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
