Re: [RFC PATCH] f*xattr: allow O_PATH descriptors

On Fri, Aug 19, 2022 at 9:05 PM Christian Göttsche
[off-list ref] wrote:

On Wed, 22 Jun 2022 at 04:57, Aleksa Sarai [off-list ref] wrote:

quoted

On 2022-06-20, Amir Goldstein [off-list ref] wrote:

quoted

To be a reasonable reaction to the currently broken API is
to either accept the patch as is or request that setxattrat()
will be added to provide the new functionality.

Since the current functionality cannot be retroactively disabled as it
is being used already through /proc/self/fd/$n, adding
*xattrat(AT_EMPTY_PATH) doesn't really change what is currently possible
by userspace.

I would say we should add *xattrat(2) and then we can add an upgrade
mask blocking it (and other operations) later.

It seems setxattrat() is the preferred way to continue.
fsetxattr() would have one advantage though (w.r.t. SELinux):

The steps to label a file are:
  1. get the type of the file (via stat(2) family)
  2. lookup the desired label from the label database via selabel_lookup(3)
  3. assign the retrieved label to the file

The label is sensitive to the file type, e.g.

    $ matchpathcon -m file /etc/shadow
    /etc/shadow     system_u:object_r:shadow_t:s0
    $ matchpathcon -m lnk_file /etc/shadow
    /etc/shadow     system_u:object_r:etc_t:s0

Using the *at() family the file type could change between step 1. and 3.,
which operating on an O_PATH file descriptor would prevent.

I don't understand the problem.

If you have an O_PATH fd, the object it represents does not change.
If you use fstatat(fd, ..., AT_EMPTY_PATH) (or fstat) and
setxattrat(fd, ..., AT_EMPTY_PATH), it prevents the race.

Thanks,
Amir.