Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
-Original Message- From: Gerd v. Egidy [mailto:li...@egidy.de] Sent: Thursday, May 27, 2010 3:13 PM To: Rose, Gregory V Cc: e1000-devel@lists.sourceforge.net Subject: Re: [E1000-devel] guest with igbvf on 82576 can't talk to host Hi Greg, Very good. Hope you find something soon. Alright we've found something and it seems to be a pretty solid lead. What I'd like for you to do if you can find the time is repeat the tests you've been running but put up a window that watches the packet stats per queue. Something like this: watch -n 2 'ethtool -S ethx | grep packets' Then watch the tx queue packet stats. We're thinking that when the problem occurs that the VF is not able to communicate with the PF you will see packets from the PF to the VF going out on tx queue 1. When it works you'll see packets going out on tx queue 0. If you could confirm that for us it would help out a great deal. Thanks for your help and your patience. - Greg -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
Hi Greg, So I guess it is some kind of race condition. I just rebooted the vm several times but the behavior didn't change. So I guess it has to do with the initialization of the host driver or setting the macs for the vms. Unfortunately the host is already in production use so I can't just reboot the host serveral times now to verify that the behavior changes on host reboots. But I think I'll be able to do that within the next days. I just rebooted the host 10 times. The hardware, kernel, setup and test procedure was identical over these 10 reboots. I got 5 times: everything working 4 times: arp broadcast ok, arp unicast ok, regular (non-arp) packets not working 1 time: arp (broadcast and unicast) not working, regular (non-arp) packets ok Rebooting the vms within one host-boot didn't change behavior: once the host was started, the behavior stayed the same. VM restarts did not alter anything. So to me this seems to be a race condition in the host driver initialization. The outcome can be as expected or different erratic behavior. I'm still able to reliably reproduce this bug with the upstream 2.6.34 kernel drivers. If I use the drivers from sourceforge then the problem doesn't occur so there is definitely some driver issue that I have yet to find. This is a race condition. So if you can't reproduce it with the sourceforge driver, it doesn't mean the bug is not there. It could be just timing issues that prevent the bug to appear on your hardware. I'll keep working the problem. Very good. Hope you find something soon. Kind regards, Gerd -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
Hi Greg, thanks for your quick reply. But the host can't send packets to the guests (or these packets are never received on the guests). When using tcpdump, I can see these packets leaving the device on the host, but they never arrive on the guest. So you're saying that the guests don't even receive packets from the host with broadcast as the destination address? that is correct. I just tried it again to verify: eth1 on the host is connected to an otherwise unused switch. The host is at 192.168.100.1, the vm at 192.168.100.2. The vm is not connected to any other network and has only the vf as nic (eth0). [r...@host]# arping -I eth1 192.168.100.2 - no response during this I ran on the vm: [r...@vm]# tcpdump -i eth0 -n - no packets When looking at the packet counter on the vm, the RX-counter is still 0. So no broadcasts from the host can reach the vm. When doing it the other way round: [r...@vm]# arping -I eth0 192.168.100.1 I still get no response. But that is because the host can't answer. The host sees the packets and replies: [r...@host]# tcpdump -i eth1 -n 14:05:48.473303 ARP, Request who-has 192.168.100.1 (Broadcast) tell 192.168.100.2, length 28 14:05:48.473329 ARP, Reply 192.168.100.1 is-at 00:1b:21:60:xx:xx, length 28 14:05:49.473478 ARP, Request who-has 192.168.100.1 (Broadcast) tell 192.168.100.2, length 28 14:05:49.473492 ARP, Reply 192.168.100.1 is-at 00:1b:21:60:xx:xx, length 28 [...] This evening I will upgrade the vm to 2.6.34 too. But aside from that I don't know what else to try. Any ideas what I could do to further trace this down? Kind regards, Gerd -- Address (better: trap) for people I really don't want to get mail from: jo...@cactusamerica.com -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
-Original Message- From: Gerd v. Egidy [mailto:li...@egidy.de] Sent: Friday, May 21, 2010 5:24 AM To: e1000-devel@lists.sourceforge.net Cc: Rose, Gregory V Subject: Re: [E1000-devel] guest with igbvf on 82576 can't talk to host Hi Greg, thanks for your quick reply. But the host can't send packets to the guests (or these packets are never received on the guests). When using tcpdump, I can see these packets leaving the device on the host, but they never arrive on the guest. So you're saying that the guests don't even receive packets from the host with broadcast as the destination address? that is correct. I just tried it again to verify: eth1 on the host is connected to an otherwise unused switch. The host is at 192.168.100.1, the vm at 192.168.100.2. The vm is not connected to any other network and has only the vf as nic (eth0). [r...@host]# arping -I eth1 192.168.100.2 - no response during this I ran on the vm: [r...