Re: [patch] mm, compaction: avoid isolating pinned pages fix

[Hugh Dickins]

On Mon, 3 Feb 2014, David Rientjes wrote:
> On Tue, 4 Feb 2014, Joonsoo Kim wrote:
> 
> > > > Okay. It can't fix your situation. Anyway, *normal* anon pages may be mapped
> > > > and have positive page_count(), so your code such as
> > > > '!page_mapping(page) && page_count(page)' makes compaction skip these *normal*
> > > > anon pages and this is incorrect behaviour.
> > > > 
> > > 
> > > So how does that work with migrate_page_move_mapping() which demands 
> > > page_count(page) == 1 and the get_page_unless_zero() in 
> > > __isolate_lru_page()?
> > 
> > Before doing migrate_page_move_mapping(), try_to_unmap() is called so that all
> > mapping is unmapped. Then, remained page_count() is 1 which is grabbed by
> > __isolate_lru_page(). Am I missing something?
> > 
> 
> Ah, good point.  I wonder if we can get away with 
> page_count(page) - page_mapcount(page) > 1 to avoid the get_user_pages() 
> pin?

Something like that.  But please go back to migrate_page_move_mapping()
to factor in what it's additionally considering.  Whether you can share
code with it, I don't know - it has to do some things under a lock you
cannot take at the preliminary stage - you haven't isolated or locked
the page yet.

There is a separate issue, that a mapping may supply its own non-default
mapping->a_ops->migratepage(): can we assume that the page_counting is
the same whatever migratepage() is in use?  I'm not sure.

If you stick to special-casing PageAnon pages, you won't face that
issue; but your proposed change would be a lot more satisfying if we
can convince ourselves that it's good for !PageAnon too.  May need a
trawl through the different migratepage() methods that exist in tree.

Hugh
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/