Re: nss-systemd D-Bus call caused by getpwent

[Dominick Grift <dac.override@gmail.com>]

Chris PeBenito <pebenito@ieee.org> writes:

> On 1/6/19 2:33 AM, Dominick Grift wrote:
>> Nicolas Iooss <nicolas.iooss@m4x.org> writes:
>>
>>> Hi,
>>> While testing the current master branch of refpolicy on Arch Linux, I
>>> encountered the following denial:
>>>
>>> type=USER_AVC msg=audit(1546729287.319:440): pid=312 uid=81
>>> auid=4294967295 ses=4294967295 subj=system_u:system_r:system_dbusd_t
>>> msg='avc:  denied  { send_msg } for msgtype=method_call
>>> interface=org.freedesktop.systemd1.Manager member=GetDynamicUsers
>>> dest=org.freedesktop.systemd1 spid=14828 tpid=1
>>> scontext=system_u:system_r:sshd_t tcontext=system_u:system_r:init_t
>>> tclass=dbus permissive=0  exe="/usr/bin/dbus-daemon" sauid=81
>>> hostname=? addr=? terminal=?'
>>>
>>> My OpenSSH server is calling GetDynamicUsers() exposed by systemd over
>>> D-Bus. This call comes from systemd's NSSwitch module and occurs when
>>> OpenSSH calls setpwent() to get information about a user
>>> (https://github.com/systemd/systemd/blob/v240/src/nss-systemd/nss-systemd.c#L676).
>>> How should this be handled by refpolicy? For example, would adding a
>>> call to init_dbus_chat(nsswitch_domain) in a ifdef(`init_systemd')
>>> block be acceptable? This would allow any callers of
>>> auth_use_nsswitch() to be able to communicate with systemd's PID 1
>>> over D-Bus.
>>
>> FWIW I have this in my nss macro too, However I have two nss macros, one
>> base macro and one superset that has this call amongst others
>> (mymachines resolve etc) I only give nss base access to my confined
>> users since they will never have access to any objects associated with
>> userns uids/gids anyways so they shouldnt get into a position where they
>> need to resolve them (except confined sysadm)
>
> I've been dissatisfied with what auth_use_nsswitch() and
> auth_use_pam() have turned into, as I think they're too big.  It's not
> an easy thing to define due them being inherently extensible.  What
> you describe is one possible good direction to go towards.  I was also
> concerned about all of the network access that is allowed by it and
> thought about splitting out the local accesses into a base interface.

I agree, but it gets hard to maintain if you split all the individual
nss modules.

The solution i implemented in my policy also has its limitations and
assumptions, and is pretty much all or almost nothing.

Atleast you have the init_systemd tunable which atleast addresses the
various nss_systemd modules to some degree.

I only allow my confined unpriv users to read passwd and nss config, the
drawback of this is that these shells cannot use /proc/net/"protocol"
which is nice on the one hand for confined shells but it "breaks" bash

Its a tough issue

-- 
Key fingerprint = 5F4D 3CDB D3F8 3652 FBD8 02D5 3B6C 5F1D 2C7B 6B02
https://sks-keyservers.net/pks/lookup?op=get&search=0x3B6C5F1D2C7B6B02
Dominick Grift