I had the exact same issue since I also use interface queue (with fq_codel in my case). There is an easy solution. First, you need to apply the queues only to unmarked packets, which is what you are already doing (“packet-mark=no-mark”).
Then you need to mark your inter-VLAN traffic (adjust lists and mangle passthrough as needed):
/interface list
add name=local
/interface list member
add interface=vlan101 list=local
add interface=vlan102 list=local
/ip firewall mangle
add action=mark-packet chain=forward comment=inter-vlan in-interface-list=local new-packet-mark=intervlan out-interface-list=local passthrough=no
Packet mark can be anything, we are not matching on it. We just need some mark to satisfy the “no-mark” filters.
But the mangle rule won’t work because of fasttrack, so you need to adjust the latter to exclude all marked traffic.
/ip firewall filter
add action=fasttrack-connection chain=forward comment="fasttrack for established,related" connection-state=established,related hw-offload=yes packet-mark=no-mark
Although marking packets is heavy on CPU, this approach is still better over having to completely disable fasttrack in order to use simple queue where you don’t involve the bridge at all. And of course, it’s better than dealing with limited inter-VLAN bandwidth.
With this configuration on a hAP ac2 I can achieve 530 Mbit/s single stream iperf inter-VLAN bandwidth, 680 Mbit/s with 4 streams. I’m sure you won’t have any troubles reaching full gigabit on RB5009.