Problems with swapping in v4.5-rc on POWER

[Aneesh Kumar K.V]

Hugh Dickins <hughd at google.com> writes:

> I've plagiarized the subject from Paulus's "Problems with THP" mail
> last weekend; but my similar problems are on PowerMac G5 baremetal,
> with 4kB pages, not capable of THP and no THP configured in.
>
> Under heavily swapping load, running kernel builds on tmpfs in limited
> memory, I've been seeing random segfaults too, internal compiler errors
> etc.  Not easily reproduced: sometimes happens in minutes, sometimes
> not for several hours.
>
> I tried and failed to construct a reproducer for you: my lack of a good
> recipe has deterred me from reporting it, and seeing Paulus's mail on
> THP gave me hope that the answer would come up in that thread; but no,
> that was quickly resolved as a THP issue, since fixed.
>
> (Mine had appeared to be fixed in v4.5-rc4 anyway; but I guess I
> just didn't try hard enough, it resurfaced on -rc5 immediately.)
>
> I've seen no sign of such problems on x86.  And I saw no sign of such
> problems on v4.4-rc8-mm1, when I included the fixes to the _PAGE_PTE
> and _PAGE_SWP_SOFT_DIRTY swapoff issues we discussed back then (in
> 33 hours of load, should be good enough; but did see such problems
> a couple of times before including those fixes - I took them to be
> a side-effect of the page flags issue, but now rather doubt that).
>

Can you test the impact of the merge listed below ?(ie, revert the merge and see if
we can reproduce and also verify with merge applied). This will give us a
set of commits to look closer. We had quiet a lot of page table
related changes going in this merge window. 

f689b742f217b2ffe7 ("Pull powerpc updates from Michael Ellerman:")

That is the merge commit that added _PAGE_PTE. 


> The minutes or hours thing: I wonder if that indicates a missing
> initialization somewhere: that can easily show up soon after booting,
> but then the machine settles into a steady state of reusing the same
> structures, now initialized; until much later something disturbs the
> state and it has to allocate more.  Sheer speculation, but I wonder.
>


-aneesh