[ovs-dev] OVN: Compromised Chassis Mitigation

[Mickey Spiegel]

On Tue, Mar 14, 2017 at 12:01 PM, Lance Richardson <lrichard at redhat.com>
wrote:

>
>
> ----- Original Message -----
> > From: "Mickey Spiegel" <mickeys.dev at gmail.com>
> > To: "Lance Richardson" <lrichard at redhat.com>
> > Cc: "Russell Bryant" <russell at ovn.org>, "devovs" <dev at openvswitch.org>
> > Sent: Tuesday, March 14, 2017 2:27:30 PM
> > Subject: Re: [ovs-dev] OVN: Compromised Chassis Mitigation
> >
> > On Tue, Mar 14, 2017 at 11:14 AM, Lance Richardson <lrichard at redhat.com>
> > wrote:
> >
> > > > From: "Russell Bryant" <russell at ovn.org>
> > > > To: "Mickey Spiegel" <mickeys.dev at gmail.com>
> > > > Cc: "Lance Richardson" <lrichard at redhat.com>, "devovs" <
> > > dev at openvswitch.org>
> > > > Sent: Tuesday, March 14, 2017 1:48:55 PM
> > > > Subject: Re: [ovs-dev] OVN: Compromised Chassis Mitigation
> > > >
> > > > On Tue, Mar 14, 2017 at 5:08 AM, Mickey Spiegel <
> mickeys.dev at gmail.com>
> > > > wrote:
> > > > >>       - An "authorization" column containing a set of "string"
> type,
> > > where
> > > > >>         each string is the name of a column (or column:key) that
> must
> > > > >> contain
> > > > >>         the ID of client attempting the transaction (CN field from
> > > client
> > > > >>         certificate). If this set is empty, all IDs are
> considered to
> > > be
> > > > >>         authorized.  If this set contains more than one string, at
> > > least
> > > > >> one
> > > > >>         must contain the client ID in order to be considered
> > > authorized.
> > > > >
> > > > >
> > > > > This is the "where" column in the RBAC approach, where the "CN = "
> > > > > part of the logical expression is implied, and only the column
> part is
> > > > > specified.
> > > > >
> > > > > When access is attempted from an ovn-controller, then the
> > > > > authorization rule applies. However, there are other things
> > > > > accessing OVN SB DB that are allowed to change these things
> > > > > even though they do not have a CN, or at least not one that
> > > > > looks like a chassis name. For example, ovn-northd.
> > > > > How is this controlled?
> > > > > This is exactly why the indirection through role is suggested, to
> > > > > allow access to various things without hardcoding specific logic
> > > > > to determine what is subject to the rules.
> > > >
> > > >
> > > > I had imagined we'd deploy OVN such that ovn-northd would connect to
> a
> > > > different ovsdb remote that didn't have ACLs turned on.
> > > >
> > >
> > > I think we do need a way to specify whether ACLs should be applied to
> > > transactions on a given client connection.
> > >
> > > We could have ACLs implicitly apply only to read-only connections, but
> > > I think it would be better to make this an explicit configuration item.
> > >
> > > One way to do this would be to add a "role" attribute to each remote
> and
> > > have a table to map "role" to a specific ACL table.
> > >
> > > We could also consider adding an ACL table reference column to the
> > > OVN_Southbound "Connection" table and make ACLs part of remote
> > > configuration; if an ACL table is configured for a particular remote,
> > > ACLs from that table would be used for transactions on that remote.
> > >
> >
> > I don't follow. What are the definitions of ACL table and the reference?
> >
> > I thought there is only one ACL / Permissions table. The reason for
> > the indirection through "role" is to identify the rows in the table that
> > apply. There may be multiple rows that apply.
> >
>
> I was suggesting a simplification that would eliminate the need for that
> indirection by having the remote refer directly to an ACL/Permissions
> table. The drawback to this simplification would be having to modify
> the db schema in order to add a new role (a new ACL/Permissions table
> would be needed for each added role), versus being able to define
> new roles and rule sets dynamically.  For OVN_Southbound, this doesn't
> seem to be much of a drawback since we're currently only envisioning
> one role needing ACLs (ovn-sbctl, ovn-northd, etc. would continue to
> use read-write sessions without ACLs.
>
>    lance
>
> > I guess another alternative is to have row just be a column in the
> > ACL / Permissions table.
>

s/row/role

This would be almost the same as what you have suggested,
except no need to change the schema when a new role is
added.

The difference between this and the indirection that I suggested
earlier is that ACL/Permission rows could not be shared between
roles.

Mickey


> >
> > Mickey
> >
> >
> > > I think either approach would provide similar flexibility, the
> > > second seems a little more straightforward.
> > >
> > > Thoughts?
> > >
> > > Thanks,
> > >
> > >    Lance
> > >
> > > > Currently, our OpenStack deployment always co-locates ovn-northd with
> > > > the OVN databases, so it just connects over a unix socket.
> > > >
> > > > --
> > > > Russell Bryant
> > > >
> > >
> > >
> >
>