Re: [PATCH 18/21] blkcg: move blkio_group_conf->weight to cfq

[Tao Ma <tm-d1IQDZat3X0@public.gmane.org>]

On 04/03/2012 06:17 AM, Tejun Heo wrote:
> On Tue, Apr 03, 2012 at 06:03:03AM +0800, Tao Ma wrote:
>> Currently weight is just used to calculate the time slice of different
>> cfq group, right? So why can't it be used to indicate other weight? So
>> say, if we are just want to use iops to indicate the difference between
>> different cgroups(100 weight vs 200 weight), so one process will send
>> 100 ios while the other will send 200 ios just for example.
> 
> Because it's configuring stuff which is completely unrelated.  Let's
> say you added a new elevator w/ iops based proportional IO which
> shares blkio.weight configuration with cfq but nothing else and in
> turn your new thing would probably need some other config parameters
> which don't make much sense for cfq, right?
> 
> Now, let's say there's a system which has two hard drives and sda is
> using cfq and sdb is using your new elevator and you're trying to
> configure cgroup blkio limits.  Now, you have blkio.weight which
> applies to both elevators and other configurations which aren't and
> from the looks of it there's no way to tell which configuration
> controls what.
> 
> It also makes the configuration implementation hairier.  We'll need
> callbacks from blkcg core layer to all policies to notify changes to
> per-cgroup configuration and from there policies would have to decide
> whether it has overriding per-cgroup-device configuration.  It's not
> even clear we even want per-cgroup configuration.  blk-throttle only
> has per-cgroup-device configuration after all.
Fair enough.
> 
> So, again, no.  blkcg.weight isn't and won't be generic.
> 
>> We will need a new iops_weight in your option to be exported?
> 
> Yeah, just add config and stat files prefixed with the name of the new
> blkcg policy.
OK, I will add a new config file for it.

Thanks
Tao