Re: TCP syn flood handling regression

Le samedi 10 mars 2012 à 09:44 -0800, Eric Dumazet a écrit :

Le samedi 10 mars 2012 à 04:27 -0800, Simon Kirby a écrit :

quoted

Hello!

A typical port 80 SYN flood started up to one of our clusters, but this
time, it didn't work so well. Legitimate connections and trying to fetch
server-status via localhost would hang for ~30 seconds before responding,
even though though the box had plenty of spare cycles. An strace of all
Apache processes showed quite a bit of sleeping in accept4().

This was with 3.2.9, so I went back in kernel builds and found that 3.1
and 3.0 were also broken, while 2.6.39 works as I remember -- when syn
cookies are enabled, everything just works and is fast. The DoS kept up,
so I was able to feed a bit to a node to do some bisection.

Of course, the DoS stopped literally seconds before the last bisection
test, but I got it down to:

# good: [0e734419923bd8e599858f8fc196c7804bb85564] ipv4: Use inet_csk_route_child_sock() in DCCP and TCP.
# bad: [ea4fc0d6193ff56fcef39b0d2210d402a7acb5f0] ipv4: Don't use rt->rt_{src,dst} in ip_queue_xmit().

...leaving ea4fc0d6193ff56fcef39b0d2210d402a7acb5f0 and
d9d8da805dcb503ef8ee49918a94d49085060f23 as culprits.

I've stared at them but can't see what could be doing this.

Hi Simon, thanks a lot for this report.

This sounds as a SYNCOOKIE side effect

It seems we miss a rebuild_header() indeed in the case of a syncookie
initiated socket.

So inet->cork.fl.u.ip4 is not correctly setup and
ea4fc0d6193ff56fcef39b0d2210d402a7acb5f relied on this.

I'll send a patch once tested.