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/list.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/anon_inodes.h>
  17 #include <linux/eventfd.h>
  18 #include <linux/syscalls.h>
  19 #include <linux/module.h>
  20
  21 struct eventfd_ctx {
  22         wait_queue_head_t wqh;
  23         /*
  24          * Every time that a write(2) is performed on an eventfd, the
  25          * value of the __u64 being written is added to "count" and a
  26          * wakeup is performed on "wqh". A read(2) will return the "count"
  27          * value to userspace, and will reset "count" to zero. The kernel
  28          * size eventfd_signal() also, adds to the "count" counter and
  29          * issue a wakeup.
  30          */
  31         __u64 count;
  32         unsigned int flags;
  33 };
  34
  35 /*
  36  * Adds "n" to the eventfd counter "count". Returns "n" in case of
  37  * success, or a value lower then "n" in case of coutner overflow.
  38  * This function is supposed to be called by the kernel in paths
  39  * that do not allow sleeping. In this function we allow the counter
  40  * to reach the ULLONG_MAX value, and we signal this as overflow
  41  * condition by returining a POLLERR to poll(2).
  42  */
  43 int eventfd_signal(struct file *file, int n)
  44 {
  45         struct eventfd_ctx *ctx = file->private_data;
  46         unsigned long flags;
  47
  48         if (n < 0)
  49                 return -EINVAL;
  50         spin_lock_irqsave(&ctx->wqh.lock, flags);
  51         if (ULLONG_MAX - ctx->count < n)
  52                 n = (int) (ULLONG_MAX - ctx->count);
  53         ctx->count += n;
  54         if (waitqueue_active(&ctx->wqh))
  55                 wake_up_locked_poll(&ctx->wqh, POLLIN);
  56         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  57
  58         return n;
  59 }
  60 EXPORT_SYMBOL_GPL(eventfd_signal);
  61
  62 static int eventfd_release(struct inode *inode, struct file *file)
  63 {
  64         kfree(file->private_data);
  65         return 0;
  66 }
  67
  68 static unsigned int eventfd_poll(struct file *file, poll_table *wait)
  69 {
  70         struct eventfd_ctx *ctx = file->private_data;
  71         unsigned int events = 0;
  72         unsigned long flags;
  73
  74         poll_wait(file, &ctx->wqh, wait);
  75
  76         spin_lock_irqsave(&ctx->wqh.lock, flags);
  77         if (ctx->count > 0)
  78                 events |= POLLIN;
  79         if (ctx->count == ULLONG_MAX)
  80                 events |= POLLERR;
  81         if (ULLONG_MAX - 1 > ctx->count)
  82                 events |= POLLOUT;
  83         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  84
  85         return events;
  86 }
  87
  88 static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count,
  89                             loff_t *ppos)
  90 {
  91         struct eventfd_ctx *ctx = file->private_data;
  92         ssize_t res;
  93         __u64 ucnt = 0;
  94         DECLARE_WAITQUEUE(wait, current);
  95
  96         if (count < sizeof(ucnt))
  97                 return -EINVAL;
  98         spin_lock_irq(&ctx->wqh.lock);
  99         res = -EAGAIN;
 100         if (ctx->count > 0)
 101                 res = sizeof(ucnt);
 102         else if (!(file->f_flags & O_NONBLOCK)) {
 103                 __add_wait_queue(&ctx->wqh, &wait);
 104                 for (res = 0;;) {
 105                         set_current_state(TASK_INTERRUPTIBLE);
 106                         if (ctx->count > 0) {
 107                                 res = sizeof(ucnt);
 108                                 break;
 109                         }
 110                         if (signal_pending(current)) {
 111                                 res = -ERESTARTSYS;
 112                                 break;
 113                         }
 114                         spin_unlock_irq(&ctx->wqh.lock);
 115                         schedule();
 116                         spin_lock_irq(&ctx->wqh.lock);
 117                 }
 118                 __remove_wait_queue(&ctx->wqh, &wait);
 119                 __set_current_state(TASK_RUNNING);
 120         }
 121         if (likely(res > 0)) {
