Re: [PATCH] Accounting rework: ct_extend + 64bit counters

[Patrick McHardy <kaber@trash.net>]

Krzysztof Oledzki wrote:
>> Mhh good point :) I was thinking of calling it from the raw table,
>> but of course we don't have a conntrack at that point. So the
>> information would have to be propagated from the raw table somehow.
>> Maybe something like the untracked conntrack? IIRC someone posted
>> a patch for something similar (propagation of parameters to helpers)
>> some time ago.
> 
> OK, I'll look at this. Can we push the current version (plus discussed 
> changes) now and tag if for 2.6.27 and try to solve above problem later 
> (2.6.28)?

I would prefer to see a final solution before pushing
it upstream. Having it only implemented half-way forces
an additional allocation on everyone (even those not
even compiling the feature in now) for now gain.

>>> Do you mean an iptables target (-j ...)? IMHO a kernel/module option 
>>> plus a sysctl/sysfs interface should be enough.
>>
>> Having it controlled through an iptables target would be preferrable
>> because you can then do selective accounting.
> 
> OK, but this will make everything slower and may be often unnecessary, 
> so I still think that setting a default mode should be possible. It is 
> something like "iptables -P", BTW.

I'm guessing the allocation is where the real cost is,
but I'm not opposed to a default (that will get changed
to off after some period).