[RFC] Performance improvements

[Jeremy Allison]

On Wed, Jun 20, 2018 at 10:51:58PM +0200, Swen Schillig wrote:

> ... and each packet is handled by an individually triggered callback.

OK, if this is the case then it's OK. I'm up
at the Redmond plugfest right now with limited
review bandwidth.

> > if part of the event processing schedules an
> > immediate event (or socket activation etc.) then you can
> > accidently trigger out-of-order event handling, which can
> > lead to horrible to find bugs.
> 
> ...and there is NO out of order handling.

If you say so :-).

> It is totally transparent to tevent....one event ... one invocation,
> no messing with multiple events in one go.

Again, if you say so :-).

> > hat has great potential
> > to cause future bugs, so on those grounds I would NAK
> > this.
> I think you're a bit too quick with your judgement, is there 
> a test-case which shows/proves the issue you saw in the past ?

No, no test case - just experience with tevent.

So long as you're not processing multiple events from
one callback you're OK. As soon as you do, you will
*certainly* get bugs.

I don't need a test case to prove this any more than
I need a test case to know that unchecked arithmetic
on network packets can lead to integer wrap security
bugs, or talloc_reference() leads to memory graphs
instead of trees.

Some things are so bad you just have to avoid them,
they'll mess you up good :-).

> well, since it's the "last" patch out of the 10 I posted, 
> I guess it is not so difficult to have a start without it.
> ...I think I could separate out just this part from the last patch and
> keep the other updates again in an individual patch.
> 
> Even though I cannot quantify the improvement of that patch I would
> still believe that it has its share on the reduction of the variation.
> 
> Therefore, I would really like you to have a look at the code again and
> "verify" if the changes are as evil as you claim.

You do have some integer wrap and variable type
work to do first - so why not do that and then resubmit
with the "questionable" patch as the last entry
so it can be reviewed carefully individually ?

Cheers,

	Jeremy.