Re: [Qemu-devel] vfio-pci issues with multiple devices on the same root port

[Peter Lieven <pl@kamp.de>]

Am 12.12.2014 um 23:21 schrieb Alex Williamson:
> On Fri, 2014-12-12 at 22:38 +0100, Peter Lieven wrote:
>> Hi,
>>
>> we have a Cisco UCS infrastructure where we have fnic Fibre-Channel Adapters that we expose to guests. The UCS
>> infrastruture allows to create virtual HBAs that can be exposed to a host so its possible to have quite a lot of them.
>>
>> We ran into a strange issue when we started having more than one vServer with a FibreChannel Adapter passed
>> thru with vfio-pci.
>>
>> When a hypervisor shuts down it the kernel sees the following error:
>>
>>  pcieport 0000:00:07.0: AER: Uncorrected (Non-Fatal) error received: id=0038
>>  pcieport 0000:00:07.0: PCIe Bus Error: severity=Uncorrected (Non-Fatal), type=Transaction Layer, id=0038(Receiver ID)
>>  pcieport 0000:00:07.0:   device [8086:340e] error status/mask=00200000/00100000
>>  pcieport 0000:00:07.0:    [21] Unknown Error Bit (First)
>>  pcieport 0000:00:07.0: broadcast error_detected message
>>  pcieport 0000:00:07.0: AER: Device recovery failed
>>
>> Bit 21 seems to be ACS Violation. And 0000:00:07.0 is the PCIE Root Port on that System.
>>
>> This wouldn't be a big problem, altough I would like to find out what the ACS Violation causes.
>>
>> The real problem is that all other vfio-pci cards on that root port get notified of this error and the connected vServers are suspended
>> with RUN_STATE_INTERNAL_ERROR.
>>
>> Any ideas to work around this other than hacking qemu to not register an error handler or modifying vfio_err_notifier_handler
>> to not suspend the vServer?
> You could set bit 21 in the AER uncorrected error mask register to avoid
> the root port signaling the error.  Is bit 21 already clear in the
> severity register to make this non-fatal?
>
>> Is it correct that all children of a root port are notified? Should qemu distinguish between fatal and non-fatal errors when
>> suspending a vServer?
> Yes, each child is notified.  QEMU only gets an eventfd signal, which is
> supposed to occur only for fatal errors.  I don't quite understand why
> this apparently non-fatal error is getting through.  The kernel-side
> VFIO code is where filtering of fatal vs non-fatal should occur.

Had a look at vfio-pci.c from master. I can't see where there is a filtering of fatal vs. non-fatal

static pci_ers_result_t vfio_pci_aer_err_detected(struct pci_dev *pdev,
						  pci_channel_state_t state)
{
	struct vfio_pci_device *vdev;
	struct vfio_device *device;

	device = vfio_device_get_from_dev(&pdev->dev);
	if (device == NULL)
		return PCI_ERS_RESULT_DISCONNECT;

	vdev = vfio_device_data(device);
	if (vdev == NULL) {
		vfio_device_put(device);
		return PCI_ERS_RESULT_DISCONNECT;
	}

	mutex_lock(&vdev->igate);

	if (vdev->err_trigger)
		eventfd_signal(vdev->err_trigger, 1);

	mutex_unlock(&vdev->igate);

	vfio_device_put(device);

	return PCI_ERS_RESULT_CAN_RECOVER;
}

static struct pci_error_handlers vfio_err_handlers = {
	.error_detected = vfio_pci_aer_err_detected,
};

static struct pci_driver vfio_pci_driver = {
	.name		= "vfio-pci",
	.id_table	= NULL, /* only dynamic ids */
	.probe		= vfio_pci_probe,
	.remove		= vfio_pci_remove,
	.err_handler	= &vfio_err_handlers,
};

Peter