Re: [PATCH v6 0/8] KVM: mm: fd-based approach for supporting KVM guest private memory

[Chao Peng]

On Thu, Jun 09, 2022 at 08:29:06PM +0000, Sean Christopherson wrote:
> On Wed, Jun 08, 2022, Vishal Annapurve wrote:
> > ...
> > > With this patch series, it's actually even not possible for userspace VMM
> > > to allocate private page by a direct write, it's basically unmapped from
> > > there. If it really wants to, it should so something special, by 
> > > intention,
> > > that's basically the conversion, which we should allow.
> > >
> > 
> > A VM can pass GPA backed by private pages to userspace VMM and when
> > Userspace VMM accesses the backing hva there will be pages allocated
> > to back the shared fd causing 2 sets of pages backing the same guest
> > memory range.
> > 
> > > Thanks for bringing this up. But in my mind I still think userspace VMM
> > > can do and it's its responsibility to guarantee that, if that is hard
> > > required.
> 
> That was my initial reaction too, but there are unfortunate side effects to 
> punting
> this to userspace. 
> 
> > By design, userspace VMM is the decision-maker for page
> > > conversion and has all the necessary information to know which page is
> > > shared/private. It also has the necessary knobs to allocate/free the
> > > physical pages for guest memory. Definitely, we should make userspace
> > > VMM more robust.
> > 
> > Making Userspace VMM more robust to avoid double allocation can get
> > complex, it will have to keep track of all in-use (by Userspace VMM)
> > shared fd memory to disallow conversion from shared to private and
> > will have to ensure that all guest supplied addresses belong to shared
> > GPA ranges.
> 
> IMO, the complexity argument isn't sufficient justfication for introducing new
> kernel functionality.  If multiple processes are accessing guest memory then 
> there
> already needs to be some amount of coordination, i.e. it can't be _that_ 
> complex.
> 
> My concern with forcing userspace to fully handle unmapping shared memory is 
> that
> it may lead to additional performance overhead and/or noisy neighbor issues, 
> even
> if all guests are well-behaved.
> 
> Unnmapping arbitrary ranges will fragment the virtual address space and 
> consume
> more memory for all the result VMAs.  The extra memory consumption isn't that 
> big
> of a deal, and it will be self-healing to some extent as VMAs will get merged 
> when
> the holes are filled back in (if the guest converts back to shared), but it's 
> still
> less than desirable.
> 
> More concerning is having to take mmap_lock for write for every conversion, 
> which
> is very problematic for configurations where a single userspace process maps 
> memory
> belong to multiple VMs.  Unmapping and remapping on every conversion will 
> create a
> bottleneck, especially if a VM has sub-optimal behavior and is converting 
> pages at
> a high rate.
> 
> One argument is that userspace can simply rely on cgroups to detect 
> misbehaving
> guests, but (a) those types of OOMs will be a nightmare to debug and (b) an 
> OOM
> kill from the host is typically considered a _host_ issue and will be treated 
> as
> a missed SLO.
> 
> An idea for handling this in the kernel without too much complexity would be 
> to
> add F_SEAL_FAULT_ALLOCATIONS (terrible name) that would prevent page faults 
> from
> allocating pages, i.e. holes can only be filled by an explicit fallocate().  
> Minor
> faults, e.g. due to NUMA balancing stupidity, and major faults due to swap 
> would
> still work, but writes to previously unreserved/unallocated memory would get a
> SIGSEGV on something it has mapped.  That would allow the userspace VMM to 
> prevent
> unintentional allocations without having to coordinate unmapping/remapping 
> across
> multiple processes.

Since this is mainly for shared memory and the motivation is catching
misbehaved access, can we use mprotect(PROT_NONE) for this? We can mark
those range backed by private fd as PROT_NONE during the conversion so
subsequence misbehaved accesses will be blocked instead of causing double
allocation silently.

Chao