On Tue, Aug 27, 2019 at 8:03 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Tue 27-08-19 19:56:16, Yafang Shao wrote:
> > On Tue, Aug 27, 2019 at 7:50 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > >
> > > On Tue 27-08-19 19:43:49, Yafang Shao wrote:
> > > > On Tue, Aug 27, 2019 at 6:43 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > >
> > > > > If there are no objection to the patch I will post it as a standalong
> > > > > one.
> > > >
> > > > I have no objection to your patch. It could fix the issue.
> > > >
> > > > I still think that it is not proper to use a new scan_control here as
> > > > it breaks the global reclaim context.
> > > >
> > > > This context switch from global reclaim to memcg reclaim is very
> > > > subtle change to the subsequent processing, that may cause some
> > > > unexpected behavior.
> > >
> > > Why would it break it? Could you be more specific please?
> > >
> >
> > Hmm, I have explained it when replying to Hillf's patch.
> > The most suspcious one is settting target_mem_cgroup here, because we
> > only use it to judge whether it is in global reclaim.
> > While the memcg softlimit reclaim is really in global reclaims.
>
> But we are reclaim the target_mem_cgroup hierarchy. This is the whole
> point of the soft reclaim. Push down that hierarchy below the configured
> limit. And that is why we absolutely have to switch the reclaim context.
>
One obvious issue is the reclaim couters may not correct.
See shrink_inactive_list().
The pages relcaimed in memcg softlimit will not be counted to
PGSCAN_{DIRECT, KSWAPD} and
PGSTEAL_{DIRECT, KSWAPD}.
That may cause some misleading. For example, if these counters are not
changed, we will think that direct relcaim doesn't occur, while it
really occurs.
May issues are also in some other code around the usage of
global_reclaim(). I'm not sure of it.
> > Another example the reclaim_idx, if is not same with reclaim_idx in
> > page allocation context, the reclaimed pages may not be used by the
> > allocator, especially in the direct reclaim.
>
> Again, we do not care about that as well. All we care about is to
> reclaim _some_ memory to get below the soft limit. This is the semantic
> that is not really great but this is how the Soft reclaim has
> traditionally worked and why we keep claiming that people shouldn't
> really use it. It does lead to over reclaim and that is a design rather
> than a bug.
>
> > And some other things in scan_control.
>
> Like?
> --
> Michal Hocko
> SUSE Labs
>
