On Mon 29-10-18 08:59:34, Alexander Duyck wrote:
On Mon, 2018-10-29 at 15:12 +0100, Michal Hocko wrote:
> On Wed 17-10-18 08:02:20, Alexander Duyck wrote:
> > On 10/17/2018 12:52 AM, Michal Hocko wrote:
> > > On Thu 11-10-18 10:38:39, Alexander Duyck wrote:
> > > > On 10/11/2018 1:55 AM, Michal Hocko wrote:
> > > > > On Wed 10-10-18 20:52:42, Michal Hocko wrote:
> > > > > [...]
> > > > > > My recollection was that we do clear the reserved bit in
> > > > > > move_pfn_range_to_zone and we indeed do in
__init_single_page. But then
> > > > > > we set the bit back right afterwards. This seems to be the
> > > > > > d0dc12e86b319 which reorganized the code. I have to study
this some more
> > > > > > obviously.
> > > > >
> > > > > so my recollection was wrong and d0dc12e86b319 hasn't really
> > > > > much because __init_single_page wouldn't zero out the struct
> > > > > the hotplug contex. A comment in move_pfn_range_to_zone explains
> > > > > want the reserved bit because pfn walkers already do see the pfn
> > > > > and the page is not fully associated with the zone until it is
> > > > >
> > > > > I am thinking that we might be overzealous here. With the full
> > > > > initialized we shouldn't actually care. pfn_to_online_page
> > > > > NULL regardless of the reserved bit and normal pfn walkers
> > > > > touch pages they do not recognize and a plain page with ref.
> > > > > doesn't tell much to anybody. So I _suspect_ that we can
simply drop the
> > > > > reserved bit setting here.
> > > >
> > > > So this has me a bit hesitant to want to just drop the bit entirely.
> > > > nothing else I think I may wan to make that a patch onto itself so
> > > > we aren't going to set it we just drop it there. That way if it
> > > > issues we can bisect it to that patch and pinpoint the cause.
> > >
> > > Yes a patch on its own make sense for bisectability.
> > For now I think I am going to back off of this. There is a bunch of other
> > changes that need to happen in order for us to make this work. As far as I
> > can tell there are several places that are relying on this reserved bit.
> Please be more specific. Unless I misremember, I have added this
> PageReserved just to be sure (f1dd2cd13c4bb) because pages where just
> half initialized back then. I am not aware anybody is depending on this.
> If there is somebody then be explicit about that. The last thing I want
> to see is to preserve a cargo cult and build a design around it.
It is mostly just a matter of going through and auditing all the
places that are using PageReserved to identify pages that they aren't
supposed to touch for whatever reason.
From what I can tell the issue appears to be the fact that the reserved
bit is used to identify if a region of memory is "online" or
No, this is wrong. pfn_to_online_page does that. PageReserved has
nothing to do with online vs. offline status. It merely says that you
shouldn't touch the page unless you own it. Sure we might have few
places relying on it but nothing should really depend the reserved bit
check from the MM hotplug POV.
So for example the call "online_pages_range" doesn't
online_page_callback unless the first pfn in the range is marked as
Yes and there is no fundamental reason to do that. We can easily check
the online status without that.
Another example Dan had pointed out was the saveable_page function
Use pfn_to_online_page there.
> > > > > Regarding the post initialization required
by devm_memremap_pages and
> > > > > potentially others. Can we update the altmap which is already a
> > > > > to get alternative struct pages by a constructor which we could
> > > > > from memmap_init_zone and do the post initialization? This would
> > > > > the additional loop in the caller while it would still fit the
> > > > > design of the altmap and the core hotplug doesn't have to
> > > > > about DAX or whatever needs a special treatment.
> > > > >
> > > > > Does that make any sense?
> > > >
> > > > I think the only thing that is currently using the altmap is the
> > > > memory init. Specifically I think it is only really used by the
> > > > devm_memremap_pages version of things, and then only under certain
> > > > circumstances. Also the HMM driver doesn't pass an altmap. What
> > > > really need is a non-ZONE_DEVICE users of the altmap to really
> > > > sticking with that as the preferred argument to pass.
> > >
> > > I am not aware of any upstream HMM user so I am not sure what are the
> > > expectations there. But I thought that ZONE_DEVICE users use altmap. If
> > > that is not generally true then we certainly have to think about a
> > > better interface.
> > I'm just basing my statement on the use of the move_pfn_range_to_zone
> > The only caller that is actually passing the altmap is devm_memremap_pages
> > and if I understand things correctly that is only used when we want to stare
> > the vmmemmap on the same memory that we just hotplugged.
> Yes, and that is what I've called as allocator callback earlier.
I am really not a fan of the callback approach. It just means we will
have to do everything multiple times in terms of initialization.
I do not follow. Could you elaborate?
> > That is why it made more sense to me to just create a
> > function for handling the page initialization because the one value I do
> > have to pass is the dev_pagemap in both HMM and memremap case, and that has
> > the altmap already embedded inside of it.
> And I have argued that this is a wrong approach to the problem. If you
> need a very specific struct page initialization then create a init
> (constructor) callback.
The callback solution just ends up being more expensive because we lose
multiple layers of possible optimization. By putting everything into on
initization function we are able to let the compiler go through and
optimize things to the point where we are essentially just doing
something akin to one bit memcpy/memset where we are able to construct
one set of page values and write that to every single page we have to
initialize within a given page block.
You are already doing per-page initialization so I fail to see a larger
unit to operate on.
My concern is that we are going to see a 2-4x regression if I were
update the current patches I have to improve init performance in order
to achieve the purity of the page initilization functions that you are
looking for. I feel we are much better off having one function that can
handle all cases and do so with high performance, than trying to
construct a set of functions that end up having to reinitialize the
same memory from the previous step and end up with us wasting cycles
and duplicating overhead in multiple spots.
The memory hotplug is just one pile of unmaintainable mess mostly because
of this kind of attitude. You just care about _your_ particular usecase
and not a wee bit beyond that.