Re: [PATCH net-next v3 2/2] vxlan: allow specifying multiple default destinations

[Mike Rapoport <mike.rapoport@ravellosystems.com>]

(added David Stevens to CC)

On Mon, Jun 03, 2013 at 08:57:37AM -0700, Stephen Hemminger wrote:
> On Sun, 2 Jun 2013 13:29:42 +0300
> Mike Rapoport <mike.rapoport@ravellosystems.com> wrote:
> 
> > On Fri, May 31, 2013 at 7:17 PM, Stephen Hemminger <stephen@networkplumber.org> wrote:
> > > Looking at this code in more detail, I see a slew of problems.
> > >
> > > First the list of destinations isn't really a list. The default one is still
> > > embedded in the fdb entry. This means you can't change it safely.
> > >
> > > Also the notification via netlink only sends back a single destination
> > > value.
> > >
> > > And the lack of locking on the open coded link list means it is not safe since
> > > the forwarding table is used with RCU. In order to be safe, proper RCU
> > > barriers would be needed or better yet convert to list_rcu..
> > >
> > > Overall, I feel guilty for not inspecting this more closely and am surprised
> > > that others did not catch the lack of locking.
> > 
> > I've tried to convert remotes list to use hlist_rcu. The patch below implements the conversion, but it does not address the netlink notification issue.
> 
> 
> Ok, let me have a go at this.
> 1. Using list rather than hlist makes more sense for handling simple list

I can replace  hlist with list.

> 2. The complexity is in how do do the netlink API.
 
As far as I can tell, there are two places in the driver that should
report multiple destinations via netlink. They are vxlan_fdb_dump and
vxlan_fdb_restore. These methods can traverse the remote destinations
list and send netlink notification for each list item. 

> 3. If we can't work this out for 3.11, the multiple remotes may have to be reverted;
>    I don't want the existing API to be exposed in a release
> 

I'd really like to help to sort this out ASAP, so that I could update and resend
my patches that, by coinsidence, require list of remote destinations.
I'm not the right person to redefine the fdb API, though.

--
Sincerely yours,
Mike.