Re: high overhead of functions blkg_*stats_* in bfq

[Tejun Heo <tj@kernel.org>]

On Mon, Nov 06, 2017 at 05:33:45PM +0100, Paolo Valente wrote:
> 
> > Il giorno 06 nov 2017, alle ore 17:30, Tejun Heo <tj@kernel.org> ha scritto:
> > 
> > Hello,
> > 
> > On Mon, Nov 06, 2017 at 05:26:38PM +0100, Paolo Valente wrote:
> >> Yes.  DEBUG_BLK_CGROUP will just serve to keep the for-debugging
> >> subset switched off even when the dynamic parameter is on.
> >> 
> >> We'll wait for possible counter-arguments from Tejun, and then start
> >> to work on what you propose.
> > 
> > So, I think the only *really* required ones are io_serviced and the
> > corresponding byte counts, which are tracked by blk-cgroup core
> > anyway.  Everything else can be hidden behind an obviously debug boot
> > param, say, bfq.__DEBUG__extra_stats or whatever.
> > 
> 
> ok, can we change it, or do you want us to change it, for cfq too?

I don't have a strong opinion there but it coudl be more hassle than
its worth given how long the stats have been out there.  idk.

Please also note that you can still allow the extra stats to be
enabled run-time through /sys/kernel/module moduleparams and gate them
with the static_branch.  No idea whether making it that way is
worthwhile tho.  Creating / removing the files dynamically is
supported but given that some stats need to be tracking all the time
to be correct (like # in flights), it seems kinda silly.

Thanks.

-- 
tejun