[Bug 235524] igb(4): Ethernet interface loses active link state

[b***@freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=235524

***@ifis.cs.tu-bs.de changed:

What |Removed |Added
----------------------------------------------------------------------------
CC| |***@ifis.cs.tu-bs.de

--- Comment #10 from ***@ifis.cs.tu-bs.de ---
Same problem here from time to time on a four port realtek card (re0, re1
currently used).

Machine is used as a router and it now happened three times that the interfaces
somehow get shut off - not seen, what the real problem is, as we had to have it
up and running very quickly, so an immediate reboot helped in all three cases.

link state changes to DOWN / UP happens once or twice every minute, when we
cannot receive any packets on both interfaces, even though only one interface
is logged with this problem.

This switching down and up again happens every now and then without further
impact. Time between DOWN and UP message is four seconds quite consistently. Is
seems to be triggered from a watchdog timeout (kernel: re0: wathcdog timeout).

Maybe the problem gets triggered by (too?) much network traffic, as the network
traffic was probably higher during those three cases.

Btw.: someone mentioned of similar problems when pf was enabled and no
problems, when pf was disabled. We are running ipfilter on that machine, so it
might be possible that filtering and high traffic promotes these problems,
regardless of the network card used.

[b***@freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=235524

Kevin Bowling <***@freebsd.org> changed:

What |Removed |Added
----------------------------------------------------------------------------
CC| |***@freebsd.org

--- Comment #11 from Kevin Bowling <***@freebsd.org> ---
(In reply to mack from comment #10)
This is a different issue. The realtek driver in FreeBSD is missing a lot of
critical workarounds for hardware issues, which is why you see the watchdog.
You can try the "net/realtek-re-kmod" driver in ports which contains some of
the workarounds needed and is tested by the vendor.

I am interested in improving re(4) but to set expectations these are colossal
tasks that cover 60 chipset variations and none of them are flawless, I don't
have docs, and have to sneak this in between a dayjob. If you can provide more
info like the pciconf output that would make for a useful bug report (in a new
PR not this one which is not related).

If you rely on the system in a material way I would switch to a vendor
supported NIC, like Chelsio T4+, Mellanox CX4+, or intel ixgbe+. Chelsio has
twisted pair options.

[b***@freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=235524

--- Comment #12 from ***@ifis.cs.tu-bs.de ---
Created a new bug report, see
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=255319