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

[David Hildenbrand]

On 19.09.22 09:53, David Hildenbrand wrote:
> On 18.09.22 18:13, Emanuele Giuseppe Esposito wrote:
> > Am 09/09/2022 um 16:30 schrieb Sean Christopherson:
> > > On Fri, Sep 09, 2022, Emanuele Giuseppe Esposito wrote:
> > > > KVM is currently capable of receiving a single memslot update through
> > > > the KVM_SET_USER_MEMORY_REGION ioctl.
> > > > The problem arises when we want to atomically perform multiple updates,
> > > > so that readers of memslot active list avoid seeing incomplete states.
> > > >
> > > > For example, in RHBZ https://bugzilla.redhat.com/show_bug.cgi?id=1979276
> > >
> > > I don't have access.  Can you provide a TL;DR?
> >
> > You should be able to have access to it now.
> >
> > > > we see how non atomic updates cause boot failure, because vcpus
> > > > will se a partial update (old memslot delete, new one not yet created)
> > > > and will crash.
> > >
> > > Why not simply pause vCPUs in this scenario?  This is an awful lot of a complexity
> > > to take on for something that appears to be solvable in userspace.
> >
> > I think it is not that easy to solve in userspace: see
> > https://lore.kernel.org/qemu-devel/20200312161217.3590-1-david@xxxxxxxxxx/
> >
> >
> > "Using pause_all_vcpus()/resume_all_vcpus() is not possible, as it will
> > temporarily drop the BQL - something most callers can't handle (esp.
> > when called from vcpu context e.g., in virtio code)."
>
> Can you please comment on the bigger picture? The patch from me works
> around *exactly that*, and for that reason, contains that comment.

FWIW, I hacked up my RFC to perform atomic updates on any memslot 
transactions (not just resizes) where ranges do add overlap with ranges 
to remove.

https://github.com/davidhildenbrand/qemu/tree/memslot


I only performed simple boot check under x86-64 (where I can see region 
resizes) and some make checks -- pretty sure it has some rough edges; 
but should indicate what's possible and what the possible price might 
be. [one could wire up a new KVM ioctl and call it conditionally on 
support if really required]



--
Thanks,

David / dhildenb