Re: virtio_net failover and initramfs

[Harald Hoyer <harald@redhat.com>]

On 17.08.2018 21:09, Samudrala, Sridhar wrote:
> On 8/17/2018 2:56 AM, Harald Hoyer wrote:
>> 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).
> In the above example eth0 is the net_failover device and eth1 is the lower virtio_net device.
> "ip" output of eth1 shows "master eth0". It indicates that eth0 is its upper/master device.
> This information can also be obtained via sysfs too. /sys/class/net/eth1/upper_eth0
>>
>> Can the master enslave the "eth0", if it is already "UP" and busy later on?
> eth0 is the master/failover device and eth1 gets registered as its slave via NETDEV_REGISTER event.
> dracut should ignore eth1 in this setup.


Care to test, if that fixes your case?
https://github.com/dracutdevs/dracut/pull/450/files