On Thu, Apr 08, 2010 at 11:23:32PM +0200, Andrea Arcangeli wrote:
> On Thu, Apr 08, 2010 at 08:48:42PM +0200, Johannes Weiner wrote:
> > On Thu, Apr 08, 2010 at 07:58:47PM +0200, Andrea Arcangeli wrote:
> > > On Thu, Apr 08, 2010 at 07:56:04PM +0200, Johannes Weiner wrote:
> > > > Humm, maybe the start pfn could be huge page aligned? That would make
> > > > it possible to check for PageTransHuge() and skip over compound_order()
> > > > pages. This way, we should never actually run into PG_tail pages.
> > >
> > > The problem here are random compound pages that aren't owned by the
> > > transparent hugepage subsystem. If we can't identify those, it's
> > > unsafe to call compound_order (like it's unsafe to call page_order for
> > > pagebuddy pages).
> >
> > But transparent huge pages are the only compound pages on the LRU, so
> > we should be able to identify them.
> >
> > The lru_lock excludes isolation, splitting and collapsing, so I think
> > this is safe:
> >
> > if (PageLRU() && PageTransCompound()) {
> > low_pfn += (1 << compound_order()) - 1
> > continue
> > }
> >
> > if (__isolate_lru_page())
> > continue
> >
> > ...
>
> I don't see anything wrong with this. You're right lru_lock excludes
> isolation, splitting and collapsing (collapsing if it's pagelru it
> means it already happened).
>
> Thanks for thinking this optimization in detail. I guess retaining the
> other optimization will be harder. It depends how costly it is to take
> the zone->lock, the main annoyance is that we can only do that if we
> release the lru_lock first or we get lock inversion deadlocks.
I would find it very difficult to justify the cost of dropping one lock
and taking the other myself. It's potentially stalling other allocators for
order-0 pages (i.e. per-cpu refill) so that direct compaction for high-order
pages can go very slightly faster.
> So it
> costs 4 locked ops to skip max 1024 pages (but in average it'll be
> much less than 1024 pages, more like 128 [no math just random guess]
> when there's quite some ram free).
>
4 irq-safe locked ops. I don't know off-hand what the cycle cost of disabling
and enabling IRQs is but I'd expect it to be longer than what it takes to
scan over a few pages.
--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>