My comment for the fix would be something like below.

/* The unity-mapped region is mapped using 1MB pages, hence 1-level translation
    if bit 20 is set, we are 1MB apart physically, hence we move the page_dir
    in case bit 20 is set. */
if (bit(vaddr,20))
    page_dir = page_dir + 1;

does it make sense ?

Regards,
Oza.

From: Mika Westerberg <mika.westerberg@iki.fi>
To: Dave Anderson <anderson@redhat.com>
Cc: paawan oza <paawan1982@yahoo.com>; "Discussion list for crash utility usage, maintenance and development" <crash-utility@redhat.com>; Thomas Fänge <thomas.fange@sonymobile.com>; jan karlsson <jan.karlsson@sonymobile.com>
Sent: Thursday, 4 October 2012 11:56 AM
Subject: Re: [Crash-utility] using crash for ARM

On Wed, Oct 03, 2012 at 09:45:51AM -0400, Dave Anderson wrote:
>
> Right -- it looks to be a bug, presuming that ARM is using 1MB pages
> for the unity-mapped region:
> 
>  crash> vtop c0000000 | grep PAGE:
>  PAGE:    11000  (1MB)
>  crash> vtop c0100000 | grep PAGE:
>  PAGE:    11000  (1MB)
>  crash> vtop c0200000 | grep PAGE:
>  PAGE:  211000  (1MB)
>  crash> vtop c0300000 | grep PAGE:
>  PAGE:  211000  (1MB)
>  crash> vtop c0400000 | grep PAGE:
>  PAGE:  411000  (1MB)
>  crash> vtop c0500000 | grep PAGE:
>  PAGE:  411000  (1MB)
>  crash>

The unity-mapped region is mapped using 1MB pages. However, we actually have
(when using the Linux ARM 2-level translation scheme):

see arch/arm/include/asm/pgtable-2level.h:

#define PMD_SHIFT              21
#define PGDIR_SHIFT            21

#define PTRS_PER_PGD            2048

So we have 2048 entries in a PGD instead of 4096 making a PGD entry an array
of "two pointers".

Anyway as you and Paawan suggested it looks like a bug - we always use the
first entry instead of the second given that bit 20 is set in the virtual
address.

Paawan, your fix looks sane to me but can you add a small comment describing
why this is done?