[libvirt] [PATCH] Updated to deal with specifying user IDs to that do not map to usernames

[Laine Stump]

On 06/07/2016 11:15 AM, Peter Krempa wrote:
> On Tue, Jun 07, 2016 at 15:50:38 +0100, Daniel Berrange wrote:
>> On Tue, Jun 07, 2016 at 08:24:14AM +0200, Peter Krempa wrote:
>>> On Mon, Jun 06, 2016 at 14:25:23 -0500, Roy Keene wrote:
>>>> Patch to libvirt master to avoid failing when a user ID is specified,
>>>> e.g. for <seclabel type='dac'>, that does not map to a user name.
>>>>
>>>> This is useful if you want to run each VM as a separate user and not
>>>> bother creating an /etc/passwd entry for each UID.
>>> For this use case you shall prefix the name with a +. Please refer to
>>> the documentation on seclabels.
>> Empirically that does not currently work:
>>
>> # virsh dumpxml serial | grep --after 2 '<seclabel'
>>    <seclabel type='static' model='dac' relabel='no'>
>>      <label>+21421:+12421421</label>
>>    </seclabel>
>>
>> # virsh start serial
>> error: Failed to start domain serial
>> error: Cannot access storage file '/var/lib/libvirt/images/demo.qcow2' (as uid:21421, gid:12421421): Success
>>
>> # ls -al /var/lib/libvirt/images/demo.qcow2
>> -rw-r--r--. 1 21421 12421421 197120 Apr 30  2015 /var/lib/libvirt/images/demo.qcow2
>>
>>
>> Looking at the libvirtd logs we see
>>
>> 2016-06-07 14:49:13.724+0000: 13490: debug : qemuDomainCheckDiskPresence:3954 : Checking for disk presence
>> 2016-06-07 14:49:16.551+0000: 13490: error : virGetUserEnt:801 : Failed to find user record for uid '21421'
>> 2016-06-07 14:49:16.551+0000: 13490: error : virStorageFileGetMetadataRecurse:3114 : Cannot access storage file '/var/lib/libvirt/images/demo.qcow2' (as uid:21421, gid:12421421): Success
>>
>>
>> So even though the QEMU driver has honoured the '+' syntax, some of the
>> things QEMU is calling appears to be trying to resolve the UID back into
>> a user password record and failing.
> Indeed.  This is caused by calling virFileAccessibleAs which calls
> virGetGroupList so that it can add all groups for the given UID which is
> very strange. The group list is then set after forking at attempting to
> check file presence. I hate root squashed NFS.
>
> I guess we could just skip reporting the error if we can't get the
> group list in that case and just set the provided numerical UID/GID and
> try it that way.

Without taking the time to wind through the code - are you saying that 
we're attempting to get the grouplist for a different uid while running 
as root, then doing a fork/seteuid, and then adding the groups that we 
learned pre-fork?

If that's the case, then I think the same results can be achieved by 
getting the grouplist *after* seteuid(). It looks like getgroups() gets 
the groups for the current uid, while mgetgroups() gets the groups for a 
specified username, so getgroups() would avoid the need for a username.