Xen with libvirt and SR-IOV
Jim Fehlig
jfehlig at suse.com
Wed Mar 22 20:50:54 UTC 2023
On 3/22/23 04:23, nospam at godawa.de wrote:
> Jim Fehlig schrieb:
>>
>> What is the libvirt version?
>
> It's the "latest and greatest" I get from this source:
>
> [root at xen1 ~]# virsh --version
> 6.6.0
>
> [root at xen1 ~]# libvirtd --version
> libvirtd (libvirt) 6.6.0
>
> [root at xengfs1f ~]# yum list | grep libvirt
> libvirt.x86_64 1:6.6.0-6.xen415.el7 @centos-virt-xen-415
> libvirt-bash-completion.x86_64 1:6.6.0-6.xen415.el7 @centos-virt-xen-415
> libvirt-client.x86_64 1:6.6.0-6.xen415.el7 @centos-virt-xen-415
> libvirt-daemon.x86_64 1:6.6.0-6.xen415.el7 @centos-virt-xen-415
That's an old libvirt, but there hasn't been a lot of changes to the PCI
passthrough code. One notable change that came with libvirt 6.8.0
https://gitlab.com/libvirt/libvirt/-/commit/9d15647dcb96831c93ac8c1d67c47265b5ed9072
However, that wouldn't be needed unless you were using the 'permissive' setting
in your xl config.
>> I know it's not complete, but https://libvirt.org/formatdomain.html makes an
>> attempt to identify hypervisor support for the various configuration settings.
>
> I will "recheck" it, but I don't see anything new, that might be missing.
>
>
>> I've used both configurations successfully. Has the sriov virtual function
>> been unbound from the native driver and bound to xen-pciback?
>
> Should be, otherwise it would not work with XL?
>
>
>> You can check which driver is in use with something like
>>
>> ls /sys/bus/pci/devices/0000\:81\:02.6/driver
>
> I can't see the exact definition, but again, same output when starting with xl
> or virsh:
>
> [root at xen1 ~]# ls /sys/bus/pci/devices/0000\:81\:02.6/driver
> 0000:81:02.0 0000:81:02.7 0000:81:03.6 0000:81:04.5 0000:81:05.4
> 0000:81:06.5 0000:81:07.4 0000:81:08.3 0000:81:09.2 module slots
> 0000:81:02.1 0000:81:03.0 0000:81:03.7 0000:81:04.6 0000:81:05.5
> 0000:81:06.6 0000:81:07.5 0000:81:08.4 0000:81:09.3 new_id uevent
> 0000:81:02.2 0000:81:03.1 0000:81:04.0 0000:81:04.7 0000:81:06.0
> 0000:81:06.7 0000:81:07.6 0000:81:08.5 0000:81:09.4 new_slot unbind
> 0000:81:02.3 0000:81:03.2 0000:81:04.1 0000:81:05.0 0000:81:06.1
> 0000:81:07.0 0000:81:07.7 0000:81:08.6 0000:81:09.5 permissive
> 0000:81:02.4 0000:81:03.3 0000:81:04.2 0000:81:05.1 0000:81:06.2
> 0000:81:07.1 0000:81:08.0 0000:81:08.7 bind quirks
> 0000:81:02.5 0000:81:03.4 0000:81:04.3 0000:81:05.2 0000:81:06.3
> 0000:81:07.2 0000:81:08.1 0000:81:09.0 irq_handlers remove_id
> 0000:81:02.6 0000:81:03.5 0000:81:04.4 0000:81:05.3 0000:81:06.4
> 0000:81:07.3 0000:81:08.2 0000:81:09.1 irq_handler_state remove_slot
>
> [root at xen1 ~]# cat /sys/bus/pci/devices/0000\:81\:02.6/device
> 0x154c
> [root at xen1 ~]# cat /sys/bus/pci/devices/0000\:81\:02.6/vendor
> 0x8086
> [root at xen1 ~]# cat /sys/bus/pci/devices/0000\:81\:02.6/subsystem_vendor
> 0x8086
>
> [root at xen1 ~]# ls /sys/bus/pci/devices/0000\:81\:02.6/driver/module/drivers/
> pci:pciback xen-backend:xen-pciback
>
> [root at xen1 ~]# ls /sys/bus/pci/devices/0000\:81\:02.6/driver/0000:81:02.6
> broken_parity_status consistent_dma_mask_bits dma_mask_bits enable
> local_cpus numa_node reset resource0_wc subsystem uevent
> class d3cold_allowed driver irq
> modalias physfn resource resource3 subsystem_device vendor
> config device driver_override local_cpulist
> msi_bus power resource0 resource3_wc subsystem_vendor
>
>
>> If bound to xen-pciback, it vf should also appear in the output of 'xl
>> pci-assignable-list'.
>
> Before starting the VM, it's there:
>
> [root at xengfs1f ~]# xl pci-assignable-list | grep "0000:81:02.6"
> 0000:81:02.6
Ah, xen-pciback is already bound to the device, so no need for managed='yes' in
your device config. The 'managed' attribute tells libvirt whether or not to
manage attaching/detaching drivers to/from the device. If you've done that
elsewhere, set managed='no', or simply drop the attribute since 'no' is the
default. It's explained in the 'pci' attribute of the hostdev element
https://libvirt.org/formatdomain.html#usb-pci-scsi-devices
But I wouldn't expect that minor config optimization to fix your issue. I tried
starting a VM with managed='yes' and the device already assigned to xen-pciback,
but it still worked for me. The sriov vf was visible and usable in the VM.
Are there any errors from xen-pciback in dom0, or any hints in the output of 'xl
dmesg'? Also, it might be worth comparing the relevant nodes in the output of
xenstore-ls between VM started with xl and libvirt.
Regards,
Jim
More information about the libvirt-users
mailing list