Re: [RFC PATCH 4/6] PCI/AER: Extend RCH RAS error handling to support VH topology case

["Li, Ming" <ming4.li@intel.com>]

On 3/15/2024 12:05 PM, Dan Williams wrote:
> Li, Ming wrote:
>> On 3/15/2024 10:30 AM, Dan Williams wrote:
>>> Li Ming wrote:
>>>> When RCEC captures CXL.cachemem protocol errors detected by CXL root
>>>> port, the recommendation from CXL r3.1 9.18.1.5 is :
>>>>
>>>> 	"Probe all CXL Downstream Ports and determine whether they have logged an
>>>> 	error in the CXL.io or CXL.cachemem status registers."
>>>>
>>>> The flow is similar with RCH RAS error handling, so reuse it to support
>>>> above case.
>>>>
>>>> Signed-off-by: Li Ming <ming4.li@intel.com>
>>>> ---
>>>>  drivers/pci/pcie/aer.c | 20 ++++++++++++--------
>>>>  1 file changed, 12 insertions(+), 8 deletions(-)
>>>>
>>>> diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
>>>> index 364c74e47273..79bfa5fb78f4 100644
>>>> --- a/drivers/pci/pcie/aer.c
>>>> +++ b/drivers/pci/pcie/aer.c
>>>> @@ -996,11 +996,15 @@ static bool is_internal_error(struct aer_err_info *info)
>>>>  	return info->status & PCI_ERR_UNC_INTN;
>>>>  }
>>>>  
>>>> -static int cxl_rch_handle_error_iter(struct pci_dev *dev, void *data)
>>>> +static int cxl_handle_error_iter(struct pci_dev *dev, void *data)
>>>>  {
>>>>  	struct aer_err_info *info = (struct aer_err_info *)data;
>>>>  	const struct pci_error_handlers *err_handler;
>>>>  
>>>> +	/* Skip the RCiEP devices not associating with RCEC */
>>>> +	if ((pci_pcie_type(dev) == PCI_EXP_TYPE_RC_END) &&
>>>> +	    !dev->rcec)
>>>> +		return 0;
>>>>  	if (!is_cxl_mem_dev(dev) || !cxl_error_is_native(dev))
>>>>  		return 0;
>>>
>>> is_cxl_mem_dev(dev) will always be false in the VH case, so how does
>>> this change help the VH case?
>>
>> Hi Dan,
>>
>> I think it won't be false if the CXL memory device is an endpoint.
>> pcie_walk_rcec_all() will walk all pci_dev in RCEC assocaited bus ranges. So these two checkings can help us to filter:
>> 1. CXL memory device is an RCiEP associated with RCEC in the RCH case
>> 2. CXL memory device is not an RCiEP, so it should be an endpoint in the VH case.
> 
> It will be an endpoint, but I though cxl_handle_error_iter() is only
> called for RCIEPs and RPs that are share a bus range with the RCEC. The
> endpoint in the VH case is downstream of the RP.
> 
> I had been assuming that pci_walk_bus() limits itself to buses within
> the Root Complex however it descends the entire bus hierarchy so this
> implementation will walk the entire topology on all root ports
> associated with the RCEC looking for any CXL device. That feels wrong.
> 
> I would expect that this limits it self to only finding root ports and
> then only proceeding if that root port has a directly attached CXL
> device.
> 
Got it, will change it in v2, thank you.

> Note, when you send a v2 of this RFC be sure to copy linux-pci for these
> core changes to PCI error handling.
Sure, I made a mistake here.