Date:   Wed, 20 Mar 2019 18:10:37 -0400
From:   "Michael S. Tsirkin" <mst@...hat.com>
To:     Liran Alon <liran.alon@...cle.com>
Cc:     Stephen Hemminger <stephen@...workplumber.org>,
        Si-Wei Liu <si-wei.liu@...cle.com>,
        Sridhar Samudrala <sridhar.samudrala@...el.com>,
        Alexander Duyck <alexander.duyck@...il.com>,
        Jakub Kicinski <kubakici@...pl>, Jiri Pirko <jiri@...nulli.us>,
        David Miller <davem@...emloft.net>,
        Netdev <netdev@...r.kernel.org>,
        virtualization@...ts.linux-foundation.org,
        boris.ostrovsky@...cle.com, vijay.balakrishna@...cle.com,
        jfreimann@...hat.com, ogerlitz@...lanox.com, vuhuong@...lanox.com
Subject: Re: [summary] virtio network device failover writeup

On Wed, Mar 20, 2019 at 11:43:41PM +0200, Liran Alon wrote:
> 
> 
> > On 20 Mar 2019, at 16:09, Michael S. Tsirkin <mst@...hat.com> wrote:
> > 
> > On Wed, Mar 20, 2019 at 02:23:36PM +0200, Liran Alon wrote:
> >> 
> >> 
> >>> On 20 Mar 2019, at 12:25, Michael S. Tsirkin <mst@...hat.com> wrote:
> >>> 
> >>> On Wed, Mar 20, 2019 at 01:25:58AM +0200, Liran Alon wrote:
> >>>> 
> >>>> 
> >>>>> On 19 Mar 2019, at 23:19, Michael S. Tsirkin <mst@...hat.com> wrote:
> >>>>> 
> >>>>> On Tue, Mar 19, 2019 at 08:46:47AM -0700, Stephen Hemminger wrote:
> >>>>>> On Tue, 19 Mar 2019 14:38:06 +0200
> >>>>>> Liran Alon <liran.alon@...cle.com> wrote:
> >>>>>> 
> >>>>>>> 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).
> >>>>>> 
> >>>>>> Cloud-init should really just ignore all devices that have a master device.
> >>>>>> That would have been more general, and safer for other use cases.
> >>>>> 
> >>>>> Given lots of userspace doesn't do this, I wonder whether it would be
> >>>>> safer to just somehow pretend to userspace that the slave links are
> >>>>> down? And add a special attribute for the actual link state.
> >>>> 
> >>>> I think this may be problematic as it would also break legit use case
> >>>> of userspace attempt to set various config on VF slave.
> >>>> In general, lying to userspace usually leads to problems.
> >>> 
> >>> I hear you on this. So how about instead of lying,
> >>> we basically just fail some accesses to slaves
> >>> unless a flag is set e.g. in ethtool.
> >>> 
> >>> Some userspace will need to change to set it but in a minor way.
> >>> Arguably/hopefully failure to set config would generally be a safer
> >>> failure.
> >> 
> >> Once userspace will set this new flag by ethtool, all operations done by other userspace components will still work.
> > 
> > Sorry about being unclear, the idea would be to require the flag on each ethtool operation.
> 
> Oh. I have indeed misunderstood your previous email then. :)
> Thanks for clarifying.
> 
> > 
> >> E.g. Running dhclient without parameters, after this flag was set, will still attempt to perform DHCP on it and will now succeed.
> > 
> > I think sending/receiving should probably just fail unconditionally.
> 
> You mean that you wish that somehow kernel will prevent Tx on net-failover slave netdev
> unless skb is marked with some flag to indicate it has been sent via the net-failover master?

We can maybe avoid binding a protocol socket to the device?

> This indeed resolves the group of userspace issues around performing DHCP on net-failover slaves directly (By dracut/initramfs, dhclient and etc.).
> 
> However, I see a couple of down-sides to it:
> 1) It doesn’t resolve all userspace issues listed in this email thread. For example, cloud-init will still attempt to perform network config on net-failover slaves.
> It also doesn’t help with regard to Ubuntu’s netplan issue that creates udev rules that match only by MAC.


How about we fail to retrieve mac from the slave?

> 2) It brings non-intuitive customer experience. For example, a customer may attempt to analyse connectivity issue by checking the connectivity
> on a net-failover slave (e.g. the VF) but will see no connectivity when in-fact checking the connectivity on the net-failover master netdev shows correct connectivity.
> 
> The set of changes I vision to fix our issues are:
> 1) Hide net-failover slaves in a different netns created and managed by the kernel. But that user can enter to it and manage the netdevs there if wishes to do so explicitly.
> (E.g. Configure the net-failover VF slave in some special way).
> 2) Match the virtio-net and the VF based on a PV attribute instead of MAC. (Similar to as done in NetVSC). E.g. Provide a virtio-net interface to get PCI slot where the matching VF will be hot-plugged by hypervisor.
> 3) Have an explicit virtio-net control message to command hypervisor to switch data-path from virtio-net to VF and vice-versa. Instead of relying on intercepting the PCI master enable-bit
> as an indicator on when VF is about to be set up. (Similar to as done in NetVSC).
> 
> Is there any clear issue we see regarding the above suggestion?
> 
> -Liran

The issue would be this: how do we avoid conflicting with namespaces
created by users?

> > 
> >> Therefore, this proposal just effectively delays when the net-failover slave can be operated on by userspace.
> >> But what we actually want is to never allow a net-failover slave to be operated by userspace unless it is explicitly stated
> >> by userspace that it wishes to perform a set of actions on the net-failover slave.
> >> 
> >> Something that was achieved if, for example, the net-failover slaves were in a different netns than default netns.
> >> This also aligns with expected customer experience that most customers just want to see a 1:1 mapping between a vNIC and a visible netdev.
> >> But of course maybe there are other ideas that can achieve similar behaviour.
> >> 
> >> -Liran
> >> 
> >>> 
> >>> Which things to fail? Probably sending/receiving packets?  Getting MAC?
> >>> More?
> >>> 
> >>>> If we reach
> >>>> to a scenario where we try to avoid userspace issues generically and
> >>>> not on a userspace component basis, I believe the right path should be
> >>>> to hide the net-failover slaves such that explicit action is required
> >>>> to actually manipulate them (As described in blog-post). E.g.
> >>>> Automatically move net-failover slaves by kernel to a different netns.
> >>>> 
> >>>> -Liran
> >>>> 
> >>>>> 
> >>>>> -- 
> >>>>> MST

Powered by blists - more mailing lists