Le jeudi 02 février 2012 à 13:55 +0100, Stefan Priebe - Profihost AG a
écrit :

quoted

Such a non-trivial backport has the danger of adding as many bugs as
it is intended to fix.

Therefore I will not be attempting to make this fix feasible for
3.0.x-stable, sorry.

I seriously suggest that if you really are hitting this race (the
exposure is very small, it only happens if you get redirects and
redirects are not common for most hosts) that you move to a more
recent kernel such as 3.1.x or 3.2.x, both of which have all the
dependencies as well as the bug fix applied.

I fully understand that - i'm seeing this bug 2-3 times a week while
running a bunch of > 1000 servers.

I don't want to go to 3.1 or 3.2 as they're not long term.

Hmm, thats a problem, we need a stable 3.0 kernel for sure.

Should distros handle this tricky problem themselves ?

I tried to find a solution, but we had so many changes in this part of
the code that its really hard.

I see 3 ways :

1) Warn the world 3.0 is unfixable and should be avoided.
   Ask stable team to instead support 3.1 or 3.2

2) Disable redirects on 3.0

3) Maybe try to revert f39925dbde77

(ipv4: Cache learned redirect information in inetpeer.)

Unfortunately this was introduced in 2.6.39, and it seems trying to
revert it on 3.0 tree is not trivial.


David, what do you think ?