Re: [PATCH -tip 26/32] sched: Add a second-level tag for nested CGroup usecase

[Joel Fernandes <joel@joelfernandes.org>]

Hi Peter,

On Wed, Nov 25, 2020 at 02:42:37PM +0100, Peter Zijlstra wrote:
> On Tue, Nov 17, 2020 at 06:19:56PM -0500, Joel Fernandes (Google) wrote:
> > From: Josh Don <joshdon@google.com>
> > 
> > Google has a usecase where the first level tag to tag a CGroup is not
> > sufficient. So, a patch is carried for years where a second tag is added which
> > is writeable by unprivileged users.
> > 
> > Google uses DAC controls to make the 'tag' possible to set only by root while
> > the second-level 'color' can be changed by anyone. The actual names that
> > Google uses is different, but the concept is the same.
> > 
> > The hierarchy looks like:
> > 
> > Root group
> >    / \
> >   A   B    (These are created by the root daemon - borglet).
> >  / \   \
> > C   D   E  (These are created by AppEngine within the container).
> > 
> > The reason why Google has two parts is that AppEngine wants to allow a subset of
> > subcgroups within a parent tagged cgroup sharing execution. Think of these
> > subcgroups belong to the same customer or project. Because these subcgroups are
> > created by AppEngine, they are not tracked by borglet (the root daemon),
> > therefore borglet won't have a chance to set a color for them. That's where
> > 'color' file comes from. Color could be set by AppEngine, and once set, the
> > normal tasks within the subcgroup would not be able to overwrite it. This is
> > enforced by promoting the permission of the color file in cgroupfs.
> 
> Why can't the above work by setting 'tag' (that's a terrible name, why
> does that still live) in CDE? Have the most specific tag live. Same with
> that thread stuff.

There's 2 parts that Google's usecase has. The first part is set by a
privileged process, and the second part (color) is set within the container.
Maybe we can just put the "color" feature behind a CONFIG option for Google
to enable?

> All this API stuff here is a complete and utter trainwreck. Please just
> delete the patches and start over. Hint: if you use stop_machine(),
> you're doing it wrong.

Ok, the idea was to use stop_machine() as in your initial patch. It works
quite well in testing. However I agree with its horrible we ought to do
better (or at least try).

Maybe we can do a synchronize_rcu() after changing cookie, to ensure we are
no longer using the old cookie value in the scheduler.

> At best you now have the requirements sorted.

Yes.

thanks,

 - Joel