I can't remember udp socket will cache it but you have to know that normal route cache is srcip+dstip+tos or more, and 
udp socket you can send to anywhere you want as it is connectionless so it may be useless but tcp is useful definitely.

yao
> Thanks Yao for your inputs!
> 
> Another point here is that this issue is seen only for connection oriented
> protocols and not connectionless ones.
> 
> VG
> 
> 
> On Tue, Jul 6, 2010 at 7:32 PM, Yao Zhao <community-noreply@qnx.com> wrote:
> 
> > That is the way how route cache works.
> > New route cache is generated when first traffic is hit after the route is
> > marked down.
> >
> > Linux will flush route cache when this happened.
> > Look at io-pkt source I didn't see where it flush the flow caches when
> > routing table changed.
> >
> > Yao
> >
> > > Hi Patrik
> > >
> > > Thanks for replying and it does make sense
> > >
> > > However, couple of hours ago, I figured out the issue.
> > >
> > > My route table had following two entries:
> > >
> > > Dest               GW
> > > 14.14.14.2     15.15.15.1
> > > default           10.118.209.1
> > >
> > > So, when I deleted the first entry and before adding the new entry,
> > packets
> > > started taking the default route. On adding a new route saying
> > >
> > > 14.14.14.2    15.15.15.2
> > >
> > > Traffic continued to take the default route instead of choosing the now
> > > available longest match.
> > >
> > > I guess this means that route selection is done only when cached route is
> > no
> > > longer in the table.
> > >
> > > Regarding source code, this is on io-net, so not sure if source code is
> > > public for that too...
> > >
> > >
> > > Regards
> > > VG
> > >
> > > On Tue, Jul 6, 2010 at 7:19 AM, Patrik Lahti <community-noreply@qnx.com
> > >wrote:
> > >
> > > > I really have no time to verify/investigate, but I seem to recall that
> > > > sockets cache routes not just for speed, but also for stability. The
> > > > intent of the latter being to stick to the routing information while
> > > > forward progress can be made and there is no indication that the path
> > is
> > > > broken (e.g. ICMP errors).
> > > >
> > > > The reference you saw is just to the regular reference counting used
> > > > when a socket has a pointer to a route.
> > > >
> > > > I remember you writing something about 'route change' working and
> > that's
> > > > probably because the routing entry itself is actually changed in that
> > > > case. As opposed to the delete case where it is just marked as down,
> > > > which effectively means all who have it cached will have to re-lookup
> > > > the route. The re-lookup is probably triggered only in some situations.
> > > >
> > > > Again, sorry I don't have the time to look into it at the moment but
> > you
> > > > can check the source code if you want, that's the best advice I have at
> > > > the moment...
> > > >
> > > > Cheers!
> > > > /P
> > > >
> > > > On 10-07-02 11:58 PM, Vineet Garg wrote:
> > > > > Found some reference to this issue on qnx site: "If a route is in use
> > > > > when it's deleted, the routing entry is marked down and removed from
> > > > > the routing table, but the resources...