Re: [PATCH v3 0/3] ACPI: APEI: GHES: Add NVIDIA Vera CPER decoding

From: Kai-Heng Feng

Date: Thu Oct 01 2026 - 02:14:15 EST


On Tue Aug 4, 2026 at 8:23 PM CST, Kai-Heng Feng wrote:
> The NVIDIA GHES driver currently decodes the vendor CPER section used by
> Grace systems. Vera uses a different section GUID and event layout.
>
> Refactor the Grace path into testable decode and print helpers, add the Vera
> event decoder, and cover both formats with KUnit. The Vera parser follows
> the extensible CPU information and event-context wire contracts: it accepts
> new minor versions and extended information headers, validates every one of
> the wire format's 0 to 255 contexts without allocation, and retains a valid
> prefix when a later context is malformed.
>
> Printing walks the validated raw context span a second time. It limits
> verbose output to 16 contexts and 32 key/value entries per section while
> still validating the complete record. Existing one-context output remains
> unchanged.
>
> The companion ras-tools series adds a fail-closed NVIDIA EINJ validator for
> the non-fatal Grace CMET-INFO and Vera CMET-NULL samples.
>
> Changes since v2:
> - Move EINJ tooling out of the kernel series and into ras-tools.
> - Follow the Vera CtxSize padding and alignment contract while retaining
> the legacy opaque-context encoding.
> - Accept compatible CPU information extensions and validate the complete
> 8-bit context count with an allocation-free cursor.
> - Bound verbose context and entry output independently, while reporting
> omitted contexts and entries.
> - Add wire-layout assertions, partial-context handling, severity-aware
> diagnostics, and 52 Grace/Vera KUnit cases.
>
> Verification:
> - x86_64 KUnit: 52/52 passed.
> - ARM64 KASAN+UBSAN KUnit under QEMU/HVF: 52/52 passed.
> - ARM64 GCC and Clang, built-in and module configurations, W=1: passed.
> - ARM64 GCC full Image link: passed.
> - ARM64 sparse C=2 with __CHECK_ENDIAN__: passed (one unrelated existing
> arm64 vDSO declaration warning).
> - Strict source checkpatch and git diff --check: passed.
> - The raw live sections were decoded independently by libcper at commit
> b83878a4aca9: Vera as an 80-byte CPU CMET-NULL event with one empty
> type-1 context, and Grace as a 544-byte CMET-INFO section with 32
> registers.
> - Vera live EINJ: a real 80-byte CMET-NULL CPER from GHES source 8193,
> event type/subtype 0, one 16-byte context, and zero entries (not the
> firmware-rejection record seen with the old selector).
> - Grace live EINJ: a real 544-byte CMET-INFO CPER from GHES source 896,
> severity 3, with all 32 register addresses decoded in order.
> - Both live runs reached the runner's PASS gate with empty tool stderr,
> captured the raw tracepoint record, and restored every EINJ control
> byte-for-byte.
>
> Kai-Heng Feng (3):
> ACPI: APEI: GHES: Refactor Grace decoder helpers
> ACPI: APEI: GHES: Add NVIDIA Vera decoder
> ACPI: APEI: GHES: Add Grace and Vera KUnit coverage

A gentle ping on this series...

Thanks,
Kai-Heng

>
> MAINTAINERS | 2 +-
> drivers/acpi/apei/Kconfig | 24 +-
> drivers/acpi/apei/Makefile | 1 +
> drivers/acpi/apei/ghes-nvidia-test-fixtures.h | 233 +++
> drivers/acpi/apei/ghes-nvidia-test.c | 1274 +++++++++++++++++
> drivers/acpi/apei/ghes-nvidia.c | 767 +++++++++-
> drivers/acpi/apei/ghes-nvidia.h | 109 ++
> 7 files changed, 2364 insertions(+), 46 deletions(-)
> create mode 100644 drivers/acpi/apei/ghes-nvidia-test-fixtures.h
> create mode 100644 drivers/acpi/apei/ghes-nvidia-test.c
> create mode 100644 drivers/acpi/apei/ghes-nvidia.h
>
>
> base-commit: 848acc8ffe1b7cd5f1bf427b93069becfebc2c9d