Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 4 Aug 2022 15:54:20 -0700
From:   David Matlack <dmatlack@google.com>
To:     isaku.yamahata@intel.com
Cc:     kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
        isaku.yamahata@gmail.com, Paolo Bonzini <pbonzini@redhat.com>,
        erdemaktas@google.com, Sean Christopherson <seanjc@google.com>,
        Sagi Shahar <sagis@google.com>
Subject: Re: [RFC PATCH v6 037/104] KVM: x86/mmu: Allow non-zero value for
 non-present SPTE
Message-ID: <YuxOHPpkhKnnstqw@google.com>
References: <cover.1651774250.git.isaku.yamahata@intel.com>
 <bfa4f7415a1d059bd3a4c6d14105f2baf2d03ba6.1651774250.git.isaku.yamahata@intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <bfa4f7415a1d059bd3a4c6d14105f2baf2d03ba6.1651774250.git.isaku.yamahata@intel.com>
Precedence: bulk

On Thu, May 05, 2022 at 11:14:31AM -0700, isaku.yamahata@intel.com wrote:
> From: Sean Christopherson <sean.j.christopherson@intel.com>
> 
> TDX introduced a new ETP, Secure-EPT, in addition to the existing EPT.
> Secure-EPT maps protected guest memory, which is called private. Since
> Secure-EPT page tables is also protected, those page tables is also called
> private.  The existing EPT is often called shared EPT to distinguish from
> Secure-EPT.  And also page tables for share EPT is also called shared.
> 
> Virtualization Exception, #VE, is a new processor exception in VMX non-root
> operation.  In certain virtualizatoin-related conditions, #VE is injected
> into guest instead of exiting from guest to VMM so that guest is given a
> chance to inspect it.  One important one is EPT violation.  When
> "ETP-violation #VE" VM-execution is set, "#VE suppress bit" in EPT entry
> is cleared, #VE is injected instead of EPT violation.
> 
> Because guest memory is protected with TDX, VMM can't parse instructions
> in the guest memory.  Instead, MMIO hypercall is used for guest to pass
> necessary information to VMM.
> 
> To make unmodified device driver work, guest TD expects #VE on accessing
> shared GPA.  The #VE handler converts MMIO access into MMIO hypercall with
> the EPT entry of enabled "#VE" by clearing "suppress #VE" bit.  Before VMM
> enabling #VE, it needs to figure out the given GPA is for MMIO by EPT
> violation.  So the execution flow looks like
> 
> - Allocate unused shared EPT entry with suppress #VE bit set.
> - EPT violation on that GPA.
> - VMM figures out the faulted GPA is for MMIO.
> - VMM clears the suppress #VE bit.
> - Guest TD gets #VE, and converts MMIO access into MMIO hypercall.
> - If the GPA maps guest memory, VMM resolves it with guest pages.
> 
> For both cases, SPTE needs suppress #VE" bit set initially when it
> is allocated or zapped, therefore non-zero non-present value for SPTE
> needs to be allowed.
> 
> This change requires to update FNAME(sync_page) for shadow EPT.
> "if(!sp->spte[i])" in FNAME(sync_page) means that the spte entry is the
> initial value.  With the introduction of shadow_nonpresent_value which can
> be non-zero, it doesn't hold any more. Replace zero check with
> "!is_shadow_present_pte() && !is_mmio_spte()".
> 
> When "if (!spt[i])" doesn't hold, but the entry value is
> shadow_nonpresent_value, the entry is wrongly synchronized from non-present
> to non-present with (wrongly) pfn changed and tries to remove rmap wrongly
> and BUG_ON() is hit.
> 
> TDP MMU uses REMOVED_SPTE = 0x5a0ULL as special constant to indicate the
> intermediate value to indicate one thread is operating on it and the value
> should be semi-arbitrary value.  For TDX (more correctly to use #VE), the
> value should include suppress #VE value which is SHADOW_NONPRESENT_VALUE.
> Rename REMOVED_SPTE to __REMOVED_SPTE and define REMOVED_SPTE as
> SHADOW_NONPRESENT_VALUE | REMOVED_SPTE to set suppress #VE bit.
> 
> Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
> Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
>  arch/x86/kvm/mmu/mmu.c         | 55 ++++++++++++++++++++++++++++++----
>  arch/x86/kvm/mmu/paging_tmpl.h |  3 +-
>  arch/x86/kvm/mmu/spte.c        |  5 +++-
>  arch/x86/kvm/mmu/spte.h        | 37 ++++++++++++++++++++---
>  arch/x86/kvm/mmu/tdp_mmu.c     | 23 +++++++++-----
>  5 files changed, 105 insertions(+), 18 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> index 4a12d862bbb6..324ea25ee0c7 100644
> --- a/arch/x86/kvm/mmu/mmu.c
> +++ b/arch/x86/kvm/mmu/mmu.c
> @@ -693,6 +693,44 @@ static void walk_shadow_page_lockless_end(struct kvm_vcpu *vcpu)
>  	}
>  }
>  
> +static inline void kvm_init_shadow_page(void *page)
> +{
> +#ifdef CONFIG_X86_64
> +	int ign;
> +
> +	WARN_ON_ONCE(shadow_nonpresent_value != SHADOW_NONPRESENT_VALUE);
> +	asm volatile (
> +		"rep stosq\n\t"
> +		: "=c"(ign), "=D"(page)
> +		: "a"(SHADOW_NONPRESENT_VALUE), "c"(4096/8), "D"(page)
> +		: "memory"
> +	);

