[PATCH 0/3] x86/sev: Make guest termination reasons identifiable
From: Kim Phillips
Date: Tue Nov 17 2026 - 17:33:22 EST
When an SEV-ES/SNP guest dies early, the only thing that reliably escapes
the guest is the GHCB MSR termination request: a 4-bit reason set and an
8-bit reason code written to the GHCB MSR before the final VMGEXIT. There
is no console, no dmesg and no crash dump to fall back on - in a hosted
environment the guest owner often has nothing but that one code to work
with.
Today a large fraction of the termination sites in the guest-side SEV code
collapse onto a handful of values that say almost nothing about what went
wrong:
- Six sites terminate with the generic (SEV_TERM_SET_GEN,
GHCB_SEV_ES_GEN_REQ) pair. A #VC the decompressor could not handle, a
#VC the running kernel could not handle, a GHCB that could not be set
up, and a GHCB protocol negotiation that failed are all reported
identically.
- do_vc_no_ghcb(), the MSR-protocol #VC handler used before a GHCB
exists, funnels every one of its failure paths through a single label,
so a CPUID table lookup failure, a failed hypervisor CPUID fallback and
a hypervisor lying about SEV being active are indistinguishable.
- Both find_cc_blob() implementations report GHCB_SNP_UNSUPPORTED when
the Confidential Computing blob they found has a bad header magic. That
code otherwise means "the hypervisor does not support SNP", so a
firmware or bootloader bug inside the guest's own boot chain looks
exactly like an unsupported host - and points debugging at the wrong
party.
The cost of fixing this is low: the Linux-specific reason set
(SEV_TERM_SET_LINUX) exists precisely so the guest can report codes of its
own choosing, and the existing GHCB_TERM_* codes already follow this
pattern for more recently added features such as Secure TSC, SVSM and
Secure AVIC. These patches extend that convention to the older, more
generic sites.
Patch 1 replaces the six generic callsites, adding GHCB_TERM_VC_FAIL,
GHCB_TERM_GHCB_SETUP and GHCB_TERM_NEGOTIATE. The four #VC handling
failures deliberately share one code, as they are the same failure class
observed at different stages of boot.
Patch 2 splits do_vc_no_ghcb()'s failure paths, reusing the existing
GHCB_TERM_CPUID and GHCB_TERM_CPUID_HV codes - no new codes needed. The
two checks that reject the #VC before any CPUID work is attempted keep
reporting GHCB_TERM_VC_FAIL.
Patch 3 adds GHCB_TERM_CC_BLOB for the invalid-blob-magic checks, leaving
the sites that genuinely mean "hypervisor does not support SNP",
including the CC-blob/SNP-enablement discrepancy check in sme_enable(),
on GHCB_SNP_UNSUPPORTED.
There is no functional change for a healthy guest: every modified site
already terminated the guest unconditionally, and only the reported reason
changes. Hypervisors that decode termination requests will see new codes
in the Linux reason set; that set is explicitly reserved for exactly this,
so no hypervisor can have been relying on the old values to mean anything
specific.
Kim Phillips (3):
x86/sev: Report distinct termination reasons for generic requests
x86/sev: Report distinct termination reasons in do_vc_no_ghcb()
x86/sev: Distinguish CC-blob failures from generic SNP-unsupported
arch/x86/boot/compressed/sev-handle-vc.c | 4 ++--
arch/x86/boot/compressed/sev.c | 2 +-
arch/x86/boot/startup/sev-shared.c | 14 +++++++++-----
arch/x86/boot/startup/sev-startup.c | 2 +-
arch/x86/coco/sev/core.c | 2 +-
arch/x86/coco/sev/vc-handle.c | 4 ++--
arch/x86/include/asm/sev-common.h | 4 ++++
7 files changed, 20 insertions(+), 12 deletions(-)
--
2.43.0