Re: [PATCH] kvm: Cap halt polling at kvm->max_halt_poll_ns

[Sean Christopherson <seanjc@google.com>]

On Thu, May 06, 2021, David Matlack wrote:
> On Thu, May 6, 2021 at 9:13 AM Sean Christopherson <seanjc@google.com> wrote:
> >
> > Prefer capitalizing KVM in the shortlog, if only because I'm lazy with grep :-)
> >
> > On Thu, May 06, 2021, Venkatesh Srinivas wrote:
> > > From: David Matlack <dmatlack@google.com>
> > >
> > > When growing halt-polling, there is no check that the poll time exceeds
> > > the per-VM limit. It's possible for vcpu->halt_poll_ns to grow past
> > > kvm->max_halt_poll_ns and stay there until a halt which takes longer
> > > than kvm->halt_poll_ns.
> > >
> >
> > Fixes: acd05785e48c ("kvm: add capability for halt polling")
> >
> > and probably Cc: stable@ too.
> >
> >
> > > Signed-off-by: David Matlack <dmatlack@google.com>
> > > Signed-off-by: Venkatesh Srinivas <venkateshs@chromium.org>
> > > ---
> > >  virt/kvm/kvm_main.c | 4 ++--
> > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > > index 2799c6660cce..120817c5f271 100644
> > > --- a/virt/kvm/kvm_main.c
> > > +++ b/virt/kvm/kvm_main.c
> > > @@ -2893,8 +2893,8 @@ static void grow_halt_poll_ns(struct kvm_vcpu *vcpu)
> > >       if (val < grow_start)
> > >               val = grow_start;
> > >
> > > -     if (val > halt_poll_ns)
> > > -             val = halt_poll_ns;
> > > +     if (val > vcpu->kvm->max_halt_poll_ns)
> > > +             val = vcpu->kvm->max_halt_poll_ns;
> >
> > Hmm, I would argue that the introduction of the capability broke halt_poll_ns.
> > The halt_poll_ns module param is writable after KVM is loaded.  Prior to the
> > capability, that meant the admin could adjust the param on the fly and all vCPUs
> > would honor the new value as it was changed.
> >
> > By snapshotting the module param at VM creation, those semantics were lost.
> > That's not necessarily wrong/bad, but I don't see anything in the changelog for
> > the capability that suggests killing the old behavior was intentional/desirable.
> 
> api.rst does say the capability overrides halt_poll_ns.

Ya, I'm more concerned about old userspace that isn't aware of the capability,
e.g. an old userspace that relies on tuning halt_poll_ns while a VM is running
would break when upgrading to a version of KVM that supports the capability.

> see value in changing the semantics to something like:
> 
> - halt_poll_ns sets machine-wide maximum halt poll time.
> - kvm->max_halt_poll_ns sets VM-wide maximum halt poll time.
> - A vCPU will poll for at most min(halt_poll_ns,
> kvm->max_halt_poll_ns) (aside from an in-progress poll when either
> parameter is changed).

Agreed.  That would also provide a good opportunity to clean up the benign
races in kvm_vcpu_block(); since both the module param and the per-VM variable
can be modified at will, they really should only be read once per instance of
kvm_vcpu_block().

> On a related note, the capability and these subtle details should be
> documented in Documentation/virtual/kvm/halt-polling.txt.
> 
> 
> >
> > >
> > >       vcpu->halt_poll_ns = val;
> > >  out:
> > > --
> > > 2.31.1.607.g51e8a6a459-goog
> > >