Re: mm: kernel BUG at mm/huge_memory.c:1829!

[Kirill A. Shutemov]

On Thu, Apr 10, 2014 at 06:27:50PM +0200, Andrea Arcangeli wrote:
> Hi,
> 
> On Thu, Apr 10, 2014 at 04:44:36PM +0300, Kirill A. Shutemov wrote:
> > Okay, below is my attempt to fix the bug. I'm not entirely sure it's
> > correct. Andrea, could you take a look?
> 
> The possibility the interval tree implicitly broke the walk order of
> the anon_vma list didn't cross my mind, that's very good catch!
> Breakage of the rmap walk order definitely can explain that BUG_ON in
> split_huge_page that signals a pte was missed by the rmap walk.

I've spent few day trying to understand rmap code. And now I think my
patch is wrong.

I actually don't see where walk order requirement comes from. It seems all
operations (insert, remove, foreach) on anon_vma is serialized with
anon_vma->root->rwsem. Andrea, could you explain this for me?

The actual bug was introduced by me with split PMD page table lock
patchset. The patch below should fix it. Please review.

It also means it can't be the root cause of other bug report[1] since
split PMD lock was introduced in v3.13 and bug report is about v3.8.

[1] https://bugzilla.redhat.com/show_bug.cgi?id=923817