[PATCH] mbuf: dump Tx offload metadata

[PATCH] mbuf: dump Tx offload metadata

[David Marchand]

When debugging TSO and other checksum offloads, dumping the various
l*_len fields and TSO segmentation size can be helpful.

Example in OVS that dumps a mbuf on rte_eth_tx_prepare failure:

Before:
netdev_dpdk(pmd-c30/id:11)|DBG|dpdk0: First invalid packet:
dump mbuf at 0x2201a916c0, iova=0x2200800580, buf_len=6864
  pkt_len=6804, ol_flags=0x3114800000000102, nb_segs=1, port=65535, ptype=0
  segment at 0x2201a916c0, data=0x22008005b2, len=6804, off=50, refcnt=1
  Dump data at [0x22008005b2], len=6804
...

After:
netdev_dpdk(pmd-c30/id:11)|DBG|dpdk0: First invalid packet:
dump mbuf at 0x2201a916c0, iova=0x2200800580, buf_len=6864, pkt_len=6804
  outer_l2_len=14, outer_l3_len=40, l2_len=38, l3_len=40, l4_len=32, tso_segsz=0
  ol_flags=0x3114800000000102, nb_segs=1, port=65535, ptype=0
  segment at 0x2201a916c0, data=0x22008005b2, len=6804, off=50, refcnt=1
  Dump data at [0x22008005b2], len=6804
...

Signed-off-by: David Marchand <david.marchand@redhat.com>
---
 lib/mbuf/rte_mbuf.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/lib/mbuf/rte_mbuf.c b/lib/mbuf/rte_mbuf.c
index 0d931c7a15..74dcb6c934 100644
--- a/lib/mbuf/rte_mbuf.c
+++ b/lib/mbuf/rte_mbuf.c
@@ -750,10 +750,12 @@ rte_pktmbuf_dump(FILE *f, const struct rte_mbuf *m, unsigned dump_len)
 
 	__rte_mbuf_sanity_check(m, 1);
 
-	fprintf(f, "dump mbuf at %p, iova=%#" PRIx64 ", buf_len=%u\n", m, rte_mbuf_iova_get(m),
-		m->buf_len);
-	fprintf(f, "  pkt_len=%u, ol_flags=%#"PRIx64", nb_segs=%u, port=%u",
-		m->pkt_len, m->ol_flags, m->nb_segs, m->port);
+	fprintf(f, "dump mbuf at %p, iova=%#" PRIx64 ", buf_len=%u, pkt_len=%u\n",
+		m, rte_mbuf_iova_get(m), m->buf_len, m->pkt_len);
+	fprintf(f, "  outer_l2_len=%u, outer_l3_len=%u, l2_len=%u, l3_len=%u, l4_len=%u, tso_segsz=%u\n",
+		m->outer_l2_len, m->outer_l3_len, m->l2_len, m->l3_len, m->l4_len, m->tso_segsz);
+	fprintf(f, "  ol_flags=%#"PRIx64", nb_segs=%u, port=%u",
+		m->ol_flags, m->nb_segs, m->port);
 
 	if (m->ol_flags & (RTE_MBUF_F_RX_QINQ | RTE_MBUF_F_TX_QINQ))
 		fprintf(f, ", vlan_tci_outer=%u", m->vlan_tci_outer);
-- 
2.51.0

RE: [PATCH] mbuf: dump Tx offload metadata

[Morten Brørup]

> From: David Marchand [mailto:david.marchand@redhat.com]
> Sent: Friday, 31 October 2025 14.12
> 
> When debugging TSO and other checksum offloads, dumping the various
> l*_len fields and TSO segmentation size can be helpful.
> 
> Example in OVS that dumps a mbuf on rte_eth_tx_prepare failure:
> 
> Before:
> netdev_dpdk(pmd-c30/id:11)|DBG|dpdk0: First invalid packet:
> dump mbuf at 0x2201a916c0, iova=0x2200800580, buf_len=6864
>   pkt_len=6804, ol_flags=0x3114800000000102, nb_segs=1, port=65535,
> ptype=0
>   segment at 0x2201a916c0, data=0x22008005b2, len=6804, off=50,
> refcnt=1
>   Dump data at [0x22008005b2], len=6804
> ...
> 
> After:
> netdev_dpdk(pmd-c30/id:11)|DBG|dpdk0: First invalid packet:
> dump mbuf at 0x2201a916c0, iova=0x2200800580, buf_len=6864,
> pkt_len=6804
>   outer_l2_len=14, outer_l3_len=40, l2_len=38, l3_len=40, l4_len=32,
> tso_segsz=0
>   ol_flags=0x3114800000000102, nb_segs=1, port=65535, ptype=0
>   segment at 0x2201a916c0, data=0x22008005b2, len=6804, off=50,
> refcnt=1
>   Dump data at [0x22008005b2], len=6804
> ...
> 
> Signed-off-by: David Marchand <david.marchand@redhat.com>

Good idea.

