Re: [PATCH 2/3] mm,migration: Prevent rmap_walk_[anon|ksm] seeing the wrong VMA information

[Andrea Arcangeli <aarcange@redhat.com>]

On Wed, Apr 28, 2010 at 06:34:17PM +0100, Mel Gorman wrote:
> Well, in the easiest case, the details of the VMA (particularly vm_start
> and vm_pgoff) can confuse callers of vma_address during rmap_walk. In the
> case of migration, it will return other false positives or negatives.

false positives are fine ;). Only problems are false negatives...

> > After you fix vma_adjust to be as safe as expand_downards you've also
> > to take care of the rmap_walk that may run on a page->mapping =
> > anon_vma that isn't the vma->anon_vma and you're not taking that
> > anon_vma->lock of the shared page, when you change the vma
> > vm_pgoff/vm_start.
> 
> Is this not what the try-lock-different-vmas-or-backoff-and-retry logic
> in patch 2 is doing or am I missing something else?

yes exactly. This is why patch 2 can't be dropped, both for the
vma_adjust and the rmap_walk that are really two separate issues.

> How so? The old PTE should have been left in place, the page count of
> the page remain positive and migration not occur.

Right only problem is for remove_migration_ptes (and for both
split_huge_page rmap_walks). For migrate the only issue is the second
rmap_walk.

> Because the list could be very large, it would make more sense to
> introduce the shared lock if this is what was required.

Kind of agree, we'll see...