Re: [PATCH 1/3] When migrate_pages returns 0, all pages must have been released

[Andrea Arcangeli <aarcange@redhat.com>]

On Fri, Jan 21, 2011 at 10:11:03AM -0600, Christoph Lameter wrote:
> On Thu, 20 Jan 2011, Andrea Arcangeli wrote:
> 
> > Which following putback_lru_page()?  You mean
> > putback_lru_page(newpage)? That is for the newly allocated page
> > (allocated at the very top, so always needed), it's not relevant to
> > the page_count(page) = 1. The page_count 1 is hold by the caller, so
> > it's leaking memory right now (for everything but compaction).
> 
> Ahh yes we removed the putback_lru_pages call from migrate_pages()
> and broke the existing release logic. The caller has to call
> putback_release_pages() as per commit

putback_lru_paeges

> cf608ac19c95804dc2df43b1f4f9e068aa9034ab

That is the very commit that introduced the two bugs that I've fixed
by code review.

> 
> If that is still the case then we still have the double free.

The caller only calls putback_lru_pages if ret != 0 (the two cases you
refer to happen with ret = 0).

Even if caller unconditionally calls putback_lru_pages (kind of what
compaction did), it can't double free because migrate_pages already
unlinked the pages before calling putback_lru_page(page), so there's
no way to do a double free (however if the caller unconditionally
called putback_lru_pages there would be no memleak to fix, but it
doesn't).

> Could we please document the calling conventions exactly in the source?
> Right now it says that the caller should call putback_lru_pages().

The caller should call putback_lru_pages only if ret != 0. Minchan
this is your commit we're discussing can you check the commentary?

Thanks!
Andrea