Packet loss with new hw-offload bridge feature

since I’ve upgraded my RB750GL from 6.3x to 6.4x I experience sporadic packet loss through the automatic generated bridge with hw-offload.
Before upgrade and master-slave switch port config there wasn’t any of that kind of problems.

There was ether3 and ether4 in a switch config with ether4 as master.

Now with bridge and hw-offload there is packet loss of following behavior:

  • Packets through bridge (between ether3 and ether 4) makes packet loss not to all hosts behind the RB750 at the same time, sometimes host a, sometimes host b, sometimes a and b…
  • Packets (icmp ping tested) from RB750 itself outgoing to different hosts via ether 3 and ether 4 makes packet loss to all hosts at the same time

Packet loss happens randomly with different outages: sometimes one icmp ping only, sometimes more even up to 20 lost icmp pings.
This is disrupting my service to clients,especially when it comes to VoIP.
When returning to bridge config without hw-offload it also degrades performance.

Did somebody else had that experience, too?

What sounds odd to me is that packets forwarded by the switch chip itself would be lost due to software upgrade, as the switch chip is not affected by the upgrade. The master-slave configuration is an equivalent of HW accelerated bridge - in both cases, the result is that packets are forwarded by the switch chip directly between its Ethernet ports, bypassing the CPU.

I suppose that you have just a single bridge configured so the hardware acceleration is really activated for that bridge. If it is true, and as you say that the issue happens on packets sent between the CPU and any of (ether3, ether4), I would be afraid of a hardware issue of the switch chip, the PCB, or the power supply.