[ovs-dev] DNS support options

[Ben Pfaff]

On Tue, Aug 08, 2017 at 04:16:21PM -0400, Lance Richardson wrote:
> > From: "Ben Pfaff" <blp at ovn.org>
> > To: "Lance Richardson" <lrichard at redhat.com>
> > Cc: dev at openvswitch.org
> > Sent: Tuesday, August 8, 2017 3:48:15 PM
> > Subject: Re: [ovs-dev] DNS support options
> > 
> > On Tue, Aug 08, 2017 at 02:59:21PM -0400, Lance Richardson wrote:
> > > > From: "Ben Pfaff" <blp at ovn.org>
> > > > To: dev at openvswitch.org
> > > > Sent: Tuesday, August 8, 2017 2:43:17 PM
> > > > Subject: [ovs-dev] DNS support options
> > > > 
> > > > Open vSwitch currently supports only IP addresses for specifying
> > > > OpenFlow and OVSDB remotes.  It would be nice to support DNS names.
> > > > Open vSwitch can't afford to let DNS resolution block processing, so it
> > > > needs some way to resolve DNS asynchronously.  I've come up with the
> > > > following solutions so far:
> > > > 
> > > >         * Implement it ourselves using threads.  Probably, this would be
> > > >           a single worker thread that loops through calls to
> > > >           getaddrinfo().
> > > > 
> > > >         * glibc has getaddrinfo_a(), an asynchronous version of
> > > >           getaddrinfo().  *BSD and Windows do not have this function, so
> > > >           we would still need to implement it ourselves (probably using
> > > >           threads) if we adopted this approach, so it might be more work
> > > >           than the first solution.
> > > > 
> > > >         * Use an external asynchronous DNS library.  GNU adns looks like
> > > >           a good choice: http://www.chiark.greenend.org.uk/~ian/adns/.
> > > >           It has a port to Windows but I don't know whether it's
> > > >           up-to-date.
> > > > 
> > > 
> > > With adns_beforepoll()/adns_afterpoll(), it looks like adns might be easy
> > > to integrate on top of lib/poll-loop.c.  Not requiring a separate thread
> > > seems like a plus.
> > 
> > Oops, I just realized that adns is GPL'd, which means that linking it
> > against OVS would force OVS to be GPL'd, so it's probably not a good
> > choice.
> > 
> > Other possible choices:
> > 
> >         - c-ares (https://c-ares.haxx.se/).  Main disadvantage seems to
> >           be that it's very "select" oriented.
> > 
> 
> c-ares has an API to retrieve the set of socket descriptors that need to
> be monitored for readability/writability, so integration shouldn't be much
> work. A socket creation callback function can be registered as well, not
> sure if it would be needed but it's nice that it's there. Looks like a
> fairly active project with some contributors I recognize.

This API?  https://c-ares.haxx.se/ares_fds.html
It seems inconvenient because it returns the fds as a fd_set, which the
caller then has to scan for 1-bits.

This is however a fairly petty grievance if it's the only downside.