With or without suggested improvement inline below,
Acked-by: Morten Brørup <mb@smartsharesystems.com>

> ---
>  lib/mbuf/rte_mbuf.c | 10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/lib/mbuf/rte_mbuf.c b/lib/mbuf/rte_mbuf.c
> index 0d931c7a15..74dcb6c934 100644
> --- a/lib/mbuf/rte_mbuf.c
> +++ b/lib/mbuf/rte_mbuf.c
> @@ -750,10 +750,12 @@ rte_pktmbuf_dump(FILE *f, const struct rte_mbuf
> *m, unsigned dump_len)
> 
>  	__rte_mbuf_sanity_check(m, 1);
> 
> -	fprintf(f, "dump mbuf at %p, iova=%#" PRIx64 ", buf_len=%u\n", m,
> rte_mbuf_iova_get(m),
> -		m->buf_len);
> -	fprintf(f, "  pkt_len=%u, ol_flags=%#"PRIx64", nb_segs=%u,
> port=%u",
> -		m->pkt_len, m->ol_flags, m->nb_segs, m->port);
> +	fprintf(f, "dump mbuf at %p, iova=%#" PRIx64 ", buf_len=%u,
> pkt_len=%u\n",
> +		m, rte_mbuf_iova_get(m), m->buf_len, m->pkt_len);
> +	fprintf(f, "  outer_l2_len=%u, outer_l3_len=%u, l2_len=%u,
> l3_len=%u, l4_len=%u, tso_segsz=%u\n",
> +		m->outer_l2_len, m->outer_l3_len, m->l2_len, m->l3_len, m-
> >l4_len, m->tso_segsz);

Printing of outer_l2_len and outer_l3_len could be omitted if not present, like the VLAN IDs, e.g.:

fprintf(f, "  ");
if (m->outer_l2_len != 0)
	fprintf(f, "outer_l2_len=%u, ", m->outer_l2_len);
if (m->outer_l3_len != 0)
	fprintf(f, "outer_l3_len=%u, ", m->outer_l3_len);
fprintf(f, "l2_len=%u, l3_len=%u, l4_len=%u, tso_segsz=%u\n",
	m->l2_len, m->l3_len, m->l4_len, m->tso_segsz);

> +	fprintf(f, "  ol_flags=%#"PRIx64", nb_segs=%u, port=%u",
> +		m->ol_flags, m->nb_segs, m->port);
> 
>  	if (m->ol_flags & (RTE_MBUF_F_RX_QINQ | RTE_MBUF_F_TX_QINQ))
>  		fprintf(f, ", vlan_tci_outer=%u", m->vlan_tci_outer);
> --
> 2.51.0

Re: [PATCH] mbuf: dump Tx offload metadata

[David Marchand]

On Fri, 31 Oct 2025 at 14:28, Morten Brørup <mb@smartsharesystems.com> wrote:
> Printing of outer_l2_len and outer_l3_len could be omitted if not present, like the VLAN IDs, e.g.:
>
> fprintf(f, "  ");
> if (m->outer_l2_len != 0)
>         fprintf(f, "outer_l2_len=%u, ", m->outer_l2_len);
> if (m->outer_l3_len != 0)
>         fprintf(f, "outer_l3_len=%u, ", m->outer_l3_len);
> fprintf(f, "l2_len=%u, l3_len=%u, l4_len=%u, tso_segsz=%u\n",
>         m->l2_len, m->l3_len, m->l4_len, m->tso_segsz);

Mm, why specifically those two only?
If the intention is to make this debug output shorter, I would apply
the same to all those fields, and omit the whole (new) line if no
field is set.


-- 
David Marchand

RE: [PATCH] mbuf: dump Tx offload metadata

[Morten Brørup]

> From: David Marchand [mailto:david.marchand@redhat.com]
> Sent: Friday, 31 October 2025 15.08
> 
> On Fri, 31 Oct 2025 at 14:28, Morten Brørup <mb@smartsharesystems.com>
> wrote:
> > Printing of outer_l2_len and outer_l3_len could be omitted if not
> present, like the VLAN IDs, e.g.:
> >
> > fprintf(f, "  ");
> > if (m->outer_l2_len != 0)
> >         fprintf(f, "outer_l2_len=%u, ", m->outer_l2_len);
> > if (m->outer_l3_len != 0)
> >         fprintf(f, "outer_l3_len=%u, ", m->outer_l3_len);
> > fprintf(f, "l2_len=%u, l3_len=%u, l4_len=%u, tso_segsz=%u\n",
> >         m->l2_len, m->l3_len, m->l4_len, m->tso_segsz);
> 
> Mm, why specifically those two only?
> If the intention is to make this debug output shorter, I would apply
> the same to all those fields, and omit the whole (new) line if no
> field is set.

The intention is not do make the output shorter, but to reduce the amount of "noise" (useless data) in the output.
My suggestion omits noise for non-tunneled packets.

Maybe tso_segsz could be conditional on non-zero too.

It depends on what the user/developer is debugging, whether l2/l3/l4_len are considered noise. I don't object to making them conditional too, if you prefer.

Anyway, I just realized that the tx_offload field is not cleared on RX, so the discussed zero-check is useless (checking uninitialized data) when dumping an mbuf at an early stage in the application's packet processing pipeline.
I have an idea to get around this... The TX flags in ol_flags are cleared at RX, so we can consider tx_offload uninitialized until one of the TX flags in ol_flags has been set.