Message-ID: <2c924ef5-2ff9-c9a8-f28c-8bedaf8b1af6@redhat.com>
Date:   Thu, 11 Nov 2021 08:06:31 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.2.0
Subject: Re: [RFC 11/19] KVM: x86/mmu: Factor shadow_zero_check out of
 make_spte
Content-Language: en-US
To:     Sean Christopherson <seanjc@google.com>,
        Ben Gardon <bgardon@google.com>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        Peter Xu <peterx@redhat.com>, Peter Shier <pshier@google.com>,
        David Matlack <dmatlack@google.com>,
        Mingwei Zhang <mizhang@google.com>,
        Yulei Zhang <yulei.kernel@gmail.com>,
        Wanpeng Li <kernellwp@gmail.com>,
        Xiao Guangrong <xiaoguangrong.eric@gmail.com>,
        Kai Huang <kai.huang@intel.com>,
        Keqian Zhu <zhukeqian1@huawei.com>,
        David Hildenbrand <david@redhat.com>
References: <20211110223010.1392399-1-bgardon@google.com>
 <20211110223010.1392399-12-bgardon@google.com>
 <80407e4a-36e1-e606-ed9f-74429f850e77@redhat.com>
 <CANgfPd8hzDU+v52t9Kr=b48utC1p_j3yJ8gHzo-uifAxHbh-eQ@mail.gmail.com>
 <YYxvSfUPTXbclpSa@google.com>
From:   Paolo Bonzini <pbonzini@redhat.com>
In-Reply-To: <YYxvSfUPTXbclpSa@google.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 11/11/21 02:18, Sean Christopherson wrote:
> But what would you actually move?  Even shadow_zero_check barely squeaks by,
> e.g. if NX is ever used to for NPT, then maybe it stops being a per-VM setting.

Hmm, I think it would still be per-VM, just like 32-bit shadow page tables
are always built for EFER.NXE=CR4.PAE=1.  Anyway, the rough sketch is to have
three structs:

* struct kvm_mmu_kind has the function pointers and the state that is
needed to operate on page tables

* struct kvm_mmu has the function pointers and the state that is
needed while the vCPU runs, including the role

* struct kvm_paging_context has the stuff related to emulation;
shadow page tables of course needs it but EPT/NPT do not (with
either the old or the new MMU)

So you'd have a "struct kvm_mmu_kind direct_mmu" in struct kvm_arch (for
either legacy EPT/NPT or the new MMU), and

	struct kvm_mmu_kind shadow_mmu;
	struct kvm_mmu root_mmu;		/* either TDP or shadow */
	struct kvm_mmu tdp12_mmu;		/* always shadow */
	struct kvm_mmu *mmu;			/* either &kvm->direct_mmu or &vcpu->shadow_mmu */
	struct kvm_paging_context root_walk;	/* maybe unified with walk01 below? dunno yet */
	struct kvm_paging_context walk01;
	struct kvm_paging_context walk12;
	struct kvm_paging_context *walk;	/* either &vcpu->root_walk or &vcpu->walk12 */

in struct kvm_vcpu_arch.  struct kvm_mmu* has a pointer to
struct kvm_mmu_kind*; however, if an spte.c function does not need
the data in struct kvm_mmu_state*, it can take a struct kvm_mmu_kind*
and it won't need a vCPU.  Likewise the TDP MMU knows its kvm_mmu_kind
is always in &kvm->direct_mmu so it can take a struct kvm* if the struct
kvm_mmu_state* is not needed.

The first part of the refactoring would be to kill the nested_mmu
and walk_mmu, replacing them by &vcpu->walk12 and vcpu->walk
respectively.  The half-useless nested_mmu has always bothered me,
I was going to play with it anyway because I want to remove the
kvm_mmu_reset_context from CR0.WP writes, I'll see if I get
something useful out of it.

Paolo