On Fri, Mar 26, 2021 at 09:35:03AM +0100, Michal Hocko wrote:
> No problem there. I will not insist on my approach unless I can convince
> you that it is a better solution. It seems I have failed and I can live
> with that.
Well, I am glad we got to discuss it at least.
> > +static int memory_block_online(unsigned long start_pfn, unsigned long nr_pages,
> > + unsigned long nr_vmemmap_pages, int online_type,
> > + int nid)
> > +{
> > + int ret;
> > + /*
> > + * Despite vmemmap pages having a different lifecycle than the pages
> > + * they describe, initialiating and accounting vmemmap pages at the
> > + * online/offline stage eases things a lot.
>
> This requires quite some explaining.
Definitely, I will expand on that and provide some context.
> Yes this is much better! Just a minor suggestion would be to push
> memory_block all the way to memory_block_online (it oline a memory
> block). I would also slightly prefer to provide 2 helpers that would make
> it clear that this is to reserve/cleanup the vmemamp space (defined in
> the memory_hotplug proper).
Glad to hear that!
By pushing memory_block all the way to memory_block_{online,offline}, you
mean passing the memblock struct together with nr_vmemmap_pages,
only_type and nid to memory_block_{offline,online}, and derive in there
the start_pfn and nr_pages?
Wrt. to the two helpers, I agree with you.
Actually, I would find quite disturbing to deal with zones in that code
domain.
I will add two proper helpers in memory_hotplug to deal with vmemmap.
If it comes out the way I envision, it could end up quite clean, and much
less disturbing.
Thanks Michal
--
Oscar Salvador
SUSE L3