Use memset64()?

> +#else
> +	BUG();
> +#endif
> +}
> +
> +static int mmu_topup_shadow_page_cache(struct kvm_vcpu *vcpu)
> +{
> +	struct kvm_mmu_memory_cache *mc = &vcpu->arch.mmu_shadow_page_cache;
> +	int start, end, i, r;
> +	bool is_tdp_mmu = is_tdp_mmu_enabled(vcpu->kvm);
> +
> +	if (is_tdp_mmu && shadow_nonpresent_value)
> +		start = kvm_mmu_memory_cache_nr_free_objects(mc);
> +
> +	r = kvm_mmu_topup_memory_cache(mc, PT64_ROOT_MAX_LEVEL);
> +	if (r)
> +		return r;
> +
> +	if (is_tdp_mmu && shadow_nonpresent_value) {
> +		end = kvm_mmu_memory_cache_nr_free_objects(mc);
> +		for (i = start; i < end; i++)
> +			kvm_init_shadow_page(mc->objects[i]);
> +	}

Doing this during top-up is probably the right decision since we're
outside the MMU lock. In v8 you'll need to also cover the eager page
splitting code paths, which go through a different allocation path for
the shadow and TDP MMU.

> +	return 0;
> +}
> +
>  static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
>  {
>  	int r;
> @@ -702,8 +740,7 @@ static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu, bool maybe_indirect)
>  				       1 + PT64_ROOT_MAX_LEVEL + PTE_PREFETCH_NUM);
>  	if (r)
>  		return r;
> -	r = kvm_mmu_topup_memory_cache(&vcpu->arch.mmu_shadow_page_cache,
> -				       PT64_ROOT_MAX_LEVEL);
> +	r = mmu_topup_shadow_page_cache(vcpu);
>  	if (r)
>  		return r;
>  	if (maybe_indirect) {
> @@ -5510,9 +5547,16 @@ void kvm_configure_mmu(bool enable_tdp, int tdp_forced_root_level,
>  	 * what is used by the kernel for any given HVA, i.e. the kernel's
>  	 * capabilities are ultimately consulted by kvm_mmu_hugepage_adjust().
>  	 */
> -	if (tdp_enabled)
> +	if (tdp_enabled) {
> +		/*
> +		 * For TDP MMU, always set bit 63 for TDX support. See the
> +		 * comment on SHADOW_NONPRESENT_VALUE.
> +		 */
> +#ifdef CONFIG_X86_64
> +		shadow_nonpresent_value = SHADOW_NONPRESENT_VALUE;
> +#endif
>  		max_huge_page_level = tdp_huge_page_level;
> -	else if (boot_cpu_has(X86_FEATURE_GBPAGES))
> +	} else if (boot_cpu_has(X86_FEATURE_GBPAGES))
>  		max_huge_page_level = PG_LEVEL_1G;
>  	else
>  		max_huge_page_level = PG_LEVEL_2M;
> @@ -5643,7 +5687,8 @@ int kvm_mmu_create(struct kvm_vcpu *vcpu)
>  	vcpu->arch.mmu_page_header_cache.kmem_cache = mmu_page_header_cache;
>  	vcpu->arch.mmu_page_header_cache.gfp_zero = __GFP_ZERO;
>  
> -	vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;
> +	if (!(is_tdp_mmu_enabled(vcpu->kvm) && shadow_nonpresent_value))
> +		vcpu->arch.mmu_shadow_page_cache.gfp_zero = __GFP_ZERO;

