Re: kernel oops on mmotm-2015-10-15-15-20

[Hugh Dickins]

On Wed, 21 Oct 2015, Hugh Dickins wrote:
> On Thu, 22 Oct 2015, Minchan Kim wrote:
> > Hello Hugh,
> > 
> > On Wed, Oct 21, 2015 at 05:59:59PM -0700, Hugh Dickins wrote:
> > > On Thu, 22 Oct 2015, Minchan Kim wrote:
> > > > 
> > > > I added the code to check it and queued it again but I had another oops
> > > > in this time but symptom is related to anon_vma, too.
> > > > (kernel is based on recent mmotm + unconditional mkdirty for bug fix)
> > > > It seems page_get_anon_vma returns NULL since the page was not page_mapped
> > > > at that time but second check of page_mapped right before try_to_unmap seems
> > > > to be true.
> > > > 
> > > > Adding 4191228k swap on /dev/vda5.  Priority:-1 extents:1 across:4191228k FS
> > > > Adding 4191228k swap on /dev/vda5.  Priority:-1 extents:1 across:4191228k FS
> > > > page:ffffea0001cfbfc0 count:3 mapcount:1 mapping:ffff88007f1b5f51 index:0x600000aff
> > > > flags: 0x4000000000048019(locked|uptodate|dirty|swapcache|swapbacked)
> > > > page dumped because: VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma)
> > > 
> > > That's interesting, that's one I added in my page migration series.
> > > Let me think on it, but it could well relate to the one you got before.

I think I have introduced a bug there; or rather, made more evident
a pre-existing bug.  But I'm not sure yet: the stacktrace was from
compaction (called by khugepaged, but that may not be relevant at all),
and thinking through the races with isolate_migratepages_block() is
never easy.

What's certain is that I was not giving any thought to
isolate_migratepages_block() when I added that VM_BUG_ON_PAGE():
I was thinking about "stable" anonymous pages, and how they get
faulted back in from swapcache while holding page lock.

It looks to me now as if a page might not yet be PageAnon when it's
first tested in __unmap_and_move(), when going to page_get_anon_vma();
but is page_mapped() and PageAnon() by time of calling try_to_unmap(),
where I inserted the VM_BUG_ON_PAGE().

If so, the code would always have been wrong (trying to unmap the
anonymous page, and later remap its replacement, without a hold on
the anon_vma needed to guide both lookups); but I'll have made it
more glaringly wrong with the VM_BUG_ON_PAGE() - let me pretend
that's a good step forward :)

There's a reference count check in isolated_migratepages_block()
before this, which would make it unlikely, but I doubt rules it out.

However... you did hit an anon_vma reference counting problem before
my migration changes went in, and Kirill had a vague suspicion that
he might be screwing up anon_vma refcounting in split_huge_page():
if he confirms that, I'd say it's more likely to be the cause of
your crash on this occasion.

Not hard to fix mine (though we'll probably have to lose the
VM_BUG_ON_PAGE on the way, so the real fix will be hidden by that
trivial fix), I just want to give the races more thought.

However it turns out, I think you have a very useful test there.

(And I've observed no PageDirty problems with your recent patchsets,
though I don't use MADV_FREE at all myself.)

Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/