Re: [HMM 01/16] mm/memory/hotplug: add memory type parameter to arch_add/remove_memory

[Jerome Glisse]

On Fri, Apr 07, 2017 at 05:11:05PM +0200, Michal Hocko wrote:
> On Fri 07-04-17 10:57:43, Jerome Glisse wrote:
> > On Fri, Apr 07, 2017 at 04:45:04PM +0200, Michal Hocko wrote:
> > > On Fri 07-04-17 10:32:49, Jerome Glisse wrote:
> > > > On Fri, Apr 07, 2017 at 02:13:49PM +0200, Michal Hocko wrote:
> > > > > On Wed 05-04-17 16:40:11, Jérôme Glisse wrote:
> > > > > > When hotpluging memory we want more information on the type of memory.
> > > > > > This is to extend ZONE_DEVICE to support new type of memory other than
> > > > > > the persistent memory. Existing user of ZONE_DEVICE (persistent memory)
> > > > > > will be left un-modified.
> > > > > 
> > > > > My current hotplug rework [1] is touching this path as well. It is not
> > > > > really clear from the chage why you are changing this and what are the
> > > > > further expectations of MEMORY_DEVICE_PERSISTENT. Infact I have replaced
> > > > > for_device with want__memblock [2]. I plan to repost shortly but I would
> > > > > like to understand your modifications more to reduce potential conflicts
> > > > > in the code. Why do you need to distinguish different types of memory
> > > > > anyway.
> > > > > 
> > > > > [1] http://lkml.kernel.org/r/20170330115454.32154-1-mhocko@xxxxxxxxxx
> > > > > [2] the current patchset is in git://git.kernel.org/pub/scm/linux/kernel/git/mhocko/mm.git
> > > > >     branch attempts/rewrite-mem_hotplug-WIP
> > > > 
> > > > This is needed for UNADDRESSABLE memory type introduced in patch 3 and
> > > > the arch specific bits are in patch 4. Basicly for UNADDRESSABLE memory
> > > > i do not want the arch code to create a linear mapping for the range
> > > > being hotpluged. Adding memory_type in this patch allow to distinguish
> > > > between different type of ZONE_DEVICE.
> > > 
> > > Why don't you use __add_pages directly then?
> > 
> > That's a possibility, i wanted to keep the arch code in the loop in case
> > some arch wanted to do something specific. But it is unlikely to ever be
> > use outside x86 and i don't think we will want to do anything more than
> > skipping linear mapping.
> 
> Hmm, I am looking closer and x86 stil updates max_pfn. Is this needed
> or you are guaranteed to not cross the max_pfn?

No guaranteed so yes i somewhat care about max_pfn, i do not care about
any of its existing user last time i check but it might matter for some
new user.

>  
> > Note that however for CDM when doing it with ZONE_DEVICE (i am gonna
> > post an RFC for that) maybe the powerpc folks will want to know the
> > memory type ie what kind of ZONE_DEVICE this is. I don't think they need
> > it but i am not sure if there is anything specific needed for their
> > next gen power 9 in respect of device memory.
> 
> Well, I really want to get rid of anything zone specific down the
> arch_add_memory. So whatever they want to do with zone they will have to
> do it after arch_add_memory.

I am ok with that except that the linear mapping thing is in arch_add_memory
and this is the thing i want to skip.

> 
> > Andrew if Michal think it is better to not use arch_add_memory directly
> > for my case than i can respin HMM patchset. Let me know.
> 
> Well, I can drop
> https://git.kernel.org/pub/scm/linux/kernel/git/mhocko/mm.git/commit/?h=attempts/rewrite-mem_hotplug-WIP&id=bb1657d823a85bca045712467f517980650652ca
> and arch_add_memory can have memory type argument as you suggest. I am
> just trying to understand what are the expectations here. If you only care
> about x86 then it sounds a bit too much to tweak all arches.

Well i care about powerpc and i know some arm folks are interested in HMM
but it is hard to know when this will materialize. So long run i expect
more arch but it might not happen soon.

Cheers,
Jérôme