Re: bisected boot regression post 2.6.25-rc3.. please revert

["Nish Aravamudan" <nish.aravamudan@gmail.com>]

On 3/3/08, Ingo Molnar <mingo@elte.hu> wrote:
>
>  * Arjan van de Ven <arjan@linux.intel.com> wrote:
>
>
> > ------------[ cut here ]------------
>  > WARNING: at arch/x86/mm/pgtable_32.c:387 pmd_bad+0x44/0x53()
>  > Modules linked in:
>  > Pid: 1, comm: swapper Not tainted 2.6.25-rc3 #14
>  >  [<c0424ba5>] warn_on_slowpath+0x41/0x67
>  >  [<c0408c5c>] ? native_sched_clock+0x94/0xa6
>  >  [<c043f432>] ? lock_release_holdtime+0x1a/0x115
>  >  [<c04702d4>] ? handle_mm_fault+0x297/0x7e2
>  >  [<c063eee6>] ? _spin_unlock+0x1d/0x20
>  >  [<c04707f0>] ? handle_mm_fault+0x7b3/0x7e2
>  >  [<c04851c1>] ? do_sync_read+0xab/0xe9
>  >  [<c0417223>] pmd_bad+0x44/0x53
>  >  [<c046f37f>] follow_page+0x8b/0x1f2
>  >  [<c0470aa0>] get_user_pages+0x281/0x2ef
>
>
> hm. I suspect some gcc related difference related to the handling of
>  this masking:
>
>
>    pmd_val(x) & ~(PAGE_MASK | _PAGE_USER | _PAGE_PSE | _PAGE_NX)
>
>
> versus:
>
>
>    pmd_val(x) & (~PAGE_MASK & ~_PAGE_USER)
>
>
> perhaps it will work if you change it to:
>
>
>    pmd_val(x) & (~PAGE_MASK & ~_PAGE_USER & ~_PAGE_PSE & ~_PAGE_NX)
>
>
> ?
>
>  in any case, the commit has to be reverted as it clearly isnt a NOP on
>  your box as it was intended to be. (it should only have made a
>  difference in a rare hugetlbfs case)

On x86/{,_64}, _PAGE_PSE and _PAGE_PROTNONE are the same bit. Would
that have any effect here? We encountered that collision when adding
mprotect() support for hugepages.

Thanks,
Nish