> [PATCH v8 4/4] net/zxdh: optimize Tx xmit pkts performance
>
> Error: the simple Tx burst signals a bad packet with a short return,
> which the application cannot distinguish from backpressure.
> for (i = 0; i < nb_pkts; i++) {
> rte_prefetch0(tx_pkts[i]);
> if (unlikely(tx_pkts[i]->data_off < hdr_len)) {
> txvq->stats.errors += nb_pkts - i;
> nb_pkts = i;
> break;
> }
> }
>
> A short return from tx_burst is the backpressure signal (transmit ring
> full, retry later). Here it is also used to mean "packet i is bad",
> and the bad mbuf is left owned by the caller. The application has no
> way to tell the two apart: the usual
>
> for (sent = 0; sent < n; )
> sent += rte_eth_tx_burst(port, q, &pkts[sent], n - sent);
>
> loop treats the short return as backpressure and resubmits pkts[i],
> which fails again every time -- head-of-line blocking, and the good
> packets after i (which had ring space) never go out.
>
> A packet that cannot be sent must be consumed by the driver, not
> handed back. Free it in tx_burst, increment the tx error counter, and
> continue with the rest of the burst. For a burst of 16 where only
> index 3 is bad and the ring has room, tx_burst should return 16, with
> stats showing 15 transmitted and 1 tx error. A short return is then
> reserved for the one case the application is entitled to retry: ring
> full.
The design intent of zxdh_xmit_pkts_simple is a fast path with
a usage restriction, not a general Tx entry point.
The restriction is data_off >= hdr_len on single-segment
mbufs. When a packet violates the restriction, the function
exits without sending — that is the contract, not a bug.
Applications that may produce reduced-headroom mbufs are
expected to keep RTE_ETH_TX_OFFLOAD_MULTI_SEGS enabled, in
which case zxdh_xmit_pkts_packed is selected instead and
handles the full set of inputs. The split between the two
paths is intentional.
how about add a paragraph to the "Limitations or Known issues"
section of doc/guides/nics/zxdh.rst describing the fast path's
requirement and pointing at the packed path as the alternative.
Thanks.
