Message-ID: <20090415155702.GC21448@hmsreliant.think-freely.org>
Date:	Wed, 15 Apr 2009 11:57:02 -0400
From:	Neil Horman <nhorman@...driver.com>
To:	Vlad Yasevich <vladislav.yasevich@...com>
Cc:	Christoph Lameter <cl@...ux.com>, netdev@...r.kernel.org,
	David Miller <davem@...emloft.net>,
	David Stevens <dlstevens@...ibm.com>
Subject: Re: [PATCH] Multicast: Avoid useless duplication of multicast
	messages

On Wed, Apr 15, 2009 at 10:41:03AM -0400, Vlad Yasevich wrote:
> Neil Horman wrote:
> > On Wed, Apr 15, 2009 at 08:51:35AM -0400, Christoph Lameter wrote:
> >> On Wed, 15 Apr 2009, Neil Horman wrote:
> >>
> >>>> How are you going to do this?
> >>> I'm going to write each application to use 2 sockets, one bound to each
> >>> multicast group.  Thats the way it works now.  I think you missed the obvious in
> >>> your construction of this example.
> >> Ok it could be done with binding. But you would need 3 sockets. One per MC
> >> groups bound to a MC group each and then one for the replies (hmmm...
> >> looks like you could use SO_BINDTODEVICE on one socket to get around the
> > Depending on your setup, 2 is perfectly sufficient.  In fact 1 can be sufficient
> > if you want to filter in your application, but we've been over that.
> > 
> >> third one --there is even an exception case for this in inet_bind causing
> >> more weird semantics-- but then the application needs to know the device
> >> name of the NIC, argh)
> > Of course it does, but thats zero incremental cost, since you need to know the
> > device name anyway, when specifying the ifindex to the join request.
> > 
> >>>> Its trivial to do with this patch and one
> >>>> socket in each process listening to port 4711 that subscribes to the
> >>>> necessary multicast groups.
> >>> Its trivial without the patch as well.
> >> I do not see how you can justify making such a statement.
> >>
> > I find it justified because I don't see an application using 2 or 3 sockets and a poll or
> > select call as anything more than trivial.  If you find that to be non-trivial,
> > perhaps a refresher programming course might be in order for you?
> > 
> >>> It boils down to this:  This is the way multicast subscriptions have worked in
> >>> bsd, linux, and presumably various other unix and non-unix operating systems for
> >>> lord only knows how long.  Provide some documentation that shows its in
> >>> violation of a newer standard, or that it is common practice to behave
> >>> differently on another OS (such that including this directive would make porting
> >>> easier).  As it stands currently, this patch only serves to create a crutch to
> >>> perpetuate misundersandings about how the behavior currently works.
> >> The way things work is counterintuitive and leads to weird code constructs
> >> with the application having to manage multiple sockets because weird
> >> semantics have developed over the years.
> > 
> > The way things work is counterintuitive to _you_ (is it was to me a few months
> > ago).  That asside, I came to understand how this actually works, how it has
> > worked for decades, and how programmers have successfully written applications
> > that use this model over that time period.  Can we modify the model?  Sure.
> > Should we?  I certainly don't see any need, given that it does little except
> > change the model.  For those who understand it, its compltely useless.  I'm
> > willing to concede that I'm wrong, but not without some modicum of evidence that
> > this change will benefit existing applications.  If some other operating system
> > adheres to the model you expect it to, perhaps this has legs, but I don't know
> > of any that do.  The current model, even if counter intuitive, is well defined,
> > well understood, and documented.  I fail to see how adding an alternate,
> > undocumented model (that may itself be counterintuitive to all the developers
> > who have developed under the current model) adds anything significant.
> > 
> > Neil
> 
> Hi Neil
> 
> This has been somewhat bugging me for a while, so I went digging.
> 
> Here is a rather pertinent text that points out that we "might" have a bug.
> RFC 4607:
> 
> 4.2.  Requirements on the Host IP Module
> 
>    An incoming datagram destined to an SSM address MUST be delivered by
>    the IP module to all sockets that have indicated (via Subscribe) a
>    desire to receive data that matches the datagram's source address,
>    destination address, and arriving interface.  It MUST NOT be
>    delivered to other sockets.
> 
I'll let David respond more fully, since I'm not familiar with this RFC, but a
quick read would suggest that (from the abstract), this only applies to a subset
of addresses, which are not being used in the application in question here.
>From what I read, the RFC defines an extenstion to the sockets api which allows
you to subscribe to a multicast group from a specific source, using one of the
reserved muticast ranges provided in the abstract.  It appears that we support
this RFC via the IP_ADD_SOURCE_MEMBERSHIP socket option.  Now, if we allow
sockets that issue IP_ADD_SOURCE_MEMBERSHIP calls to receive datagrams from
multicast addresses within the range defined by the rfc from other sources that
they have not subscribed to, yes we have a bug, but thats not overly relevant I
think to Christophs problem, since he's using the any-source model, and its
corresponding addresses.  Switching to the specific-source model would solve his
immeidate problem here that we've been debating, but would likely introduce a
new set, in that he would then have to write his app to subscribe to the myrriad
of sources that are sending to that multicast group.

> 
> Additionally, RFC 3678 describes IP_ADD_MEMBERSHIP as an 'any-source group'
> and is allowed by the SSM spec.  This is also how it is implemented in the kernel.
> However, we do not appear to perform the filtering required by the above quoted
> section 4.2.  
Very true, so we may have a bug in the SSM model, but again, thats not what
Christoph is using, its the any-source model, using group address unrelated to
the ssm RFC.

In particular, if we fail to match the 'datagram's destination address',
> we deliver the packet, which I believe is in violation of the "MUST NOT" above.
> 
I think only if the SSM model is used via the socket extensions the RFC
describes.  If Christophs app is subscribing via IP_ADD_SOURCE_MEMBERSHIP, then
yes, we have a problem.  But everything I've read says he uses the standard, any-source
IP_ADD_MEMBERSHIP option which I think makes assertions from RFC 4607 void.
Christoph, are you using IP_ADD_SOURCE_MEMBERSHIP?

Neil

> I've CC'd Dave Stevens, since I'd like to hear his opinion regarding this text.
> 
> Thanks
> -vlad
> 
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists