On Mon, Jun 23, 2014 at 1:55 AM, dormando [off-list ref] wrote:

On Mon, 23 Jun 2014, Eric Dumazet wrote:

quoted

On Sun, 2014-06-22 at 12:07 -0700, dormando wrote:

quoted

Update on testing:

I only have two machines that crash on their own frequently (more like
one, even). Unfortunately something happened to the datacenter it's in and
it was offline for a week. The machine normally crashes after 1.5-4d,
averaging 2d.

It's done about three days total time without a new crash. I also have the
kernel running in another datacenter for ~10 days.. but it takes 30-150
days to crash in that one.

So, inconclusive, but still promising. If the machine survives the week it
probably means it's fixed, or at least greatly reduced.

I saw that one of your patches got queued for stable, but all three were
necessary to fix udpkill. What's your plan for cleanup/upstreaming?

Did you folks end up running udpkill under the tester thing?

I did not test udpkill, as the known problem is the DST_NOCACHE flag.

We end up calling sk_dst_set(sk, dst) with a dst having this flag set.

So maybe DST_NOCACHE should be renamed, if we _can _ cache a dst like
this. Its meaning is really that dst_release() has to track when
refcount reaches 0 so that last owner fress dst, but we need to respect
rcu grace period.

Fixing sk_dst_set() as I did is not enough, as it is only reducing race
window.

Hrm. I'll have to spend more time trying to understand how to test this
(beyond just putting a kernel into production and seeing if it crashes).
Outside of udpkill it's been slightly hard to reproduce, and udpkill ran
for over five hours with your previous patches.

Do you or other folks have any methods for testing this?

Well, running with kasan should reduce time to crash by an order of magnitude.
Alexey, have we tried running udpkill with kasan?