2020년 4월 7일 (화) 오전 9:22, Yang Shi <shy828301@xxxxxxxxx>님이 작성:
>
> On Sun, Apr 5, 2020 at 6:03 PM Joonsoo Kim <js1304@xxxxxxxxx> wrote:
> >
> > 2020년 4월 4일 (토) 오전 3:29, Yang Shi <shy828301@xxxxxxxxx>님이 작성:
> > >
> > > On Thu, Apr 2, 2020 at 10:41 PM <js1304@xxxxxxxxx> wrote:
> > > >
> > > > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx>
> > > >
> > > > Currently, some swapped-in pages are not charged to the memcg until
> > > > actual access to the page happens. I checked the code and found that
> > > > it could cause a problem. In this implementation, even if the memcg
> > > > is enabled, one can consume a lot of memory in the system by exploiting
> > > > this hole. For example, one can make all the pages swapped out and
> > > > then call madvise_willneed() to load the all swapped-out pages without
> > > > pressing the memcg. Although actual access requires charging, it's really
> > > > big benefit to load the swapped-out pages to the memory without pressing
> > > > the memcg.
> > > >
> > > > And, for workingset detection which is implemented on the following patch,
> > > > a memcg should be committed before the workingset detection is executed.
> > > > For this purpose, the best solution, I think, is charging the page when
> > > > adding to the swap cache. Charging there is not that hard. Caller of
> > > > adding the page to the swap cache has enough information about the charged
> > > > memcg. So, what we need to do is just passing this information to
> > > > the right place.
> > > >
> > > > With this patch, specific memcg could be pressured more since readahead
> > > > pages are also charged to it now. This would result in performance
> > > > degradation to that user but it would be fair since that readahead is for
> > > > that user.
> > >
> > > If I read the code correctly, the readahead pages may be *not* charged
> > > to it at all but other memcgs since mem_cgroup_try_charge() would
> > > retrieve the target memcg id from the swap entry then charge to it
> > > (generally it is the memcg from who the page is swapped out). So, it
> > > may open a backdoor to let one memcg stress other memcgs?
> >
> > It looks like you talk about the call path on CONFIG_MEMCG_SWAP.
> >
> > The owner (task) for a anonymous page cannot be changed. It means that
> > the previous owner written on the swap entry will be the next user. So,
> > I think that using the target memcg id from the swap entry for readahead pages
> > is valid way.
> >
> > As you concerned, if someone can control swap-readahead to readahead
> > other's swap entry, one memcg could stress other memcg by using the fact above.
> > However, as far as I know, there is no explicit way to readahead other's swap
> > entry so no problem.
>
> Swap cluster readahead would readahead in pages on consecutive swap
> entries which may belong to different memcgs, however I just figured
> out patch #8 ("mm/swap: do not readahead if the previous owner of the
> swap entry isn't me") would prevent from reading ahead pages belonging
> to other memcgs. This would kill the potential problem.
Yes, that patch kill the potential problem. However, I think that swap cluster
readahead would not open the backdoor even without the patch #8 in
CONFIG_MEMCG_SWAP case, because:
1. consecutive swap space is usually filled by the same task.
2. swap cluster readahead needs a large I/O price to the offender and effect
isn't serious to the target.
3. those pages would be charged to their previous owner and it is valid.
Thanks.
