On Wed, Aug 03, 2011 at 05:59:03PM +0530, Kautuk Consul wrote:

Hi Mel,

Sorry for the formatting.

I forgot to include the following entire backtrace:
#> cp test_huge_file nfsmnt
kernel BUG at mm/page_alloc.c:849!
Unable to handle kernel NULL pointer dereference at virtual address 00000000
pgd = ce9f0000
<SNIP>
Backtrace:
[<c00269ac>] (__bug+0x0/0x30) from [<c008e8b0>]
(move_freepages_block+0xd4/0x158)

It's still horribly mangled and pretty much unreadable but at least we
know where the bug is hitting.

<SNIP>

Since I was testing on linux-2.6.35.9, line 849 in page_alloc.c is the
same line as you have mentioned:
BUG_ON(page_zone(start_page) != page_zone(end_page))

I reproduce this crash by altering the memory banks' memory ranges
such that they are not aligned to the SECTION_SIZE_BITS size.

How are you altering the ranges? Are you somehow breaking
the checks based on the information in stuct zone that is in
move_freepages_block()?

It no longer seems like a punching-hole-in-memmap problem. Can
you investigate how and why the range of pages passed in to
move_freepages() belong to different zones?

Thanks.

-- 
Mel Gorman
SUSE Labs