port-xen/57743: ARP lossage with xennet

[marcotte%panix.com@localhost]

>Number:         57743
>Category:       port-xen
>Synopsis:       ARP lossage with xennet
>Confidential:   no
>Severity:       serious
>Priority:       high
>Responsible:    port-xen-maintainer
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Sat Dec 02 13:10:00 +0000 2023
>Originator:     Brian Marcotte
>Release:        10.0 or -current
>Organization:
Public Access Networks, Corp
>Environment:
NetBSD xxx 10.0_RC1 NetBSD 10.0_RC1 (XEN3_DOMU) #1: Fri Dec  1 19:02:12 EST 2023  root@xxx:/misc/bug/src/sys/arch/amd64/compile/XEN3_DOMU amd64
>Description:
I've found a problem with ARP and the xennet driver.

This is happening with NetBSD-10 (or -current) domUs ONLY when they're
running on the same Linux dom0 host in the same subnet. If the NetBSD-10
domUs are running on different hosts or on different subnets, there is
no problem.

What I'm seeing is that ARP replies between NetBSD-10 domUs are not
reaching each other. Tcpdump in the NetBSD domUs show the replies going
out, but tcpdump on the Linux dom0 doesn't see them at all.


I also found that if I initiate contact on both domUs to the other, this
causes an "ARP storm" bogging down the whole machine. They both send out
ARP requests very fast never receiving replies.

Thanks. I've been running NetBSD-10 for months. Sorry for only noticing
this now.

>How-To-Repeat:
My environment:

dom0: Xen 4.17, Linux 5.10, bridging
domUs: NetBSD-10 or -current; PV, PVH, or PVHVM
       same subnet

Have one domU ping the other so they will ARP.

>Fix:
Unsure, but I found that this first broke on March 23, 2020.

The xennet changes on that date were related to checksums committed by
Jaromir Dolecek.