Re: [PATCH] mm/vmstat: Add events for PMD based THP migration without split

From: Matthew Wilcox
Date: Wed Jun 03 2020 - 07:09:52 EST


On Wed, Jun 03, 2020 at 10:28:41AM +0530, Anshuman Khandual wrote:
> On 06/03/2020 08:27 AM, Matthew Wilcox wrote:
> > On Wed, Jun 03, 2020 at 06:56:57AM +0530, Anshuman Khandual wrote:
> >> On 06/02/2020 08:31 PM, Matthew Wilcox wrote:
> >>> On Fri, May 22, 2020 at 09:04:04AM +0530, Anshuman Khandual wrote:
> >>>> This adds the following two new VM events which will help in validating PMD
> >>>> based THP migration without split. Statistics reported through these events
> >>>> will help in performance debugging.
> >>>>
> >>>> 1. THP_PMD_MIGRATION_SUCCESS
> >>>> 2. THP_PMD_MIGRATION_FAILURE
> >>>
> >>> There's nothing actually PMD specific about these events, is there?
> >>> If we have a THP of a non-PMD size, you'd want that reported through the
> >>> same statistic, wouldn't you?
> >>
> >> Yes, there is nothing PMD specific here and we would use the same statistics
> >> for non-PMD size THP migration (if any) as well. But is THP migration really
> >> supported for non-PMD sizes ? CONFIG_ARCH_ENABLE_THP_MIGRATION depends upon
> >> CONFIG_TRANSPARENT_HUGEPAGE without being specific or denying about possible
> >> PUD level support. Fair enough, will drop the PMD from the events and their
> >> functions.
> >
> > I guess you haven't read my large pages patchset?
>
> I believe you are referring this "[PATCH v5 00/39] Large pages in the page cache"
> (https://lkml.org/lkml/2020/5/28/1755). Unfortunately, I have not been following
> the series. But is there something else in particular that needs to be taken care
> of as well ?

I don't think so, but I haven't looked at the migration path at all.
I'm hoping it "just works", but experience with the rest of the mm has
taught me there's probably an assumption in there that THP => PMD that
will need to be fixed. I'm not currently testing on a NUMA machine, and
I'm still debugging other paths.