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

On Sat, 18 Jun 2022 at 11:11, Amir Goldstein [off-list ref] wrote:

On Sat, Jun 18, 2022 at 6:18 AM Aleksa Sarai [off-list ref] wrote:

quoted

On 2022-06-08, Amir Goldstein [off-list ref] wrote:

quoted

On Wed, Jun 8, 2022 at 3:48 PM Christian Brauner [off-list ref] wrote:

quoted

On Wed, Jun 08, 2022 at 03:28:52PM +0300, Amir Goldstein wrote:

quoted

On Wed, Jun 8, 2022 at 2:57 PM Christian Brauner [off-list ref] wrote:

quoted

On Tue, Jun 07, 2022 at 05:31:39PM +0200, Christian Göttsche wrote:

quoted

From: Miklos Szeredi <redacted>

Support file descriptors obtained via O_PATH for extended attribute
operations.

Extended attributes are for example used by SELinux for the security
context of file objects. To avoid time-of-check-time-of-use issues while
setting those contexts it is advisable to pin the file in question and
operate on a file descriptor instead of the path name. This can be
emulated in userspace via /proc/self/fd/NN [1] but requires a procfs,
which might not be mounted e.g. inside of chroots, see[2].

[1]: https://github.com/SELinuxProject/selinux/commit/7e979b56fd2cee28f647376a7233d2ac2d12ca50
[2]: https://github.com/SELinuxProject/selinux/commit/de285252a1801397306032e070793889c9466845

Original patch by Miklos Szeredi [off-list ref]
https://patchwork.kernel.org/project/linux-fsdevel/patch/20200505095915.11275-6-mszeredi@redhat.com/

quoted

While this carries a minute risk of someone relying on the property of
xattr syscalls rejecting O_PATH descriptors, it saves the trouble of
introducing another set of syscalls.

Only file->f_path and file->f_inode are accessed in these functions.

Current versions return EBADF, hence easy to detect the presense of
this feature and fall back in case it's missing.

CC: linux-api@vger.kernel.org
CC: linux-man@vger.kernel.org
Signed-off-by: Christian Göttsche <redacted>
---

I'd be somewhat fine with getxattr and listxattr but I'm worried that
setxattr/removexattr waters down O_PATH semantics even more. I don't
want O_PATH fds to be useable for operations which are semantically
equivalent to a write.

It is not really semantically equivalent to a write if it works on a
O_RDONLY fd already.

The fact that it works on a O_RDONLY fd has always been weird. And is
probably a bug. If you look at xattr_permission() you can see that it

Bug or no bug, this is the UAPI. It is not fixable anymore.

quoted

checks for MAY_WRITE for set operations... setxattr() writes to disk for
real filesystems. I don't know how much closer to a write this can get.

In general, one semantic aberration doesn't justify piling another one
on top.

(And one thing that speaks for O_RDONLY is at least that it actually
opens the file wheres O_PATH doesn't.)

Ok. I care mostly about consistent UAPI, so if you want to set the
rule that modify f*() operations are not allowed to use O_PATH fd,
I can live with that, although fcntl(2) may be breaking that rule, but
fine by me.
It's good to have consistent rules and it's good to add a new UAPI for
new behavior.

However...

quoted

quoted

quoted

In sensitive environments such as service management/container runtimes
we often send O_PATH fds around precisely because it is restricted what
they can be used for. I'd prefer to not to plug at this string.

But unless I am mistaken, path_setxattr() and syscall_fsetxattr()
are almost identical w.r.t permission checks and everything else.

So this change introduces nothing new that a user in said environment
cannot already accomplish with setxattr().

Besides, as the commit message said, doing setxattr() on an O_PATH
fd is already possible with setxattr("/proc/self/$fd"), so whatever security
hole you are trying to prevent is already wide open.

That is very much a something that we're trying to restrict for this
exact reason and is one of the main motivator for upgrade mask in
openat2(). If I want to send a O_PATH around I want it to not be
upgradable. Aleksa is working on upgrade masks with openat2() (see [1]
and part of the original patchset in [2]. O_PATH semantics don't need to
become weird.

