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