@vm]# tcpdump -i eth0 -n - no packets When looking at the packet counter on the vm, the RX-counter is still 0. So no broadcasts from the host can reach the vm. When doing it the other way round: [r...@vm]# arping -I eth0 192.168.100.1 I still get no response. But that is because the host can't answer. The host sees the packets and replies: [r...@host]# tcpdump -i eth1 -n 14:05:48.473303 ARP, Request who-has 192.168.100.1 (Broadcast) tell 192.168.100.2, length 28 14:05:48.473329 ARP, Reply 192.168.100.1 is-at 00:1b:21:60:xx:xx, length 28 14:05:49.473478 ARP, Request who-has 192.168.100.1 (Broadcast) tell 192.168.100.2, length 28 14:05:49.473492 ARP, Reply 192.168.100.1 is-at 00:1b:21:60:xx:xx, length 28 [...] This evening I will upgrade the vm to 2.6.34 too. But aside from that I don't know what else to try. Any ideas what I could do to further trace this down? Pretty strange but one thing I can think of to do is make sure that the source MAC address in the ARP packets the host is receiving is the same one you assigned via the ip link set... commands. Other than that I'm in the process of setting something up to reproduce the problem. - Greg -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
-Original Message- From: Rose, Gregory V Sent: Friday, May 21, 2010 12:29 PM To: 'Gerd v. Egidy'; e1000-devel@lists.sourceforge.net Subject: RE: [E1000-devel] guest with igbvf on 82576 can't talk to host -Original Message- From: Gerd v. Egidy [mailto:li...@egidy.de] Sent: Friday, May 21, 2010 5:24 AM To: e1000-devel@lists.sourceforge.net Cc: Rose, Gregory V Subject: Re: [E1000-devel] guest with igbvf on 82576 can't talk to host [snip] This evening I will upgrade the vm to 2.6.34 too. But aside from that I don't know what else to try. Any ideas what I could do to further trace this down? Pretty strange but one thing I can think of to do is make sure that the source MAC address in the ARP packets the host is receiving is the same one you assigned via the ip link set... commands. Other than that I'm in the process of setting something up to reproduce the problem. Good news of a sort. I can reproduce the problem. Which means I can debug it and hopefully fix it. I'll get back to you when I have some results. - Greg -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
But the host can't send packets to the guests (or these packets are never received on the guests). When using tcpdump, I can see these packets leaving the device on the host, but they never arrive on the guest. This evening I will upgrade the vm to 2.6.34 too. But aside from that I don't know what else to try. host and vm are on 2.6.34 now. But it didn't change anything, the problem still persists. Don't know if this helps: I took a look at the rx-interrupts for the virtual function on the vm. The number of interrupts increases by one every two seconds (estimated). The number of rx-interrupts is not in any way dependant on the number of packets sent from the host, even when sending a hundret packets the interrupts increase only every two seconds. Kind regards, Gerd -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
Re: [E1000-devel] guest with igbvf on 82576 can't talk to host
-Original Message- From: Gerd v. Egidy [mailto:li...@egidy.de] Sent: Friday, May 21, 2010 1:08 PM To: e1000-devel@lists.sourceforge.net Cc: Rose, Gregory V Subject: Re: [E1000-devel] guest with igbvf on 82576 can't talk to host But the host can't send packets to the guests (or these packets are never received on the guests). When using tcpdump, I can see these packets leaving the device on the host, but they never arrive on the guest. This evening I will upgrade the vm to 2.6.34 too. But aside from that I don't know what else to try. host and vm are on 2.6.34 now. But it didn't change anything, the problem still persists. Yes, I reproduced the problem on 2.6.34. I'm debugging it now. Will let you know of results ASAP. - Greg -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired
[E1000-devel] guest with igbvf on 82576 can't talk to host
Hi, I'm using a dualport ET (82576) running with SR-IOV on kernel 2.6.34. The guests are managed with kvm (qemu-kvm, current git, but I tried 0.12.4 release too). After some work the whole vt-d and sr-iov stuff is working. The host can use the main card without problems to communicate with other hosts on the lan. The guests (using igbvf with 2.6.33.4) can communicate with other hosts on the lan. Two guests can talk to each other on the same virtual link. The guests can send packets to the host. But the host can't send packets to the guests (or these packets are never received on the guests). When using tcpdump, I can see these packets leaving the device on the host, but they never arrive on the guest. This does not seem to be an arp-related problem as I have tried to manually insert the macs into the arp table. Now it's my ping-packets that are lost, not the arp ones. I have set the mac of all using iproute2 2.6.24 and ip link set eth0 vf 0 mac 02:00:00:00:00:00 ip link set eth0 vf 1 mac 02:00:00:00:00:01 ... I do not have set any vlan-related options, ip link show doesn't write anything about vlans. So I don't think it is a vlan-related problem. Any ideas what could cause this? Any ideas what I could try to fix it? Thanks a lot. Kind regards, Gerd -- ___ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel#174; Ethernet, visit http://communities.intel.com/community/wired