Re: [PATCH] mm: thp: Acquire the anon_vma rwsem for lock during split

[Michel Lespinasse <walken@google.com>]

On Fri, Jan 4, 2013 at 6:08 AM, Mel Gorman <mgorman@suse.de> wrote:
> Despite the reason for these commits, NUMA balancing is not the direct
> source of the problem. split_huge_page() expected the anon_vma lock to be
> exclusive to serialise the whole split operation. Ordinarily it is expected
> that the anon_vma lock would only be required when updating the avcs but
> THP also uses it. The locking requirements for THP are complex and there
> is some overlap but broadly speaking they include the following
>
> 1. mmap_sem for read or write prevents THPs being created underneath
> 2. anon_vma is taken for write if collapsing a huge page
> 3. mm->page_table_lock should be taken when checking if pmd_trans_huge as
>    split_huge_page can run in parallel
> 4. wait_split_huge_page uses anon_vma taken for write mode to serialise
>    against other THP operations
> 5. compound_lock is used to serialise between
>    __split_huge_page_refcount() and gup
>
> split_huge_page takes anon_vma for read but that does not serialise against
> parallel split_huge_page operations on the same page (rule 2). One process
> could be modifying the ref counts while the other modifies the page tables
> leading to counters not being reliable. This patch takes the anon_vma
> lock for write to serialise against parallel split_huge_page and parallel
> collapse operations as it is the most fine-grained lock available that
> protects against both.

Your comment about this being the most fine-grained lock made me
think, couldn't we use lock_page() on the THP page here ?

Now I don't necessarily want to push you that direction, because I
haven't fully thought it trough and because what you propose brings us
closer to what happened before anon_vma became an rwlock, which is
more obviously safe. But I felt I should still mention it, since we're
really only trying to protect from concurrent operations on the same
THP page, so locking at just that granularity would seem desirable.

-- 
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.