Message-ID: <87171e54-1d4f-f755-ab90-7c1373268950@redhat.com>
Date:   Wed, 22 Aug 2018 09:23:05 +0200
From:   Harald Hoyer <harald@...hat.com>
To:     Siwei Liu <loseweigh@...il.com>
Cc:     "Samudrala, Sridhar" <sridhar.samudrala@...el.com>,
        Jiri Pirko <jiri@...nulli.us>, initramfs@...r.kernel.org,
        "Michael S. Tsirkin" <mst@...hat.com>,
        Netdev <netdev@...r.kernel.org>, vijay.balakrishna@...cle.com,
        si-wei liu <si-wei.liu@...cle.com>, liran.alon@...cle.com
Subject: Re: virtio_net failover and initramfs

On 22.08.2018 09:17, Siwei Liu wrote:
> On Tue, Aug 21, 2018 at 6:44 AM, Harald Hoyer <harald@...hat.com> wrote:
>> 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@...el.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
> 
> Sorry, I forgot to mention that some of our setups do not have
> 'ip=dhcp' explicitly specified, but are still able to boot from iSCSI
> target using "netroot=iscsi:...". Move the iface_is_enslaved check out
> from the parsing of ip= lines?
> 
> -Siwei
> 

Does that mean, you want to use the enslaved interface?

Powered by blists - more mailing lists