Re: [PATCH v2 3/3] acpi/apei: Add NVIDIA GHES vendor CPER record handler

[Rafael J. Wysocki]

On Wed, Mar 25, 2026 at 6:08 PM Jonathan Cameron
<jonathan.cameron@xxxxxxxxxx> wrote:
>
> On Wed, 25 Mar 2026 10:36:28 -0500
> Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
>
> > On Wed, Mar 25, 2026 at 07:34:50PM +0800, Kai-Heng Feng wrote:
> > > On Wed Mar 25, 2026 at 12:15 AM CST, Bjorn Helgaas wrote:
> > > > On Tue, Mar 24, 2026 at 05:33:06PM +0800, Kai-Heng Feng wrote:
> > > >> On 2026-03-20 09:52, Bjorn Helgaas wrote:
> > > >> > On Thu, Mar 19, 2026 at 07:13:09PM +0800, Kai-Heng Feng wrote:
> > > >> > > Add support for decoding NVIDIA-specific CPER sections delivered via
> > > >> > > the APEI GHES vendor record notifier chain. NVIDIA hardware generates
> > > >> > > vendor-specific CPER sections containing error signatures and diagnostic
> > > >> > > register dumps. This implementation registers a notifier_block with the
> > > >> > > GHES vendor record notifier and decodes these sections, printing error
> > > >> > > details via dev_info().
> > > >> > >
> > > >> > > The driver binds to ACPI device NVDA2012, present on NVIDIA server
> > > >> > > platforms. The NVIDIA CPER section contains a fixed header with error
> > > >> > > metadata (signature, error type, severity, socket) followed by
> > > >> > > variable-length register address-value pairs for hardware diagnostics.
> > > >> > >
> > > >> > > This work is based on libcper [0].
> > > >> > >
> > > >> > > Example output:
> > > >> > > nvidia-ghes NVDA2012:00: NVIDIA CPER section, error_data_length: 544
> > > >> > > nvidia-ghes NVDA2012:00: signature: CMET-INFO
> > > >> > > nvidia-ghes NVDA2012:00: error_type: 0
> > > >> > > nvidia-ghes NVDA2012:00: error_instance: 0
> > > >> > > nvidia-ghes NVDA2012:00: severity: 3
> > > >> > > nvidia-ghes NVDA2012:00: socket: 0
> > > >> > > nvidia-ghes NVDA2012:00: number_regs: 32
> > > >> > > nvidia-ghes NVDA2012:00: instance_base: 0x0000000000000000
> > > >> > > nvidia-ghes NVDA2012:00: register[0]: address=0x8000000100000000 value=0x0000000100000000
> > > >> >
> > > >> > Is there a convenient way to connect NVDA2012:00 with the actual
> > > >> > device?  I assume this is typically a PCIe device?  How would we
> > > >> > relate this with PCIe errors?
> > > >>
> > > >> The CPER report is from ARM RAS firmware and not neccessarily be
> > > >> related to a PCIe device.
> > > >
> > > > Right, I know CPER is more general than just PCI/PCIe.
> > > >
> > > > But in this case, I think NVDA2012 probably *is* a PCIe device.  How
> > > > would we figure out which one?  If we have to manually do an acpidump,
> > > > figure out which NVDA2012 is :00, and look for an _ADR or something,
> > > > that doesn't really seem convenient for multi-NVDA2012 situations.
> > >
> > > It's actually just an ACPI device:
> > > Device (CPER)
> > > {
> > >   Name (_HID, "NVDA2012")  // _HID: Hardware ID
> > >   Name (_UID, 0x00)  // _UID: Unique ID
> > >   Method (_DSM, 4, Serialized) // _DSM: Device-Specific Method
> > > }
> > >
> > > And that's it.
> >
> > Weird.  There's nothing for a driver to operate the device with except
> > _DSM?  The device doesn't need any MMIO resources?  I would expect some
> > resources described by a _CRS method or some native enumeration protocol
> > like PCI BARs.
> >
> > The _UID 0x00 matches the "00" in "NVDA2012:00", but I think that's a
> > coincidence; I think the "00" in the device name came from the ida_alloc()
> > in acpi_device_set_name(), not from _UID.
> >
> > So I still don't know how you would identify the correct part in a system
> > with multiple NVDA2012 devices.  I do see the "socket" and "instance_base"
> > in the output.  Maybe that would help, but those seem to be
> > device-specific, and it seems like we should have a generic mechanism.
>
> It's not unique in ACPI terms.  There are a few cases even in the ACPI spec
> of IDs that exist just to say some feature is there.
>
> ACPI0017 is an example. Simply says, there be CXL here, go look for the
> tables.
>
> Here this device is used to indicate that a platform should be ready to handle
> a particular type of error record.  If it happened to expose any other
> interfaces, then I agree it would need resources or a _DSM etc.
>
> Basically it's a workaround for the lack of discoverability in APEI /
> ACPI error reporting. Could use an _OSC bit for the same job but then
> we'd run out of those fast.  Device IDs are near free.

Well, in principle, an auxiliary device could be registered when a
given ACPI table was present.  It would then trigger a driver load and
the driver would probe against the table in question.