[PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Eric Northup]

I've been reading the x86's mmu.c recently and had been wondering
about something.  Avi's recent mmu documentation (thanks!) seems to
have confirmed my understanding of how the shadow paging is supposed
to be working.  In TDP mode, when mmu_alloc_roots() calls
kvm_mmu_get_page(), why does it pass (vcpu->arch.cr3 >> PAGE_SHIFT) or
(vcpu->arch.mmu.pae_root[i]) as gfn?

It seems to me that in TDP mode, gfn should be either zero for the
root page table, or 0/1GB/2GB/3GB (for PAE page tables).

The existing behavior can lead to multiple, semantically-identical TDP
roots being created by mmu_alloc_roots, depending on the VCPU's CR3 at
the time that mmu_alloc_roots was called.  But the nested page tables
should be* independent of the VCPU state. That wastes some memory and
causes extra page faults while populating the extra copies of the page
tables.

*assuming that we aren't modeling per-VCPU state that might change the
physical address map as seen by that VCPU, such as setting the APIC
base to an address overlapping RAM.

All feedback would be welcome, since I'm new to this system!  A
strawman patch follows.

thanks,
-Eric

--

For TDP mode, avoid creating multiple page table roots for the single
guest-to-host physical address map by fixing the inputs used for the
shadow page table hash in mmu_alloc_roots().

Signed-off-by: Eric Northup <digitaleric@google.com>
---
 arch/x86/kvm/mmu.c |   12 ++++++++----
 1 files changed, 8 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index ddfa865..9696d65 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -2059,10 +2059,12 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
                hpa_t root = vcpu->arch.mmu.root_hpa;

                ASSERT(!VALID_PAGE(root));
-               if (tdp_enabled)
-                       direct = 1;
                if (mmu_check_root(vcpu, root_gfn))
                        return 1;
+               if (tdp_enabled) {
+                       direct = 1;
+                       root_gfn = 0;
+               }
                sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
                                      PT64_ROOT_LEVEL, direct,
                                      ACC_ALL, NULL);
@@ -2072,8 +2074,6 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
                return 0;
        }
        direct = !is_paging(vcpu);
-       if (tdp_enabled)
-               direct = 1;
        for (i = 0; i < 4; ++i) {
                hpa_t root = vcpu->arch.mmu.pae_root[i];

@@ -2089,6 +2089,10 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
                        root_gfn = 0;
                if (mmu_check_root(vcpu, root_gfn))
                        return 1;
+               if (tdp_enabled) {
+                       direct = 1;
+                       root_gfn = i << 30;
+               }
                sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
                                      PT32_ROOT_LEVEL, direct,
                                      ACC_ALL, NULL);
--

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Avi Kivity]

On 04/23/2010 12:15 AM, Eric Northup wrote:
> I've been reading the x86's mmu.c recently and had been wondering
> about something.  Avi's recent mmu documentation (thanks!) seems to
> have confirmed my understanding of how the shadow paging is supposed
> to be working.

Wasn't it a lot more fun to understand it from the code?  reading the 
documentation is way to easy.

> In TDP mode, when mmu_alloc_roots() calls
> kvm_mmu_get_page(), why does it pass (vcpu->arch.cr3>>  PAGE_SHIFT) or
> (vcpu->arch.mmu.pae_root[i]) as gfn?
>    

Historical accident.  Luckily, no harmful effects other than wasting a 
bit of cpu cycles every now and then.

> It seems to me that in TDP mode, gfn should be either zero for the
> root page table, or 0/1GB/2GB/3GB (for PAE page tables).
>
> The existing behavior can lead to multiple, semantically-identical TDP
> roots being created by mmu_alloc_roots, depending on the VCPU's CR3 at
> the time that mmu_alloc_roots was called.  But the nested page tables
> should be* independent of the VCPU state. That wastes some memory and
> causes extra page faults while populating the extra copies of the page
> tables.
>    

Yeah.

