A switch socket's speed and duplex set from the second XGS 3300's side, 2026-11-15
The front ports offered autoselect alone, and a change of medium was accepted and ignored. The switch sockets'
speed and duplex are set by the network agent's KSETTINGS attribute (0x55), the one the SDK's own host sends for
an ethtool -s, with what autonegotiation advertises in attribute 0x43
(the attributes). Since the change
a socket offers 10 and 100 half and full, 1000baseT full and autoselect (octep_dp_media_change), and a cage keeps
autoselect.
How it was measured
2026-11-15 from 05:09 to 05:13 UTC: the second XGS 3300 (vendor version 22, octep.ko 2cec49f9), its Port3 cabled
to a USB 2.5 Gbit/s adapter on the FreeBSD test machine (ue2), which negotiates.
- Each step: the medium asked for with
ifconfig oxp4 media ...(from autoselect, since ifconfig adds new options to the ones in force); 25 seconds later, what the port reports, what the switch holds (a GET of0x55: autonegotiation, Mbit/s, duplex; and of0x43), and the adapter's view of the link. - Traffic at 100 and 10 half: TCP, four connections for 30 seconds, routed from a jail behind Port4 out of Port3, as in the PPPoE measurement.
- A cage: F1 asked for 100baseTX full.
Results
| Asked for | The port reports | The switch: KSETTINGS, advertises | The adapter | TCP through it |
|---|---|---|---|---|
100baseTX |
100baseTX half | 0 / 100 / 0, 0x2f |
100baseTX half | 88.0 Mbit/s, 99 % wire to wire |
10baseT/UTP |
10baseT half | 0 / 10 / 0, 0x2f |
10baseT half | 8.76 Mbit/s, 98 % wire to wire |
1000baseT full |
1000baseT full | 1 / 1000 / 1, 0x60 |
1000baseT full | - |
autoselect |
1000baseT full | 1 / 1000 / 1, 0x6f |
1000baseT full | - |
F1, 100baseTX full |
refused: the cage offers autoselect alone | - | - | - |
- Each medium took effect on the wire: the adapter, negotiating, came to the speed set on the XGS 3300.
- A forced 10 or 100 is half at a partner that negotiates: Ethernet's parallel detection, not this driver.
Forced 100 full against the adapter forced to 100 full ran at 1.11 Mbit/s on 2026-11-14: the switch port read
100 full (status register
0x0d0f) and counted 817 fragments from the adapter, which behaved as half duplex. Not tried with a partner that keeps a forced full duplex. - A gigabit copper link is never forced: 1000baseT is autonegotiation with 1000 full alone advertised.
- XGS: Ports reads the speed from what the link runs at, and says "half duplex" and "fixed" when the port was set by hand.

A partner forced to half, 2026-11-15 from 11:42 to 11:52 UTC
The XGS 3300 read the duplex only of a medium set on its own side, and took a negotiated link to be full: with the
XGS 136's Port8 forced to 100 half, the second XGS 3300's Port8 facing it, autoselect, said 100baseTX
<full-duplex> (octep.ko 92ee29bb, the XGS 136's media). The link ran half all the
same - Ethernet's parallel detection - and only the network agent knew. Since the change a switch socket's duplex is
read from the agent (DUPLEX, 0x0d) when its link comes up and when its speed changes, and the driver logs it.
| The XGS 136's Port8 | The XGS 3300's Port8, before | After (octep.ko a1e72726) |
Ping over the link |
|---|---|---|---|
100baseTX, half |
autoselect, 100baseTX full | autoselect, 100baseTX half | 5 of 5 |
10baseT/UTP, half |
autoselect, 10baseT full | autoselect, 10baseT half | 5 of 5 |
1000baseT full |
autoselect, 1000baseT full | autoselect, 1000baseT full | 5 of 5 |
| autoselect | autoselect, 1000baseT full | autoselect, 1000baseT full | 5 of 5 |
- XGS: Ports takes the duplex from what the link runs at: the Ports page's own parsing, run with php on the XGS 3300 against its Port8 line while the XGS 136 was forced to 100 half, gave "100M", "half" and not fixed. The page itself was not read in a browser: the XGS 3300's restarts for the new module had ended the GUI session.
- Not measured: a cage, whose duplex is its module's and is left full.

The discussion of this page is held in the project's GitHub Discussions. Showing it here loads giscus from giscus.app and GitHub, and writing in it needs a GitHub account - what that means.