On 22.08.2018 09:17, Siwei Liu wrote: > On Tue, Aug 21, 2018 at 6:44 AM, Harald Hoyer <harald@xxxxxxxxxx> 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@xxxxxxxxx> 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?
