Date:   Thu, 21 Mar 2019 12:07:57 +0200
From:   Liran Alon <liran.alon@...cle.com>
To:     "Michael S. Tsirkin" <mst@...hat.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 21 Mar 2019, at 10:58, Michael S. Tsirkin <mst@...hat.com> wrote:
> 
> On Thu, Mar 21, 2019 at 12:19:22AM +0200, Liran Alon wrote:
>> 
>> 
>>> On 21 Mar 2019, at 0:10, Michael S. Tsirkin <mst@...hat.com> wrote:
>>> 
>>> 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?
>> 
>> That is indeed another possibility that would work to avoid the DHCP issues.
>> And will still allow checking connectivity. So it is better.
>> However, I still think it provides an non-intuitive customer experience.
>> In addition, I also want to take into account that most customers are expected a 1:1 mapping between a vNIC and a netdev.
>> i.e. A cloud instance should show 1-netdev if it has one vNIC attached to it defined.
>> Customers usually don’t care how they get accelerated networking. They just care they do.
>> 
>>> 
>>>> 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?
>> 
>> That would work but I think it is cleaner to just not bind PV and VF based on having the same MAC.
> 
> There's a reference to that under "Non-MAC based pairing".
> 
> I'll look into making it more explicit.

Yes I know. I was referring to what you described in that section.

> 
>>> 
>>>> 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?
>> 
>> This is kinda controversial, but maybe separate netns names into 2 groups: hidden and normal.
>> To reference a hidden netns, you need to do it explicitly. 
>> Hidden and normal netns names can collide as they will be maintained in different namespaces (Yes I’m overloading the term namespace here…).
> 
> Maybe it's an unnamed namespace. Hidden until userspace gives it a name?

This is also a good idea that will solve the issue. Yes.

> 
>> Does this seems reasonable?
>> 
>> -Liran
> 
> Reasonable I'd say yes, easy to implement probably no. But maybe I
> missed a trick or two.

BTW, from a practical point of view, I think that even until we figure out a solution on how to implement this,
it was better to create an kernel auto-generated name (e.g. “kernel_net_failover_slaves")
that will break only userspace workloads that by a very rare-chance have a netns that collides with this then
the breakage we have today for the various userspace components.

-Liran

> 
>>> 
>>>>> 
>>>>>> 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