Re: [PATCH v3 1/2] cgroup: Show # of subsystem CSSes in cgroup.stat

[Waiman Long]

On 7/12/24 12:29, Johannes Weiner wrote:
> On Thu, Jul 11, 2024 at 05:00:41PM -0400, Waiman Long wrote:
> > On 7/11/24 15:59, Johannes Weiner wrote:
> > > On Thu, Jul 11, 2024 at 03:13:12PM -0400, Waiman Long wrote:
> > > > On 7/11/24 14:59, Tejun Heo wrote:
> > > > > On Thu, Jul 11, 2024 at 02:51:38PM -0400, Waiman Long wrote:
> > > > > > On 7/11/24 14:44, Tejun Heo wrote:
> > > > > > > Hello,
> > > > > > >
> > > > > > > On Thu, Jul 11, 2024 at 01:39:38PM -0400, Waiman Long wrote:
> > > > > > > > On 7/11/24 13:18, Tejun Heo wrote:
> > > > > > > ...
> > > > > > > > Currently, I use the for_each_css() macro for iteration. If you mean
> > > > > > > > displaying all the possible cgroup subsystems even if they are not enabled
> > > > > > > > for the current cgroup, I will have to manually do the iteration.
> > > > > > > Just wrapping it with for_each_subsys() should do, no? for_each_css() won't
> > > > > > > iterate anything if css doesn't exist for the cgroup.
> > > > > > OK, I wasn't sure if you were asking to list all the possible cgroup v2
> > > > > > cgroup subsystems even if they weren't enabled in the current cgroup.
> > > > > > Apparently, that is the case. I prefer it that way too.
> > > > > Yeah, I think listing all is better. If the list corresponded directly to
> > > > > cgroup.controllers, it may make sense to only show enabled ones but we can
> > > > > have dying ones and implicitly enabled memory and so on, so I think it'd be
> > > > > cleaner to just list them all.
> > > > That will means cgroup subsystems that are seldomly used like rdma, misc
> > > > or even hugetlb will always be shown in all the cgroup.stat output. I
> > > > actually prefer just showing those that are enabled. As for dying memory
> > > > cgroups, they will only be shown in its online ancestors. We currently
> > > > don't know how many level down are each of the dying ones.
> > > It seems odd to me to not show dead ones after a cgroup has disabled
> > > the controller again. They still consume memory, after all, and so
> > > continue to be property of that cgroup afterwards.
> > >
> > > Instead of doing for_each_css(), would it make more sense to have
> > >
> > > 	struct cgroup {
> > > 		...
> > > 		int nr_dying_subsys[CGROUP_SUBSYS_COUNT];
> > What exactly does new this array for?
> For keeping the counts. Instead of inside the css.
>
> AFAICS, with your current patch, if somebody were to disable the
> controller in cgroup.subtree_control, it would offline the css at that
> level, become unreachable from cgroup->subsys[], and you'd lose
> remaining counts of dead css that are still associated with that
> cgroup. Re-enabling the controller would create a new css with new
> descendant counts, and now the reported numbers are actively misleading.
>
> That seems undesirable.
>
> If you track the counts in the cgroup itself, then cgroup.stat would
> reliably show the total counts of dead controllers that are associated
> with the subtree, even after disabling or toggling controllers.
>
> The hooks in online, offline, release should be the same, just update
> css->cgroup->nr_dying_subsys[id] instead of css->nr_dying_descendants.

That does make sense. Thank for the suggestion. I will update the patch 
accordingly.

Cheers,
Longman