fanotify: fix -EOVERFLOW with large files on 64-bit

From: Will Woods <[email protected]>

commit 1e2ee49f7f1b79f0b14884fe6a602f0411b39552 upstream.

On 64-bit systems, O_LARGEFILE is automatically added to flags inside the open() syscall (also openat(), blkdev_open(), etc). Userspace therefore defines O_LARGEFILE to be 0 - you can use it, but it's a no-op. Everything should be O_LARGEFILE by default.

But: when fanotify does create_fd() it uses dentry_open(), which skips all that. And userspace can't set O_LARGEFILE in fanotify_init() because it's defined to 0. So if fanotify gets an event regarding a large file, the read() will just fail with -EOVERFLOW.

This patch adds O_LARGEFILE to fanotify_init()'s event_f_flags on 64-bit systems, using the same test as open()/openat()/etc.

Addresses https://bugzilla.redhat.com/show_bug.cgi?id=696821

[snu: Backported to 3.2 / 3.10: adjusted context]

---
 fs/notify/fanotify/fanotify_user.c | 3 +++
 1 file changed, 3 insertions(+)

index d57995e1cfd7..dd3d9d13a4c4 100644
@@ -712,6 +712,9 @@ SYSCALL_DEFINE2(fanotify_init, unsigned int, flags, unsigned int, event_f_flags)
 	group->fanotify_data.user = user;
 	atomic_inc(&user->fanotify_listeners);
 
+	if (force_o_largefile())
+		event_f_flags |= O_LARGEFILE;
+
 	group->fanotify_data.f_flags = event_f_flags;
 #ifdef CONFIG_FANOTIFY_ACCESS_PERMISSIONS
 	mutex_init(&group->fanotify_data.access_mutex);

2.15.0

3.10.y is end-of-life :(

See the front page of www.kernel.org for that list if you are curious.

thanks,

greg k-h