[1]: https://lore.kernel.org/all/20220526130355.fo6gzbst455fxywy@senku (local)
[2]: https://patchwork.ozlabs.org/project/linuxppc-dev/patch/20190728010207.9781-8-cyphar@cyphar.com

... thinking forward, if this patch is going to be rejected, the patch that
will follow is *xattrat() syscalls.

What will you be able to argue then?

There are several *at() syscalls that modify metadata.
fchownat(.., AT_EMPTY_PATH) is intentionally designed for this.

Do you intend to try and block setxattrat()?
Just try and block setxattrat(.., AT_EMPTY_PATH)?
those *at() syscalls have real use cases to avoid TOCTOU races.
Do you propose that applications will have to use fsetxattr() on an open
file to avert races?

I completely understand the idea behind upgrade masks
for limiting f_mode, but I don't know if trying to retroactively
change semantics of setxattr() in the move to setxattrat()
is going to be a good idea.

The goal would be that the semantics of fooat(<fd>, AT_EMPTY_PATH) and
foo(/proc/self/fd/<fd>) should always be identical, and the current
semantics of /proc/self/fd/<fd> are too leaky so we shouldn't always
assume that keeping them makes sense (the most obvious example is being
able to do tricks to open /proc/$pid/exe as O_RDWR).

Please make a note that I have applications relying on current magic symlink
semantics w.r.t setxattr() and other metadata operations, and the libselinux
commit linked from the patch commit message proves that magic symlink
semantics are used in the wild, so it is not likely that those semantics could
be changed, unless userspace breakage could be justified by fixing a serious
security issue (i.e. open /proc/$pid/exe as O_RDWR).

quoted

I suspect that the long-term solution would be to have more upgrade
masks so that userspace can opt-in to not allowing any kind of
(metadata) write access through a particular file descriptor. You're
quite right that we have several metadata write AT_EMPTY_PATH APIs, and
so we can't retroactively block /everything/ but we should try to come
up with less leaky rules by default if it won't break userspace.

Ok, let me try to say this in my own words using an example to see that
we are all on the same page:

- lsetxattr(PATH_TO_FILE,..) has inherent TOCTOU races
- fsetxattr(fd,...) is not applicable for symbolic links

fsetxattr(2) works on symbolic links, e.g. for "security.selinux",
except for the user namespace:

https://github.com/torvalds/linux/blob/4b35035bcf80ddb47c0112c4fbd84a63a2836a18/fs/xattr.c#L124-L136
/*
* In the user.* namespace, only regular files and directories can have
* extended attributes. For sticky directories, only the owner and
* privileged users can write attributes.
*/
if (!strncmp(name, XATTR_USER_PREFIX, XATTR_USER_PREFIX_LEN)) {
    if (!S_ISREG(inode->i_mode) && !S_ISDIR(inode->i_mode))
        return (mask & MAY_WRITE) ? -EPERM : -ENODATA;
    if (S_ISDIR(inode->i_mode) && (inode->i_mode & S_ISVTX) &&
       (mask & MAY_WRITE) &&
        !inode_owner_or_capable(mnt_userns, inode))
        return -EPERM;
}

Currently it just does not support O_PATH file descriptors.
And with O_RDONLY setting extended attributes is supported as well
(fsetxattr(2) does not require O_RDWR or O_WRONLY).

- setxattr("/proc/self/fd/<fd>",...) is the current API to avoid TOCTOU races
  when setting xattr on symbolic links
- setxattrat(o_path_fd, "", ..., AT_EMPTY_PATH) is proposed as a the
  "new API" for setting xattr on symlinks (and special files)
- The new API is going to be more strict than the old magic symlink API
- *If* it turns out to not break user applications, old API can also become
  more strict to align with new API (unlikely the case for setxattr())
- This will allow sandboxed containers to opt-out of the "old API", by
  restricting access to /proc/self/fd and to implement more fine grained
  control over which metadata operations are allowed on an O_PATH fd

Did I understand the plan correctly?
Do you agree with me that the plan to keep AT_EMPTY_PATH and
magic symlink semantics may not be realistic?

Thanks,
Amir.