Re: [bug report] resctrl high memory comsumption

From: Shakeel Butt
Date: Mon Jan 13 2020 - 13:38:59 EST


On Wed, Jan 8, 2020 at 1:54 PM Reinette Chatre
<reinette.chatre@xxxxxxxxx> wrote:
>
> Hi Fenghua,
>
> On 1/8/2020 1:42 PM, Fenghua Yu wrote:
> > On Wed, Jan 08, 2020 at 12:42:17PM -0800, Reinette Chatre wrote:
> >> Hi Fenghua,
> >> On 1/8/2020 12:23 PM, Fenghua Yu wrote:
> >>> On Wed, Jan 08, 2020 at 09:07:41AM -0800, Shakeel Butt wrote:
> >>>> Recently we had a bug in the system software writing the same pids to
> >>>> the tasks file of resctrl group multiple times. The resctrl code
> >>> Subject: [RFC PATCH] x86/resctrl: Fix redundant task movements
> >> I think your fix would address this specific use case but a slightly
> >> different use case will still encounter the problem of high memory
> >> consumption. If for example, sleeping tasks are moved (many times)
> >> between resource or monitoring groups then their task_works queue would
> >> just keep growing. It seems that a call to task_work_cancel() before
> >> adding a new work item should address all these cases?
> >
> > The checking code in this patch is also helpful to avoid redundant
> > task move preparation (kzalloc(), task_work_add(), etc) in the same
> > rdtgroup.
>
> Indeed.
>
> >
> > How about adding both the checking code and task_work_cancel()?
>
> That does sound good to me.
>

Hi Fenghua, any updates here?

> There is something in the current implementation that I would appreciate
> your feedback on: Currently the task's closid and rmid are initialized
> _after_ the call to task_work_add() succeeds. Should these not be
> initialized before the call to task_work_add()?
>

This seems like a potential race.

thanks,
Shakeel