Re: [PATCH] KVM: x86/mmu: Remove KVM MMU write lock when accessing indirect_shadow_pages

From: Sean Christopherson
Date: Mon Jun 26 2023 - 16:43:30 EST


On Mon, Jun 26, 2023, Jim Mattson wrote:
> On Thu, Jun 15, 2023 at 4:58 PM Mingwei Zhang <mizhang@xxxxxxxxxx> wrote:
> >
> > On Tue, Jun 6, 2023 at 5:28 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > >
> > > On Tue, Jun 06, 2023, Mingwei Zhang wrote:
> > > > > > Hmm. I agree with both points above, but below, the change seems too
> > > > > > heavyweight. smp_wb() is a mfence(), i.e., serializing all
> > > > > > loads/stores before the instruction. Doing that for every shadow page
> > > > > > creation and destruction seems a lot.
> > > > >
> > > > > No, the smp_*b() variants are just compiler barriers on x86.
> > > >
> > > > hmm, it is a "lock addl" now for smp_mb(). Check this: 450cbdd0125c
> > > > ("locking/x86: Use LOCK ADD for smp_mb() instead of MFENCE")
> > > >
> > > > So this means smp_mb() is not a free lunch and we need to be a little
> > > > bit careful.
> > >
> > > Oh, those sneaky macros. x86 #defines __smp_mb(), not the outer helper. I'll
> > > take a closer look before posting to see if there's a way to avoid the runtime
> > > barrier.
> >
> > Checked again, I think using smp_wmb() and smp_rmb() should be fine as
> > those are just compiler barriers. We don't need a full barrier here.
>
> That seems adequate.

Strictly speaking, no, because neither FNAME(fetch) nor kvm_mmu_pte_write() are
pure readers or writers. FNAME(fetch) reads guest memory (guest PTEs) and writes
indirect_shadow_pages. kvm_mmu_pte_write() writes guest memory (guest PTEs) and
reads indirect_shadow_pages (it later writes indirect_shadow_pages too, but that
write isn't relevant to the ordering we care about here).