On Tue, Mar 19, 2019 at 02:38:06PM +0200, Liran Alon wrote:
> Hi Michael,
>
> Great blog-post which summarise everything very well!
>
> Some comments I have:
Thanks!
I'll try to update everything in the post when I'm not so jet-lagged.
> 1) I think that when we are using the term “1-netdev model” on community
> discussion, we tend to refer to what you have defined in blog-post as
> "3-device model with hidden slaves”.
> Therefore, I would suggest to just remove the “1-netdev model” section and
> rename the "3-device model with hidden slaves” section to “1-netdev model”.
>
> 2) The userspace issues result both from using “2-netdev model” and “3-netdev
> model”. However, they are described in blog-post as they only exist on
> “3-netdev model”.
> The reason these issues are not seen in Azure environment is because these
> issues were partially handled by Microsoft for their specific 2-netdev model.
> Which leads me to the next comment.
>
> 3) I suggest that blog-post will also elaborate on what exactly are the
> userspace issues which results in models different than “1-netdev model”.
> The issues that I’m aware of are (Please tell me if you are aware of others!):
> (a) udev rename race-condition: When net-failover device is opened, it also
> opens it's slaves. However, the order of events to udev on KOBJ_ADD is first
> for the net-failover netdev and only then for the virtio-net netdev. This
> means that if userspace will respond to first event by open the net-failover,
> then any attempt of userspace to rename virtio-net netdev as a response to
> the second event will fail because the virtio-net netdev is already opened.
> Also note that this udev rename rule is useful because we would like to add
> rules that renames virtio-net netdev to clearly signal that it’s used as the
> standby interface of another net-failover netdev.
> The way this problem was workaround by Microsoft in NetVSC is to delay the
> open done on slave-VF from the open of the NetVSC netdev. However, this is
> still a race and thus a hacky solution. It was accepted by community only
> because it’s internal to the NetVSC driver. However, similar solution was
> rejected by community for the net-failover driver.
> The solution that we currently proposed to address this (Patch by Si-Wei) was
> to change the rename kernel handling to allow a net-failover slave to be
> renamed even if it is already opened. Patch is still not accepted.
> (b) Issues caused because of various userspace components DHCP the
> net-failover slaves: DHCP of course should only be done on the net-failover
> netdev. Attempting to DHCP on net-failover slaves as-well will cause
> networking issues. Therefore, userspace components should be taught to avoid
> doing DHCP on the net-failover slaves. The various userspace components
> include:
> b.1) dhclient: If run without parameters, it by default just enum all netdevs
> and attempt to DHCP them all.
> (I don’t think Microsoft has handled this)
> b.2) initramfs / dracut: In order to mount the root file-system from iSCSI,
> these components needs networking and therefore DHCP on all netdevs.
> (Microsoft haven’t handled (b.2) because they don’t have images which perform
> iSCSI boot in their Azure setup. Still an open issue)
> b.3) cloud-init: If configured to perform network-configuration, it attempts
> to configure all available netdevs. It should avoid however doing so on
> net-failover slaves.
> (Microsoft has handled this by adding a mechanism in cloud-init to blacklist
> a netdev from being configured in case it is owned by a specific PCI driver.
> Specifically, they blacklist Mellanox VF driver. However, this technique
> doesn’t work for the net-failover mechanism because both the net-failover
> netdev and the virtio-net netdev are owned by the virtio-net PCI driver).
> b.4) Various distros network-manager need to be updated to avoid DHCP on
> net-failover slaves? (Not sure. Asking...)
>
> 4) Another interesting use-case where the net-failover mechanism is useful is
> for handling NIC firmware failures or NIC firmware Live-Upgrade.
> In both cases, there is a need to perform a full PCIe reset of the NIC. Which
> lose all the NIC eSwitch configuration of the various VFs.
In this setup, how does VF keep going? If it doesn't keep going, why is
it helpful?
> To handle these cases gracefully, one could just hot-unplug all VFs from
> guests running on host (which will make all guests now use the virtio-net
> netdev which is backed by a netdev that eventually is on top of PF).
> Therefore, networking will be restored to guests once the PCIe reset is
> completed and the PF is functional again. To re-acceelrate the guests
> network, hypervisor can just hot-plug new VFs to guests.
>
> P.S:
> I would very appreciate all this forum help in closing on the pending items
> written in (3). Which currently prevents using this net-failover mechanism in
> real production use-cases.
>