[dpdk-dev] [PATCH] net/sfc: add 8000-series EF10 hardware TSO bug workaround

[Ivan Malov <ivan.malov@oktetlabs.ru>]

Innermost IP length and outer UDP datagram length must be
greater than or equal to the corresponding values derived
from the MSS; otherwise, the checksum offloads will break.

Fixes: c1ce2ba218f8 ("net/sfc: support tunnel TSO on EF10 native Tx datapath")
Fixes: 6bc985e41155 ("net/sfc: support TSO in EF10 Tx datapath")
Fixes: fec33d5bb3eb ("net/sfc: support firmware-assisted TSO")
Cc: stable@dpdk.org

Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Reviewed-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Reviewed-by: Andy Moreton <amoreton@xilinx.com>
---
 drivers/net/sfc/sfc_ef10_tx.c | 19 +++++++++++++++++++
 drivers/net/sfc/sfc_tso.c     |  7 +++++++
 drivers/net/sfc/sfc_tso.h     | 30 ++++++++++++++++++++++++++++++
 3 files changed, 56 insertions(+)

diff --git a/drivers/net/sfc/sfc_ef10_tx.c b/drivers/net/sfc/sfc_ef10_tx.c
index 87fa40f3e..33d2d637c 100644
--- a/drivers/net/sfc/sfc_ef10_tx.c
+++ b/drivers/net/sfc/sfc_ef10_tx.c
@@ -481,6 +481,25 @@ sfc_ef10_xmit_tso_pkt(struct sfc_ef10_txq * const txq, struct rte_mbuf *m_seg,
 			needed_desc--;
 	}
 
+	/*
+	 * 8000-series EF10 hardware requires that innermost IP length
+	 * be greater than or equal to the value which each segment is
+	 * supposed to have; otherwise, TCP checksum will be incorrect.
+	 *
+	 * The same concern applies to outer UDP datagram length field.
+	 */
+	switch (m_seg->ol_flags & PKT_TX_TUNNEL_MASK) {
+	case PKT_TX_TUNNEL_VXLAN:
+		/* FALLTHROUGH */
+	case PKT_TX_TUNNEL_GENEVE:
+		sfc_tso_outer_udp_fix_len(first_m_seg, hdr_addr);
+		break;
+	default:
+		break;
+	}
+
+	sfc_tso_innermost_ip_fix_len(first_m_seg, hdr_addr, iph_off);
+
 	/*
 	 * Tx prepare has debug-only checks that offload flags are correctly
 	 * filled in in TSO mbuf. Use zero IPID if there is no IPv4 flag.
diff --git a/drivers/net/sfc/sfc_tso.c b/drivers/net/sfc/sfc_tso.c
index d6f111989..b090ef14d 100644
--- a/drivers/net/sfc/sfc_tso.c
+++ b/drivers/net/sfc/sfc_tso.c
@@ -140,6 +140,13 @@ sfc_efx_tso_do(struct sfc_efx_txq *txq, unsigned int idx,
 		tsoh = rte_pktmbuf_mtod(m, uint8_t *);
 	}
 
+	/*
+	 * 8000-series EF10 hardware requires that innermost IP length
+	 * be greater than or equal to the value which each segment is
+	 * supposed to have; otherwise, TCP checksum will be incorrect.
+	 */
+	sfc_tso_innermost_ip_fix_len(m, tsoh, nh_off);
+
 	/*
 	 * Handle IP header. Tx prepare has debug-only checks that offload flags
 	 * are correctly filled in in TSO mbuf. Use zero IPID if there is no
diff --git a/drivers/net/sfc/sfc_tso.h b/drivers/net/sfc/sfc_tso.h
index 8597c2868..361aa2219 100644
--- a/drivers/net/sfc/sfc_tso.h
+++ b/drivers/net/sfc/sfc_tso.h
@@ -38,6 +38,36 @@ sfc_tso_ip4_get_ipid(const uint8_t *pkt_hdrp, size_t ip_hdr_off)
 	return rte_be_to_cpu_16(ipid);
 }
 
+static inline void
+sfc_tso_outer_udp_fix_len(const struct rte_mbuf *m, uint8_t *tsoh)
+{
+	rte_be16_t len = rte_cpu_to_be_16(m->l2_len + m->l3_len + m->l4_len +
+					  m->tso_segsz);
+
+	rte_memcpy(tsoh + m->outer_l2_len + m->outer_l3_len +
+		   offsetof(struct rte_udp_hdr, dgram_len),
+		   &len, sizeof(len));
+}
+
+static inline void
+sfc_tso_innermost_ip_fix_len(const struct rte_mbuf *m, uint8_t *tsoh,
+			     size_t iph_ofst)
+{
+	size_t ip_payload_len = m->l4_len + m->tso_segsz;
+	size_t field_ofst;
+	rte_be16_t len;
+
+	if (m->ol_flags & PKT_TX_IPV4) {
+		field_ofst = offsetof(struct rte_ipv4_hdr, total_length);
+		len = rte_cpu_to_be_16(m->l3_len + ip_payload_len);
+	} else {
+		field_ofst = offsetof(struct rte_ipv6_hdr, payload_len);
+		len = rte_cpu_to_be_16(ip_payload_len);
+	}
+
+	rte_memcpy(tsoh + iph_ofst + field_ofst, &len, sizeof(len));
+}
+
 unsigned int sfc_tso_prepare_header(uint8_t *tsoh, size_t header_len,
 				    struct rte_mbuf **in_seg, size_t *in_off);
 
-- 
2.20.1