Comment 98 for bug 1753662

I would have thought this would be the relevant patch:

bonding: speed/duplex update at NETDEV_UP event
 Mahesh Bandewar authored and davem330 committed on Sep 28, 2017
1 parent b5c7d4e commit 4d2c0cda07448ea6980f00102dc3964eb25e241c

However, it was first available in v4.15-rc1.

At least as far as bonding kernel changes go, there does not
seem another obvious candidate that might have fixed this problem
between 4.12 and 4.13 (first skim).

At least for one scenario I looked at, we got a bad speed/duplex
setting, which eventually ended up with the bond interface
aggregating on a separate port, and/or ending up in LACP DISABLED
state which it never got out of. We only checked correct/latest
device speed/duplex settings via the NETDEV_CHANGE path, where
we called _ethtool_get_settings(). If we don't receive a change
event again to correct the speed/duplex, we never recover.

There are some other patches which help address this at different
points, but are either before or later (see above) the window.

I'll take a look at code outside the bonding dir which might
impact this.

Joseph, could you provide the raw config files you used as well?
It was not super clear in the png image if those were the only
diffs. They did not seem very relevant diffs either.