Intel 82599 Rx counter not working with higher than 1514 bytes

[hopoaat]

Hi!

Did you observe absolute packet drop (all packets on Rx side) on Intel 82599ES NIC when receiving a higher frame than 1514 bytes with latest MoonGen version (commit eb3ee3222cfe7af1e940950ad5a905dc4f813962)?

• We use 1 port both for Tx, and Rx.

• Jumbo frame is enabled on eXtreme switch, server, and in vm too.

• On Tx side we observed that, the traffic goes out from the NIC even if we transmit over 2000 bytes/frame.

• The traffic goes a round trip, like this: Nic(tx)->eXtreme SW->Server->Vm(loopback)->Server->eXtreme SW->Nic(rx) Until Nic(rx) state I can see the counters(in eXtreme too), and seems like they are ok. The eXtreme forward the traffic back to the NIC, but we can't detect packets arriving to the NIC with MoonGen.

• Hardware rate controll used.

• Port setup

    local txDev = device.config{ port = args.txDev, rxQueues = args.queues, txQueues = args.queues}
    local rxDev
    if args.rxDev == args.txDev then
        rxDev = txDev
    else
        rxDev = device.config{ port = args.rxDev, rxQueues = args.queues }
    end

• The counting task something like this:

  function counterSlave(queue)
      local rxStats = stats:newDevRxCounter(queue, "plain")
      local bufs = memory.bufArray()
      while mg.running and runtime:running() do
          local rx = queue : tryRecv ( bufs , 100 )
          --Algorithm on packets here
          rxStats:update()
          bufs:freeAll()
      end
      rxStats:finalize()
  end

• Tried with ''dropEnable = false'' - not helped Any suggestion?

Thanks for your help!

Br,

Peter

[fernnf]

Hi,

This happen because libmoon is configuring dpdk port without jumbo frame feature, I was a patch that enables this. You need apply this patch on device.c of libmoon sources.

--- device.c 2017-03-29 14:27:18.657159902 -0300 +++ device.c 2017-03-22 14:49:13.263529000 -0300 @@ -118,8 +118,9 @@ .header_split = 0, .hw_ip_checksum = !cfg->disable_offloads, .hw_vlan_filter = 0, - .jumbo_frame = 0, + .jumbo_frame = 1, .hw_strip_crc = 1, + .max_rx_pkt_len= 9218, .hw_vlan_strip = cfg->strip_vlan ? 1 : 0, }, .txmode = {

Try this !

[hopoaat]

Hi!

Thanks for the patch! Worked!

Br,

Peter

[emmericp]

thanks!

I'll add something in the port configuration to enable jumbo frames manually.

[emmericp]

(re-opening because I'd like to create a proper fix for this some time later)

[hopoaat]

Ok. I also suggest to extend your MoonGen/examples/benchmark/packet-sizes.lua script to handle Jumbo frames, so others may find example of the usage. (Or write another example for it.)

Br,

Peter

[adolia]

Hi! I have the same issue on 82599ES 10-Gigabit SFI/SFP+ Network Connection NIC. I've set .jumbo_frame = 1, .hw_strip_crc = 1, .max_rx_pkt_len = 15872, to port_conf, but all packets > 10K are being dropped. Any idea? Also, is it possible to set MTU value? Thanks in advance. Br, Alex

[emmericp]

I don't think the ixgbe driver supports frames > 10k? (while the hardware should technically support it)