returns %0 if successful, or the following error codes:
 150  *
 151  * -EAGAIN      : The operation would have blocked.
 152  *
 153  * This is used to atomically remove a wait queue entry from the eventfd wait
 154  * queue head, and read/reset the counter value.
 155  */
 156 int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_t *wait,
 157                                   __u64 *cnt)
 158 {
 159         unsigned long flags;
 160
 161         spin_lock_irqsave(&ctx->wqh.lock, flags);
 162         eventfd_ctx_do_read(ctx, cnt);
 163         __remove_wait_queue(&ctx->wqh, wait);
 164         if (*cnt != 0 && waitqueue_active(&ctx->wqh))
 165                 wake_up_locked_poll(&ctx->wqh, POLLOUT);
 166         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 167
 168         return *cnt != 0 ? 0 : -EAGAIN;
 169 }
 170 EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue);
 171
 172 /**
 173  * eventfd_ctx_read - Reads the eventfd counter or wait if it is zero.
 174  * @ctx: [in] Pointer to eventfd context.
 175  * @no_wait: [in] Different from zero if the operation should not block.
 176  * @cnt: [out] Pointer to the 64-bit counter value.
 177  *
 178  * Returns %0 if successful, or the following error codes:
 179  *
 180  * -EAGAIN      : The operation would have blocked but @no_wait was non-zero.
 181  * -ERESTARTSYS : A signal interrupted the wait operation.
 182  *
 183  * If @no_wait is zero, the function might sleep until the eventfd internal
 184  * counter becomes greater than zero.
 185  */
 186 ssize_t eventfd_ctx_read(struct eventfd_ctx *ctx, int no_wait, __u64 *cnt)
 187 {
 188         ssize_t res;
 189         DECLARE_WAITQUEUE(wait, current);
 190
 191         spin_lock_irq(&ctx->wqh.lock);
 192         *cnt = 0;
 193         res = -EAGAIN;
 194         if (ctx->count > 0)
 195                 res = 0;
 196         else if (!no_wait) {
 197                 __add_wait_queue(&ctx->wqh, &wait);
 198                 for (;;) {
 199                         set_current_state(TASK_INTERRUPTIBLE);
 200                         if (ctx->count > 0) {
 201                                 res = 0;
 202                                 break;
 203                         }
 204                         if (signal_pending(current)) {
 205                                 res = -ERESTARTSYS;
 206                                 break;
 207                         }
 208                         spin_unlock_irq(&ctx->wqh.lock);
 209                         schedule();
 210                         spin_lock_irq(&ctx->wqh.lock);
 211                 }
 212                 __remove_wait_queue(&ctx->wqh, &wait);
 213                 __set_current_state(TASK_RUNNING);
 214         }
 215         if (likely(res == 0)) {
 216                 eventfd_ctx_do_read(ctx, cnt);
 217                 if (waitqueue_active(&ctx->wqh))
 218                         wake_up_locked_poll(&ctx->wqh, POLLOUT);
 219         }
 220         spin_unlock_irq(&ctx->wqh.lock);
 221
 222         return res;
 223 }
 224 EXPORT_SYMBOL_GPL(eventfd_ctx_read);
 225
 226 static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count,
 227                             loff_t *ppos)
 228 {
 229         struct eventfd_ctx *ctx = file->private_data;
 230         ssize_t res;
 231         __u64 cnt;
 232
 233         if (count < sizeof(cnt))
 234                 return -EINVAL;
 235         res = eventfd_ctx_read(ctx, file->f_flags & O_NONBLOCK, &cnt);
 236         if (res < 0)
 237                 return res;
 238
 239         return put_user(cnt, (__u64 __user *) buf) ? -EFAULT : sizeof(cnt);
 240 }
 241
 242 static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
 243                              loff_t *ppos)
 244 {
 245         struct eventfd_ctx *ctx = file->private_data;
 246         ssize_t res;
 247         __u64 ucnt;
 248         DECLARE_WAITQUEUE(wait, current);
 249
 250         if (count < sizeof(ucnt))
 251                 return -EINVAL;
 252         if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
 253                 return -EFAULT;
 254         if (ucnt == ULLONG_MAX)
 255                 return -EINVAL;
 256         spin_lock_irq(&ctx->wqh.lock);
 257         res = -EAGAIN;
 258         if (ULLONG_MAX - ctx->count > ucnt)
 259                 res = sizeof(ucnt);
 260         else if (!(file->f_flags & O_NONBLOCK)) {
 261                 __add_wait_queue(&ctx->wqh, &wait);
 262                 for (res = 0;;) {
 263                         set_current_state(TASK_INTERRUPTIBLE);
 264                         if (ULLONG_MAX - ctx->count > ucnt) {
 265                                 res = sizeof(ucnt);
 266                                 break;
 267                         }
 268                         if (signal_pending(current)) {
 269                                 res = -ERESTARTSYS;
 270                                 break;
 271                         }
 272                         spin_unlock_irq(&ctx->wqh.lock);
 273                         schedule();
 274                         spin_lock_irq(&ctx->wqh.lock);
 275                 }
 276                 __remove_wait_queue(&ctx->wqh, &wait);
 277                 __set_current_state(TASK_RUNNING);
 278         }
 279         if (likely(res > 0)) {
 280                 ctx->count += ucnt;
 281                 if (waitqueue_active(&ctx->wqh))
 282                         wake_up_locked_poll(&ctx->wqh, POLLIN);
 283         }
 284         spin_unlock_irq(&ctx->wqh.lock);
 285
 286         return res;
 287 }
 288
 289 #ifdef CONFIG_PROC_FS
 290 static void eventfd_show_fdinfo(struct seq_file *m, struct file *f)
 291 {
 292         struct eventfd_ctx *ctx = f->private_data;
 293
 294         spin_lock_irq(&ctx->wqh.lock);
 295         seq_printf(m, "eventfd-count: %16llx\n",
 296                    (unsigned long long)ctx->count);
 297         spin_unlock_irq(&ctx->wqh.lock);
 298 }
 299 #endif
 300
 301 static const struct file_operations eventfd_fops = {
 302 #ifdef CONFIG_PROC_FS
 303         .show_fdinfo    = eventfd_show_fdinfo,
 304 #endif
 305         .release        = eventfd_release,
 306         .poll           = eventfd_poll,
 307         .read           = eventfd_read,
 308         .write          = eventfd_write,
 309         .llseek         = noop_llseek,
 310 };
 311
 312 /**
 313  * eventfd_fget - Acquire a reference of an eventfd file descriptor.
