[Hyper-V] Transparent SR-IOV solves bonding race conditions

[Joshua R. Poulson]

Description of problem:
Because of Azure provisioning of VF interfaces taking several seconds after boot, a number of race conditions were found in testing. By putting the effect of SR-IOV bonding into netvsc we avoid these race conditions.

Upstream commit:
https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git/commit/?id=f6775a28c43375945ba39581bdf8e90d0061f350

(because of merge requirements, there may be other dependencies that need to be merged, even for linux-azure's 4.11+ base for Hyper-V)

Background

In Hyper-V SR-IOV can be enabled (and disabled) by changing guest settings
on host. When SR-IOV is enabled a matching PCI device is hot plugged and
visible on guest. The VF device is an add-on to an existing netvsc
device, and has the same MAC address.

How is this different?

The original support of VF relied on using bonding driver in active
standby mode to handle the VF device.

With the new netvsc VF logic, the Linux hyper-V network
virtual driver will directly manage the link to SR-IOV VF device.
When VF device is detected (hot plug) it is automatically made a
slave device of the netvsc device. The VF device state reflects
the state of the netvsc device; i.e. if netvsc is set down, then
VF is set down. If netvsc is set up, then VF is brought up.

Packet flow is independent of VF status; all packets are sent and
received as if they were associated with the netvsc device. If VF is
removed or link is down then the synthetic VMBUS path is used.

What was wrong with using bonding script?

A lot of work went into getting the bonding script to work on all
distributions, but it was a major struggle. Linux network devices
can be configured many, many ways and there is no one solution from
userspace to make it all work. What is really hard is when
configuration is attached to synthetic device during boot (eth0) and
then the same addresses and firewall rules needs to also work later if
doing bonding. The new code gets around all of this.

How does VF work during initialization?

Since all packets are sent and received through the logical netvsc
device, initialization is much easier. Just configure the regular
netvsc Ethernet device; when/if SR-IOV is enabled it just
works. Provisioning and cloud init only need to worry about setting up
netvsc device (eth0). If SR-IOV is enabled (even as a later step), the
address and rules stay the same.

What devices show up?

Both netvsc and PCI devices are visible in the system. The netvsc
device is active and named in usual manner (eth0). The PCI device is
visible to Linux and gets renamed by udev to a persistent name
(enP2p3s0). The PCI device name is now irrelevant now.

The logic also sets the PCI VF device SLAVE flag on the network
device so network tools can see the relationship if they are smart
enough to understand how layered devices work.

This is a lot like how I see Windows working.
The VF device is visible in Device Manager, but is not configured.

Is there any performance impact?
There is no visible change in performance. The bonding
and netvsc driver both have equivalent steps.

Is it compatible with old bonding script?

It turns out that if you use the old bonding script, then everything
still works but in a sub-optimum manner. The previous bonding script package can be deprecated when this kernel patch is present.

What if I add address or firewall rule onto the VF?

Same problems occur with now as already occur with bonding, bridging,
teaming on Linux if user incorrectly does configuration onto
an underlying slave device. It will sort of work, packets will come in
and out but the Linux kernel gets confused and things like ARP don’t
work right. There is no way to block manipulation of the slave
device, and I am sure someone will find some special use case where
they want it.

How reproducible:
A discussion of potential race conditions with the bonding script can be had, but this change gets past that behavior. Please work with Microsoft on reproduction scenarios.

Steps to Reproduce:
1. Configure a VM in Azure with SR-IOV enabled.
2. Synthetic path will provision first
3. SR-IOV path will provision but will be traffic will be routed to the VF by the netvsc module and will not otherwise be visible in the guest

Actual results:
When the VF is up, lower latency and improved throughput.

Expected results:
With the bonding script method, potential races leading to no network interface being left up when the VF comes up.

[Joshua R. Poulson (jrp)]

I have created https://bugs.launchpad.net/ubuntu/+source/linux-azure/+bug/1708995 to fix a deadlock found upstream and fixed in https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git/commit/?id=fb84af8a4397ee664a51c2da1dd64fb3d582ee24

[Marcelo Cerri (mhcerri)]

Thanks. I already picked and submitted the fix.

[Marcelo Cerri (mhcerri)]

Original patchset: https://lists.ubuntu.com/archives/kernel-team/2017-August/086197.html

[Joshua R. Poulson (jrp)]

Another upstream fix... https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git/commit/?id=7b83f52047e8a3d551a9495b0267df5d0754c5bf

[Leann Ogasawara (leannogasawara)]

Per discussion with Microsoft, the backport to the stock Xenial 4.4 kernel is too risky to consider for SRU. Users will need to use the linux-azure kernel for this. The linux-azure kernel has been promoted to -updates so I will mark that task as Fix Released and set the linux task to Won't Fix. Thanks.