On Mon, Dec 06, 2021 at 11:14:58PM -0800, Shakeel Butt wrote:
> On Mon, Dec 6, 2021 at 3:25 AM Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sun, Dec 05, 2021 at 10:06:27PM -0800, Shakeel Butt wrote:
> > > On Fri, Dec 3, 2021 at 11:08 AM Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > [...]
> > > > > I am in agreement with the motivation of the whole series. I am just
> > > > > making sure that the motivation of VMSCAN_THROTTLE_NOPROGRESS based
> > > > > throttle is more than just the congestion_wait of
> > > > > mem_cgroup_force_empty_write.
> > > > >
> > > >
> > > > The commit that primarily targets congestion_wait is 8cd7c588decf
> > > > ("mm/vmscan: throttle reclaim until some writeback completes if
> > > > congested"). The series recognises that there are other reasons why
> > > > reclaim can fail to make progress that is not directly writeback related.
> > > >
> > >
> > > I agree with throttling for VMSCAN_THROTTLE_[WRITEBACK|ISOLATED]
> > > reasons. Please explain why we should throttle for
> > > VMSCAN_THROTTLE_NOPROGRESS? Also 69392a403f49 claims "Direct reclaim
> > > primarily is throttled in the page allocator if it is failing to make
> > > progress.", can you please explain how?
> >
> > It could happen if the pages on the LRU are being reactivated continually
> > or holding an elevated reference count for some reason (e.g. gup,
> > page migration etc). The event is probably transient, hence the short
> > throttling.
> >
>
> What's the worst that can happen if the kernel doesn't throttle at all
> for these transient scenarios? Premature oom-kills?
Excessive CPU usage in reclaim, potential premature OOM kills.
> The kernel already
> has some protection against such situations with retries i.e.
> consecutive 16 unsuccessful reclaim tries have to fail to give up the
> reclaim.
>
The retries mitigate the premature OOM kills but not the excessive
CPU usage.
> Anyways, I have shared my view which is 'no need to throttle at all
> for no-progress reclaims for now and course correct if there are
> complaints in future' but will not block the patch.
>
We've gone through periods of bugs that had either direct reclaim or
kswapd pegged at 100% CPU usage. While kswapd now just stops, the patch
still minimises the risk of excessive CPU usage bugs due to direct reclaim.
--
Mel Gorman
SUSE Labs
