Re: [Ksummit-discuss] [TrouSerS-tech] TPM MiniSummit @ LinuxCon Europe

[Andy Lutomirski <luto@amacapital.net>]

On Tue, Oct 7, 2014 at 10:54 AM, Stefan Berger
<stefanb@linux.vnet.ibm.com> wrote:
> On 09/23/2014 12:42 PM, Andy Lutomirski wrote:
>>
>> On Sep 22, 2014 2:07 AM, "Peter Huewe" <PeterHuewe@gmx.de> wrote:
>>>
>>> Hi,
>>>
>>> I would like to 'invite' all interested parties in a short TPM minisummit
>>> where we can discuss the following hot topics of the TPM subsystem over a
>>> beer or two:
>>>   - State of the TPM Subsystem
>>>   - De-/Initialization Mess
>>>   - Devm'ification
>>>   - Testing
>>>   - TPM 2.0 Support
>>>   - Dependencies / interaction with other subsystems (e.g. keyring / IMA)
>>>   - Status of old 1.1b TPM drivers, deprecation plans
>>>   - ...
>>>
>> I am unlikely to be there, but I have a feature request / food for
>> thought:
>>
>> Using a mandatory userspace daemon (e.g. trousers) for TPM access
>> sucks.  Might it be possible to teach the kernel to handle context
>> save and restore and let multiple processes open the device at once?
>> Then a daemon wouldn't be necessary.
>
>
> Why add the complexity of swapping of authenticated sessions and keys into
> the kernel if you can handle this in userspace? You need a library that is
> aware of the number of key slots and slots for sessions in the TPM and swaps
> them in at out when applications need them. Trousers is such a library that
> was designed to cope with the limitations of the device and make its
> functionality available to all applications that want to access it.
>

Trousers is a daemon, not a library, and it's really quite scary.

Admittedly, my information may be a bit out of date, but trousers
contains way too much code (it has layers in the server!), it has
parsers and serializers of questionable safety (I found one critical
bug *by accident* a couple years ago), it listens on a TCP socket
(this should really be a UNIX socket under /run), it's heavy-weight,
and it does far more than necessary (all it needs to do is context
switching).

Also, Trousers is quite unfriendly to non-Trousers-using programs
(e.g. chapsd/trunks).

If the kernel helped with context switching, then user programs that
actually want Trousers' functionality could run their own copies.

--Andy