Re: [PATCH v4 00/13] mm: sub-section memory hotplug support

[Michal Hocko]

On Thu 16-03-17 12:04:48, Dan Williams wrote:
> On Thu, Mar 16, 2017 at 10:48 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > Hi,
> > I didn't get to look through the patch series yet and I might not be
> > able before LSF/MM. How urgent is this? I am primarily asking because
> > the memory hotplug is really convoluted right now and putting more on
> > top doesn't really sound like the thing we really want. I have tried to
> > simplify the code [1] already but this is an early stage work so I do
> > not want to impose any burden on you. So I am wondering whether this
> > is something that needs to be merged very soon or it can wait for the
> > rework and hopefully end up being much simpler in the end as well.
> >
> > What do you think?
> 
> In general, I think it's better to add new features after
> reworks/cleanup, but it's not clear to me (yet) that the problem you
> are trying to solve makes this sub-section enabling for ZONE_DEVICE
> any simpler.
> 
> > [1] http://lkml.kernel.org/r/20170315091347.GA32626@xxxxxxxxxxxxxx
> 
> ZONE_DEVICE pages are never "online". The patch says "Instead we do
> page->zone association from move_pfn_range which is called from
> online_pages." which means the new scheme currently doesn't comprehend
> the sprinkled ZONE_DEVICE hacks in the memory hotplug code.

I hope we can get rid of those...
 
> However, that said, I might take a look at whether the hacks belong in
> the auto-online code so that we can share the delayed zone
> initialization, but still skip marking the memory online per the
> expectations of ZONE_DEVICE.

I think this should be trivial. AFAIU it should be sufficient to split
my move_pfn_range into online_pfn_range which would do the MMOP_ONLINE*
handling and the real move_pfn_range which would do the zone specific
association. Your devm_memremap_pages would then call this
move_pfn_range after arch_add_memory. Or am I overlooking something?

I would still have to addapt your changes to remove hardcoded section
aligned expectations but that shouldn't be a big problem I guess. I
still haven't looked into those deeply to fully understand them.

> I expect it would be confusing to have
> memblock devices in sysfs for ranges that can't be marked online?

Well, if their only valid_zone would be ZONE_DEVICE then I believe it
shouldn't be confusing much.
-- 
Michal Hocko
SUSE Labs