linux-kernel - Re: [PATCH v2 00/23] Extend test coverage for nested SVM

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <vh3yjo36ortltqjrcsegllzbpkmum2c5ywna25q3ah25txlv74@4edzsqjjs73c>
Date: Wed, 19 Nov 2025 00:01:50 +0000
From: Yosry Ahmed <yosry.ahmed@...ux.dev>
To: Sean Christopherson <seanjc@...gle.com>
Cc: Jim Mattson <jmattson@...gle.com>, Paolo Bonzini <pbonzini@...hat.com>, 
	kvm@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 00/23] Extend test coverage for nested SVM

On Tue, Nov 18, 2025 at 03:49:55PM -0800, Sean Christopherson wrote:
> On Tue, Nov 18, 2025, Yosry Ahmed wrote:
> > On Tue, Nov 18, 2025 at 03:00:26PM -0800, Jim Mattson wrote:
> > > On Tue, Nov 18, 2025 at 2:26 PM Yosry Ahmed <yosry.ahmed@...ux.dev> wrote:
> > > >
> > > > On Tue, Oct 21, 2025 at 07:47:13AM +0000, Yosry Ahmed wrote:
> > > > > There are multiple selftests exercising nested VMX that are not specific
> > > > > to VMX (at least not anymore). Extend their coverage to nested SVM.
> > > > >
> > > > > This version is significantly different (and longer) than v1 [1], mainly
> > > > > due to the change of direction to reuse __virt_pg_map() for nested EPT/NPT
> > > > > mappings instead of extending the existing nested EPT infrastructure. It
> > > > > also has a lot more fixups and cleanups.
> > > > >
> > > > > This series depends on two other series:
> > > > > - "KVM: SVM: GIF and EFER.SVME are independent" [2]
> > > > > - "KVM: selftests: Add test of SET_NESTED_STATE with 48-bit L2 on 57-bit L1" [3]
> > > >
> > > > v2 of Jim's series switches all tests to use 57-bit by default when
> > > > available:
> > > > https://lore.kernel.org/kvm/20251028225827.2269128-4-jmattson@google.com/
> > > >
> > > > This breaks moving nested EPT mappings to use __virt_pg_map() because
> > > > nested EPTs are hardcoded to use 4-level paging, while __virt_pg_map()
> > > > will assume we're using 5-level paging.
> > > >
> > > > Patch #16 ("KVM: selftests: Use __virt_pg_map() for nested EPTs") will
> > > > need the following diff to make nested EPTs use the same paging level as
> > > > the guest:
> > > >
> > > > diff --git a/tools/testing/selftests/kvm/lib/x86_64/vmx.c b/tools/testing/selftests/kvm/lib/x86_64/vmx.c
> > > > index 358143bf8dd0d..8bacb74c00053 100644
> > > > --- a/tools/testing/selftests/kvm/lib/x86/vmx.c
> > > > +++ b/tools/testing/selftests/kvm/lib/x86/vmx.c
> > > > @@ -203,7 +203,7 @@ static inline void init_vmcs_control_fields(struct vmx_pages *vmx)
> > > >                 uint64_t ept_paddr;
> > > >                 struct eptPageTablePointer eptp = {
> > > >                         .memory_type = X86_MEMTYPE_WB,
> > > > -                       .page_walk_length = 3, /* + 1 */
> > > > +                       .page_walk_length = get_cr4() & X86_CR4_LA57 ? 4 : 3, /* + 1 */
> > > 
> > > LA57 does not imply support for 5-level EPT. (SRF, IIRC)
> 
> Yuuuup.  And similarly, MAXPHYADDR=52 doesn't imply 5-level EPT (thank you TDX!).
> 
> > Huh, that's annoying. We can keep the EPTs hardcoded to 4 levels and
> > pass in the max level to __virt_pg_map() instead of hardcoding
> > vm->pgtable_levels.
> 
> I haven't looked at the series in-depth so I don't know exactly what you're trying
> to do, but why not check MSR_IA32_VMX_EPT_VPID_CAP for PWL5?

The second part of the series reuses __virt_pg_map() to be used for
nested EPTs (and NPTs). __virt_pg_map() uses vm->pgtable_levels to find
out how many page table levels we have.

So we need to either:

(a) Always use the same number of levels for page tables and EPTs.

(b) Make __virt_pg_map() take the number of page table levels as a
  parameter, and always pass 4 for EPTs (for now).

I suggested (a) initially, but it doesn't work because we can
technically have LA57 but not MSR_IA32_VMX_EPT_VPID_CAP, so we need to
do (b). We can still check MSR_IA32_VMX_EPT_VPID_CAP and use PWL5 for
EPTs, but that's an orthogonal change at this point.

Anyway, do you prefer that I resend the series on top of Jim's v2, or do
you want to wait and see if you'll fix it up (or apply a part of it
before I rebase the rest)?