Re: [PATCH 1/3 v2][NET] gen_estimator: faster gen_kill_estimator

[jamal <hadi@cyberus.ca>]

On Tue, 2008-22-01 at 13:29 +0100, Jarek Poplawski wrote:
> On Tue, Jan 22, 2008 at 06:42:07AM -0500, jamal wrote:
> ...
> > Jarek,
> > 
> > That looks different from the suggestion from Dave.
> 
> Hmm..., I'm not sure you mean my or your suggestion here, but you
> are right anyway...

Your idea to grab a pointer to the estimator so you can quickly find it
upon destruction is a good one.
The challenge was not to break the ABI to user space.
Dave suggested to use a different struct for the kernel side and
maintain the user one as is[1]. Your patch didnt do this, hence my
statement;->

> Maybe I miss something, but there still could be a lot of this walking

Indeed, that is possible in the case of many estimators configured with
the same interval - because they will all fall in the same table bucket
and the idx is not that useful to begin with.

I was wrong given the nature of interval - the majority of the users
will have an estimator interval of say 1 sec which will put everything
in one bucket still. 
We could introduce a proper index that will allow proper distribution
and have that stored by the class. I am not sure i made sense.
But you are coding - and your idea sounds better.

cheers,
jamal


[1] This is _not uncommon_ (note the usage of double negation here for
emphasis;->) technique actually; ones that go further for example can be
found all over the net/sched code (struct tcf_police vs tc_police) etc.