[PATCH] Tight check of pfn_valid on sparsemem - v2

Johannes Weiner hannes at cmpxchg.org
Tue Jul 20 06:15:58 EDT 2010


Hi,

On Sun, Jul 18, 2010 at 07:18:31PM +0900, Minchan Kim wrote:
> Kukjin reported oops happen while he change min_free_kbytes
> http://www.spinics.net/lists/arm-kernel/msg92894.html
> It happen by memory map on sparsemem.
> 
> The system has a memory map following as. 
>      section 0             section 1              section 2
> 0x20000000-0x25000000, 0x40000000-0x50000000, 0x50000000-0x58000000
> SECTION_SIZE_BITS 28(256M)
> 
> It means section 0 is an incompletely filled section.
> Nontheless, current pfn_valid of sparsemem checks pfn loosely. 
> It checks only mem_section's validation but ARM can free mem_map on hole 
> to save memory space. So in above case, pfn on 0x25000000 can pass pfn_valid's 
> validation check. It's not what we want.
>
> We can match section size to smallest valid size.(ex, above case, 16M)
> But Russell doesn't like it due to mem_section's memory overhead with different
> configuration(ex, 512K section).
>
> I tried to add valid pfn range in mem_section but everyone doesn't like it 
> due to size overhead. This patch is suggested by KAMEZAWA-san. 
> I just fixed compile error and change some naming.

I did not like it, because it messes up the whole concept of a
section.

But most importantly, we already have a crutch for ARM in place,
namely memmap_valid_within().  Looking at Kukjin's bug report,
wouldn't it be enough to use that check in
setup_zone_migrate_reserve()?

Your approach makes every pfn_valid() more expensive, although the
extensive checks are not not needed everywhere (check the comment
above memmap_valid_within): vm_normal_page() for example can probably
assume that a PTE won't point to a hole within the memory map.

OTOH, if the ARM people do not care, we could probably go with your
approach, encode it all into pfn_valid(), and also get rid of
memmap_valid_within() completely.  But I would prefer doing a bugfix
first and such a conceptual change in a different patch, would you
agree?

Kukjin, does the appended patch also fix your problem?

	Hannes

---
From: Johannes Weiner <hannes at cmpxchg.org>
Subject: mm: check mem_map backing in setup_zone_migrate_reserve

Kukjin encountered kernel oopsen when changing
/proc/sys/vm/min_free_kbytes.  The problem is that his sparse memory
layout on ARM is the following:

     section 0             section 1              section 2
0x20000000-0x25000000, 0x40000000-0x50000000, 0x50000000-0x58000000
SECTION_SIZE_BITS 28(256M)

where there is a memory hole at the end of section 0.

Since section 0 has _some_ memory, pfn_valid() will return true for
all PFNs in this section.  But ARM releases the mem_map pages of this
hole and pfn_valid() alone is not enough anymore to ensure there is a
valid page struct behind a PFN.

We acknowledged that ARM does this already and have a function to
double-check for mem_map in cases where we do PFN range walks (as
opposed to coming from a page table entry, which should not point to a
memory hole in the first place e.g.).

setup_zone_migrate_reserve() contains one such range walk which does
not have the extra check and was also the cause of the oopsen Kukjin
encountered.

This patch adds the needed memmap_valid_within() check.

Reported-by: Kukjin Kim <kgene.kim at samsung.com>
Signed-off-by: Johannes Weiner <hannes at cmpxchg.org>
---

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0b0b629..cb6d6d3 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3168,6 +3168,10 @@ static void setup_zone_migrate_reserve(struct zone *zone)
 			continue;
 		page = pfn_to_page(pfn);
 
+		/* Watch out for holes in the memory map */
+		if (!memmap_valid_within(pfn, page, zone))
+			continue;
+
 		/* Watch out for overlapping nodes */
 		if (page_to_nid(page) != zone_to_nid(zone))
 			continue;



More information about the linux-arm-kernel mailing list