Is there any reason to prefer using __GFP_ZERO? I suspect the code would
be simpler if KVM unconditionally initialized shadow pages.

>  
>  	vcpu->arch.mmu = &vcpu->arch.root_mmu;
>  	vcpu->arch.walk_mmu = &vcpu->arch.root_mmu;
> diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h
> index b025decf610d..058efd4bbcbc 100644
> --- a/arch/x86/kvm/mmu/paging_tmpl.h
> +++ b/arch/x86/kvm/mmu/paging_tmpl.h
> @@ -1030,7 +1030,8 @@ static int FNAME(sync_page)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *sp)
>  		gpa_t pte_gpa;
>  		gfn_t gfn;
>  
> -		if (!sp->spt[i])
> +		if (!is_shadow_present_pte(sp->spt[i]) &&
> +		    !is_mmio_spte(sp->spt[i]))
>  			continue;
>  
>  		pte_gpa = first_pte_gpa + i * sizeof(pt_element_t);
> diff --git a/arch/x86/kvm/mmu/spte.c b/arch/x86/kvm/mmu/spte.c
> index 75c9e87d446a..1bf934f64b6f 100644
> --- a/arch/x86/kvm/mmu/spte.c
> +++ b/arch/x86/kvm/mmu/spte.c
> @@ -36,6 +36,9 @@ u64 __read_mostly shadow_present_mask;
>  u64 __read_mostly shadow_me_value;
>  u64 __read_mostly shadow_me_mask;
>  u64 __read_mostly shadow_acc_track_mask;
> +#ifdef CONFIG_X86_64
> +u64 __read_mostly shadow_nonpresent_value;
> +#endif
>  
>  u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
>  u64 __read_mostly shadow_nonpresent_or_rsvd_lower_gfn_mask;
> @@ -330,7 +333,7 @@ void kvm_mmu_set_mmio_spte_mask(u64 mmio_value, u64 mmio_mask, u64 access_mask)
>  	 * not set any RWX bits.
>  	 */
>  	if (WARN_ON((mmio_value & mmio_mask) != mmio_value) ||
> -	    WARN_ON(mmio_value && (REMOVED_SPTE & mmio_mask) == mmio_value))
> +	    WARN_ON(mmio_value && (__REMOVED_SPTE & mmio_mask) == mmio_value))

Why use __REMOVED_SPTE here and not REMOVED_SPTE?

