Messages in this thread | | | Date | Fri, 27 Aug 2010 09:43:43 -0700 | Subject | Re: [PATCH] mm: fix hang on anon_vma->root->lock | From | Hugh Dickins <> |
| |
On Fri, Aug 27, 2010 at 2:55 AM, Andrea Arcangeli <aarcange@redhat.com> wrote: > On Thu, Aug 26, 2010 at 06:43:31PM -0700, Hugh Dickins wrote: >> some light., I think you're mistaking the role that RCU plays here. > > That's exactly correct, I thought it prevented reuse of the slab > entry, not only of the whole slab... SLAB_DESTROY_BY_RCU is a lot more > tricky to use than I though... > > However at the light of this, I think page_lock_anon_vma could have > returned a freed and reused anon_vma well before the anon-vma changes. > > The anon_vma could have been freed after the first page_mapped check > succeed but before taking the spinlock. I think, it worked fine > because the rmap walks are robust enough just not to fall apart on a > reused anon_vma while the lock is hold. It become a visible problem > now because we were unlocking the wrong lock leading to a > deadlock. But I guess it wasn't too intentional to return a reused > anon_vma out of page_lock_anon_vma.
What you say there is all exactly right, except for "I guess it wasn't too intentional": it was intentional, and known that it all worked out okay in the rare case when a reused anon_vma got fed into the loops - the anon_vma, after all, is nothing more than a list of places where you may find the page mapped, it has never asserted that a page will be found everywhere that the anon_vma lists.
I would have liked to say "well known" above, but perhaps well known only to me: you're certainly not the first to be surprised by this. IIRC both Christoph and Peter have at different times proposed patches to tighten up page_lock_anon_vma() to avoid returning a stale/reused anon_vma, probably both were dropped because neither was actually necessary, until now: I guess it's a good thing for understandability that anon_vma->root->lock now requires that we weed out that case.
Hugh
| |