Re: [PATCH v7 3/5] mm/hugetlb: add support for mempolicy MPOL_PREFERRED_MANY

[Feng Tang]

Hi Huge,

On Tue, Aug 10, 2021 at 02:35:05PM -0700, Hugh Dickins wrote:
> On Tue, 10 Aug 2021, Feng Tang wrote:
> > On Mon, Aug 09, 2021 at 03:19:32PM +0200, Michal Hocko wrote:
> > [snip]
> > > > > Do you think you can provide same helpers for other policies as well?
> > > > > Maybe we can get rid of some other ifdefery as well.
> > > > 
> > > > Sure. I can make separate patch(es) for that.
> > > > 
> > > > And you mean helper like mpol_is_bind/default/local/preferred? 
> > > > 
> > > > I just run 'git-grep MPOL', and for places using "mode == MPOL_XXX",
> > > > mostly they are in mempolicy.[ch], the only another place is in
> > > > shmem.c, do we need to create all the helpers for it and the
> > > > potential future users? 
> > > 
> > > I would just go with those instances which need to ifdef for NUMA.
> > > Thanks!
> > 
> > Yes, following is a patch to remove one CONFIG_NUMA check, though
> > an bolder idea to extend the patch by removing the CONFIG_TMPFS
> > check in the same line.
> > 
> > Thanks,
> > Feng
> > 
> > ---------8<---------------------------------
> > 
> > From 1a5858721ac8ce99c27c13d310bba2983dc73d97 Mon Sep 17 00:00:00 2001
> > From: Feng Tang <feng.tang@xxxxxxxxx>
> > Date: Tue, 10 Aug 2021 17:00:59 +0800
> > Subject: [PATCH] mm: shmem: avoid open coded check for mempolicy's mode
> > 
> > Add a mempolicy helper to do the check, which can also remove
> > a CONFIG_NUMA option check.
> > 
> > Suggested-by: Michal Hocko <mhocko@xxxxxxxx>
> > Signed-off-by: Feng Tang <feng.tang@xxxxxxxxx>
> 
> No thanks: this is not an improvement.
> 
> The "#if defined(CONFIG_NUMA) && defined(CONFIG_TMPFS)" is there to
> eliminate dead code that would not be automatically eliminated by the
> optimizer: it's not there just to avoid MPOL_DEFAULT, and it's there
> to cover shmem_get_sbmpol() along with shmem_show_mpol().

Thanks for the explaination! I did some tests that in !NUMA case, the
'sbinfo->mpol' is always NULL (I could be wrong) which makes the
2 functions almost non-ops.

> I know we tend to avoid #ifdefs in .c files, and that's good; and
> I know you could find other code in mm/shmem.c which might also be
> #ifdef'ed to eliminate other dead code in other configs; but unless
> there's a new drive to purge our .c source of all #ifdefs,
> please just leave this as is.
 
Ok, and sorry for the noise.

Thanks,
Feng

> Thanks,
> Hugh
>