seccomp: fix UAF in user-trap code

On the failure path, we do an fput() of the listener fd if the filter fails
to install (e.g. because of a TSYNC race that's lost, or if the thread is
killed, etc.). fput() doesn't actually release the fd, it just ads it to a
work queue. Then the thread proceeds to free the filter, even though the
listener struct file has a reference to it.

To fix this, on the failure path let's set the private data to null, so we
know in ->release() to ignore the filter.

Reported-by: syzbot+981c26489b2d1c6316ba@syzkaller.appspotmail.com
Fixes: 6a21cc50 ("seccomp: add a return code to trap to userspace")
Signed-off-by: Tycho Andersen <tycho@tycho.ws>
Acked-by: Kees Cook <keescook@chromium.org>
Signed-off-by: James Morris <james.morris@microsoft.com>

kernel/seccomp.c

@@ -976,6 +976,9 @@ static int seccomp_notify_release(struct inode *inode, struct file *file)
 	struct seccomp_filter *filter = file->private_data;
 	struct seccomp_knotif *knotif;
 
+	if (!filter)
+		return 0;
+
 	mutex_lock(&filter->notify_lock);
 
 	/*
@@ -1300,6 +1303,7 @@ static long seccomp_set_mode_filter(unsigned int flags,
 out_put_fd:
 	if (flags & SECCOMP_FILTER_FLAG_NEW_LISTENER) {
 		if (ret < 0) {
+			listener_f->private_data = NULL;
 			fput(listener_f);
 			put_unused_fd(listener);
 		} else {