[lxc-devel] [PATCH 6/6] cgroup: Major rewrite of cgroup logic
Serge Hallyn
serge.hallyn at ubuntu.com
Tue Sep 10 22:28:10 UTC 2013
Quoting Christian Seiler (christian at iwakd.de):
> Hi Serge,
>
> Thanks for testing / reviewing!
>
> > I'd like to just get rid of mountcgroups and make this a
> > configurable option straight in lxc, which when set will cause lxc,
> > for every mountpoint which is in handler->cgroup, bind-mount the
> > the container init's directory into the container. The question is,
> > would $rootfs/sys/fs/cgroup/$controllername be an ok assumption for
> > all distributions? If not maybe we can't do this universally...
> > But I should think it'd be ok.
>
> As far as I know /sys/fs/cgroup/$controller with appropriate symlinks
> (i.e. if cpu and cpuacct are mounted together, create a directory
> cpu,cpuacct and symlink cpu and cpuacct to that directory) has become
> standard. So I'd suggest using that as the current path to go forward,
> and see later if we have to adjust that due to distribution's needs -
> it's definitely better than the current state where cgroup support is
> not available at all in containers out of the box. (Also, if somebody
> needs something else, they can write an own hook and disable this
> automation.)
>
> Side note: But instead of mounting the container's directory directly to
> /sys/fs/cgroup/$controller: as I said in an earlier thread it would
> probably be better to mount the container's cgroup directory into
> /sys/fs/cgroup/$controller/$cgrouppath and create that recursively
> inside a tmpfs. Otherwise, this might really confuse other software that
> looks for cgroups in specific locations.
Ah, right. Agreed.
> If you're interested, I could implement that.
Very interested if you have time!
> > In the meantime I'm going to push your patch to staging.
>
> Great, thanks!
>
> Final thought: Generally speaking, it'd probably be a good idea to have
> some kind of support in LXC to mount all standard file systems (/proc,
> /sys, tmpfs-/run, cgroups, etc.) which may be specified as a
> configuration option. For example: 'lxc.mount.auto = proc sys run
> cgroups' or so. That would save a lot of useless fstab entries in the
> lxc configuration, ideally you'd need nothing for the simplest
> configurations.
Hm. Definately worth thinking about.
Now there are still lots of people who want the bare minimum, and all
mounted read-only... but so long as we don't make the auto-mounts the
default we shouldn't be bothering them.
thanks,
-serge
More information about the lxc-devel
mailing list