On Fri, Oct 15, 2021, Vitaly Kuznetsov wrote:
> Several selftests (memslot_modification_stress_test, kvm_page_table_test,
> dirty_log_perf_test,.. ) which rely on vm_get_max_gfn() started to fail
> since commit ef4c9f4f65462 ("KVM: selftests: Fix 32-bit truncation of
> vm_get_max_gfn()") on AMD EPYC 7401P:
>
> ./tools/testing/selftests/kvm/demand_paging_test
> Testing guest mode: PA-bits:ANY, VA-bits:48, 4K pages
> guest physical test memory offset: 0xffffbffff000
This look a lot like the signature I remember from the original bug[1]. I assume
you're hitting the magic HyperTransport region[2]. I thought that was fixed, but
the hack-a-fix for selftests never got applied[3].
[1] https://lore.kernel.org/lkml/20210623230552.4027702-4-seanjc@xxxxxxxxxx/
[2] https://lkml.kernel.org/r/7e3a90c0-75a1-b8fe-dbcf-bda16502ace9@xxxxxxx
[3] https://lkml.kernel.org/r/20210805105423.412878-1-pbonzini@xxxxxxxxxx
> Finished creating vCPUs and starting uffd threads
> Started all vCPUs
> ==== Test Assertion Failure ====
> demand_paging_test.c:63: false
> pid=47131 tid=47134 errno=0 - Success
> 1 0x000000000040281b: vcpu_worker at demand_paging_test.c:63
> 2 0x00007fb36716e431: ?? ??:0
> 3 0x00007fb36709c912: ?? ??:0
> Invalid guest sync status: exit_reason=SHUTDOWN
>
> The commit, however, seems to be correct, it just revealed an already
> present issue. AMD CPUs which support SEV may have a reduced physical
> address space, e.g. on AMD EPYC 7401P I see:
>
> Address sizes: 43 bits physical, 48 bits virtual
>
> The guest physical address space, however, is not reduced as stated in
> commit e39f00f60ebd ("KVM: x86: Use kernel's x86_phys_bits to handle
> reduced MAXPHYADDR"). This seems to be almost correct, however, APM has one
> more clause (15.34.6):
>
> Note that because guest physical addresses are always translated through
> the nested page tables, the size of the guest physical address space is
> not impacted by any physical address space reduction indicated in CPUID
> 8000_001F[EBX]. If the C-bit is a physical address bit however, the guest
> physical address space is effectively reduced by 1 bit.
>
> Implement the reduction.
>
> Fixes: e39f00f60ebd (KVM: x86: Use kernel's x86_phys_bits to handle reduced MAXPHYADDR)
> Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx>
> ---
> - RFC: I may have misdiagnosed the problem as I didn't dig to where exactly
> the guest crashes.
> ---
> arch/x86/kvm/cpuid.c | 13 ++++++++++---
> 1 file changed, 10 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
> index 751aa85a3001..04ae280a0b66 100644
> --- a/arch/x86/kvm/cpuid.c
> +++ b/arch/x86/kvm/cpuid.c
> @@ -923,13 +923,20 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
> *
> * If TDP is enabled but an explicit guest MAXPHYADDR is not
> * provided, use the raw bare metal MAXPHYADDR as reductions to
> - * the HPAs do not affect GPAs.
> + * the HPAs do not affect GPAs. The value, however, has to be
> + * reduced by 1 in case C-bit is a physical bit (APM section
> + * 15.34.6).
> */
> - if (!tdp_enabled)
> + if (!tdp_enabled) {
> g_phys_as = boot_cpu_data.x86_phys_bits;
> - else if (!g_phys_as)
> + } else if (!g_phys_as) {
> g_phys_as = phys_as;
>
> + if (kvm_cpu_cap_has(X86_FEATURE_SEV) &&
> + (cpuid_ebx(0x8000001f) & 0x3f) < g_phys_as)
> + g_phys_as -= 1;
This is incorrect, non-SEV guests do not see a reduced address space. See Tom's
explanation[*]
[*] https://lkml.kernel.org/r/324a95ee-b962-acdf-9bd7-b8b23b9fb991@xxxxxxx
> + }
> +
> entry->eax = g_phys_as | (virt_as << 8);
> entry->edx = 0;
> cpuid_entry_override(entry, CPUID_8000_0008_EBX);
> --
> 2.31.1
>
