[BUG] DPDK VLAN-push flow is rejected by mlx5 under default HWS mode

[dleshchev]

Describe the bug

A raw DPDK configuration containing a TX vlan_push action cannot initialize on
ConnectX-7 under DAQIRI's default dv_flow_en=2 hardware-steering mode.

DAQIRI completes configuration parsing, memory registration, ethdev setup, and TX
queue creation. The mlx5 driver then rejects the VLAN-push action while creating the
TX flow. daqiri_init() returns failure and the benchmark cannot start.

Steps/Code to reproduce bug

Build a local container from main commit
57f3d8ebdd2e5410f654edea2e25b211cbca17de:

IMAGE_TAG=daqiri:vlan-repro \
BASE_TARGET=dpdk \
DAQIRI_ENGINE="dpdk ibverbs" \
scripts/build-container.sh

Start it with access to the GPU, NIC, and hugepages:

docker run --rm -it \
  --privileged \
  --gpus all \
  --network host \
  -v /dev/hugepages:/dev/hugepages \
  -v "$PWD:/workspace/daqiri" \
  -w /workspace/daqiri \
  daqiri:vlan-repro

Inside the container, save this as /tmp/vlan-dpdk.yaml:

%YAML 1.2
---
daqiri:
  cfg:
    version: 1
    stream_type: raw
    engine: dpdk
    master_core: 3
    debug: false
    log_level: info
    loopback: ''
    memory_regions:
    - name: Data_TX_GPU
      kind: device
      affinity: 0
      num_bufs: 16384
      buf_size: 1042
    - name: Data_RX_GPU
      kind: device
      affinity: 0
      num_bufs: 16384
      buf_size: 1042
    interfaces:
    - name: tx_port
      address: '0005:03:00.0'
      tx:
        queues:
        - name: tx_q_0
          id: 0
          batch_size: 1024
          cpu_core: 4
          memory_regions: [Data_TX_GPU]
          offloads: [tx_eth_src]
        flows:
        - name: vlan_encap
          id: 101
          actions:
          - type: vlan_push
            vlan_id: 100
            pcp: 0
            dei: 0
            ethertype: 33024
          match:
            udp_src: 48100
            udp_dst: 48100
    - name: rx_port
      address: '0005:03:00.1'
      rx:
        flow_isolation: true
        queues:
        - name: rx_q_0
          id: 0
          cpu_core: 5
          batch_size: 1024
          memory_regions: [Data_RX_GPU]
        flows:
        - name: vlan_decap
          id: 101
          match: {}
          actions:
          - type: vlan_pop
          - type: queue
            id: 0
bench_rx:
- interface_name: rx_port
  queue_id: 0
  cpu_core: 7
bench_tx:
- interface_name: tx_port
  queue_id: 0
  cpu_core: 6
  batch_size: 1024
  payload_size: 1000
  header_size: 42
  eth_dst_addr: 48:b0:2d:f4:04:24
  ip_src_addr: 1.1.1.1
  ip_dst_addr: 2.2.2.2
  udp_src_port: 48100
  udp_dst_port: 48100

Run:

/opt/daqiri/bin/daqiri_bench_raw_gpudirect \
  /tmp/vlan-dpdk.yaml --seconds 2

The DPDK arguments confirm that hardware steering is selected:

-a 0005:03:00.0,txq_inline_max=0,dv_flow_en=2
-a 0005:03:00.1,txq_inline_max=0,dv_flow_en=2

Flow creation then fails:

[INFO] ... Adding TX flow vlan_encap
mlx5_net: [mlx5dr_action_root_build_attr]: Found unsupported action type: 14
mlx5_net: [mlx5dr_bwc_rule_create_root]: BWC rule: failed creating rule on root tbl
[CRITICAL] ... Failed to create TX flow 'vlan_encap' on port 0:
                 fail to create rte flow (Invalid argument)
[CRITICAL] ... Failed to initialize DPDK
daqiri_init failed

Expected behavior

DAQIRI documents VLAN push/pop as supported by the raw DPDK engine, so the VLAN
flow should install under the engine's default steering configuration.

If mlx5 requires a different steering mode for this action, DAQIRI should either:

1. select the compatible mode for interfaces containing VLAN actions, or
2. reject the configuration before partial initialization with a clear message
   identifying the required mode.

Whether dv_flow_en=1 works on this hardware has not yet been tested.

Environment overview (please complete the following information)

• Environment location: Docker on bare-metal IGX
• Method of DAQIRI install: locally built Docker image
  • No docker pull command was used.
  • Build and run commands are included above.

Environment details

• Host OS: Ubuntu 22.04.5 LTS
• Host kernel: 5.15.0-1039-nvidia-tegra-igx, aarch64
• Container base: CUDA 13.1 development image on Ubuntu 24.04
• DAQIRI version: 2026.7.0
• DAQIRI commit: 57f3d8ebdd2e5410f654edea2e25b211cbca17de
• DPDK version: 25.11
• Hardware: NVIDIA IGX with discrete RTX 6000 Ada GPU
• NIC: NVIDIA ConnectX-7, ports 0005:03:00.0 and 0005:03:00.1
• NIC firmware: 28.39.3004
• mlx5 kernel driver: 24.07-0.6.1
• MLNX_OFED/rdma-core: 24.07
• NVIDIA GPU driver: 535.274.02

Additional context

VLAN/tunnel actions were introduced in PR #194 as the implementation for #196.
The recorded PR checks did not compile or exercise the VLAN flow on DPDK hardware.

The rejected action is emitted as RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN. Memory
registration and queue setup finish before this error, isolating the observed failure
to TX flow creation.