Re: [PATCH v4] app/testpmd: support runt frames in txonly

[Stephen Hemminger <stephen@networkplumber.org>]

On Fri, 7 Aug 2026 14:17:27 +0800
Xingui Yang <yangxingui@huawei.com> wrote:

> Allow setting transmit size to be a small value which has Ethernet
> header but no IP or UDP header, since control level protocols can be
> very short.
> 
> Checksum offloads are disabled when headers are incomplete.
> copy_buf_to_pkt_segs stops at the last segment to prevent OOB access.
> 
> Suggested-by: Stephen Hemminger <stephen@networkplumber.org>
> Signed-off-by: Xingui Yang <yangxingui@huawei.com>
> ---
> Changes in v4:
> - Removed ultra-small frame support (< 14 bytes) per Stephen's review.
> ---

Still some issues reported by AI.
Also not sure why your email keeps ending up in Spam folder.

Reviewed v4 against main (c1a46b9). Applies cleanly, builds clean with
-Dwerror=true.

Error
-----

app/test-pmd/config.c: set_tx_pkt_segments() is shared with the flowgen
forwarding engine ("set txpkts" is documented as setting the length for
FLOWGEN too), and flowgen still assumes the old 42 byte minimum:

	unsigned pkt_size = tx_pkt_length - 4;	/* Adjust FCS */
	...
	pkt->data_len = pkt_size;
	ip_hdr->total_length = RTE_CPU_TO_BE_16(pkt_size - sizeof(*eth_hdr));
	udp_hdr->dgram_len = RTE_CPU_TO_BE_16(pkt_size - sizeof(*eth_hdr) -
					      sizeof(*udp_hdr));

With "set txpkts 14", pkt_size is 10. Both length fields underflow in
unsigned arithmetic and are truncated to 16 bits, and flowgen emits
10 byte frames - shorter than the Ethernet header this patch is trying
to guarantee. It also writes 42 bytes of header into an mbuf whose
data_len is 10.

Confirmed:

  dpdk-testpmd --vdev='net_pcap0,rx_iface=lo,tx_pcap=/tmp/fg.pcap' -- \
	--txpkts=14 --forward-mode=flowgen

  every captured frame is 10 bytes: 00010203040100010203

flowgen_begin() should reject tx_pkt_length below
sizeof(rte_ether_hdr) + sizeof(rte_ipv4_hdr) + sizeof(rte_udp_hdr) + 4,
the same way tx_only_begin() guards the split and multi-flow cases.

Warning
-------

doc/guides/testpmd_app_ug/testpmd_funcs.rst: the note listing what
cannot be combined with runt lengths omits "set txtimes".
tx_only_begin() additionally requires sizeof(struct tx_timestamp)
beyond the 42 byte header stack when timestamping is enabled, so runt
lengths fail with -EINVAL there as well.

Info
----

app/test-pmd/txonly.c: the l3_len clamp and the ol_flags masking are
recomputed per packet in the fast path, but the result is invariant.
TX_PKT_SPLIT_RND is already rejected when the first segment is short,
so pkt_len does not vary across packets. Both could be decided once in
tx_only_begin().

app/test-pmd/config.c: the first sentence of the rewritten comment is
still wrong - the loop rejects segments *larger* than
mbuf_data_size[0], not smaller. Worth fixing while touching it.

doc/guides/testpmd_app_ug/testpmd_funcs.rst: spelling out "14 bytes"
reads better than sizeof(struct rte_ether_hdr) in a user guide.