Re: [PATCH 2/2] mm, memory_hotplug: remove zone restrictions

[Michal Hocko <mhocko@kernel.org>]

On Mon 10-07-17 13:11:29, Vlastimil Babka wrote:
> On 07/10/2017 08:45 AM, Michal Hocko wrote:
> > On Fri 07-07-17 17:02:59, Vlastimil Babka wrote:
> >> [+CC linux-api]
> >>
> >> On 06/29/2017 09:35 AM, Michal Hocko wrote:
> >>> From: Michal Hocko <mhocko@suse.com>
> >>>
> >>> Historically we have enforced that any kernel zone (e.g ZONE_NORMAL) has
> >>> to precede the Movable zone in the physical memory range. The purpose of
> >>> the movable zone is, however, not bound to any physical memory restriction.
> >>> It merely defines a class of migrateable and reclaimable memory.
> >>>
> >>> There are users (e.g. CMA) who might want to reserve specific physical
> >>> memory ranges for their own purpose. Moreover our pfn walkers have to be
> >>> prepared for zones overlapping in the physical range already because we
> >>> do support interleaving NUMA nodes and therefore zones can interleave as
> >>> well. This means we can allow each memory block to be associated with a
> >>> different zone.
> >>>
> >>> Loosen the current onlining semantic and allow explicit onlining type on
> >>> any memblock. That means that online_{kernel,movable} will be allowed
> >>> regardless of the physical address of the memblock as long as it is
> >>> offline of course. This might result in moveble zone overlapping with
> >>> other kernel zones. Default onlining then becomes a bit tricky but still
> >>> sensible. echo online > memoryXY/state will online the given block to
> >>> 	1) the default zone if the given range is outside of any zone
> >>> 	2) the enclosing zone if such a zone doesn't interleave with
> >>> 	   any other zone
> >>>         3) the default zone if more zones interleave for this range
> >>> where default zone is movable zone only if movable_node is enabled
> >>> otherwise it is a kernel zone.
> >>>
> >>> Here is an example of the semantic with (movable_node is not present but
> >>> it work in an analogous way). We start with following memblocks, all of
> >>> them offline
> >>> memory34/valid_zones:Normal Movable
> >>> memory35/valid_zones:Normal Movable
> >>> memory36/valid_zones:Normal Movable
> >>> memory37/valid_zones:Normal Movable
> >>> memory38/valid_zones:Normal Movable
> >>> memory39/valid_zones:Normal Movable
> >>> memory40/valid_zones:Normal Movable
> >>> memory41/valid_zones:Normal Movable
> >>>
> >>> Now, we online block 34 in default mode and block 37 as movable
> >>> root@test1:/sys/devices/system/node/node1# echo online > memory34/state
> >>> root@test1:/sys/devices/system/node/node1# echo online_movable > memory37/state
> >>> memory34/valid_zones:Normal
> >>> memory35/valid_zones:Normal Movable
> >>> memory36/valid_zones:Normal Movable
> >>> memory37/valid_zones:Movable
> >>> memory38/valid_zones:Normal Movable
> >>> memory39/valid_zones:Normal Movable
> >>> memory40/valid_zones:Normal Movable
> >>> memory41/valid_zones:Normal Movable
> >>
> >> Hm so previously, blocks 37-41 would only allow Movable at this point, right?
> > 
> > yes
> > 
> >> Shouldn't we still default to Movable for them? We might be breaking some
> >> existing userspace here.
> > 
> > I do not think so. Prior to this merge window f1dd2cd13c4b ("mm,
> > memory_hotplug: do not associate hotadded memory to zones until online")
> > we allowed only the last offline or the adjacent to existing movable
> > memory block to be onlined movable. So the above wasn't possible.
> 
> Not exactly the above, but let's say 1-34 is onlined as Normal, 35-37 is
> Movable. Then the only possible action before would be online 38 as
> Movable? Now it defaults to Normal?

Yes. And let me repeat you couldn't onlne 35-37 as movable before. So no
userspace could depend on that before the rework. Or do I still miss
your point?
-- 
Michal Hocko
SUSE Labs