Re: [PATCH] mm, numa: Fix bad pmd by atomically check for pmd_trans_huge when marking page tables prot_numa

From: Mel Gorman
Date: Mon Apr 10 2017 - 14:07:30 EST


On Mon, Apr 10, 2017 at 12:49:40PM -0500, Zi Yan wrote:
> On 10 Apr 2017, at 12:20, Mel Gorman wrote:
>
> > On Mon, Apr 10, 2017 at 11:45:08AM -0500, Zi Yan wrote:
> >>> While this could be fixed with heavy locking, it's only necessary to
> >>> make a copy of the PMD on the stack during change_pmd_range and avoid
> >>> races. A new helper is created for this as the check if quite subtle and the
> >>> existing similar helpful is not suitable. This passed 154 hours of testing
> >>> (usually triggers between 20 minutes and 24 hours) without detecting bad
> >>> PMDs or corruption. A basic test of an autonuma-intensive workload showed
> >>> no significant change in behaviour.
> >>>
> >>> Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
> >>> Cc: stable@xxxxxxxxxxxxxxx
> >>
> >> Does this patch fix the same problem fixed by Kirill's patch here?
> >> https://lkml.org/lkml/2017/3/2/347
> >>
> >
> > I don't think so. The race I'm concerned with is due to locks not being
> > held and is in a different path.
>
> I do not agree. Kirill's patch is fixing the same race problem but in
> zap_pmd_range().
>
> The original autoNUMA code first clears PMD then sets it to protnone entry.
> pmd_trans_huge() does not return TRUE because it saw cleared PMD, but
> pmd_none_or_clear_bad() later saw the protnone entry and reported it as bad.
> Is this the problem you are trying solve?
>
> Kirill's patch will pmdp_invalidate() the PMD entry, which keeps _PAGE_PSE bit,
> so pmd_trans_huge() will return TRUE. In this case, it also fixes
> your race problem in change_pmd_range().
>
> Let me know if I miss anything.
>

Ok, now I see. I think you're correct and I withdraw the patch.

--
Mel Gorman
SUSE Labs