Re: [PATCH] PCI/ERR: Fix run error recovery callbacks for all affected devices

[Sinan Kaya <okaya@kernel.org>]

On 1/25/2019 12:37 PM, Sinan Kaya wrote:
> On 1/25/2019 12:17 PM, Keith Busch wrote:
>> On Fri, Jan 25, 2019 at 06:28:03AM -0800, Dongdong Liu wrote:
>>> I want to fix 2 points by the patch.
>>>
>>> 1. For EP devices (such as multi-function EP device) under the same bus,
>>> when one of the EP devices met non-fatal error, should report non-fatal
>>> error only to the error endpoint device, no need to broadcast all of them.
>>> That is the patch (PCI/AER: Report non-fatal errors only to the affected 
>>> endpoint  #4.15)
>>> have done, but current code PATCH [1] broken this.
>>
>> How do you know a non-fatal affects only the reporting end point? These can
>> certainly be bus errors, and it's not the first to detect may be affected.
>>
>> In any case, what harm does the broadcast cause?
>>
> 
> What is the PCIe spec rule about AER errors for multi-function devices?
> 
> Does it say it needs to be propagated to all functions or each function has
> its own unique AER error handler?
> 

Thinking more...

I think there is value in probing all devices for errors like today because
multiple errors bit can be set. Since root port's AER register only captures the
first error, the rest of the errors requires OS to poll each device to see what
is going on.

In this case the AER error status of other functions should not report any
outstanding event. Please verify this. Otherwise, you are looking at a device quirk.