[PATCH] ARM: map_init_section flushes incorrect pmd

[Po-Yu Chuang]

On Wed, May 29, 2013 at 2:48 AM, Sricharan R <r.sricharan at ti.com> wrote:
>
> On Tuesday 28 May 2013 07:37 PM, Will Deacon wrote:
> > On Tue, May 28, 2013 at 03:03:36PM +0100, Sricharan R wrote:
> >> On Tuesday 28 May 2013 06:35 PM, Will Deacon wrote:
> >>> On Tue, May 28, 2013 at 11:48:20AM +0100, Po-Yu Chuang wrote:
> >>>> This bug was introduced in commit e651eab0.
> >>>> Some v4/v5 platforms failed to boot due to this.
> >>>>
> >>>> Signed-off-by: Po-Yu Chuang <ratbert.chuang at gmail.com>
> >>>> ---
> >>>>  arch/arm/mm/mmu.c |    4 +++-
> >>>>  1 file changed, 3 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/arch/arm/mm/mmu.c b/arch/arm/mm/mmu.c
> >>>> index e0d8565..19a43f8 100644
> >>>> --- a/arch/arm/mm/mmu.c
> >>>> +++ b/arch/arm/mm/mmu.c
> >>>> @@ -620,6 +620,8 @@ static void __init map_init_section(pmd_t *pmd, unsigned long addr,
> >>>>                    unsigned long end, phys_addr_t phys,
> >>>>                    const struct mem_type *type)
> >>>>  {
> >>>> +  pmd_t *p = pmd;
> >>>> +
> >>>>  #ifndef CONFIG_ARM_LPAE
> >>>>    /*
> >>>>     * In classic MMU format, puds and pmds are folded in to
> >>>> @@ -638,7 +640,7 @@ static void __init map_init_section(pmd_t *pmd, unsigned long addr,
> >>>>            phys += SECTION_SIZE;
> >>>>    } while (pmd++, addr += SECTION_SIZE, addr != end);
> >>>>
> >>>> -  flush_pmd_entry(pmd);
> >>>> +  flush_pmd_entry(p);
> >>> Wait, shouldn't this flush be *inside* the loop anyway? Otherwise we just
> >>> flush the cacheline containing the first pmd. The flushing code could also
> >>> flush to PoU instead of PoC for UP ARMv7, but that's an unrelated optimisation.
> >>    I think in LPAE this loop iterates once and non LPAE twice.
> >>    So both the entries should be contained in same cache line right ?
> > Dunno, are there any guarantees about alignment of the starting pmd? Even
> > so, the function takes the range as parameters, so I don't think we
> > should tailor it to the caller. It may explain why this hasn't come up
> > sooner though.
> >
> > Will
>
>  This function is not exposed outside. And the ranges passed to this is going
> to not more than 2 entries in any case. If we put the flush inside the loop,
> then we will end up doing an extra flush for the same line. Regarding the
> alignment, I think if the pgd base is aligned, then rest should be fine.
> Will have to check this.
>
> Regards,
>  Sricharan
>

Hi ,

Catalin and Sricharan,
Thanks for your explanation.

Will,
I guess nobody noticed this because the MMU of later v7 processors
fetches page table
from D-cache. It even doesn't need to clean pmd to PoU.

Regards,
Po-Yu