Bug 7535

On Thu, 16 Nov 2006 01:11:20 -0800
bugme-daemon@bugzilla.kernel.org wrote:

> http://bugzilla.kernel.org/show_bug.cgi?id=7535
> 
>            Summary: ARP notifications to NETLINK_ROUTE
>     Kernel Version: 2.6.16, 2.6.17
>             Status: NEW
>           Severity: normal
>              Owner: shemminger@osdl.org
>          Submitter: lyricsfan@lycos.com
> 
> 
> Most recent kernel where this bug did *NOT* occur:
> 2.6.8 (maybe also some later version, but < 2.6.16)
> 
> Distribution:
> ubuntu 6.06 desktop/server (also with standard kernel 2.6.16), knoppix 5.0.1
> (kernel 2.6.17)
> 
> Hardware Environment:
> irrelevant
> 
> Software Environment:
> irrelevant
> 
> Problem Description:
> While listening to the multicast group RTMGRP_NEIGH on a NETLINK_ROUTE socket,
> on kernel 2.6.8 a RTM_NEWNEIGH message is received whenever an entry is added or
> deleted to the ARP table. On 2.6.16, no such message is received, except for
> IPv6 events.
> Setting "/proc/sys/net/ipv4/neigh/*/app_solicit" to a value > 0 doesn't solve
> the problem.
> 
> Steps to reproduce:
> Open a NETLINK_ROUTE socket and listen to RTMGRP_NEIGH. Try to add a new entry
> to the ARP table, either manually or by pinging a neighbour on the same subnet.
> Try also to manually delete an entry, or simply let it expire. No RTM_NEWNEIGH
> message is triggered when any of these events occurs (it IS in kernel 2.6.8).
> 

Did you forget to set CONFIG_ARPD?
Unless kernel is configured to support ARP daemon, no notification happens.

I had already compiled with ARPD set, but I found out this was not enough. I
think the problem arises in net/core/neighbour.c. Two times, in functions
neigh_update() and neigh_timer_handler(), just inside a
  #ifdef CONFIG_ARPD ... #endif
block, there's the check (line 823 and 1056, in my case)
  if (notify && neigh->parms->app_probes)
after which neigh_app_notify() is called. This last function triggers the
message over the netlink socket. To make it work, I couldn't find anything
better than comment the check on app_probes. I also found out that the 2.6.8
kernel I previously tested presents the same adjustment... I tried to trace
changes to the app_probes value, but maybe someone with a deeper knowledge of
the code can help. I use a kernel patched with MIPL 2.0.2 (Mobile IP for Linux),
but this doesn't appear to be an issue.

Created attachment 10062 [details]
Demo of how to receive neighbor events

Extracted netlink code from arpd source and made a demo program.
Running on 2.6.20-rc4 this code correctly receives neighbor events.

Created attachment 10063 [details]
netlink library 

netlink library from iproute2 source

Created attachment 10064 [details]
netlink library header

See attached program, this is not a kernel problem.  The existing netlink
interface works in the manner arpd expects.