> *assuming that we aren't modeling per-VCPU state that might change the
> physical address map as seen by that VCPU, such as setting the APIC
> base to an address overlapping RAM.
>    

We don't actually support that.

> All feedback would be welcome, since I'm new to this system!  A
> strawman patch follows.
>
>    

Patch is correct, but I have already fixed this in a more extensive 
patch set (that also folds the 32-bit and 64-bit cases together, etc.) 
and I'm too lazy to rebase on top of yours.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Avi Kivity]

On 04/23/2010 02:11 PM, Avi Kivity wrote:
>
>> All feedback would be welcome, since I'm new to this system!  A
>> strawman patch follows.
>>
>
> Patch is correct, but I have already fixed this in a more extensive 
> patch set (that also folds the 32-bit and 64-bit cases together, etc.) 
> and I'm too lazy to rebase on top of yours.

Well my patchset is broken.  Marcelo, please apply this and I'll rebase 
and fix.

-- 
error compiling committee.c: too many arguments to function

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Marcelo Tosatti]

On Thu, Apr 22, 2010 at 02:15:14PM -0700, Eric Northup wrote:
> I've been reading the x86's mmu.c recently and had been wondering
> about something.  Avi's recent mmu documentation (thanks!) seems to
> have confirmed my understanding of how the shadow paging is supposed
> to be working.  In TDP mode, when mmu_alloc_roots() calls
> kvm_mmu_get_page(), why does it pass (vcpu->arch.cr3 >> PAGE_SHIFT) or
> (vcpu->arch.mmu.pae_root[i]) as gfn?
> 
> It seems to me that in TDP mode, gfn should be either zero for the
> root page table, or 0/1GB/2GB/3GB (for PAE page tables).
> 
> The existing behavior can lead to multiple, semantically-identical TDP
> roots being created by mmu_alloc_roots, depending on the VCPU's CR3 at
> the time that mmu_alloc_roots was called.  But the nested page tables
> should be* independent of the VCPU state. That wastes some memory and
> causes extra page faults while populating the extra copies of the page
> tables.
> 
> *assuming that we aren't modeling per-VCPU state that might change the
> physical address map as seen by that VCPU, such as setting the APIC
> base to an address overlapping RAM.
> 
> All feedback would be welcome, since I'm new to this system!  A
> strawman patch follows.
> 
> thanks,
> -Eric
> 
> --
> 
> For TDP mode, avoid creating multiple page table roots for the single
> guest-to-host physical address map by fixing the inputs used for the
> shadow page table hash in mmu_alloc_roots().
> 
> Signed-off-by: Eric Northup <digitaleric@google.com>
> ---
>  arch/x86/kvm/mmu.c |   12 ++++++++----
>  1 files changed, 8 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index ddfa865..9696d65 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -2059,10 +2059,12 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
>                 hpa_t root = vcpu->arch.mmu.root_hpa;
> 
>                 ASSERT(!VALID_PAGE(root));
> -               if (tdp_enabled)
> -                       direct = 1;
>                 if (mmu_check_root(vcpu, root_gfn))
>                         return 1;
> +               if (tdp_enabled) {
> +                       direct = 1;
> +                       root_gfn = 0;
> +               }
>                 sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
>                                       PT64_ROOT_LEVEL, direct,
>                                       ACC_ALL, NULL);
> @@ -2072,8 +2074,6 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
>                 return 0;
>         }
>         direct = !is_paging(vcpu);
> -       if (tdp_enabled)
> -               direct = 1;
>         for (i = 0; i < 4; ++i) {
>                 hpa_t root = vcpu->arch.mmu.pae_root[i];
> 
> @@ -2089,6 +2089,10 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
>                         root_gfn = 0;
>                 if (mmu_check_root(vcpu, root_gfn))
>                         return 1;
> +               if (tdp_enabled) {
> +                       direct = 1;
> +                       root_gfn = i << 30;
> +               }
>                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
>                                       PT32_ROOT_LEVEL, direct,
>                                       ACC_ALL, NULL);

There is no need to allocate 4 different roots for TDP tables if
kvm_x86_ops->get_tdp_level() == PT64_ROOT_LEVEL.

