On Mon, Dec 12, 2011 at 03:51:55PM +0100, Michal Nazarewicz wrote:

quoted

On Fri, Nov 18, 2011 at 05:43:10PM +0100, Marek Szyprowski wrote:

quoted

From: Michal Nazarewicz <redacted>

diff --git a/mm/compaction.c b/mm/compaction.c
index 6afae0e..09c9702 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c

@@ -111,7 +111,10 @@ skip:

next:
		pfn += isolated;
-		page += isolated;
+		if (zone_pfn_same_memmap(pfn - isolated, pfn))
+			page += isolated;
+		else
+			page = pfn_to_page(pfn);
	}

On Mon, 12 Dec 2011 15:19:53 +0100, Mel Gorman [off-list ref] wrote:

quoted

Is this necessary?

We are isolating pages, the largest of which is a MAX_ORDER_NR_PAGES
page.  [...]

On Mon, 12 Dec 2011 15:40:30 +0100, Mel Gorman [off-list ref] wrote:

quoted

To be clear, I'm referring to a single page being isolated here. It may
or may not be a high-order page but it's still going to be less then
MAX_ORDER_NR_PAGES so you should be able check when a new block is
entered and pfn_to_page is necessary.

Do you mean something like:

if (same pageblock)
	just do arithmetic;
else
	use pfn_to_page;

something like the following untested snippet.

/*
 * Resolve pfn_to_page every MAX_ORDER_NR_PAGES to handle the case where
 * memmap is not contiguous such as with SPARSEMEM memory model without
 * VMEMMAP
 */
pfn += isolated;
page += isolated;
if ((pfn & ~(MAX_ORDER_NR_PAGES-1)) == 0)
	page = pfn_to_page(pfn);

That would be closer to what other PFN walkers do

?

I've discussed it with Dave and he suggested that approach as an
optimisation since in some configurations zone_pfn_same_memmap()
is always true thus compiler will strip the else part, whereas
same pageblock test will be false on occasions regardless of kernel
configuration.

Ok, while I recognise it's an optimisation, it's a very small
optimisation and I'm not keen on introducing something new for
CMA that has been coped with in the past by always walking PFNs in
pageblock-sized ranges with pfn_valid checks where necessary.

See setup_zone_migrate_reserve as one example where pfn_to_page is
only called once per pageblock and calls pageblock_is_reserved()
for examining pages within a pageblock. Still, if you really want
the helper, at least keep it in compaction.c as there should be no
need to have it in mmzone.h

-- 
Mel Gorman
SUSE Labs