On Tue, Oct 29, 2024 at 10:05:21AM +0000, Stuart Henderson wrote:
> So to summarize what is clear from the recent findings:
>
> - em(4) is doing something that trips problems
> - either the path from LRO->TSO, or something with LRO itself,
> is doing something that trips problems
>
> and (mentioned elsewhere but including it for tech@ readers) -
> mbuf handling in wg(4) is broken.
>
> So on some systems disabling tcplro on interfaces ("ifconfig
> vio0 -tcplro", "ifconfig ix0 -tcplro" etc) may help work around
> the problem. On others (like the em case) that won't help.
>
> With the em case with no TSO, possibly running wg(4) in a VM and
> disabling tcplro on vio in the guest may help sidestep the wg(4)
> problems. (And if it's not that, net.inet.tcp.tso=0 might be worth
> a try).
To add another data point (not sure if it helps):
On a vmd host which uses ix(4), a vmd guest running wg(4) has
been stable for half a year of uptime. No crashes at all.
Both systems are still running 7.5. Haven't upgraded yet.
