Re: chroot's userspec option

[Pádraig Brady]

On 02/28/2014 10:34 AM, Ken Werner wrote:
> On 02/27/2014 05:45 PM, Pádraig Brady wrote:
>> On 02/27/2014 03:48 PM, Ken Werner wrote:
>>> Hi,
>>>
>>> I noticed when using chroot's --userspec option the Gnulib's 
>>> parse_user_spec function gets called that leads the glibc to dlopen 
>>> libnss_compat.so.2 (probably getpwnam() that triggers the libc's NSS 
>>> mechanism). Since parse_user_spec is called after the chroot system call 
>>> the new root directory will be searched. I guess this means that the chroot 
>>> utility attempts to parse the user spec in the "guest" environment. Is this 
>>> behavior intended? In my case the chroot environment contains a 
>>> libnss_compat.so.2 that's not compatible and the chroot utility fails with:
>>>
>>> /usr/bin/chroot: relocation error: /lib/libnss_compat.so.2: symbol 
>>> _nss_files_parse_pwent, version GLIBC_2.0 not defined in file libc.so.6 
>>> with link time reference
>>>
>>> As soon as I LD_PRELOAD libnss_compat.so.2 the "host" environment is used 
>>> to parse the user spec. If this is the intended behavior it would be better 
>>> if chroot calls the parse_user_spec prior issuing the chroot syscall. Any 
>>> thoughts? :)
>>
>> This issue was noted previously with an explicit workaround:
>>    http://lists.gnu.org/archive/html/coreutils/2011-07/msg00057.html
>>
>> Then again with an implicit workaround:
>>    http://lists.gnu.org/archive/html/coreutils/2012-05/msg00018.html
>>
>> I had mentioned an amendment to that but there was no response.
>>
>> I'll look at a fix now to do:
>>
>> t_ids = parse_user_spec(); //outside chroot
>> ids = parse_user_spec(); //inside chroot
>> if (!ids)
>>    ids = t_ids;
> 
> Thank you for providing those pointers! I have to admit that it's still not 
> clear to me whether the userspec option is supposed to lookup the user/group 
> using the A) the old or B) the new root. In case of A) the fix would be call 
> parse_user_spec prior switching to the new root. While B) is not trivial to 
> support imho. The way it's implemented by now assumes the libc's NSS plugins 
> of the new root are compatible to the libc of the old root. As you noticed 
> that's not the case when chrooting into a 32bit userland on a 64bit system 
> (and there are many more cases).
> 
> Since I do not really depend on the uid/gid lookup I wondered why getpwnam() 
> and getgrnam() are still called even if numeric IDs are provided rather than 
> the names. It turns out the code [1] only skips the lookup if the IDs are 
> prefixed with '+'. For example:
>   chroot --userspec=+1234:+1234 /path/to/new/root
> 
> Unfortunately the --groups option doesn't have a way to skip the lookup 
> currently [2]. It calls getgrgid/getgrnam that probably trigger libc's NSS 
> plugins as well.
> 
> I guess the first thing would be to discuss and decide which approach is the 
> desired one - then to post patches that changes the docs+code accordingly. 
> /me ducks ;)
> 
> [1] Gnulib's userspec.c:parse_user_spec calls parse_with_separator that skips 
> the lookup in case the first char is a '+':
> http://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=blob;f=lib/userspec.c;h=1be9266eb54638a2624d0a9205d8e68fd516205e;hb=HEAD#l160
> 
> [2] Coretutil's chroot.c:main calls set_additional_groups() that calls 
> getgrgid/getgrnam:
> http://git.savannah.gnu.org/gitweb/?p=coreutils.git;a=blob;f=src/chroot.c;h=50bb2537ea7df4b963e151bbcb54c217533f32d0;hb=HEAD#l66

Thanks for looking into the details.

Yes the `chroot --userspec=+1234:+1234 --groups=+123,+456 /new/root` technique 
is a good way
to specify that the lookup is done outside. I.E. one can do the name lookup 
outside like:

  group_ids() { printf '%s\n' "$@" | xargs -n1 id -gr | sed 's/^/+/' | paste -s 
-d,; }
  chroot --userspec=+$(id -ur user):+$(id -gr user) --groups=$(group_ids group1 
group2) /new/root

Now the above is a bit awkward, but also not the usual case.
I.E. one only needs to specify a +number when _enforcing_ outside lookup.
I.E. that's only needed when there are different IDs inside and out,
and we want to use the outside ones. We'll document this at least.
We should also document that this is more secure in the unusual
case where the whole chroot is untrusted, as less code is executed
between the chroot() and the setuid().

Now to support that, --groups will need to be adjusted as you mention.
Currently is does getgrgid() to better validate the passed group IDs,
by providing diagnostics for a particular invalid ID in the list.
Now that could be changed to only do the lookup if the setgroups fails.
Also I notice that --groups assumes an ID if numeric, even without a leading +.
That's inconsistent, so we should fix that up too (we already lookup the
name in all cases now, so adding the + constraint to avoid that will
not be adding new restrictions).

So with the above in place we can enforce lookup outside.
But again needing that is unusual, and we can just do name
lookups as normal like:

 t_ids = parse_user_spec(); //outside chroot
 ids = parse_user_spec(); //inside chroot in case different
 if (!ids)
    ids = t_ids;

thanks,
Pádraig.