> On Aug. 31, 2016, 10:17 p.m., Stephan Erb wrote:
> > src/main/python/apache/aurora/executor/common/sandbox.py, line 239
> > <https://reviews.apache.org/r/51564/diff/3/?file=1489394#file1489394line239>
> >
> >     This changes seems to come with a severe security risk. As an normal 
> > user, I can now gain root on any agent:
> >     
> >     * Prepare a docker/appc container with a manually crafted user with UID 
> > 0 but with my role name.
> >     * Launch the container with said role name.
> >     * The sandbox code will bail out early here and don't proceed to create 
> > an unpriviledged user
> >     * Setuid will switch from root to my prepare custom user with root 
> > permissions
> >     * Game over  
> >     
> >     Unless someone can correct me here, that would be a -1 from my end.
> 
> Joshua Cohen wrote:
>     I'm not sure about step 4 above. Are you referring to the [setuid in 
> process.py](https://github.com/apache/aurora/blob/master/src/main/python/apache/thermos/core/process.py#L369-L380)?
>  If so, that setuid shouldn't be switching to root, it will be switching to 
> the user matching the role name on the host system, the uid set in your 
> docker/appc image wouldn't have any impact on that. Am I missing something?
> 
> John Sirois wrote:
>     Joshua mentioned this in Slack/IRC, but I do think we need to ensure the 
> uid/uname and gid/gname pairs in the chroot match those of the host system 
> when we hit an exists condition in either direction.
>     
>     Given:
>     Job author only specifies a role name, in this example `jsirois`
>     
>     Scenarios:
>     1. host (uid=1000, uname=jsirois) chroot (uid=500, uname=jsirois)
>     2. host (uid=1000, uname=jsirois) chroot (uid=1000, uname=fred)
>     3. host (uid=1000, uname=jsirois) chroot (uid=1000, uname=jsirois)
>      
>     A Job author can have task code that references the role name, for 
> example it might shell out a call to `id -g jsirois` where the role name is 
> `jsirois` to find the primary group id for the current role.  It seems then 
> that we must ensure the chroot has the role name available, and fwict, 
> besides the special case of uid 0, we don't really care what the uid is.  If 
> it matches that's fine, but since the chroot environment will share nothing 
> with the host, ids need not match (IIUC).
>     
>     So it seems to me scenarios 1 and 3 are OK - the sandbox can move along.  
> Scenario 2 though should fail (we currently pass).
> 
> John Sirois wrote:
>     > Joshua mentioned this in Slack/IRC, but I do think we need to ensure 
> the uid/uname and gid/gname pairs in the chroot match those of the host 
> system when we hit an exists condition in either direction.
>     
>     Obviously I changed my thinking as I composed the 
> Given/Scenarios/Conclusion below that statement!
> 
> Stephan Erb wrote:
>     Excellent idea to spell out the scenarios explicitly.
>     
>         [W]e don't really care what the uid is.  If it matches that's fine, 
> but since the chroot environment will share nothing with the host, ids need 
> not match (IIUC).
>         
>     I would disagree here. With container mounts, the host and a container 
> can share parts of the filesystem. As filesystem permissions are based on 
> IDs, we have to make sure those match inside and outside of the container.
>     
>     This implies that 'Scenario 3' would be the only acceptable one. We have 
> to abort the container launch in Scenario 1 and 2.
> 
> John Sirois wrote:
>     I concur.  If we'll (Aurora or Mesos) supply mounts into the chroot then 
> 3 is the only acceptable scenario.

tl'dr +1 for checking uid/gid matches between the image filesystem and host for 
now.

This is a quite complex problem, and IMO one of the aspects that Linux 
containerization and cluster management layer generally haven't really solved 
well.

1. In our private use case, because we control all images running, we actually 
made special arrangements to make sure relevent users have the same uid/gid 
between image and host system;
2. The above solution clearly doesn't scale for people who don't 100% control 
their build. user namespace support should the savior here, but unfortunately 
it often conflicts with network namespace, therefore neither docker daemon nor 
mesos really support it well;
3. For your quote of `With container mounts, the host and a container can share 
parts of the filesystem. ` However, only certain parts we want the container to 
share with host system (pretty much everything mounted into the container) will 
indeed be shared. This also depending on how we want volume support for Mesos 
containerizer be.
4. Also, I think the risk you mentioned right now is also possible if people 
use DockerContainerizer and specify a malicious `--volume` argument value.


- Zhitao


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/51564/#review147497
-----------------------------------------------------------


On Aug. 31, 2016, 8:56 p.m., Zhitao Li wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/51564/
> -----------------------------------------------------------
> 
> (Updated Aug. 31, 2016, 8:56 p.m.)
> 
> 
> Review request for Aurora, Joshua Cohen, John Sirois, and Zameer Manji.
> 
> 
> Bugs: AURORA-1761
>     https://issues.apache.org/jira/browse/AURORA-1761
> 
> 
> Repository: aurora
> 
> 
> Description
> -------
> 
> Allow E_NAME_IN_USE in useradd/groupadd.
> 
> 
> Diffs
> -----
> 
>   src/main/python/apache/aurora/executor/common/sandbox.py 
> a172691e164cf64792f65f049d698f9758336542 
>   src/test/python/apache/aurora/executor/common/test_sandbox.py 
> 57ab39e2444100c3a689bb0ff745c62f7bc2f1a6 
> 
> Diff: https://reviews.apache.org/r/51564/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Zhitao Li
> 
>

Reply via email to