RE: [PATCH 00/13] lib/generic-radix-tree: genradix bug fix and optimisations.

[David Laight]

From: Kent Overstreet
> Sent: 25 August 2020 17:32
> 
> On Tue, Aug 25, 2020 at 04:00:35PM +0000, David Laight wrote:
> > From: 'Marcelo Ricardo Leitner'
> > > Sent: 25 August 2020 16:41
> > >
> > > On Tue, Aug 25, 2020 at 02:52:34PM +0000, David Laight wrote:
> > > > The genradix code is used by SCTP for accessing per-stream data.
> > > > This means there are quite a lot of lookups but the code wasn't
> > > > really optimised at all.
> > >
> > > My test box is down for the moment and will bring it on later today or
> > > tomorrow, so I can't test it yet. What should we expect as performance
> > > gains here?
> >
> > Not sure, probably not much, but it ought to show up :-)
> > There'll be bigger gains on a cpu that has software ilog2().
> >
> > I've only checked SCTP still works.
> > I've requested 32k streams on a listener - to force a level-2 tree.
> > I've also done at least one check with a massive pad in the sctp
> > stream structure.
> 
> Have you benchmarked at all? Or were you looking at the generated assembly?

I've been reading a lot of assembly.
(Some of the code generated by modern gcc is crap.)
With horrible casts it is the easiest way to check the code is right!

I'm going to try marking the lookup functions with '__attribute__ ((pure))'.
That should help the sctp code that does repeated SCTP_SI().
In reality I want to mark them __attribute__ ((const)) - including the
inline wrappers, but that isn't allowed if they read memory.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)