|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-users] all packets between certain guests on the same host being dropped
On 01/11/2017 05:26 PM, WebDawg wrote: On Wed, Jan 11, 2017 at 11:02 AM, Sherrard Burton <sburton@xxxxxxxxxxxxx <mailto:sburton@xxxxxxxxxxxxx>> wrote: On 01/10/2017 03:19 PM, Sherrard Burton wrote: TL;DR all packets are being dropped in a debian 7 (wheezy) guest only when they are coming from a debian 5 (lenny) guest on the same host. the console and kernel log report 'net eth0: Invalid extra type: 4' when packets are being dropped. the problem goes away if i change wheezy configuration from 1 vcpu to >1 vcpu. i tested all of this on fresh, minimal installs, so AFAICT there are no firewalls or other esoteric settings involved. i noticed that there is a backport kernel for lenny that includes the PV on HVM drivers. after upgrading to that kernel, the problem also goes away. i also confirmed the problem when the target is a debian 8 (jessie) guest. so it appears that the problem here is specific to HVM guests attempting to communicate with single-vcpu PV guests. if my analysis is correct, that would seem to imply that the root of the problem is in the packet-handling code in the host, no? FULL VERSION this is a strange one, so please forgive me if i omit some useful details. intro: i have a pair of xen hosts which are running pairs of guest HA pairs. for example: host1 \_apache-guest1 | \_haproxy-guest1 | \_appserver-guest1 host2 \_apache-guest2 | \_haproxy-guest2 | \_appserver-guest2 with various HA solutions implemented within the guests. this is not germane to the particular problem, but germane to how i discovered it. for the sake of balancing, i have configured the guests' HA preferences so that the active nodes tend to be on different hosts. so under normal circumstances, apache-guest1 and haproxy-guest2 would be the active nodes. no problem at all in that situation. but i discovered that i cannot communicate between apache-guest1 and haproxy-guest1, located on the same host. after much tcpdumping in the host and guests, i discovered that the problem is unidirectional and specific to a particular OS combination. a) inbound packets to a debian wheezy guest are dropped only when they originate from a debian lenny guest on the same host b) outbound packets from a wheezy guest to a lenny guest are passed correctly, even though the wheezy cannot see the return communication from the lenny guest c) there is no problem communicating to or from the wheezy guest and an identically-configured lenny guest on the other host d) there is no problem communicating to or from other combinations of guests on the same host. ie, from jessie to wheezy, lenny to lenny and wheezy to wheezy, etc. even stranger, my attempts in trying to narrow it down to the simplest possible test case led me to discover that for the same exact guest, changing the vcpu setting from 1 to >1 makes the problem go away. sburton@host:~$ virsh -c xen:/// dumpxml wheezy-guest > ~/cannot-ping.xml # test and reconfigure sburton@host:~$ virsh -c xen:/// dumpxml wheezy-guest > ~/can-ping.xml sburton@host:~$ diff ~/can-ping.xml ~/cannot-ping.xml 6c6 < <vcpu placement='static'>2</vcpu> --- <vcpu placement='static'>1</vcpu> testing methodology: simple ping between hosts. initially broken because the ARP 'is-at' traffic from the lenny guest is dropped going into the wheezy guest, and ARP 'who-has' traffic from the lenny guest is dropped going into the wheezy guest. therefore the guests cannot discover one another. after manually setting the ARP cache entries on both guests: pinging from lenny to wheezy, tcpdump shows ICMP echo requests in the lenny guest and on the VIFs for both guests in the host. but the ICMP requests are unseen in the wheezy guest. pinging from wheezy to lenny, tcpdump shows ICMP echo requests and replies in the lenny guest and on the VIFs for both guests in the host. ICMP requests are seen in the wheezy guest, since they originate there, but the replies from the lenny guest are unseen. the problem is not limited to ARP or ICMP, all other communication i have tried fails similarly. the smoking gun (i hope): when packets are being dropped in the wheezy guest, the console and various logs report [ 6977.669408] net eth0: Invalid extra type: 4 and the only reference i have found via my searching is this thread: https://lists.xenproject.org/archives/html/xen-devel/2016-09/msg00565.html <https://lists.xenproject.org/archives/html/xen-devel/2016-09/msg00565.html> which seems to be unresolved. i'm hoping that some part of this tickles someone's memory, or piques their interest, or at least that someone can point me to some more troubleshooting steps i haven't thought of. TIA setup details: HOST: sburton@host:~$ cat /etc/issue Debian GNU/Linux 8 \n \l sburton@host:~$ uname -a Linux host 4.7.0-0.bpo.1-amd64 #1 SMP Debian 4.7.8-1~bpo8+1 (2016-10-19) x86_64 GNU/Linux sburton@host:~$ dpkg -l | grep -F -e libvirt-daemon -e xen-hypervisor -e qemu-system ii libvirt-daemon 1.2.9-9+deb8u3 amd64 programs for the libvirt library ii libvirt-daemon-system 1.2.9-9+deb8u3 amd64 Libvirt daemon configuration files ii qemu-system-common 1:2.7+dfsg-3~bpo8+2 amd64 QEMU full system emulation binaries (common files) ii qemu-system-x86 1:2.7+dfsg-3~bpo8+2 amd64 QEMU full system emulation binaries (x86) ii xen-hypervisor-4.4-amd64 4.4.1-9+deb8u8 amd64 Xen Hypervisor on AMD64 sburton@host:~$ grep -F -A1 '<os>' ~/cannot-ping.xml <os> <type arch='x86_64' machine='xenfv'>hvm</type> sburton@host:~$ grep -F -C2 'xenbr0' ~/cannot-ping.xml <interface type='bridge'> <mac address='00:16:3e:fb:2e:1c'/> <source bridge='xenbr0'/> <model type='rtl8139'/> </interface> sburton@host:~$ ip addr show xenbr0 8: xenbr0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default qlen 1000 link/ether bc:30:5b:f0:32:b4 brd ff:ff:ff:ff:ff:ff inet 192.168.240.52/20 <http://192.168.240.52/20> brd 192.168.255.255 scope global xenbr0 valid_lft forever preferred_lft forever inet6 fe80::be30:5bff:fef0:32b4/64 scope link valid_lft forever preferred_lft forever GUESTS: fullvirt installs, created from netinst ISO via virt-manager running on my workstation, manipulated through some combination of virt-manager and local virsh commands. root@wheezy-guest:~# uname -a Linux wheezy-guest 3.16.0-0.bpo.4-amd64 #1 SMP Debian 3.16.36-1+deb8u2~bpo70+1 (2016-10-19) x86_64 GNU/Linux root@wheezy-guest:~# cat /etc/issue Debian GNU/Linux 7 \n \l root@lenny-guest:~# uname -a Linux lenny-guest 2.6.26-2-amd64 #1 SMP Sun Mar 4 21:48:06 UTC 2012 x86_64 GNU/Linux root@lenny-guest:~# cat /etc/issue Debian GNU/Linux 5.0 \n \l By target you mean debian jessie as a guest right? Looking at your logs, are you running etch as the dom0? yes, the last test case was debian 8 (jessie) as a guest.the host is running jessie, and the connections fail from HVM guests to single-vcpu PV guests on the same host. _______________________________________________ Xen-users mailing list Xen-users@xxxxxxxxxxxxx https://lists.xen.org/xen-users _______________________________________________ Xen-users mailing list Xen-users@xxxxxxxxxxxxx https://lists.xen.org/xen-users
|
 |
Lists.xenproject.org is hosted with RackSpace, monitoring our |