[dm-devel] [PATCH 7/7] dm: optimize clone_rq() when track_peak_rq_based_ios is disabled

[Mike Snitzer]

On Thu, Sep 12 2013 at  7:30pm -0400,
Mikulas Patocka <mpatocka at redhat.com> wrote:

> 
> 
> On Thu, 12 Sep 2013, Mike Snitzer wrote:
> 
> > On Thu, Sep 12 2013 at  7:00pm -0400,
> > Mikulas Patocka <mpatocka at redhat.com> wrote:
> > 
> > > 
> > > 
> > > On Thu, 12 Sep 2013, Mike Snitzer wrote:
> > > 
> > > > Make use of static keys to eliminate the relevant branch in clone_rq()
> > > > when dm_mod's 'track_peak_rq_based_ios' paramter is set to 0 (disabled).
> > > > 
> > > > Even if 'track_peak_rq_based_ios' is set to 0 it will not take effect
> > > > until the next request-based table reload.  Once it is disabled the
> > > > 'peak_rq_based_ios' parameter will be reset to 0.
> > > 
> > > This patch changes it so that the value track_peak_rq_based_ios is sampled 
> > > only when reloading a table. I think it will be confusing to the user if 
> > > he changes the value, but the change doesn't take effect until something 
> > > reloads some table.
> > 
> > Well we need to be able to notice the change but I do not wish to do so
> > in a performance critical path (which clone_rq is).
> 
> I think one condition isn't that bad so there is no need to use static 
> keys.

Considering we're talking about the IO path I'd rather avoid it given
'track_peak_rq_based_ios' will almost always be disabled.

> If you want to use static keys, you need to refresh them in some place 
> that it called regularly.

No you don't.  I'm assuming if someone enables 'track_peak_rq_based_ios'
they know what they are doing.

But I'm open to suggestions on a more appropriate hook to be catching
the update to track_peak_rq_based_ios.