[RFC PATCH v1 2/2] PCI/AER: report fatal errors of RCiEP and EP if link recoverd

[Shuai Xue]

在 2024/11/7 00:02, Bjorn Helgaas 写道:
> On Wed, Nov 06, 2024 at 05:03:39PM +0800, Shuai Xue wrote:
>> The AER driver has historically avoided reading the configuration space of an
>> endpoint or RCiEP that reported a fatal error, considering the link to that
>> device unreliable. Consequently, when a fatal error occurs, the AER and DPC
>> drivers do not report specific error types, resulting in logs like:
>>
>> [  245.281980] pcieport 0000:30:03.0: EDR: EDR event received
>> [  245.287466] pcieport 0000:30:03.0: DPC: containment event, status:0x0005 source:0x3400
>> [  245.295372] pcieport 0000:30:03.0: DPC: ERR_FATAL detected
>> [  245.300849] pcieport 0000:30:03.0: AER: broadcast error_detected message
>> [  245.307540] nvme nvme0: frozen state error detected, reset controller
>> [  245.722582] nvme 0000:34:00.0: ready 0ms after DPC
>> [  245.727365] pcieport 0000:30:03.0: AER: broadcast slot_reset message
>>
>> But, if the link recovered after hot reset, we can safely access AER status of
>> the error device. In such case, report fatal error which helps to figure out the
>> error root case.
> 
> Explain why we can access these registers after reset.  I think it's
> important that these registers are sticky ("RW1CS" per spec).

Yes, AER error status registers are Sticky and Write-1-to-clear. If we 
does not read them after reset_subordinates, the registers will be 
cleared in pci_error_handlers, e.g. nvme_err_handler

   slot_reset() => nvme_slot_reset()
     pci_restore_state()
       pci_aer_clear_status()

Will add the reason in commit log.

> 
>> After this patch, the logs like:
>>
>> [  414.356755] pcieport 0000:30:03.0: EDR: EDR event received
>> [  414.362240] pcieport 0000:30:03.0: DPC: containment event, status:0x0005 source:0x3400
>> [  414.370148] pcieport 0000:30:03.0: DPC: ERR_FATAL detected
>> [  414.375642] pcieport 0000:30:03.0: AER: broadcast error_detected message
>> [  414.382335] nvme nvme0: frozen state error detected, reset controller
>> [  414.645413] pcieport 0000:30:03.0: waiting 100 ms for downstream link, after activation
>> [  414.788016] nvme 0000:34:00.0: ready 0ms after DPC
>> [  414.796975] nvme 0000:34:00.0: PCIe Bus Error: severity=Uncorrectable (Fatal), type=Data Link Layer, (Receiver ID)
>> [  414.807312] nvme 0000:34:00.0:   device [144d:a804] error status/mask=00000010/00504000
>> [  414.815305] nvme 0000:34:00.0:    [ 4] DLP                    (First)
>> [  414.821768] pcieport 0000:30:03.0: AER: broadcast slot_reset message
> 
> Capitalize subject lines to match history (use "git log --oneline
> drivers/pci/pcie/aer.c" to see it).
> 
> Remove timestamps since they don't help understand the problem.
> 
> Indent the quoted material two spaces.
> 
> Wrap commit log to fit in 75 columns (except the quoted material;
> don't insert line breaks there).

Will do.

> 
>> Signed-off-by: Shuai Xue <xueshuai at linux.alibaba.com>
>> ---
>>   drivers/pci/pci.h      |  1 +
>>   drivers/pci/pcie/aer.c | 50 ++++++++++++++++++++++++++++++++++++++++++
>>   drivers/pci/pcie/err.c |  6 +++++
>>   3 files changed, 57 insertions(+)
>>
>> diff --git a/drivers/pci/pci.h b/drivers/pci/pci.h
>> index 0866f79aec54..143f960a813d 100644
>> --- a/drivers/pci/pci.h
>> +++ b/drivers/pci/pci.h
>> @@ -505,6 +505,7 @@ struct aer_err_info {
>>   };
>>   
>>   int aer_get_device_error_info(struct pci_dev *dev, struct aer_err_info *info);
>> +int aer_get_device_fatal_error_info(struct pci_dev *dev, struct aer_err_info *info);
>>   void aer_print_error(struct pci_dev *dev, struct aer_err_info *info);
>>   #endif	/* CONFIG_PCIEAER */
>>   
>> diff --git a/drivers/pci/pcie/aer.c b/drivers/pci/pcie/aer.c
>> index 13b8586924ea..0c1e382ce117 100644
>> --- a/drivers/pci/pcie/aer.c
>> +++ b/drivers/pci/pcie/aer.c
>> @@ -1252,6 +1252,56 @@ int aer_get_device_error_info(struct pci_dev *dev, struct aer_err_info *info)
>>   	return 1;
>>   }
>>   
>> +/**
>> + * aer_get_device_fatal_error_info - read fatal error status from EP or RCiEP
>> + * and store it to info
>> + * @dev: pointer to the device expected to have a error record
>> + * @info: pointer to structure to store the error record
>> + *
>> + * Return 1 on success, 0 on error.
> 
> Backwards from the usual return value convention.

Yes. As @Keith pointed, aer_get_device_fatal_error_info() is copied from 
  aer_get_device_error_info(), I will try to add a helper to reduce 
duplication.

> 
>> + * Note that @info is reused among all error devices. Clear fields properly.
>> + */
>> +int aer_get_device_fatal_error_info(struct pci_dev *dev, struct aer_err_info *info)
>> +{
>> +	int type = pci_pcie_type(dev);
>> +	int aer = dev->aer_cap;
>> +	u32 aercc;
>> +
>> +	pci_info(dev, "type :%d\n", type);
> 
> I don't see this line in the sample output in the commit log.  Is this
> debug that you intended to remove?


Sorry, I missed this line, will remove it.

> 
>> +	/* Must reset in this function */
>> +	info->status = 0;
>> +	info->tlp_header_valid = 0;
>> +	info->severity = AER_FATAL;
>> +
>> +	/* The device might not support AER */
> 
> Unnecessary comment.

Will remove it.

Thank you for valuable comments.

Best Regards,
Shuai