Re: virtio_net failover and initramfs (was: Re: [PATCH net-next v11 2/5] netvsc: refactor notifier/event handling code to use the failover framework)

[Harald Hoyer <harald@redhat.com>]

On 17.08.2018 11:51, Harald Hoyer wrote:
> On 16.08.2018 00:17, Siwei Liu wrote:
>> On Wed, Aug 15, 2018 at 12:05 PM, Samudrala, Sridhar
>> <sridhar.samudrala@intel.com> wrote:
>>> On 8/14/2018 5:03 PM, Siwei Liu wrote:
>>>>
>>>> Are we sure all userspace apps skip and ignore slave interfaces by
>>>> just looking at "IFLA_MASTER" attribute?
>>>>
>>>> When STANDBY is enabled on virtio-net, a failover master interface
>>>> will appear, which automatically enslaves the virtio device. But it is
>>>> found out that iSCSI (or any network boot) cannot boot strap over the
>>>> new failover interface together with a standby virtio (without any VF
>>>> or PT device in place).
>>>>
>>>> Dracut (initramfs) ends up with timeout and dropping into emergency shell:
>>>>
>>>> [  228.170425] dracut-initqueue[377]: Warning: dracut-initqueue
>>>> timeout - starting timeout scripts
>>>> [  228.171788] dracut-initqueue[377]: Warning: Could not boot.
>>>>           Starting Dracut Emergency Shell...
>>>> Generating "/run/initramfs/rdsosreport.txt"
>>>> Entering emergency mode. Exit the shell to continue.
>>>> Type "journalctl" to view system logs.
>>>> You might want to save "/run/initramfs/rdsosreport.txt" to a USB stick or
>>>> /boot
>>>> after mounting them and attach it to a bug report.
>>>> dracut:/# ip l sh
>>>> 1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN
>>>> mode DEFAULT group default qlen 1000
>>>>      link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
>>>> 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue
>>>> state UP mode DEFAULT group default qlen 1000
>>>>      link/ether 9a:46:22:ae:33:54 brd ff:ff:ff:ff:ff:ff\
>>>> 3: eth1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast
>>>> master eth0 state UP mode DEFAULT group default qlen 1000
>>>>      link/ether 9a:46:22:ae:33:54 brd ff:ff:ff:ff:ff:ff
>>>> dracut:/#
>>>>
>>>> If changing dracut code to ignore eth1 (with IFLA_MASTER attr),
>>>> network boot starts to work.
>>>
>>>
>>> Does dracut by default tries to use all the interfaces that are UP?
>>>
>> Yes. The specific dracut cmdline of our case is "ip=dhcp
>> netroot=iscsi:... ", but it's not specific to iscsi boot. And because
>> of same MAC address for failover and standby, while dracut tries to
>> run DHCP on all interfaces that are up it eventually gets same route
>> for each interface. Those conflict route entries kill off the network
>> connection.
>>
>>>
>>>>
>>>> The reason is that dracut has its own means to differentiate virtual
>>>> interfaces for network boot: it does not look at IFLA_MASTER and
>>>> ignores slave interfaces. Instead, users have to provide explicit
>>>> option e.g. bond=eth0,eth1 in the boot line, then dracut would know
>>>> the config and ignore the slave interfaces.
>>>
>>>
>>> Isn't it possible to specify the interface that should be used for network
>>> boot?
>> As I understand it, one can only specify interface name for running
>> DHCP but not select interface for network boot.  We want DHCP to run
>> on every NIC that is up (excluding the enslaved interfaces), and only
>> one of them can get a route entry to the network boot server (ie.g.
>> iSCSI target).
>>
>>>
>>>
>>>>
>>>> However, with automatic creation of failover interface that assumption
>>>> is no longer true. Can we change dracut to ignore all slave interface
>>>> by checking  IFLA_MASTER? I don't think so. It has a large impact to
>>>> existing configs.
>>>
>>>
>>> What is the issue with checking for IFLA_MASTER? I guess this is used with
>>> team/bonding setups.
>> That should be discussed within and determined by the dracut
>> community. But the current dracut code doesn't check IFLA_MASTER for
>> team or bonding specifically. I guess this change might have broader
>> impact to existing userspace that might be already relying on the
>> current behaviour.
>>
>> Thanks,
>> -Siwei
> 
> Is there a sysfs flag for IFF_SLAVE? Or any "ip" output I can use to detect, that it is a IFF_SLAVE?
> 

Oh, it's the other way around.. dracut should ignore "master" (eth1).

Can the master enslave the "eth0", if it is already "UP" and busy later on?