[PATCH] sparsemem swiss cheese numa layouts

The part of the sparsemem patch which modifies memmap_init_zone() has recently
become a problem.  It changes behavior so that there is a call to
pfn_to_page() for each individual page inside of a node's range:
node_start_pfn through node_end_pfn.  It used to simply do this once, at the
beginning of the node, but having sparsemem's non-contiguous mem_map[]s inside
of a node made it necessary to change.

Mike Kravetz recently wrote a patch which made the NUMA code accept some new
kinds of layouts.  The system's memory was laid out like this, with node 0's
memory in two pieces: one before and one after node 1's memory:

	Node 0: +++++     +++++
	Node 1:      +++++

Previous behavior before Mike's patch was to assign nodes like this:

	Node 0: 00000     XXXXX
	Node 1:      11111

Where the 'X' areas were simply thrown away.  The new behavior was to make the
pg_data_t span node 0 across all of its areas, including areas that are really
node 1's: Node 0: 000000000000000 Node 1: 11111

This wastes a little bit of mem_map space, but ends up being OK, and more
fully utilizes the system's memory.  memmap_init_zone() initializes all of the
"struct page"s for node 0, even for the "hole", but those never get used,
because there is no pfn_to_page() that resolves to those pages.  However, only
calling pfn_to_page() once, memmap_init_zone() always uses the pages that were
allocated for node0->node_mem_map because:

	struct page *start = pfn_to_page(start_pfn);
	// effectively start = &node->node_mem_map[0]
	for (page = start; page < (start + size); page++) {
		init_page_here();...
		page++;
	}

Slow, and wasteful, but generally harmless.

But, modify that to call pfn_to_page() for each loop iteration (like sparsemem
does):

	for (pfn = start_pfn; pfn < < (start_pfn + size); pfn++++) {
		page = pfn_to_page(pfn);
	}

And you end up trying to initialize node 1's pages too early, along with bogus
data from node 0.  This patch checks for those weird layouts and declines to
touch the pages, making the more frequent pfn_to_page() calls OK to do.

Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

3 files changed, 20 insertions, 0 deletions

diff --git a/arch/ppc64/Kconfig b/arch/ppc64/Kconfig
index 011b5c0..85f8fcf 100644
--- a/arch/ppc64/Kconfig
+++ b/arch/ppc64/Kconfig
@@ -211,6 +211,18 @@ config ARCH_FLATMEM_ENABLE
 
 source "mm/Kconfig"
 
+# Some NUMA nodes have memory ranges that span
+# other nodes.  Even though a pfn is valid and
+# between a node's start and end pfns, it may not
+# reside on that node.
+#
+# This is a relatively temporary hack that should
+# be able to go away when sparsemem is fully in
+# place
+config NODES_SPAN_OTHER_NODES
+	def_bool y
+	depends on NEED_MULTIPLE_NODES
+
 config NUMA
 	bool "NUMA support"
 	depends on DISCONTIGMEM
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 19860d3..746b57e 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -528,6 +528,12 @@ void sparse_init(void);
 #define sparse_init()	do {} while (0)
 #endif /* CONFIG_SPARSEMEM */
 
+#ifdef CONFIG_NODES_SPAN_OTHER_NODES
+#define early_pfn_in_nid(pfn, nid)	(early_pfn_to_nid(pfn) == (nid))
+#else
+#define early_pfn_in_nid(pfn, nid)	(1)
+#endif
+
 #ifndef early_pfn_valid
 #define early_pfn_valid(pfn)	(1)
 #endif
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 5c1b898..1eb683f 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1656,6 +1656,8 @@ void __init memmap_init_zone(unsigned long size, int nid, unsigned long zone,
 	for (pfn = start_pfn; pfn < end_pfn; pfn++, page++) {
 		if (!early_pfn_valid(pfn))
 			continue;
+		if (!early_pfn_in_nid(pfn, nid))
+			continue;
 		page = pfn_to_page(pfn);
 		set_page_links(page, zone, nid, pfn);
 		set_page_count(page, 0);