 122                 ucnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
 123                 ctx->count -= ucnt;
 124                 if (waitqueue_active(&ctx->wqh))
 125                         wake_up_locked_poll(&ctx->wqh, POLLOUT);
 126         }
 127         spin_unlock_irq(&ctx->wqh.lock);
 128         if (res > 0 && put_user(ucnt, (__u64 __user *) buf))
 129                 return -EFAULT;
 130
 131         return res;
 132 }
 133
 134 static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
 135                              loff_t *ppos)
 136 {
 137         struct eventfd_ctx *ctx = file->private_data;
 138         ssize_t res;
 139         __u64 ucnt;
 140         DECLARE_WAITQUEUE(wait, current);
 141
 142         if (count < sizeof(ucnt))
 143                 return -EINVAL;
 144         if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
 145                 return -EFAULT;
 146         if (ucnt == ULLONG_MAX)
 147                 return -EINVAL;
 148         spin_lock_irq(&ctx->wqh.lock);
 149         res = -EAGAIN;
 150         if (ULLONG_MAX - ctx->count > ucnt)
 151                 res = sizeof(ucnt);
 152         else if (!(file->f_flags & O_NONBLOCK)) {
 153                 __add_wait_queue(&ctx->wqh, &wait);
 154                 for (res = 0;;) {
 155                         set_current_state(TASK_INTERRUPTIBLE);
 156                         if (ULLONG_MAX - ctx->count > ucnt) {
 157                                 res = sizeof(ucnt);
 158                                 break;
 159                         }
 160                         if (signal_pending(current)) {
 161                                 res = -ERESTARTSYS;
 162                                 break;
 163                         }
 164                         spin_unlock_irq(&ctx->wqh.lock);
 165                         schedule();
 166                         spin_lock_irq(&ctx->wqh.lock);
 167                 }
 168                 __remove_wait_queue(&ctx->wqh, &wait);
 169                 __set_current_state(TASK_RUNNING);
 170         }
 171         if (likely(res > 0)) {
 172                 ctx->count += ucnt;
 173                 if (waitqueue_active(&ctx->wqh))
 174                         wake_up_locked_poll(&ctx->wqh, POLLIN);
 175         }
 176         spin_unlock_irq(&ctx->wqh.lock);
 177
 178         return res;
 179 }
 180
 181 static const struct file_operations eventfd_fops = {
 182         .release        = eventfd_release,
 183         .poll           = eventfd_poll,
 184         .read           = eventfd_read,
 185         .write          = eventfd_write,
 186 };
 187
 188 struct file *eventfd_fget(int fd)
 189 {
 190         struct file *file;
 191
 192         file = fget(fd);
 193         if (!file)
 194                 return ERR_PTR(-EBADF);
 195         if (file->f_op != &eventfd_fops) {
 196                 fput(file);
 197                 return ERR_PTR(-EINVAL);
 198         }
 199
 200         return file;
 201 }
 202 EXPORT_SYMBOL_GPL(eventfd_fget);
 203
 204 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
 205 {
 206         int fd;
 207         struct eventfd_ctx *ctx;
 208
 209         /* Check the EFD_* constants for consistency.  */
 210         BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
 211         BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
 212
 213         if (flags & ~EFD_FLAGS_SET)
 214                 return -EINVAL;
 215
 216         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 217         if (!ctx)
 218                 return -ENOMEM;
 219
 220         init_waitqueue_head(&ctx->wqh);
 221         ctx->count = count;
 222         ctx->flags = flags;
 223
 224         /*
 225          * When we call this, the initialization must be complete, since
 226          * anon_inode_getfd() will install the fd.
 227          */
 228         fd = anon_inode_getfd("[eventfd]", &eventfd_fops, ctx,
 229                               flags & EFD_SHARED_FCNTL_FLAGS);
 230         if (fd < 0)
 231                 kfree(ctx);
 232         return fd;
 233 }
 234
 235 SYSCALL_DEFINE1(eventfd, unsigned int, count)
 236 {
 237         return sys_eventfd2(count, 0);
 238 }
 239