Re: [RFC PATCH 0/9] kvm: implement atomic memslot updates

[David Hildenbrand]

> The main cases are:
>
> - for the boot case, splitting and merging existing memslots.  QEMU
> likes to merge adjacent memory regions into a single memslot, so if
> something goes from read-write to read-only it has to be split and vice
> versa.  I guess a "don't merge this memory region" flag would be the
> less hideous way to solve it in userspace.
>
> - however, there is also the case of resizing an existing memslot which
> is what David would like to have for virtio-mem.  This is not really
> fixable because part of the appeal of virtio-mem is to have a single
> huge memslot instead of many smaller ones, in order to reduce the
> granularity of add/remove (David, correct me if I'm wrong).

Yes, the most important case I am working on in that regard is reducing 
the memslot size/overhead when only exposing comparatively little memory 
towards a VM using virtio-mem (say, a virtio-mem device that could grow 
to 1 TiB, but we initially only expose 1 GiB to the VM).

One approach I prototyped in the past (where my RFC for atomic updates 
came into play because I ran into this issue) to achieve that was 
dynamically growing (and eventually shrinking) a single memslot on demand.

An alternative [1] uses multiple individual memslots, and exposes them 
on demand. There, I have to make sure that QEMU won't merge adjacent 
memslots into a bigger one -- because otherwise, we'd again need atomic 
memslot updates again, which KVM, vhost-user, ... don't support. But in 
the future, I think we want to have that: if there is no fragmentation, 
we should only have a single large memslot and all memslot consumers 
should be able to cope with atomic updates.


So in any case, I will have good use for atomic memslot updates. Just 
like other hypervisors that (will) implement virtio-mem or something 
comparable :)

[1] https://lore.kernel.org/all/20211013103330.26869-1-david@xxxxxxxxxx/T/#u

--
Thanks,

David / dhildenb