> we started reproducing the problem by flooding the NIC with multicast > packets in a directly connected setup (no switch, just device and test > device with a single cable). At some point, the driver stops receiving > packets; but we've noticed that the link level leds are dead too.
After some more testing the problem seems to occur more often when subscribing and unsubscribing from multicast addresses (I don't know if the number of userspace listeners have a real impact at the moment). We do not seem to have the issue when running the GNU/Debian stretch kernel (4.9.228) while we do see it when running a GNU/Debian buster kernel (4.19.132). Having a look at the relevant changes; I'm wondering if the patches of Paul Burton could have something to do with this, from a quick glance, these are the ones that state that changes have been made in the multicast behaviour. > Reloading the driver does not re-establish the link (no-carrier), only > a reboot when the CPU comes out of reset does. I have not yet checked if this is related with the previous: in normal circumstances there is little need to unload/load the normal kernel module and this is not a normal operational situation. g. Marc
