Re: [RFC PATCH V3 3/4] KVM: X86: Alloc role.pae_root shadow page

[Sean Christopherson]

On Wed, Mar 30, 2022, Lai Jiangshan wrote:
> From: Lai Jiangshan <jiangshan.ljs@xxxxxxxxxxxx>
> 
> Currently pae_root is special root page, this patch adds facility to
> allow using kvm_mmu_get_page() to allocate pae_root shadow page.

I don't think this will work for shadow paging.  CR3 only has to be 32-byte aligned
for PAE paging.  Unless I'm missing something subtle in the code, KVM will incorrectly
reuse a pae_root if the guest puts multiple PAE CR3s on a single page because KVM's
gfn calculation will drop bits 11:5.

Handling this as a one-off is probably easier.  For TDP, only 32-bit KVM with NPT
benefits from reusing roots, IMO and shaving a few pages in that case is not worth
the complexity.

> @@ -332,7 +337,8 @@ union kvm_mmu_page_role {
>  		unsigned ad_disabled:1;
>  		unsigned guest_mode:1;
>  		unsigned glevel:4;
> -		unsigned :2;
> +		unsigned pae_root:1;

If we do end up adding a role bit, it can simply be "root", which may or may not
be useful for other things.  is_pae_root() is then a combo of root+level.  This
will clean up the code a bit as role.root is (mostly?) hardcoded based on the
function, e.g. root allocators set it, child allocators clear it.

> +		unsigned :1;
>  
>  		/*
>  		 * This is left at the top of the word so that
> @@ -699,6 +705,7 @@ struct kvm_vcpu_arch {
>  	struct kvm_mmu_memory_cache mmu_shadow_page_cache;
>  	struct kvm_mmu_memory_cache mmu_gfn_array_cache;
>  	struct kvm_mmu_memory_cache mmu_page_header_cache;
> +	void *mmu_pae_root_cache;
>  
>  	/*
>  	 * QEMU userspace and the guest each have their own FPU state.
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index d53037df8177..81ccaa7c1165 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -694,6 +694,35 @@ static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
>  	}
>  }
>  
> +static int mmu_topup_pae_root_cache(struct kvm_vcpu *vcpu)
> +{
> +	struct page *page;
> +
> +	if (vcpu->arch.mmu->shadow_root_level != PT32E_ROOT_LEVEL)
> +		return 0;
> +	if (vcpu->arch.mmu_pae_root_cache)
> +		return 0;
> +
> +	page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO | __GFP_DMA32);
> +	if (!page)
> +		return -ENOMEM;
> +	vcpu->arch.mmu_pae_root_cache = page_address(page);
> +
> +	/*
> +	 * CR3 is only 32 bits when PAE paging is used, thus it's impossible to
> +	 * get the CPU to treat the PDPTEs as encrypted.  Decrypt the page so
> +	 * that KVM's writes and the CPU's reads get along.  Note, this is
> +	 * only necessary when using shadow paging, as 64-bit NPT can get at
> +	 * the C-bit even when shadowing 32-bit NPT, and SME isn't supported
> +	 * by 32-bit kernels (when KVM itself uses 32-bit NPT).
> +	 */
> +	if (!tdp_enabled)
> +		set_memory_decrypted((unsigned long)vcpu->arch.mmu_pae_root_cache, 1);
> +	else
> +		WARN_ON_ONCE(shadow_me_mask);
> +	return 0;
> +}
> +
>  static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
>  {
>  	int r;
> @@ -705,6 +734,9 @@ static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
>  		return r;
>  	r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
>  				       PT64_ROOT_MAX_LEVEL);
> +	if (r)
> +		return r;
> +	r = mmu_topup_pae_root_cache(vcpu);

This doesn't need to be called from the common mmu_topup_memory_caches(), e.g. it
will unnecessarily require allocating another DMA32 page when handling a page fault.
I'd rather call this directly kvm_mmu_load(), which also makes it more obvious
that the cache really is only used for roots.

>  	if (r)
>  		return r;
>  	if (maybe_indirect) {
> @@ -717,12 +749,23 @@ static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
>  					  PT64_ROOT_MAX_LEVEL);
>  }
>  

...

>  static void link_shadow_page(struct kvm_vcpu *vcpu, u64 *sptep,
>  			     struct kvm_mmu_page *sp)
>  {
> +	struct kvm_mmu_page *parent_sp = sptep_to_sp(sptep);
>  	u64 spte;
>  
>  	BUILD_BUG_ON(VMX_EPT_WRITABLE_MASK != PT_WRITABLE_MASK);
>  
> -	spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
> +	if (!parent_sp->role.pae_root)

Hmm, without role.root, this could be:

	if (sp->role.level == (PT32E_ROOT_level - 1) &&
	    ((__pa(sptep) & PT64_BASE_ADDR_MASK) == vcpu->arch.mmu->root.hpa))
		spte = make_pae_pdpte(sp->spt);
	else
		spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));

Which is gross, but it works.  We could also do FNAME(link_shadow_page) to send
PAE roots down a dedicated path (also gross).  Point being, I don't think we
strictly need a "root" flag unless the PAE roots are put in mmu_page_hash.

> +		spte = make_nonleaf_spte(sp->spt, sp_ad_disabled(sp));
> +	else
> +		spte = make_pae_pdpte(sp->spt);
>  
>  	mmu_spte_set(sptep, spte);
>  
> @@ -4782,6 +4847,8 @@ kvm_calc_tdp_mmu_root_page_role(struct kvm_vcpu *vcpu,
>  	role.base.level = kvm_mmu_get_tdp_level(vcpu);
>  	role.base.direct = true;
>  	role.base.has_4_byte_gpte = false;
> +	if (role.base.level == PT32E_ROOT_LEVEL)
> +		role.base.pae_root = 1;
>  
>  	return role;
>  }
> @@ -4848,6 +4915,9 @@ kvm_calc_shadow_mmu_root_page_role(struct kvm_vcpu *vcpu,
>  	else
>  		role.base.level = PT64_ROOT_4LEVEL;
>  
> +	if (role.base.level == PT32E_ROOT_LEVEL)
> +		role.base.pae_root = 1;
> +
>  	return role;
>  }
>  
> @@ -4893,6 +4963,8 @@ kvm_calc_shadow_npt_root_page_role(struct kvm_vcpu *vcpu,
>  
>  	role.base.direct = false;
>  	role.base.level = kvm_mmu_get_tdp_level(vcpu);
> +	if (role.base.level == PT32E_ROOT_LEVEL)
> +		role.base.pae_root = 1;
>  
>  	return role;
>  }
> diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h
> index 67489a060eba..1015f33e0758 100644
> --- a/arch/x86/kvm/mmu/paging_tmpl.h
> +++ b/arch/x86/kvm/mmu/paging_tmpl.h
> @@ -1043,6 +1043,7 @@ static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
>  		.access = 0x7,
>  		.quadrant = 0x3,
>  		.glevel = 0xf,
> +		.pae_root = 0x1,
>  	};
>  
>  	/*
> -- 
> 2.19.1.6.gb485710b
>