Re: btrfs_tree_lock & trylock

[Chris Mason <chris.mason@oracle.com>]

On Mon, 2008-09-08 at 08:50 -0700, Stephen Hemminger wrote:
> On Mon, 8 Sep 2008 17:47:14 +0200
> Andi Kleen <andi@firstfloor.org> wrote:
> 
> > On Mon, Sep 08, 2008 at 08:07:51AM -0700, Stephen Hemminger wrote:
> > > On Mon, 8 Sep 2008 16:20:52 +0200
> > > Andi Kleen <andi@firstfloor.org> wrote:
> > > 
> > > > On Mon, Sep 08, 2008 at 10:02:30AM -0400, Chris Mason wrote:
> > > > > On Mon, 2008-09-08 at 15:54 +0200, Andi Kleen wrote:
> > > > > > > The idea is to try to spin for a bit to avoid scheduling away, which is
> > > > > > > especially important for the high levels.  Most holders of the mutex
> > > > > > > let it go very quickly.
> > > > > > 
> > > > > > Ok but that surely should be implemented in the general mutex code then
> > > > > > or at least in a standard adaptive mutex wrapper? 
> > > > > 
> > > > > That depends, am I the only one crazy enough to think its a good idea?
> > > > 
> > > > Adaptive mutexes are classic, a lot of other OS have it.
> > > 
> > > The problem is that they are a nuisance. It is impossible to choose
> > > the right trade off between spin an no-spin, also they optimize for
> > > a case that doesn't occur often enough to be justified.
> > 
> > At least the numbers done by Gregory et.al. were dramatic improvements.
> > Given that was an extreme case in that the rt kernel does everything
> > with mutexes, but it was still a very clear win on a wide range
> > of workloads.
> > 
> > -Andi
> 
> My guess is that the improvement happens mostly from the first couple of tries,
> not from repeated spinning. And since it is a mutex, you could even do:

I started with lower spin counts, I really didn't want to spin at all
but the current values came from trial and error.

-chris