On Sat, 2013-08-17 at 12:56 +0200, Josip Rodin wrote: > On Sat, Aug 17, 2013 at 12:33:02PM +0200, Josip Rodin wrote: > > On Sat, Aug 17, 2013 at 12:23:21PM +0200, Ben Hutchings wrote: > > > > tc qdisc add dev $iface_Internet parent 1:2 handle 20: tbf rate 20mbit buffer 20480 limit 16384 > > > > > > > > However, the upgrade to this new kernel appears to have killed it - the tbf > > > > rule is causing outgoing HTTP connections to max out at around 8 Kbps. > > > [...] > > > > > > This might be the same as bug #708995. Does turning off GRO on the > > > internal interface (not the bridge but the physical interface) work > > > around it? > > > > Yes, it looks like ifenslave -c bond0 eth0 && ethtool -K eth0 gro off makes > > TBF precise again, and vice versa. > > That's on one machine. But on another wheezy machine with the same setup but > somewhat different hardware, turning off GRO didn't help. > > How do I debug this further? You could try using the perf dropmonitor script as I described on my bug report. The other machine might also have LRO enabled on the internal interface, although this is supposed to be disabled for bridged interfaces. If the other machine is also passing traffic from another VM on the same physical host, it might be necessary to disable TSO on the interface within the other VM. Ben. -- Ben Hutchings Teamwork is essential - it allows you to blame someone else.
Attachment:
signature.asc
Description: This is a digitally signed message part