>  		mmio_value = 0;
>  
>  	if (!mmio_value)
> diff --git a/arch/x86/kvm/mmu/spte.h b/arch/x86/kvm/mmu/spte.h
> index fbbab180395e..3319ca7f8f48 100644
> --- a/arch/x86/kvm/mmu/spte.h
> +++ b/arch/x86/kvm/mmu/spte.h
> @@ -140,6 +140,19 @@ static_assert(MMIO_SPTE_GEN_LOW_BITS == 8 && MMIO_SPTE_GEN_HIGH_BITS == 11);
>  
>  #define MMIO_SPTE_GEN_MASK		GENMASK_ULL(MMIO_SPTE_GEN_LOW_BITS + MMIO_SPTE_GEN_HIGH_BITS - 1, 0)
>  
> +/*
> + * non-present SPTE value for both VMX and SVM for TDP MMU.
> + * For SVM NPT, for non-present spte (bit 0 = 0), other bits are ignored.
> + * For VMX EPT, bit 63 is ignored if #VE is disabled.
> + *              bit 63 is #VE suppress if #VE is enabled.
> + */
> +#ifdef CONFIG_X86_64
> +#define SHADOW_NONPRESENT_VALUE	BIT_ULL(63)
> +static_assert(!(SHADOW_NONPRESENT_VALUE & SPTE_MMU_PRESENT_MASK));
> +#else
> +#define SHADOW_NONPRESENT_VALUE	0ULL
> +#endif

The terminology "shadow_nonpresent" implies it would be the opposite of
e.g.  is_shadow_present_pte(), when in fact they are completely
different concepts.

Also, this is a good opportunity to follow the same naming terminology
as REMOVED_SPTE in the TDP MMU.

How about EMPTY_SPTE?

