[PATCH 0/2] KVM: x86: Fix mostly theoretical undefined behavior

From: Sean Christopherson
Date: Wed Sep 29 2021 - 18:24:36 EST


Fix a mostly theoretical undefined behavior bug due to consuming
uninitialized data when searching for a matching CPUID entry during vCPU
RESET/INIT. The bug is mostly theoretical because it requires very
aggressive inlining from the compiler, as well as deliberate "sabotage"
from the compiler (which _is_ allowed by the C standard) in the face of
known uninitialized data.

Patch 1, the "fix" that is tagged for stable, is all kinds of gross in that
it doesn't directly address uninitialized data, and instead tweaks a low
level CPUID helper to avoid consuming the uninitialized data. I went that
route for the fix so that the fix would be easily/directy consumable
downstream, as porting the fix from v5.15-rcN to literally any other buggy
kernel would require hand coding the fix due to refactoring and code
movement across files.

Patch 2 directly addresses the uninitialized data.

If patch 1 is unpalatable, an alternative would be to do a bit of merge
magic and feed in a fix to initialize "dummy" in svm.c, which was the only
buggy path prior to v5.15-rcN. However, KVM lived from 2012-2020 with
what's effectively the behavior after applying patch 1, and no one noticed
that the behavior was broken in 2020 until v5.15-rc1 introduced the bad
behavior to VMX, i.e. opened up the validation surface to bots that
presumably run the majority of their cycles on Intel CPUs.

Sean Christopherson (2):
KVM: x86: Swap order of CPUID entry "index" vs. "significant flag"
checks
KVM: x86: Manually retrieve CPUID.0x1 when getting FMS for RESET/INIT

arch/x86/kvm/cpuid.c | 4 ++--
arch/x86/kvm/x86.c | 11 +++++------
2 files changed, 7 insertions(+), 8 deletions(-)

--
2.33.0.685.g46640cef36-goog