[PATCH] timerfd: fix nonblocking reads

[Jens Axboe <axboe@xxxxxxxxx>]

timerfd is utterly buggy wrt nonblocking reads - regardless of whether
or not there's data available, it returns -EAGAIN. This is incompatible
with how nonblocking reads should work. If there's data available, it
should be returned.

Convert it to use fops->read_iter() so it can handle both nonblocking
fds and IOCB_NOWAIT, mark it as FMODE_NOWAIT to signify that it's
compatible with nonblocking reads, and finally have timerfd_read_iter()
properly check for data availability in nonblocking mode.

Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Jens Axboe <axboe@xxxxxxxxx>

---

Can't believe it's this broken... Patch has been tested with a test case
that was reported via io_uring, and I also ran the ltp timerfd test
cases and it passes all of those too.

diff --git a/fs/timerfd.c b/fs/timerfd.c
index e9c96a0c79f1..9297b82af13d 100644
--- a/fs/timerfd.c
+++ b/fs/timerfd.c
@@ -262,20 +262,24 @@ static __poll_t timerfd_poll(struct file *file, poll_table *wait)
 	return events;
 }
 
-static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
-			    loff_t *ppos)
+static ssize_t timerfd_read_iter(struct kiocb *iocb, struct iov_iter *to)
 {
+	struct file *file = iocb->ki_filp;
 	struct timerfd_ctx *ctx = file->private_data;
 	ssize_t res;
 	u64 ticks = 0;
 
-	if (count < sizeof(ticks))
+	if (iov_iter_count(to) < sizeof(ticks))
 		return -EINVAL;
+
 	spin_lock_irq(&ctx->wqh.lock);
-	if (file->f_flags & O_NONBLOCK)
-		res = -EAGAIN;
-	else
-		res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
+	if (!ctx->ticks) {
+		if (file->f_flags & O_NONBLOCK || iocb->ki_flags & IOCB_NOWAIT)
+			res = -EAGAIN;
+		else
+			res = wait_event_interruptible_locked_irq(ctx->wqh,
+								  ctx->ticks);
+	}
 
 	/*
 	 * If clock has changed, we do not care about the
@@ -313,7 +317,7 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
 	}
 	spin_unlock_irq(&ctx->wqh.lock);
 	if (ticks)
-		res = put_user(ticks, (u64 __user *) buf) ? -EFAULT: sizeof(ticks);
+		res = copy_to_iter(&ticks, sizeof(ticks), to);
 	return res;
 }
 
@@ -384,7 +388,7 @@ static long timerfd_ioctl(struct file *file, unsigned int cmd, unsigned long arg
 static const struct file_operations timerfd_fops = {
 	.release	= timerfd_release,
 	.poll		= timerfd_poll,
-	.read		= timerfd_read,
+	.read_iter	= timerfd_read_iter,
 	.llseek		= noop_llseek,
 	.show_fdinfo	= timerfd_show,
 	.unlocked_ioctl	= timerfd_ioctl,
@@ -407,6 +411,7 @@ SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 {
 	int ufd;
 	struct timerfd_ctx *ctx;
+	struct file *file;
 
 	/* Check the TFD_* constants for consistency.  */
 	BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
@@ -443,11 +448,22 @@ SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 
 	ctx->moffs = ktime_mono_to_real(0);
 
-	ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
-			       O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
-	if (ufd < 0)
+	ufd = get_unused_fd_flags(O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
+	if (ufd < 0) {
 		kfree(ctx);
+		return ufd;
+	}
+
+	file = anon_inode_getfile("[timerfd]", &timerfd_fops, ctx,
+				    O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
+	if (IS_ERR(file)) {
+		put_unused_fd(ufd);
+		kfree(ctx);
+		return PTR_ERR(file);
+	}
 
+	file->f_mode |= FMODE_NOWAIT;
+	fd_install(ufd, file);
 	return ufd;
 }
 
-- 
Jens Axboe