Re: [PATCH v2 1/3] cxl/pci: Fix appropriate checking for _OSC while handling CXL RAS registers

[Smita Koralahalli]

On 8/7/2023 8:17 PM, Dan Williams wrote:
> Smita Koralahalli wrote:
> > According to Section 9.17.2, Table 9-26 of CXL Specification [1], owner
> > of AER should also own CXL Protocol Error Management as there is no
> > explicit control of CXL Protocol error. And the CXL RAS Cap registers
> > reported on Protocol errors should check for AER _OSC rather than CXL
> > Memory Error Reporting Control _OSC.
> >
> > The CXL Memory Error Reporting Control _OSC specifically highlights
> > handling Memory Error Logging and Signaling Enhancements. These kinds of
> > errors are reported through a device's mailbox and can be managed
> > independently from CXL Protocol Errors.
> >
> > This change fixes handling and reporting CXL Protocol Errors and RAS
> > registers natively with native AER and FW-First CXL Memory Error Reporting
> > Control.
>
> I feel like this could be said more succinctly and with an indication of
> what the end user should expect to see. Something like:
>
> "cxl_pci fails to unmask CXL protocol errors when CXL memory error
> reporting is not granted native control. Given that CXL memory error
> reporting uses the event interface and protocol errors use AER, unmask
> protocol errors based only on the native AER setting. Without this
> change end user deployments will fail to report protocol errors in the
> case where native memory error handling is not granted to Linux."

Sure, will make the change for a more clearer description. Thanks!
> > [1] Compute Express Link (CXL) Specification, Revision 3.1, Aug 1 2022.
> >
> > Fixes: 248529edc86f ("cxl: add RAS status unmasking for CXL")
> > Signed-off-by: Smita Koralahalli <Smita.KoralahalliChannabasappa@xxxxxxx>
> > ---
> > v2:
> > 	Added fixes tag.
> > 	Included what the patch fixes in commit message.
> > ---
> >   drivers/cxl/pci.c | 6 +++---
> >   1 file changed, 3 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/cxl/pci.c b/drivers/cxl/pci.c
> > index 1cb1494c28fe..2323169b6e5f 100644
> > --- a/drivers/cxl/pci.c
> > +++ b/drivers/cxl/pci.c
> > @@ -541,9 +541,9 @@ static int cxl_pci_ras_unmask(struct pci_dev *pdev)
> >   		return 0;
> >   	}
> >   
> > -	/* BIOS has CXL error control */
> > -	if (!host_bridge->native_cxl_error)
> > -		return -ENXIO;
> > +	/* BIOS has PCIe AER error control */
> > +	if (!host_bridge->native_aer)
> > +		return 0;
>
> The error code does not matter here and changing it makes the patch that
> bit much more noisier than it needs to be. So just leave it as:

Doing this will return an error from cxl_pci probe thereby failing the 
device node creation in FW-First AER/DPC. I cannot think of other places 
where we reference the device node in FW-First mode but I have a place 
where this could potentially be a roadblock.

I'm trying to add trace events support for FW-First Protocol Errors. 
https://lore.kernel.org/linux-cxl/D9381C12-A585-4089-873B-3707C17823D3@xxxxxx/T/#mcaf8a78c1295372ab811be7e1ccb6a8a4d99f3e9

And we already have an existing trace_cxl_aer_correctable_error() and 
similarly for uncorrectable error for native protocol error reporting. I 
was trying to reuse the same function for fw-first as well. This 
function references cxl memory device node which will be NULL in 
FW-First if this returns an error.

I don't mind having a separate trace event function for FW-First mode as 
it would simplify things especially when dealing with RCH DP.. But there 
may be other potential places where we might reference this device node 
in FW-First. Please advice.

Thanks,
Smita

> 	return -ENXIO;
>
> >   
> >   	rc = pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &cap);
> >   	if (rc)
> > --
> > 2.17.1