Re: [PATCH 1/9] io-throttle documentation

[Andrea Righi <righi.andrea-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

On Fri, Apr 17, 2009 at 09:56:31AM +0800, Li Zefan wrote:
> KAMEZAWA Hiroyuki wrote:
> > On Tue, 14 Apr 2009 22:21:12 +0200
> > Andrea Righi <righi.andrea-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> > 
> >> +Example:
> >> +* Create an association between an io-throttle group and a bio-cgroup group
> >> +  with "bio" and "blockio" subsystems mounted in different mount points:
> >> +  # mount -t cgroup -o bio bio-cgroup /mnt/bio-cgroup/
> >> +  # cd /mnt/bio-cgroup/
> >> +  # mkdir bio-grp
> >> +  # cat bio-grp/bio.id
> >> +  1
> >> +  # mount -t cgroup -o blockio blockio /mnt/io-throttle
> >> +  # cd /mnt/io-throttle
> >> +  # mkdir foo
> >> +  # echo 1 > foo/blockio.bio_id
> > 
> > Why do we need multiple cgroups at once to track I/O ?
> > Seems complicated to me.
> > 
> 
> IIUC, it also disallows other subsystems to be binded with blockio subsys:
>   # mount -t cgroup -o blockio cpuset xxx /mnt
>   (failed)
> 
> and if a task is moved from cg1(id=1) to cg2(id=2) in bio subsys, this task
> will be moved from CG1(id=1) to CG2(id=2) automatically in blockio subsys.
> 
> All these are odd, unexpected, complex and bug-prone I think..

Implementing bio-cgroup functionality as pure infrastructure framework
instead of a cgroup subsystem would remove all this oddity and
complexity.

For example, the actual functionality that I need for the io-throttle
controller is just an interface to set and get the cgroup owner of a
page. I think it should be the same also for other potential users of
bio-cgroup.

So, what about implementing the bio-cgroup functionality as cgroup "page
tracking" infrastructure and provide the following interfaces:

/*
 * Encode the cgrp->css.id in page_group->flags
 */
void set_cgroup_page_owner(struct page *page, struct cgroup *cgrp);

/*
 * Returns the cgroup owner of a page, decoding the cgroup id from
 * page_cgroup->flags.
 */
struct cgroup *get_cgroup_page_owner(struct page *page);

This also wouldn't increase the size of page_cgroup because we can
encode the cgroup id in the unused bits of page_cgroup->flags, as
originally suggested by Kame.

And I think it could be used also by dm-ioband, even if it's not a
cgroup-based subsystem... but I may be wrong. Ryo what's your opinion?

-Andrea