Re: [TrouSerS-tech] [tpmdd-devel] [PATCH 09/13] tpm: Pull everything related to sysfs into tpm-sysfs.c

[Jason Gunthorpe <jgunthorpe@obsidianresearch.com>]

On Thu, Oct 10, 2013 at 07:42:49AM +0000, Fuchs, Andreas wrote:

> In any case, I like your idea to split trousers IPC to two distinct
> unix sockets for localities. In this case, we could also split tcsd
> into two processes along with it for accessing the distinct
> char-devices and thereby make it more robust against bugs for
> "locality-escalation".

You still have to somehow manage cross locality state between the two
daemons..
 
> Also remember that many people have developed alternative stacks
> that don't use trousers but operate directly on the char-device.
> They would also benefit from char-device access control for localities.

I am one of those people, we actually don't use any middleware at
all. But to make that work I've had to carry the multi-open patch for
years :|

> Even with only a single trousers, I see no harm in two devices. For
> backwards compatibility, the current /dev/tpm0 could be exported (with
> highest level access control) along with tpm0l1, tpm0l2, ... and/or 
> trousers could open both char-devices if it wanted to.

Well, we could start with a 'no way out IOCTL'. So trousers can open
/dev/tpm twice and lock the two FDs to a specific locality then drop
privileges and fork priv-sep style sub processes.

The current kernel code is not ready for multiple char devices, it
will need a device class first..

> The kernel may want to use localityAtRelease OS in order to protect sealed
> data (trusted keyrings) such that user-space could not even unseal

It seems reasonable to have TPM data that will only live in the kernel
to be only releasable by the kernel..

Jason