`eth_tx_tstrb_issue_test` shows packet size mismatch: expected 78 bytes, observed 84 bytes (+6). Suggests final AXI beat byte-valid mask is not respected.

[RaiMuhammadDanish]

Expected behavior

For payload size 52:

• Final AXI beat has 2 valid bytes (tstrb=0x03)
• Only 2 bytes from final beat should be transmitted
• Frame size should match reference model (78 bytes)

Actual behavior

• Driver reports:
  • remaining bytes = 2
  • tstrb = 3
• Observed waveform:
  • exp_pkt_size=78bytes
  • act_pkt_size=84bytes
• Extra bytes observed: +6 (matches invalid lanes of final 64-bit beat)

Analysis

Difference is exactly +6 bytes, which equals invalid byte lanes on last 64-bit beat when only 2 bytes are valid (8 - 2 = 6).
Likely final-beat mask handling issue in TX datapath (tstrb/tkeep not applied correctly near EOP).

[dngr2]

Your diagnosis is right, and the cause is one level down in a dependency rather than in this repo. Fix proposed as pulp-platform/axi_stream#8.

Where the bytes come from

eth_top.sv:123 instantiates axi_stream_dw_downsizer to convert the 64-bit AXI stream to the MAC's 8-bit interface (FramingDataWidth = 8). That module emits TotalSubTransfers sub-transfers per input beat regardless of TKEEP — so a final beat with two valid bytes still produces eight.

Those extra beats then reach the wire because framing_top.sv:214-218 wires only tdata, tvalid, tready, tlast and tuser to the MAC. t.keep and t.strb are declared and carried through the port, but never connected onward, so nothing downstream can tell a null lane from a real byte.

Reproduced in Verilator against the unmodified downsizer — one beat, TKEEP=0x03, TLAST=1:

sub-beats emitted      : 8
of which keep=1        : 2
spurious bytes on wire : 6

+6, exactly the 78 → 84 you measured.

Fix

pulp-platform/axi_stream#8 makes the last sub-transfer of a beat the index of its highest asserted TKEEP bit, latched when the beat is accepted. Ten checks including a full-beat regression; four of them fail against the unfixed RTL.

Fixing it there rather than here means every consumer of the downsizer benefits, and pulp-ethernet needs no RTL change.

One thing that will block it reaching you

Bender.lock in this repo pins axi_stream at 472751f550e3918215603e21734fe0ece3c66f79. That revision is not reachable in the public pulp-platform/axi_stream repository — I cloned it and fetched all refs before saying so.

So even once the downsizer is fixed upstream, this pin will not pick it up, and today it cannot be resolved from a clean checkout at all. Worth a look independently of this bug.

I have no Questa licence, so I verified the downsizer in isolation with Verilator rather than running this repo's eth_tb. The eth_tx_tstrb_issue_test you mention is not in the tree — if you can share it, I am happy to check the fix against the full datapath.