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Marcelo Tosatti]

On Mon, Apr 26, 2010 at 06:30:00PM -0300, Marcelo Tosatti wrote:
> > @@ -2089,6 +2089,10 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
> >                         root_gfn = 0;
> >                 if (mmu_check_root(vcpu, root_gfn))
> >                         return 1;
> > +               if (tdp_enabled) {
> > +                       direct = 1;
> > +                       root_gfn = i << 30;
> > +               }
> >                 sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
> >                                       PT32_ROOT_LEVEL, direct,
> >                                       ACC_ALL, NULL);
> 
> There is no need to allocate 4 different roots for TDP tables if
> kvm_x86_ops->get_tdp_level() == PT64_ROOT_LEVEL.

Doh, and your patch does not. But it does not apply to kvm.git -next 
branch, can you regenerate please?

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Eric Northup]

On Mon, Apr 26, 2010 at 2:46 PM, Marcelo Tosatti <mtosatti@redhat.com> wrote:
> Doh, and your patch does not. But it does not apply to kvm.git -next
> branch, can you regenerate please?

--

For TDP mode, avoid creating multiple page table roots for the single
guest-to-host physical address map by fixing the inputs used for the
shadow page table hash in mmu_alloc_roots().

Signed-off-by: Eric Northup <digitaleric@google.com>
---

diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
index ddfa865..9696d65 100644
--- a/arch/x86/kvm/mmu.c
+++ b/arch/x86/kvm/mmu.c
@@ -2059,10 +2059,12 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
 		hpa_t root = vcpu->arch.mmu.root_hpa;

 		ASSERT(!VALID_PAGE(root));
-		if (tdp_enabled)
-			direct = 1;
 		if (mmu_check_root(vcpu, root_gfn))
 			return 1;
+		if (tdp_enabled) {
+			direct = 1;
+			root_gfn = 0;
+		}
 		sp = kvm_mmu_get_page(vcpu, root_gfn, 0,
 				      PT64_ROOT_LEVEL, direct,
 				      ACC_ALL, NULL);
@@ -2072,8 +2074,6 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 	direct = !is_paging(vcpu);
-	if (tdp_enabled)
-		direct = 1;
 	for (i = 0; i < 4; ++i) {
 		hpa_t root = vcpu->arch.mmu.pae_root[i];

@@ -2089,6 +2089,10 @@ static int mmu_alloc_roots(struct kvm_vcpu *vcpu)
 			root_gfn = 0;
 		if (mmu_check_root(vcpu, root_gfn))
 			return 1;
+		if (tdp_enabled) {
+			direct = 1;
+			root_gfn = i << 30;
+		}
 		sp = kvm_mmu_get_page(vcpu, root_gfn, i << 30,
 				      PT32_ROOT_LEVEL, direct,
 				      ACC_ALL, NULL);
--

Re: [PATCH RFC] KVM MMU: fix hashing for TDP and non-paging modes

[Marcelo Tosatti]

On Mon, Apr 26, 2010 at 05:00:05PM -0700, Eric Northup wrote:
> On Mon, Apr 26, 2010 at 2:46 PM, Marcelo Tosatti <mtosatti@redhat.com> wrote:
> > Doh, and your patch does not. But it does not apply to kvm.git -next
> > branch, can you regenerate please?
> 
> --
> 
> For TDP mode, avoid creating multiple page table roots for the single
> guest-to-host physical address map by fixing the inputs used for the
> shadow page table hash in mmu_alloc_roots().
> 
> Signed-off-by: Eric Northup <digitaleric@google.com>

Applied, thanks.