Skip to contentSkip to conversation

Multicast problems between 2.4.20 and 2.4.21?

DISCUSSION Frank Lind 3 participants 7 messages May 24, 2004 – May 25, 2004

lore
Share & tools
lore

Frank Lind

author
#0 lore raw

Hi,

Was there ever a resolution to the multicast issues with kernels beyond 2.4.20?

I have a very similar problem that has shown up with code which works fine prior with 2.4.20-6smp (Redhat 9) and then stops working for later kernels. I ran into this upgrading a machine to RedHat Enterprise (2.4.21-9.ELsmp) but it also happens under Fedora (2.6.5-1.358smp). I was going to try some separate kernel compiles and also a different distribution when I get a chance to try to localize the change. I haven't been able to find a clear reason why this is happening yet. I was wondering if anyone had any additional insights beyond what showed up in the kernel mailing lists so far (i.e. Any changes in Multicast code between 2.4.20 and 2.4.22/23)?

I am able to get the multicasting to work under these kernels by changing my bind calls to use the INADDR_ANY instead of binding to the address and port I'm using for the multicast group. I haven't found good documentation that this is the "correct" way to do this but I ran into some comments for multicast under ipv6 which made me think it was worth a try. This substitution gets the machines multicasting but I've had some problems that appear to be the switch selectively deciding to deliver data or not depending on if certain machines connect to the multicast group. (This needs more testing but basically if a machine using the 2.4.20 kernel joins then some of the newer kernel machines stop getting packets for reasons I don't understand yet).

I noted it was asked why the network switch might matter. There is a note from HP on their Procurve switches indicates that certain IP addresses are not IGMP filtered. In particular 224.0.0.x gets no IGMP filtering :

http://www.hp.com/rnd/pdfs/IGMP_Tech_Note.pdf

I got bit by this one a while back in that my network was getting flooded despite using IGMP in the code.

-- Frank

Signature · 4 lines
Frank D. Lind			email: [email protected]	
MIT Haystack Observatory	WWW: http://www.haystack.mit.edu
Route 40			tel: 781 981 5570
Westford, MA  01886  USA	fax: 781 981 5766
#1 lore raw
Quoted 6 lines · --- Frank Lind

Hi,

Was there ever a resolution to the multicast issues with kernels beyond
2.4.20?

Hi Frank,

no real resolution in my case, just a gross hack :-) With help from
some network folks and a lot of experiments we found out that somehow IGMPv3 did cause the problems. The new protocol was introduced in 2.4.20 or 2.4.21.

The problem only shows up when using MC groups different from
224.0.0.1. Assuming the correctness of the IGMPv3 code, the suspects were the switch (some CISCO 6000, with the latest - v3 capable- Firmware) or the network card (tg3). *I* am almost sure it is not the card. Exchanging the driver from a "working" kernel into the new kernel (and vice versa) did not change anything. The new kernel did not work with the old driver, and the old kernel worked fine with the new driver. So it is likely the switch.

Solution for me was to force IGMPv2 instead of v3. The patch for taht
looks like:

--- ../../../linux-2.4.23/net/ipv4/igmp.c Fri Nov 28 19:26:21
2003
+++ ./igmp.c Fri Jan 9 16:59:21 2004
@@ -124,8 +124,13 @@

 #define IGMP_V1_SEEN(in_dev) ((in_dev)->mr_v1_seen && \
                time_before(jiffies, (in_dev)->mr_v1_seen))
+#if 0
 #define IGMP_V2_SEEN(in_dev) ((in_dev)->mr_v2_seen && \
                time_before(jiffies, (in_dev)->mr_v2_seen))

+#else
+#define IGMP_V2_SEEN(in_dev) (1)
+#endif
+

The same problem occured on another cluster (again tg3 cards - maybe
there is something ...) and an HP/Procurve-41xx switch. Here we just switched to using 224.0.0.1 to solve the problem (we could not reboot to use a patched kernel).

B^hCheers
Martin

=====
------------------------------------------------------
Martin Knoblauch

email: k n o b i AT knobisoft DOT de
www:   http://www.knobisoft.de
#2 lore raw

You don't need a patch to force IGMPv2, there is a sysctl available now in 2.4.x for this purpose.

#3 lore raw
Quoted 5 lines · --- "David S. Miller

You don't need a patch to force IGMPv2, there is a sysctl available now in 2.4.x for this purpose.

David,

what is the name of the sysctl, and when was it added to 2.4? What
about 2.6.x?

Thanks
Martin

=====
------------------------------------------------------
Martin Knoblauch

email: k n o b i AT knobisoft DOT de
www:   http://www.knobisoft.de
author
#4 lore raw

Hi Martin,

/sbin/sysctl -w net.ipv4.conf.eth1.force_igmp_version=2

the above worked for me. I think it was added in 2.4.25 and 2.6.2.

-- Frank

Quoted 28 lines · Martin Knoblauch
Quoted 6 lines · --- "David S. Miller

You don't need a patch to force IGMPv2, there is a sysctl available now in 2.4.x for this purpose.

David,

what is the name of the sysctl, and when was it added to 2.4? What
about 2.6.x?

Thanks
Martin

=====
------------------------------------------------------
Martin Knoblauch

email: k n o b i AT knobisoft DOT de
www:   http://www.knobisoft.de

!DSPAM:40b38d3220961361122433!

Signature · 4 lines
Frank D. Lind			email: [email protected]	
MIT Haystack Observatory	WWW: http://www.haystack.mit.edu
Route 40			tel: 781 981 5570
Westford, MA  01886  USA	fax: 781 981 5766
#5 lore raw
In reply to Martin Knoblauch
Quoted 2 lines · On Tue, 25 May 2004 11:15:10 -0700 (PDT) Martin Knoblauch <[email protected]>…

what is the name of the sysctl, and when was it added to 2.4? What
about 2.6.x?

/proc/sys/net/ipv4/conf/${DEV}/force_igmp_version

Replace ${DEV} with a specific device name, "default", or "all" as desired.

It got added to 2.4.25 I believe, and yes 2.6.x has it too.

#6 lore raw
Quoted 5 lines · --- "David S. Miller
Quoted 1 line · On Tue, 25 May 2004 11:15:10 -0700 (PDT) Martin Knoblauch <[email protected]>…

what is the name of the sysctl, and when was it added to 2.4?

What

Quoted 10 lines
Quoted 1 line

about 2.6.x?

/proc/sys/net/ipv4/conf/${DEV}/force_igmp_version

Replace ${DEV} with a specific device name, "default", or "all" as desired.

It got added to 2.4.25 I believe, and yes 2.6.x has it too.

cool. What value does one put in? Is it just 0/1, or do you specify
the protocol version you want to force?

Cheers
Martin

=====
------------------------------------------------------
Martin Knoblauch

email: k n o b i AT knobisoft DOT de
www:   http://www.knobisoft.de