Re: [RFC PATCH v2 1/5] mm: make HPAGE_PxD_{SHIFT,MASK,SIZE} always available

From: Matthew Wilcox
Date: Fri Jul 10 2020 - 12:57:54 EST


On Fri, Jul 10, 2020 at 12:40:37PM -0400, Andrea Arcangeli wrote:
> Hello Hugh and Mike,
>
> On Mon, Jul 06, 2020 at 10:07:34PM -0700, Hugh Dickins wrote:
> > Adding Andrea to Cc, he's the one who structured it that way,
> > and should be consulted.
> >
> > I'm ambivalent myself. Many's the time I've been irritated by the
> > BUILD_BUG() in HPAGE_etc, and it's responsible for very many #ifdef
> > CONFIG_TRANSPARENT_HUGEPAGEs or IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)s
> > that you find uglily scattered around the source.
> >
> > But that's the point of it: it's warning when you write code peculiar
> > to THP, that is going to bloat the build of kernels without any THP.
> >
> > So although I've often been tempted to do as you suggest, I've always
> > ended up respecting Andrea's intention, and worked around it instead
> > (sometimes with #ifdef or IS_ENABLED(), sometimes with
> > PMD_{SHIFT,MASK_SIZE}, sometimes with a local definition).
>
> The only other reasons that comes to mind in addition of optimizing
> the bloat away at build time is to make it easier to identify the THP
> code and to make it explicit that hugetlbfs shouldn't us it or it
> could be wrong on some arches.
>
> However for this case the BUILD_BUG() looks right and this doesn't
> look like a false positive.
>
> This patchset has nothing to do THP, so it'd be more correct to use
> MAX_ORDER whenever the fragmentation is about the buddy (doesn't look
> the case here) or PUD_SIZE/ORDER/PMD_SIZE/ORDER if the objective is
> not to unnecessarily split extra and unrelated hugepud/hugepmds in the
> direct mapping (as in this case).
>
> The real issue exposed by the BUILD_BUG is the lack of PMD_ORDER
> definition and fs/dax.c already run into and it solved it locally in the
> dax.c file:
>
> /* The order of a PMD entry */
> #define PMD_ORDER (PMD_SHIFT - PAGE_SHIFT)
>
> The fact it's not just this patch but also dax.c that run into the
> same issue, makes me think PMD_ORDER should be defined and then you
> can use PMD_* and PUD_* for this non-THP purpose.

We'll run into some namespace issues.

arch/arm/kernel/head.S:#define PMD_ORDER 3
arch/arm/kernel/head.S:#define PMD_ORDER 2
arch/mips/include/asm/pgtable-32.h:#define PMD_ORDER aieeee_attempt_to_allocate_pmd
arch/mips/include/asm/pgtable-64.h:#define PMD_ORDER 0
arch/parisc/include/asm/pgtable.h:#define PMD_ORDER 1 /* Number of pages per pmd */

> Then the question if to remove the BUILD_BUG becomes orthogonal to
> this patchset, but I don't see much value in retaining HPAGE_PMD/PUD_*
> unless the BUILD_BUG is retained too, because this patchset already
> hints that without the BUILD_BUG() the HPAGE_PMD_* definitions would
> likely spill into non THP paths and they would lose also the only
> value left (the ability to localize the THP code paths). So I wouldn't
> be against removing the BUILD_BUG if it's causing maintenance
> overhead, but then I would drop HPAGE_PMD_* too along with it or it
> may just cause confusion.

btw, using the hpage_ prefix already caused one problem in the hugetlb
code:

https://lore.kernel.org/linux-mm/20200629185003.97202-1-mike.kravetz@xxxxxxxxxx/

I'd suggest we rename these to THP_PMD_* and THP_PUD_* to make it clear
they're only for the THP case.