On Thu 25-03-21 11:55:01, Oscar Salvador wrote:
> On Thu, Mar 25, 2021 at 10:17:33AM +0100, Michal Hocko wrote:
> > Why do you think it is wrong to initialize/account pages when they are
> > used? Keep in mind that offline pages are not used until they are
> > onlined. But vmemmap pages are used since the vmemmap is established
> > which happens in the hotadd stage.
>
> Yes, that is true.
> vmemmap pages are used right when we populate the vmemmap space.
>
> > > plus the fact that I dislike to place those pages in
> > > ZONE_NORMAL, although they are not movable.
> > > But I think the vmemmap pages should lay within the same zone the pages
> > > they describe, doing so simplifies things, and I do not see any outright
> > > downside.
> >
> > Well, both ways likely have its pros and cons. Nevertheless, if the
> > vmemmap storage is independent (which is the case for normal hotplug)
> > then the state is consistent over hotadd, {online, offline} N times,
> > hotremove cycles. Which is conceptually reasonable as vmemmap doesn't
> > go away on each offline.
> >
> > If you are going to bind accounting to the online/offline stages then
> > the accounting changes each time you go through the cycle and depending
> > on the onlining type it would travel among zones. I find it quite
> > confusing as the storage for vmemmap hasn't changed any of its
> > properties.
>
> That is a good point I guess.
> vmemmap pages do not really go away until the memory is unplugged.
>
> But I see some questions to raise:
>
> - As I said, I really dislike it tiding vmemmap memory to ZONE_NORMAL
> unconditionally and this might result in the problems David mentioned.
> I remember David and I discussed such problems but the problems with
> zones not being contiguos have also been discussed in the past and
> IIRC, we reached the conclusion that a maximal effort should be made
> to keep them that way, otherwise other things suffer e.g: compaction
> code.
Yeah, David has raised the contiguous flag for zone already. And to be
completely honest I fail to see why we should shape a design based on an
optimization. If anything we can teach set_zone_contiguous to simply
ignore zone affiliation of vmemmap pages. I would be really curious if
that would pose any harm to the compaction code as they are reserved and
compaction should simply skip them.
> So if we really want to move the initialization/account to the
> hot-add/hot-remove stage, I would really like to be able to set the
> proper zone in there (that is, the same zone where the memory will lay).
THere is nothing like a proper zone.
> - When moving the initialization/accounting to hot-add/hot-remove,
> the section containing the vmemmap pages will remain offline.
Yes this sucks! I do not have a good answer for that as the
online/offline granularity seems rather coarse on that.
> It might get onlined once the pages get online in online_pages(),
> or not if vmemmap pages span a whole section.
> I remember (but maybe David rmemeber better) that that was a problem
> wrt. pfn_to_online_page() and hybernation/kdump.
> So, if that is really a problem, we would have to care of ot setting
> the section to the right state.
>
> - AFAICS, doing all the above brings us to former times were some
> initialization/accounting was done in a previous stage, and I remember
> it was pushed hard to move those in online/offline_pages().
Not sure what you are referring to but if you have prior to f1dd2cd13c4b
("mm, memory_hotplug: do not associate hotadded memory to zones until
online") then this was entirely a different story. Users do care where
they memory goes because that depends on the usecase but do they care
about vmemmap?
--
Michal Hocko
SUSE Labs
