Re: [PATCH 6/9] KVM: arm64: Split huge pages when dirty logging is enabled

[Ricardo Koller <ricarkol@google.com>]

Hi Marc,

On Tue, Jan 31, 2023 at 2:28 AM Marc Zyngier <maz@kernel.org> wrote:
>
> On Fri, 27 Jan 2023 15:45:15 +0000,
> Ricardo Koller <ricarkol@google.com> wrote:
> >
> > > The one thing that would convince me to make it an option is the
> > > amount of memory this thing consumes. 512+ pages is a huge amount, and
> > > I'm not overly happy about that. Why can't this be a userspace visible
> > > option, selectable on a per VM (or memslot) basis?
> > >
> >
> > It should be possible.  I am exploring a couple of ideas that could
> > help when the hugepages are not 1G (e.g., 2M).  However, they add
> > complexity and I'm not sure they help much.
> >
> > (will be using PAGE_SIZE=4K to make things simpler)
> >
> > This feature pre-allocates 513 pages before splitting every 1G range.
> > For example, it converts 1G block PTEs into trees made of 513 pages.
> > When not using this feature, the same 513 pages would be allocated,
> > but lazily over a longer period of time.
>
> This is an important difference. It avoids the upfront allocation
> "thermal shock", giving time to the kernel to reclaim memory from
> somewhere else. Doing it upfront means you *must* have 2MB+ of
> immediately available memory for each GB of RAM you guest uses.
>
> >
> > Eager-splitting pre-allocates those pages in order to split huge-pages
> > into fully populated trees.  Which is needed in order to use FEAT_BBM
> > and skipping the expensive TLBI broadcasts.  513 is just the number of
> > pages needed to break a 1G huge-page.
>
> I understand that. But it also clear that 1GB huge pages are unlikely
> to be THPs, and I wonder if we should treat the two differently. Using
> HugeTLBFS pages is significant here.
>
> >
> > We could optimize for smaller huge-pages, like 2M by splitting 1
> > huge-page at a time: only preallocate one 4K page at a time.  The
> > trick is how to know that we are splitting 2M huge-pages.  We could
> > either get the vma pagesize or use hints from userspace.  I'm not sure
> > that this is worth it though.  The user will most likely want to split
> > big ranges of memory (>1G), so optimizing for smaller huge-pages only
> > converts the left into the right:
> >
> > alloc 1 page            |    |  alloc 512 pages
> > split 2M huge-page      |    |  split 2M huge-page
> > alloc 1 page            |    |  split 2M huge-page
> > split 2M huge-page      | => |  split 2M huge-page
> >                         ...
> > alloc 1 page            |    |  split 2M huge-page
> > split 2M huge-page      |    |  split 2M huge-page
> >
> > Still thinking of what else to do.
>
> I think the 1G case fits your own use case, but I doubt this covers
> the majority of the users. Most people rely on the kernel ability to
> use THPs, which are capped at the first level of block mapping.
>
> 2MB (and 32MB for 16kB base pages) are the most likely mappings in my
> experience (512MB with 64kB pages are vanishingly rare).
>
> Having to pay an upfront cost for HugeTLBFS doesn't shock me, and it
> fits the model. For THPs, where everything is opportunistic and the
> user not involved, this is a lot more debatable.
>
> This is why I'd like this behaviour to be a buy-in, either directly (a
> first class userspace API) or indirectly (the provenance of the
> memory).

This all makes sense, thanks for the explanation. I decided to implement
something for both cases: small caches (~1 page) where the PUDs are
split one PMD at a time, and bigger caches (>513) where the PUDs can
be split with a single replacement. The user specifies the size of the cache
via a capability, and size of 0 implies no eager splitting (the feature is off).

Thanks,
Ricardo

>
> Thanks,
>
>         M.
>
> --
> Without deviation from the norm, progress is not possible.