PROBLEM: Power9: kernel oops on memory hotunplug from ppc64le guest

[Nicholas Piggin]

Bharata B Rao's on May 21, 2019 12:29 am:
> On Mon, May 20, 2019 at 01:50:35PM +0530, Bharata B Rao wrote:
>> On Mon, May 20, 2019 at 05:00:21PM +1000, Nicholas Piggin wrote:
>> > Bharata B Rao's on May 20, 2019 3:56 pm:
>> > > On Mon, May 20, 2019 at 02:48:35PM +1000, Nicholas Piggin wrote:
>> > >> >> > git bisect points to
>> > >> >> >
>> > >> >> > commit 4231aba000f5a4583dd9f67057aadb68c3eca99d
>> > >> >> > Author: Nicholas Piggin <npiggin at gmail.com>
>> > >> >> > Date:   Fri Jul 27 21:48:17 2018 +1000
>> > >> >> >
>> > >> >> >     powerpc/64s: Fix page table fragment refcount race vs speculative references
>> > >> >> >
>> > >> >> >     The page table fragment allocator uses the main page refcount racily
>> > >> >> >     with respect to speculative references. A customer observed a BUG due
>> > >> >> >     to page table page refcount underflow in the fragment allocator. This
>> > >> >> >     can be caused by the fragment allocator set_page_count stomping on a
>> > >> >> >     speculative reference, and then the speculative failure handler
>> > >> >> >     decrements the new reference, and the underflow eventually pops when
>> > >> >> >     the page tables are freed.
>> > >> >> >
>> > >> >> >     Fix this by using a dedicated field in the struct page for the page
>> > >> >> >     table fragment allocator.
>> > >> >> >
>> > >> >> >     Fixes: 5c1f6ee9a31c ("powerpc: Reduce PTE table memory wastage")
>> > >> >> >     Cc: stable at vger.kernel.org # v3.10+
>> > >> >> 
>> > >> >> That's the commit that added the BUG_ON(), so prior to that you won't
>> > >> >> see the crash.
>> > >> > 
>> > >> > Right, but the commit says it fixes page table page refcount underflow by
>> > >> > introducing a new field &page->pt_frag_refcount. Now we are hitting the underflow
>> > >> > for this pt_frag_refcount.
>> > >> 
>> > >> The fixed underflow is caused by a bug (race on page count) that got 
>> > >> fixed by that patch. You are hitting a different underflow here. It's
>> > >> not certain my patch caused it, I'm just trying to reproduce now.
>> > > 
>> > > Ok.
>> > 
>> > Can't reproduce I'm afraid, tried adding and removing 8GB memory from a
>> > 4GB guest (via host adding / removing memory device), and it just works.
>> 
>> Boot, add 8G, reboot, remove 8G is the sequence to reproduce.
>> 
>> > 
>> > It's likely to be an edge case like an off by one or rounding error
>> > that just happens to trigger in your config. Might be easiest if you
>> > could test with a debug patch.
>> 
>> Sure, I will continue debugging.
> 
> When the guest is rebooted after hotplug, the entire memory (which includes
> the hotplugged memory) gets remapped again freshly. However at this time
> since no slab is available yet, pt_frag_refcount never gets initialized as we
> never do pte_fragment_alloc() for these mappings. So we right away hit the
> underflow during the first unplug itself, it looks like.

Nice catch, good debugging work.

> I will check how this can be fixed.

Tricky problem. What do you think? You might be able to make the early 
page table allocations in the same pattern as the frag allocations, and 
then fill in the struct page metadata when you have those.

Other option may be create a new set of page tables after mm comes up
to replace the early page tables with. That's a bigger hammer though.

Thanks,
Nick