Re: [PATCH 7/9] AF_UNIX: Documentation on multicast Unix Sockets

[Alban Crequy <alban.crequy@collabora.co.uk>]

Le Mon, 22 Nov 2010 21:07:40 +0200,
"Rémi Denis-Courmont" <remi@remlab.net> a écrit :

> Le lundi 22 novembre 2010 20:36:20 Alban Crequy, vous avez écrit :
> > +Multicast Unix sockets
> > +======================
> > +
> > +Multicast group memberships are stored in struct unix_mcast nodes.
> > An Unix +socket can join several multicast groups. Struct
> > unix_mcast nodes are doubly +linked:
> > +- In (struct unix_sock)->mcast_subscriptions
> > +- In (struct unix_sock)->mcast_members
> 
> I may be stupid, but I found this whole documentation very confusing,
> and so the API it tries to describe. Traditionally:
> - Senders may or not may be part of the group and are not kept track
> of.
> - Receivers join to the group then receive message sent to it.
> - Loopback defines whether a sender receives its own echo if it sends
> to a group that it has joined.
> - If connected to a multicast group, messages from the socket are
> routed to the group (in absence of a contradictoy socket address).
> This has no effect on membership to the multicast group under any
> circumstance.

I keep these traditional properties for multicast on Unix sockets.

> You cannot 'listen' or 'accept' on a multicast group.

Datagram sockets cannot listen() or accept() but seqpacket sockets can.
I would like multicast to work on seqpacket sockets too. In this case,
there is a central daemon who listen(), and accept() returns a new
socket. The central daemon controls the lifetime of the multicast
group and can receive the messages from the peers on the socket
returned by accept() if UNIX_MREQ_SEND_TO_PEER is set.

The accepted socket could join the multicast group (and then receive
messages addressed to the group) with the setsockopt() call, but then
there would be a race that it may not receive the first messages if a
peer connect() and send a message immediately afterwards. connect() can
returns on the peer process before the daemon accept() and runs
setsockopt(). I added the flag UNIX_MREQ_AUTOJOIN (to be set when
creating the multicast group) to prevent that race.

Using connected sockets (seqpacket) is useful for D-Bus because a
central daemon can know when members are connecting and disconnecting
and then emit the D-Bus signal 'NameOwnerChanged'.

> So I am not entirely clear what semantics your patchset is following.
> But it does not seem like "multicast" to me and therefore seems not
> very well documented :-(

I am willing to improve it.

-- 
Alban