> +
>  extern u64 __read_mostly shadow_host_writable_mask;
>  extern u64 __read_mostly shadow_mmu_writable_mask;
>  extern u64 __read_mostly shadow_nx_mask;
> @@ -154,6 +167,12 @@ extern u64 __read_mostly shadow_present_mask;
>  extern u64 __read_mostly shadow_me_value;
>  extern u64 __read_mostly shadow_me_mask;
>  
> +#ifdef CONFIG_X86_64
> +extern u64 __read_mostly shadow_nonpresent_value;
> +#else
> +#define shadow_nonpresent_value	0ULL
> +#endif
> +
>  /*
>   * SPTEs in MMUs without A/D bits are marked with SPTE_TDP_AD_DISABLED_MASK;
>   * shadow_acc_track_mask is the set of bits to be cleared in non-accessed
> @@ -174,9 +193,12 @@ extern u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
>  
>  /*
>   * If a thread running without exclusive control of the MMU lock must perform a
> - * multi-part operation on an SPTE, it can set the SPTE to REMOVED_SPTE as a
> + * multi-part operation on an SPTE, it can set the SPTE to __REMOVED_SPTE as a
>   * non-present intermediate value. Other threads which encounter this value
> - * should not modify the SPTE.
> + * should not modify the SPTE.  For the case that TDX is enabled,
> + * SHADOW_NONPRESENT_VALUE, which is "suppress #VE" bit set because TDX module
> + * always enables "EPT violation #VE".  The bit is ignored by non-TDX case as
> + * present bit (bit 0) is cleared.
>   *
>   * Use a semi-arbitrary value that doesn't set RWX bits, i.e. is not-present on
>   * bot AMD and Intel CPUs, and doesn't set PFN bits, i.e. doesn't create a L1TF
> @@ -184,10 +206,17 @@ extern u64 __read_mostly shadow_nonpresent_or_rsvd_mask;
>   *
>   * Only used by the TDP MMU.
>   */
> -#define REMOVED_SPTE	0x5a0ULL
> +#define __REMOVED_SPTE	0x5a0ULL
>  
>  /* Removed SPTEs must not be misconstrued as shadow present PTEs. */
> -static_assert(!(REMOVED_SPTE & SPTE_MMU_PRESENT_MASK));
> +static_assert(!(__REMOVED_SPTE & SPTE_MMU_PRESENT_MASK));
> +static_assert(!(__REMOVED_SPTE & SHADOW_NONPRESENT_VALUE));
> +
> +/*
> + * See above comment around __REMOVED_SPTE.  REMOVED_SPTE is the actual
> + * intermediate value set to the removed SPET.  it sets the "suppress #VE" bit.
> + */
> +#define REMOVED_SPTE	(SHADOW_NONPRESENT_VALUE | __REMOVED_SPTE)
>  
>  static inline bool is_removed_spte(u64 spte)
>  {
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index 4fabb2cd0ba9..383904742f44 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -673,8 +673,16 @@ static inline int tdp_mmu_zap_spte_atomic(struct kvm *kvm,
>  	 * special removed SPTE value. No bookkeeping is needed
>  	 * here since the SPTE is going from non-present
>  	 * to non-present.
> +	 *
> +	 * Set non-present value to SHADOW_NONPRESENT_VALUE, rather than 0.
> +	 * It is because when TDX is enabled, TDX module always
> +	 * enables "EPT-violation #VE", so KVM needs to set
> +	 * "suppress #VE" bit in EPT table entries, in order to get
> +	 * real EPT violation, rather than TDVMCALL.  KVM sets
> +	 * SHADOW_NONPRESENT_VALUE (which sets "suppress #VE" bit) so it
> +	 * can be set when EPT table entries are zapped.
>  	 */
> -	kvm_tdp_mmu_write_spte(iter->sptep, 0);
> +	kvm_tdp_mmu_write_spte(iter->sptep, SHADOW_NONPRESENT_VALUE);
>  
>  	return 0;
>  }
> @@ -846,8 +854,8 @@ static void __tdp_mmu_zap_root(struct kvm *kvm, struct kvm_mmu_page *root,
>  			continue;
>  
>  		if (!shared)
> -			tdp_mmu_set_spte(kvm, &iter, 0);
> -		else if (tdp_mmu_set_spte_atomic(kvm, &iter, 0))
> +			tdp_mmu_set_spte(kvm, &iter, SHADOW_NONPRESENT_VALUE);
> +		else if (tdp_mmu_set_spte_atomic(kvm, &iter, SHADOW_NONPRESENT_VALUE))
>  			goto retry;
>  	}
>  }
> @@ -903,8 +911,9 @@ bool kvm_tdp_mmu_zap_sp(struct kvm *kvm, struct kvm_mmu_page *sp)
>  	if (WARN_ON_ONCE(!is_shadow_present_pte(old_spte)))
>  		return false;
>  
> -	__tdp_mmu_set_spte(kvm, kvm_mmu_page_as_id(sp), sp->ptep, old_spte, 0,
> -			   sp->gfn, sp->role.level + 1, true, true);
> +	__tdp_mmu_set_spte(kvm, kvm_mmu_page_as_id(sp), sp->ptep, old_spte,
> +			   SHADOW_NONPRESENT_VALUE, sp->gfn, sp->role.level + 1,
> +			   true, true);
>  
>  	return true;
>  }
> @@ -941,7 +950,7 @@ static bool tdp_mmu_zap_leafs(struct kvm *kvm, struct kvm_mmu_page *root,
>  		    !is_last_spte(iter.old_spte, iter.level))
>  			continue;
>  
> -		tdp_mmu_set_spte(kvm, &iter, 0);
> +		tdp_mmu_set_spte(kvm, &iter, SHADOW_NONPRESENT_VALUE);
>  		flush = true;
>  	}
>  
> @@ -1312,7 +1321,7 @@ static bool set_spte_gfn(struct kvm *kvm, struct tdp_iter *iter,
>  	 * invariant that the PFN of a present * leaf SPTE can never change.
>  	 * See __handle_changed_spte().
>  	 */
> -	tdp_mmu_set_spte(kvm, iter, 0);
> +	tdp_mmu_set_spte(kvm, iter, SHADOW_NONPRESENT_VALUE);
>  
>  	if (!pte_write(range->pte)) {
>  		new_spte = kvm_mmu_changed_pte_notifier_make_spte(iter->old_spte,

In addition to the suggestions above, I'd suggest breaking this patch
up, since it is doing multiple things:

1. Patch initialize shadow page tables to EMPTY_SPTE (0) and
   replace TDP MMU hard-coded 0 with EMPTY_SPTE.
2. Patch to change FNAME(sync_page) to not assume EMPTY_SPTE is 0.
3. Patch to set bit 63 in EMPTY_SPTE.
4. Patch to set bit 63 in REMOVED_SPTE.

> -- 
> 2.25.1
>