One of the big mistakes of Linux is that when you create a file with
open() and O_CREAT and the file already exists as dangling symlink that
the symlink will be followed and the file created that it points to.
This has resulted in many vulnerabilities, and triggered the creation of
the O_MOFOLLOW flag, addressing the problem.
O_NOFOLLOW is less than ideal in many ways, but in particular one: when
actually creating a file it makes sense to set, because it is a problem
to follow final symlinks in that case. But if the file is already
existing, it actually does make sense to follow the symlinks. With
openat_report_new() we distinguish these two cases anyway (the whole
function exists only to distinguish the create and the exists-already
case after all), hence let's do something about this: let's simply never
create files "through symlinks".
This can be implemented very easily: just pass O_NOFOLLOW to the 2nd
openat() call, where we actually create files.
And then basically remove 0dd82dab91
again, because we don't need to care anymore, we already will see ELOOP
when we touch a symlink.
Note that this change means that openat_report_new() will thus start to
deviate from plain openat() behaviour in this one small detail: when
actually creating files we will *never* follow the symlink. That should
be a systematic improvement of security.
Fixes: #34088
For SELinux it is essential that we reset the file creation label both
in the success and in the error path, hence do so.
Moreover, when calling the label post ops do it if possible with the
opened fd of the inode itself, rather than always going via its path,
simply to reduce the attack surface.
If openat_report_new() fails, then 'made_file' will be false, as no file
was created, hence there's no need to skip the unlinkat() explicitly
early, given that we check for 'made_file' anyway in the error path. The
extra error code checks are hence entirely redundant.
We have two distinct implementations of the post hook.
1. For SELinux we just reset the selinux label we told the kernel
earlier to use for new inodes.
2. For SMACK we might apply an xattr to the specified file.
The two calls are quite different: the first call we want to call in all
cases (failure or success), the latter only if we actually managed to
create an inode, in which case it is called on the inode.
Previously, we were using touch(), which usually works fine, because the
path should always refer to an existing directory, in which case it just
updates the timestamp. However, if the dir does not exist yet (which
shouldn't happen), it would be created as regular file, which is just
wrong.
Hence, let's instead create the dir as dir if it is missing, and then
update its timestamp.
openat() will always resolve symlinks, except if O_NOFOLLOW is passed
or O_CREAT|O_EXCL is passed. This means that if a dangling symlink is
passed to openat_report_new(), the first call to openat() will always
fail with ENOENT and the second call to openat() will always fail with
EEXIST.
Let's catch this case explicitly and fallback to creating the file with
just O_CREAT and assume we're the ones that created the file. We can't
resolve the symlink with chase() because this function is itself called
by chase() so we could end up in weird recursive calls if we'd try to do
so.
Currently, if proc_mounted() != 0, some functions
propagate -ENOENT while others return -EBADF.
Let's make things consistent, by introducing
a static inline helper responsible for finding out
the appropriate errno.
From readlinkat(2):
Since Linux 2.6.39, pathname can be an empty string, in which case the
call operates on the symbolic link referred to by dirfd (which should
have been obtained using open(2) with the O_PATH and O_NOFOLLOW flags).
When systemd-firstboot (or any other unit which uses the tty) is started,
systemd will reset the terminal. If systemd-vconsole-setup happens to be
running at that time, it'll error out when it tries to use the vconsole fd and
gets an EIO from ioctl.
e019ea738d was the first fix. It added an
implicit ordering between units using the tty and systemd-vconsole-setup.
(The commit title is wrong. The approach was generalized, but the commit title
wasn't updated.)
Then cea32691c3 was added to restart
systemd-vconsole-setup.service from systemd-firstboot.service. This was OK,
with the ordering in place, systemd-vconsole-setup.service would wait until
systemd-firstboot.service exited. But this wasn't enough, because we want the
key mappings to be loaded immediately after systemd-firstboot writes the
config. 8eb668b9ab implemented that, but actually
reintroduced the original issue. I had to drop the ordering between the two
units because otherwise we'd deadlock, waiting from firstboot for
vconsole-setup which wouldn't start while firstboot was running.
Restarting vconsole-setup.service from systemd-firstboot.service works just
fine, but when vconsole-setup.service is started earlier, it may be interrupted
by systemd-firstboot.service.
To resolve the issue, let's take a lock around the tty device. The reset is
performed after fork, so the (short) delay should not matter too much.
In xopenat_lock() the assert on <path> is dropped so that we can call
xopenat(fd, NULL) to get a copy of the original fd.
Fixes#26908.
We recently codified in the coding style that for openat() style APIs,
an empty path can be passed both as the empty string and as NULL, so
let's make sure we follow that style in xopenat().
The comment makes a reference to the function fchmod_path() but this
function does not exist in the source tree.
However, the function fchmod_opath() exists; it was introduced by the
commit 4dfaa528d4.
As the comment tells, the function futimens_opath() introduced by the
commit f25bff5eaf is similar to the
function fchmod_opath(); therefore, it should reference it.
This fixes the typo in the comment by referencing the proper function
fchmod_opath().
Lennart Poettering
Mike Yuan
Adrian Vovk
Yu Watanabe
Daan De Meyer
Christian Göttsche
Arseny MaslennikovandMike Yuan
Zbigniew Jędrzejewski-Szmek
Gaël PORTAY
Antonio Alvarez Feijoo