lkml.org 
[lkml]   [2009]   [Jan]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH]Fix: 32bit binary has 64bit address of stack vma
On Mon, 12 Jan 2009, Andrew Morton wrote:
> On Mon, 12 Jan 2009 15:05:51 -0800
> Mike Waychison <mikew@google.com> wrote:
>
> > > Subject: fs/exec.c: fix value of vma->vm_pgoff for the stack VMA of 32-bit processes
> > > From: Ying Han <yinghan@google.com>
> > >
> > > With a 32 bit binary running on a 64 bit system, the /proc/pid/maps for
> > > the [stack] VMA displays a 64-bit address:
> > >
> > > ff96c000-ff981000 rwxp 7ffffffea000 00:00 0 [stack]
> > >
> > > This is because vma->vm_pgoff for that VMA is incorrectly being stored in
> > > units of offset-in-bytes. It should be stored in units of offset-in-pages.
> > >
> >
> > The problem is that the offset was stored without taking into account
> > the shift.
>
> Sigh. Is it efficient to have me sitting here reverse-engineering the
> code, writing your changelog?
>
>
>
> From: Ying Han <yinghan@google.com>
>
> With a 32 bit binary running on a 64 bit system, the /proc/pid/maps for
> the [stack] VMA displays a 64-bit address:
>
> ff96c000-ff981000 rwxp 7ffffffea000 00:00 0 [stack]
>
> This is because vma->vm_pgoff for that VMA is not being updated for the
> shift of the stack VMA.
>
> Signed-off-by: Ying Han <yinghan@google.com>
> Cc: Mike Waychison <mikew@google.com>
> Cc: Hugh Dickins <hugh@veritas.com>
> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
> ---

Sorry for all the waste of your time, but
I'm afraid this is entirely bogus, as Kamezawa-san has kindly
discovered for us (in: Re: mmotm 2009-01-12-16-53 uploaded thread).

The code was correct before, there was no shifting issue, and you
cannot muck around with the vm_pgoff of a vma like this patch does.
Thank Nick for the CONFIG_DEBUG_VM BUG_ON which warns that this
patch has inadvertently made these stack pages unreclaimable
(because their indexes no longer match up with vm_pgoff).

All there is here is a cosmetic issue: we're used to seeing 8-hex-
digit offsets in the /proc/pid/maps of a 32-bit process. Though it
could perfectly well mmap() a large file from a very high offset and
show more hex digits there, couldn't it?

Because this 32-bit process has been started by a 64-bit kernel, it
started off working with the 64-bit stack position before moving into
place, and that happens to get reflected in the final vm_pgoff.

You prefer only to see 8-hex-digit offsets for anonymous vmas of 32-bit
processes? If so, then perhaps Ying Han could supply a cosmetic patch
for /proc/pid/maps to mask off the upper bits in that case.

Hugh

>
> fs/exec.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff -puN fs/exec.c~fs-execc-fix-value-of-vma-vm_pgoff-for-the-stack-vma-of-32-bit-processes fs/exec.c
> --- a/fs/exec.c~fs-execc-fix-value-of-vma-vm_pgoff-for-the-stack-vma-of-32-bit-processes
> +++ a/fs/exec.c
> @@ -509,6 +509,7 @@ static int shift_arg_pages(struct vm_are
> unsigned long length = old_end - old_start;
> unsigned long new_start = old_start - shift;
> unsigned long new_end = old_end - shift;
> + unsigned long new_pgoff = new_start >> PAGE_SHIFT;
> struct mmu_gather *tlb;
>
> BUG_ON(new_start > new_end);
> @@ -523,7 +524,7 @@ static int shift_arg_pages(struct vm_are
> /*
> * cover the whole range: [new_start, old_end)
> */
> - vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL);
> + vma_adjust(vma, new_start, old_end, new_pgoff, NULL);
>
> /*
> * move the page tables downwards, on failure we rely on
> @@ -556,7 +557,7 @@ static int shift_arg_pages(struct vm_are
> /*
> * shrink the vma to just the new range.
> */
> - vma_adjust(vma, new_start, new_end, vma->vm_pgoff, NULL);
> + vma_adjust(vma, new_start, new_end, new_pgoff, NULL);
>
> return 0;
> }
> _


\
 
 \ /
  Last update: 2009-01-13 19:03    [from the cache]
©2003-2011 Jasper Spaans