Date:   Tue, 14 Dec 2021 23:45:13 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Ben Gardon <bgardon@google.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org,
        Hou Wenlong <houwenlong93@linux.alibaba.com>
Subject: Re: [PATCH 10/28] KVM: x86/mmu: Allow yielding when zapping GFNs for
 defunct TDP MMU root
Message-ID: <YbksiTgVdzN0Z6Dn@google.com>
References: <20211120045046.3940942-1-seanjc@google.com>
 <20211120045046.3940942-11-seanjc@google.com>
 <CANgfPd_H3CZn_rFfEZoZ7Sa==Lnwt4tXSMsO+eg5d8q9n39BSQ@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CANgfPd_H3CZn_rFfEZoZ7Sa==Lnwt4tXSMsO+eg5d8q9n39BSQ@mail.gmail.com>
Precedence: bulk

On Mon, Nov 22, 2021, Ben Gardon wrote:
> On Fri, Nov 19, 2021 at 8:51 PM Sean Christopherson <seanjc@google.com> wrote:
> >
> > Allow yielding when zapping SPTEs for a defunct TDP MMU root.  Yielding
> > is safe from a TDP perspective, as the root is unreachable.  The only
> > potential danger is putting a root from a non-preemptible context, and
> > KVM currently does not do so.
> >
> > Yield-unfriendly iteration uses for_each_tdp_mmu_root(), which doesn't
> > take a reference to each root (it requires mmu_lock be held for the
> > entire duration of the walk).
> >
> > tdp_mmu_next_root() is used only by the yield-friendly iterator.
> >
> > kvm_tdp_mmu_zap_invalidated_roots() is explicitly yield friendly.
> >
> > kvm_mmu_free_roots() => mmu_free_root_page() is a much bigger fan-out,
> > but is still yield-friendly in all call sites, as all callers can be
> > traced back to some combination of vcpu_run(), kvm_destroy_vm(), and/or
> > kvm_create_vm().
> >
> > Signed-off-by: Sean Christopherson <seanjc@google.com>
> 
> Reviewed-by: Ben Gardon <bgardon@google.com>
> 
> I'm glad to see this fixed. I assume we don't usually hit this in
> testing because most of the teardown happens in the zap-all path when
> we unregister for MMU notifiers and actually deleting a fully
> populated root while the VM is running is pretty rare.

Another *sigh*.

AFAIK, the above analysis is 100% correct, but there's a subtle problem with
yielding while putting the last reference to a root.  If the mmu_notifier runs
in parallel, it (obviously) won't be able to get a reference to the root, and so
KVM will fail to ensure all references to an unmapped range are removed prior to
returning from the mmu_notifier.

But, I have a idea.  Instead of synchronously zapping the defunct root, mark it
invalid, set the refcount back to '1', and then use a helper kthread to do the
teardown.  Assuming there is exactly one helper, that would also address my
concerns with kvm_tdp_mmu_zap_invalidated_roots() being unsafe to call in parallel,
e.g. two zappers processing an invalid root would both put the last reference to
a root and trigger use-after-free of a different kind.