Package: src:linux Version: 3.16.7-ckt2-1~bpo70+1 Severity: normal Hi! I have encountered a strange slowness on a router/packetfilter system (Wheezy 7.8 with backported kernel 3.16.7-ckt2-1~bpo70+1) of mine while forwarding GRE packets. The system in question does not terminate the GRE tunnels itself, it merely forwards those packets from one interface to another. This problem occurred after a reboot from a 3.12-bpo to the current 3.16.7-ckt2-1~bpo70+1. This problem reminded me of http://lists.openwall.net/netdev/2014/07/08/132 and http://patchwork.ozlabs.org/patch/324953/ and as described on those mailing-list exchanges, after disabling GSO and GRO on the involved interfaces, the forwarding speed for GRE packets was back to normal. What puzzles me is the fact that this bug should be fixed since 3.14.x and 3.15.x and sure enough, I can find the code changes from the patch from patchwork in the current kernel code in the Debian package. But the symptoms are the same as described: with active GRO/GSO the GRE tunnels max out at about 200kbit/s, after disabling GRO/GSO the possible bandwidth is only limited by the speed of the interfaces and connections involved. The hardware is like this: 14:00.0 Ethernet controller: Intel Corporation 82599EB 10-Gigabit SFI/SFP+ Network Connection (rev 01) 14:00.1 Ethernet controller: Intel Corporation 82599EB 10-Gigabit SFI/SFP+ Network Connection (rev 01) 0e:00.0 Ethernet controller: Intel Corporation 82599EB 10-Gigabit SFI/SFP+ Network Connection (rev 01) 0e:00.1 Ethernet controller: Intel Corporation 82599EB 10-Gigabit SFI/SFP+ Network Connection (rev 01) ixgbe 0000:14:00.0: Multiqueue Enabled: Rx Queue count = 16, Tx Queue count = 16 ixgbe 0000:14:00.0: PCI Express bandwidth of 32GT/s available ixgbe 0000:14:00.0: (Speed:5.0GT/s, Width: x8, Encoding Loss:20%) ixgbe 0000:14:00.0: MAC: 2, PHY: 14, SFP+: 5, PBA No: FFFFFF-0FF ixgbe 0000:14:00.0: 00:10:f3:33:d3:e8 ixgbe 0000:14:00.0: Intel(R) 10 Gigabit Network Connection I created two LACP bonds (bond0 and bond1) out of two interfaces and on top of those two bonds there are several VLANs. The GRE packets are forwarded from a VLAN on bond0 to a VLAN on bond1. The regain the full bandwidth I disabled GSO and GRO on each of the four slave interfaces and both bond interfaces. Grüße, Sven.
Attachment:
signature.asc
Description: OpenPGP digital signature