Comment # 43 on bug 1159882 from Michal Hocko

(In reply to Vlastimil Babka from comment #42)
> (In reply to Michal Hocko from comment #41)
> > Vlastimil, I remember we have discussed this problem in upstream some time
> > ago. You've had a patch which has disabled the heuristic (2c012a4ad1a2).
> > Testing with that reverted would sound like a good next step.
> 
> I think the past discussion was about us *not* having 2c012a4ad1a2 (in an
> older kernel) as the problem was different - file pages thrashing while
> unused anonymous pages sit idly. See
> https://lore.kernel.org/linux-mm/b7f5e356-1f0a-98be-4a32-09a766c3949b@suse.
> cz/

Ahh, I remember now.

> Anyway, what is the actual observed issue here? Is it that part of the swap
> gets used? I think Michal's analysis in comment 33 shows the swapped out
> pages are not accessed (no increase in pswpin) so it shouldn't actually
> cause excessive IO. So is it only that the swap being used looks bad?

Yes this is the case here. But I am more worried this is a more general problem
that might actually hit somewhere else. There shouldn't be really any real
reason to swap out anything with that much of easily reclaimable page cache
which doesn't refault heavily. Remember this is a simple stream writer usecase.
That shouldn't really disrupt anonymous memory users.

I am quite busy now but I will try to prepare a kernel with 2c012a4ad1a2
reverted because that might be easier to adopt in 15.2 resp SLE15-SP2 kernels
than the current upstream which is likely fixing the problem by applying the
refault logic to the anonymous memory as well.

Thanks Vlastimil!