lkml.org 
[lkml]   [2005]   [May]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [RCF] [PATCH] unprivileged mount/umount
From
Date
> >For ptrace the definition is:
> > If the tracee has different privileges, than the tracer, than it
> > can't be traced.
> > For this definition, the check is not a hack. It's the only way to go.
>
> I agree this is the proper goal for ptrace and this code is not a hack.
> It's a bug. In Linux, uid, euid, suid, gid, egid, and sgid do not by
> themselves determine privileges. And that's what ptrace checks. The main
> determiner of basic privileges is the process capabilities. The euid,
> etc. do also. I have frequently have on my system a process that runs
> with the same uid, euid, etc. as some other process, but should not be
> allowed to ptrace it because the tracee has CAP_DAC_OVERRIDE (the
> privilege to access files in spite of file permissions) and the tracer
> does not.

I'm constantly getting lost in the maze of rules, on what exactly
happens on setuid() etc, but I know that setuid() resets the
capabilities as well. What's the way of changing euid and suid back
to ruid, and yet keeping some capabilities?

> So as with the user space programs I mentioned (where the euid check is
> indeed a hack), I have to fix ptrace too. Fortunately, it looks as simple
> as comparing capability sets.
>
> > Now this definition is really what is needed for the filesystem case
> > too, so I think it's not a hack either.
>
> Maybe I got lost in the problem we were trying to solve, then. What does
> comparing the privileges of one process with those of another have to do
> with this thread about making safe unprivileged mounts via namespaces?

If process A accesses a filesystem controlled by process B, then B
gets ptrace-like ability over A.

Controlling a filesystem means, that a userspace process provides the
data/metadata for the filesystem as requested by the kernel. It could
be a v9fs, FUSE, or NFS served by a userspace entity.

So what exactly can B do with A?

- Know the exact sequence and timing of file operations being
performed. This is an information leak, though without serious
consequences in most cases

- Provide arbitrary data/metadata. Extremely large files, extremly
deep directories, etc. This can lead to DoS, by forcing a more
privileged program to process/store more data, than the user could
possibly provide otherwise (because of quota, or simply filesystem
size limits).

- Determine the time it takes for each filesystem operation. This
is the most serious problem, as delaying an operation will result
in hanging of the requester, possibly causing other hanging
resources (e.g. file locks).

For more detailed analysis of the problem see
Documentation/filesystems/fuse.txt in -mm.

> The post to which I replied said we have to deal with set-uid
> programs. Aren't we talking about the problem where someone sets a
> file's setuid flag on with the assumption that when the program
> within runs, it will see certain files at certain places? And the
> fact that if one could mount whatever he wants, that would violate
> the assumption?

No. That assumption can easily be satisfied, by limiting where the
user can mount. If mounting is only allowed on files/directories that
the user has unlimited write capability, than it cannot modify files
directories that the privileged process has assumptions about.

For example, in a private namespace, it's perfectly safe to allow bind
mounts (without any restrictions about which processes can access it)
on user writable files/directories.

> The two ways suggested of handling that are: 1) after the private
> mount, ignore all setuid flags. 2) after the private mount, don't
> let a program that has gained privileges via set-uid see the
> user-made names.
>
> My point is still that (2) can't be done because you can't know that a
> program has gained privileged via set-uid.
>
> If it's really not about set-uid, but about ptrace-like privilege
> borrowing, please enlighten me.

Yes, it's about ptrace-like privilege borrowing and not about set-uid.

And as such, it probably needs the capability fix that you propose for
ptrace as well.

Thanks,
Miklos
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-05-14 09:09    [from the cache]
©2003-2011 Jasper Spaans