 314  * @fd: [in] Eventfd file descriptor.
 315  *
 316  * Returns a pointer to the eventfd file structure in case of success, or the
 317  * following error pointer:
 318  *
 319  * -EBADF    : Invalid @fd file descriptor.
 320  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 321  */
 322 struct file *eventfd_fget(int fd)
 323 {
 324         struct file *file;
 325
 326         file = fget(fd);
 327         if (!file)
 328                 return ERR_PTR(-EBADF);
 329         if (file->f_op != &eventfd_fops) {
 330                 fput(file);
 331                 return ERR_PTR(-EINVAL);
 332         }
 333
 334         return file;
 335 }
 336 EXPORT_SYMBOL_GPL(eventfd_fget);
 337
 338 /**
 339  * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
 340  * @fd: [in] Eventfd file descriptor.
 341  *
 342  * Returns a pointer to the internal eventfd context, otherwise the error
 343  * pointers returned by the following functions:
 344  *
 345  * eventfd_fget
 346  */
 347 struct eventfd_ctx *eventfd_ctx_fdget(int fd)
 348 {
 349         struct eventfd_ctx *ctx;
 350         struct fd f = fdget(fd);
 351         if (!f.file)
 352                 return ERR_PTR(-EBADF);
 353         ctx = eventfd_ctx_fileget(f.file);
 354         fdput(f);
 355         return ctx;
 356 }
 357 EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
 358
 359 /**
 360  * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
 361  * @file: [in] Eventfd file pointer.
 362  *
 363  * Returns a pointer to the internal eventfd context, otherwise the error
 364  * pointer:
 365  *
 366  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 367  */
 368 struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
 369 {
 370         if (file->f_op != &eventfd_fops)
 371                 return ERR_PTR(-EINVAL);
 372
 373         return eventfd_ctx_get(file->private_data);
 374 }
 375 EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
 376
 377 /**
 378  * eventfd_file_create - Creates an eventfd file pointer.
 379  * @count: Initial eventfd counter value.
 380  * @flags: Flags for the eventfd file.
 381  *
 382  * This function creates an eventfd file pointer, w/out installing it into
 383  * the fd table. This is useful when the eventfd file is used during the
 384  * initialization of data structures that require extra setup after the eventfd
 385  * creation. So the eventfd creation is split into the file pointer creation
 386  * phase, and the file descriptor installation phase.
 387  * In this way races with userspace closing the newly installed file descriptor
 388  * can be avoided.
 389  * Returns an eventfd file pointer, or a proper error pointer.
 390  */
 391 struct file *eventfd_file_create(unsigned int count, int flags)
 392 {
 393         struct file *file;
 394         struct eventfd_ctx *ctx;
 395
 396         /* Check the EFD_* constants for consistency.  */
 397         BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
 398         BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
 399
 400         if (flags & ~EFD_FLAGS_SET)
 401                 return ERR_PTR(-EINVAL);
 402
 403         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 404         if (!ctx)
 405                 return ERR_PTR(-ENOMEM);
 406
 407         kref_init(&ctx->kref);
 408         init_waitqueue_head(&ctx->wqh);
 409         ctx->count = count;
 410         ctx->flags = flags;
 411
 412         file = anon_inode_getfile("[eventfd]", &eventfd_fops, ctx,
 413                                   O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS));
 414         if (IS_ERR(file))
 415                 eventfd_free_ctx(ctx);
 416
 417         return file;
 418 }
 419
 420 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
 421 {
 422         int fd, error;
 423         struct file *file;
 424
 425         error = get_unused_fd_flags(flags & EFD_SHARED_FCNTL_FLAGS);
 426         if (error < 0)
 427                 return error;
 428         fd = error;
 429
 430         file = eventfd_file_create(count, flags);
 431         if (IS_ERR(file)) {
 432                 error = PTR_ERR(file);
 433                 goto err_put_unused_fd;
 434         }
 435         fd_install(fd, file);
 436
 437         return fd;
 438
 439 err_put_unused_fd:
 440         put_unused_fd(fd);
 441
 442         return error;
 443 }
 444
 445 SYSCALL_DEFINE1(eventfd, unsigned int, count)
 446 {
 447         return sys_eventfd2(count, 0);
 448 }
 449