[Lustre-devel] Security issues

Eric Mei Eric.Mei at Sun.COM
Fri Aug 8 10:44:06 PDT 2008

Peter Braam wrote:
> On 8/8/08 11:03 AM, "Eric Barton" <eeb at sun.com> wrote:
>     1. Securing bulk data.
>     It seems to me that it _is_ appropriate to use the GSSAPI to secure the
>     transfer of bulk data between client and server since it's
>     effectively just
>     another message.  I can see (at least naively) that it would be good to
>     avoid double encryption in the case where file contents are actually
>     stored
>     encrypted on disk.  
> You are not telling me that we are going through a lot of re-design, 
> that we are encrypting data and that then we are not storing it 
> encrypted on disk?  Come on, adding an EA with a key to decrypt is not 
> so hard and one gets lots of value from it.
>     But even in this case, don't we still have to sign the
>     (encrypted) bulk so that the receiver can be sure it arrived intact?
> Well, yes, but as I indicated you can sign the hash that is stored on 
> (ZFS) disk for this.  That avoids generating the hash twice.  So I am 
> really not convinced yet.

Peter, are you saying that except using NASD-style protocol, we don't 
need to encrypt/sign bulk data at all?

> The issue is not the message mechanism, but is what identity to use for 
> GSS to authenticate and how to manage and revoke that etc.

Here we only want to protect on-wire data, the gss authentication is 
only for the "node", not particular user, as you pointed out previously.

>     2. Securing Capabilities.
>     If we want to be sure that a Capability given to client A cannot be
>     snooped and used by client B we either (a) have to make the Capability
>     secret (i.e. never passed in cleartext) or (b) have to make the
>     Capability
>     identify which client it is valid for.
>     It seems to me that (b) is preferrable since it ensures that a malicious
>     client cannot leak Capabilities to a 3rd party.  The downside is
>     that this
>     multiplies the number of unique Capabilities by the number of clients,
>     potentially increasing CPU load when 1000s of clients all open the same
>     shared file and each require unique Capabilities to access the
>     stripe objects.
>     Do we have a feel for how bad this could be?
> Yes, very bad, and it is absolutely necessary to have an option that 
> avoids this (also 1000s is out of date – it could be 100x worse).  That 
> option could be to simply not have security on the compute cluster if 
> customers agree with this.
> We also need to discuss your proposals with a review committee from LLNL 
> and Sandia, as we did during the PF discussions.

We're trying to figure out a way to generate only one capability for 
each MD object, but somehow mingled with per-export data to generate 
client-unique capability, but till now we haven't found a good solution.

The other thought is using some kind of light-weight, but still 
reasonably secure hash algorithm. By changing the KEY frequently enough 
(e.g. every 2 hours) we can still be secure. But we'v no idea what hash 
algorithm could fit our needs.


More information about the lustre-devel mailing list