Control: retitle -1 TX watchdog fires on e1000e interface with flow control enabled On Tue, 2017-03-21 at 18:36 -0400, Bruce Momjian,,, wrote: > On Tue, Mar 21, 2017 at 04:04:11PM -0400, Bruce Momjian,,, wrote: > > I think this proves my problems are related to flow control. How would > > you like to proceed? Is there a patch or change you would like me to > > test? Just close the ticket? > > > > I have a fix, but it is likely others would not know they had this > > problem unless they were monitoring their kernel logs or their network > > traffic for lag. > > Oh, I should also mention the port that is having problems is connected > to a NetGear GS108Ev3 switch, with current firmware, version 2.00.09. > The port connected to my Actiontec FIOS router is not having problems. I don't know about any specific bug, but if the switch sends flow control XOFF frames continually for long enough (usually 5 seconds) this will trigger the TX watchdog. It sounds like your switch implements flow control properly (some broken switches auto-negotiate it but actually flood flow control frames). However, if a device on some other port (that also has flow control enabled) sends XOFF frames continually *and* your server sends frames that should go to that other port, the switch will do the same to the server once the switch's internal queue has filled up. If the switch has port statistics including numbers of pause frames then you can see where they are coming from, but I think it doesn't. Without that information it's going to be hard to tell exactly where the fault lies. The e1000e driver *does* have statistics for pause frames transmitted and received (run: "ethtool -S eth0| grep flow_control"). If you log these every second then it should be possible to see what happens around the time the TX watchdog fires. That could provide some clues as to whether the NIC is behaving correctly. Ben. -- Ben Hutchings Power corrupts. Absolute power is kind of neat. - John Lehman, Secretary of the US Navy 1981-1987
Attachment:
signature.asc
Description: This is a digitally signed message part