On Wed, 30 Mar 2011, John Baldwin wrote:
On Wednesday, March 30, 2011 7:27:36 am Randall Stewart wrote:
John:
The original complaint on this came from Daniel... I believe he
claimed that up until bms's multi-cast work.. you would NOT
get a packet sent to you if you did not join the multi-cast group.
Not necessarily. :( See below..
I will also comment in-line below...
On Mar 29, 2011, at 2:01 PM, John Baldwin wrote:
On Wednesday, January 19, 2011 2:07:16 pm Randall Stewart wrote:
Author: rrs
Date: Wed Jan 19 19:07:16 2011
New Revision: 217592
URL: http://svn.freebsd.org/changeset/base/217592
Log:
Fix a bug where Multicast packets sent from a
udp endpoint may end up echoing back to the sender
even with OUT joining the multi-cast group.
Reviewed by: gnn, bms, bz?
Obtained from: deischen (with help from)
Modified:
head/sys/netinet/udp_usrreq.c
Modified: head/sys/netinet/udp_usrreq.c
==============================================================================
--- head/sys/netinet/udp_usrreq.c Wed Jan 19 18:20:11 2011
(r217591)
+++ head/sys/netinet/udp_usrreq.c Wed Jan 19 19:07:16 2011
(r217592)
@@ -479,11 +479,13 @@ udp_input(struct mbuf *m, int off)
* and source-specific multicast. [RFC3678]
*/
imo = inp->inp_moptions;
- if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) &&
- imo != NULL) {
+ if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) {
struct sockaddr_in group;
int blocked;
-
+ if(imo == NULL) {
+ INP_RUNLOCK(inp);
+ continue;
+ }
bzero(&group, sizeof(struct sockaddr_in));
group.sin_len = sizeof(struct sockaddr_in);
group.sin_family = AF_INET;
So it turns out that this is a feature, not a bug, and is how multicast has
always worked. Specifically, if you bind a UDP socket with a wildcard
address, it should receive all traffic for the bound port, unicast or
multicast. When you join a group, you have switched the socket into a mode
where it now has a whitelist of acceptable multicast groups, but if a socket
has no joined groups, it should receive all multicast traffic, not none. This
change breaks that.
I did not find this behavior intuitive at first, but it does seem to be
required. Note the description of IP_ADD_MEMBERSHIP from RFC 3678 for
example:
I agree getting a packet that is coming to your port without joining the
multi-cast group is not intuitive to me...
3. Overview of APIs
There are a number of different APIs described in this document that
are appropriate for a number of different application types and IP
versions. Before providing detailed descriptions, this section
provides a "taxonomy" with a brief description of each.
There are two categories of source-filter APIs, both of which are
designed to allow multicast receiver applications to designate the
unicast address(es) of sender(s) along with the multicast group
(destination address) to receive.
o Basic (Delta-based): Some applications desire the simplicity of
a delta-based API in which each function call specifies a
single source address which should be added to or removed from
the existing filter for a given multicast group address on
which to listen. Such applications typically fall into either
of two categories:
+ Any-Source Multicast: By default, all sources are accepted.
Individual sources may be turned off and back on as needed
over time. This is also known as "exclude" mode, since the
source filter contains a list of excluded sources.
+ Source-Specific Multicast: Only sources in a given list are
allowed. The list may change over time. This is also known
as "include" mode, since the source filter contains a list
of included sources.
This API would be used, for example, by "single-source"
applications such as audio/video broadcasting. It would
also be used for logical multi-source sessions where each
source independently allocates its own Source-Specific
Multicast group address.
Not the above document is talking about a receiver that as joined the
multicast group (or is joining it and wants some filtering)... I don't
see how that applies to a UDP socket that has NOT joined the M-cast group..
.....
4.1.1. IPv4 Any-Source Multicast API
The following socket options are defined in <netinet/in.h> for
applications in the Any-Source Multicast category:
Socket option Argument type
IP_ADD_MEMBERSHIP struct ip_mreq
IP_BLOCK_SOURCE struct ip_mreq_source
IP_UNBLOCK_SOURCE struct ip_mreq_source
IP_DROP_MEMBERSHIP struct ip_mreq
IP_ADD_MEMBERSHIP and IP_DROP_MEMBERSHIP are already implemented on
most operating systems, and are used to join and leave an any-source
group.
IP_BLOCK_SOURCE can be used to block data from a given source to a
given group (e.g., if the user "mutes" that source), and
IP_UNBLOCK_SOURCE can be used to undo this (e.g., if the user then
"unmutes" the source).
As to why the packets loop back to the receiver, I believe that is a separate
issue on the output side, not the receive side.
But that is what the commit fixes...
Except you change receive, not transmit. I think the error in Daniel's case
is on the transmit side, not receive. I would be happy to know what the true
"official" behavior is for a socket that binds to INADDR_ANY but does not join
any groups.
As far as prior to BMS's changes: in the older version of udp_input() we only
checked the membership list if inp->inp_moptions was != NULL. If it was NULL,
we would send all multicast packets to a given socket for which the address
fields matched (addr, port).
BMS preserved this behavior and your patch changes
But he changed the behavior on output. Pre-BMS and post-BMS
behave differently and not like Solaris 10 or VxWorks. Haven't
tried Linux.
it. UDP sockets start off with inp_moptions == NULL, so if you never do any
multicast-related setsockopt() you will receive all matching multicast packets.
However, once you do any multicast-related setsockopt() (IP_MULTICAST_LOOP,
IP_ADD_MEMBERSHIP, etc.) then inp_moptions is allocated and is non-NULL.
At that point it only accepts packets that match, except that even then we
used a sysctl which defaulted to off (!) to see if we should check the list of
memberships (net.inet.udp.strict_mcast_mship). This options structure was never
free'd, however, so you could get the truly bizarre behavior of:
- bind a new socket, it will not receive all matching multicast traffic
- use IP_ADD_MEMBERSHIP to add a group, it will now receive only matching
multicast
traffic for the group
- use IP_DROP_MEMBERSHIP to remove the group, it will now receive no multicast
traffic
The different behavior in states 1 and 3 I find confusing and odd. By default
all sockets just always received all matching multicast traffic though. :)
Not if a multicast group was not joined. The pre-BMS changes
did not loop back multicast packets in ip_output.c.
However, this change is not restoring "old" behavior, it is a change in behavior
compared to the pre-BMS changes.
Agreed, to preserve pre-BMS behavior, the change should be
made on output.
and as to the above, it again is
talking about Multicast members.. AFAIKT... I am actually at the IETF
so if you would like I can gladly go talk to the authors of this RFC
(if they are here) and get their opinion on this.
One other thing.. note this is NOT a standard but a informational RFC.
Informational
RFC are guidelines and NOT mandatory at all.. there will never be a MUST/SHOULD
etc
within them.
Well, for lack of anything else I was looking to that. I would really prefer
the behavior in your change as I find it far more intuitive. I wasn't able to
find anything else in Stevens or elsewhere that seemed to indicate what the
proper behavior was beyond this.
I also agree :-)
--
DE
_______________________________________________
svn-src-all@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
