Re: [PATCH v4 0/4] Error recovery for zPCI passthrough devices

[Farhan Ali <alifm@linux.ibm.com>]

On 9/3/2026 2:19 AM, Cédric Le Goater wrote:
> On 8/31/26 20:31, Farhan Ali wrote:
>> Hi,
>>
>> This patch series introduces support for error recovery for 
>> passthrough PCI
>> devices on System Z (s390x). This is the user space component for the 
>> Linux
>> kernel patches [1]. The kernel patches were merged for 7.3 and it had 
>> some
>> significant changes in VFIO API since the last QEMU series was 
>> posted. But the
>> design for QEMU component is still the same.
>>
>> For QEMU on eventfd notification for PCI error from vfio-pci driver 
>> we call the
>> vfio error handler. We can use a per device error handler callback to 
>> override
>> the default vfio error handler.  For s390x specific error handler, we 
>> retrieve
>> the architecture specific PCI error information and inject the 
>> information into
>> the guest. Once the guest receives the error information, the guest 
>> drivers will
>> drive the error recovery.  Typically recovery involves a device reset 
>> which
>> translate to CLP disable/enable cycle for the device.
>>
>> I would appreciate some feedback on this patch series.
>
Hi Cedric,

> Hi Farhan,
>
> Have you seen Shameer's kernel series for PCI error recovery [1] ?
> It uses a dedicated eventfd registered via VFIO_DEVICE_FEATURE_SET,
> with a self-contained handler independent of the legacy err_notifier.
>
> The QEMU side [2] just probes the feature, installs the eventfd, and
> suppresses the legacy vm_stop path with a bool flag : no callback in
> generic VFIO PCI code, which is cleaner.
>
> This is also the pattern used by the other s390 VFIO devices: vfio-ccw
> has three dedicated eventfds (io_notifier, crw_notifier, req_notifier)
> and vfio-ap has two (req_notifier, cfg_notifier).

My thought process was to use the existing legacy err_notifier and let 
each vfio-pci device handle the error. I thought the callback would also 
provide some flexibility on each device handling the error.

I can take a look at Shameer's approach and see if its something we can 
do. Just briefly looking at the QEMU code, it looks like it it ties the 
eventfd to the generic vfio-pci code (of course this is in flux). This 
would become problematic for s390x as unfortunately AER recovery will 
not work for zPCI devices due to platform firmware being involved in the 
recovery process.

> For zPCI, it would mean extending VFIO_DEVICE_FEATURE_ZPCI_ERROR to
> support SET with an eventfd field. The ABI is still new so the struct
> can be extended, keeping the s390 error recovery code self-contained,
> with no changes to generic VFIO  PCI code. I think this approach is
> preferable.
>
I would like to get some feedback from Alex to see if we could extend 
the ABI.

Thanks

Farhan


> Thanks,
>
> C.
>
> [1] 
> https://lore.kernel.org/all/20260901093217.8539-1-skolothumtho@nvidia.com/
> [2] 
> https://github.com/shamiali2008/qemu-master/commit/e45458710cd8509fa5bd81f4376126b77b6f40be
>
>