diff mbox series

[v4,01/22] ethdev: fix MTU size exceeds max rx packet length

Message ID 20210118070428.36998-2-stevex.yang@intel.com (mailing list archive)
State Accepted, archived
Delegated to: Ferruh Yigit
Headers
Series fix rx packets dropped issue |

Checks

Context Check Description
ci/checkpatch success coding style OK

Commit Message

Steve Yang Jan. 18, 2021, 7:04 a.m. UTC 
  Ethdev is using default Ethernet overhead to decide if provided
'max_rx_pkt_len' value is bigger than max (non jumbo) MTU value,
and limits it to MAX if it is.

Since the application/driver used Ethernet overhead is different than
the ethdev one, check result is wrong.

If the driver is using Ethernet overhead bigger than the default one,
the provided 'max_rx_pkt_len' is trimmed down, and in the driver when
correct Ethernet overhead is used to convert back, the resulting MTU is
less than the intended one, causing some packets to be dropped.

Like,
app     -> max_rx_pkt_len = 1500/*mtu*/ + 22/*overhead*/ = 1522
ethdev  -> 1522 > 1518/*MAX*/; max_rx_pkt_len = 1518
driver  -> MTU = 1518 - 22 = 1496
Packets with size 1497-1500 are dropped although intention is to be able
to send/receive them.

The fix is to make ethdev use the correct Ethernet overhead for port,
instead of default one.

Fixes: 59d0ecdbf0e1 ("ethdev: MTU accessors")

Cc: Thomas Monjalon <thomas@monjalon.net>
Cc: Ferruh Yigit <ferruh.yigit@intel.com>
Cc: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>

Signed-off-by: Steve Yang <stevex.yang@intel.com>
---
 lib/librte_ethdev/rte_ethdev.c | 25 ++++++++++++++++++++++---
 1 file changed, 22 insertions(+), 3 deletions(-)
diff mbox series

Patch

diff --git a/lib/librte_ethdev/rte_ethdev.c b/lib/librte_ethdev/rte_ethdev.c
index 17ddacc78d..71e1e9a6db 100644
--- a/lib/librte_ethdev/rte_ethdev.c
+++ b/lib/librte_ethdev/rte_ethdev.c
@@ -1292,8 +1292,10 @@  rte_eth_dev_configure(uint16_t port_id, uint16_t nb_rx_q, uint16_t nb_tx_q,
 	struct rte_eth_dev *dev;
 	struct rte_eth_dev_info dev_info;
 	struct rte_eth_conf orig_conf;
+	uint16_t overhead_len;
 	int diag;
 	int ret;
+	uint16_t old_mtu;
 
 	RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
 
@@ -1319,10 +1321,20 @@  rte_eth_dev_configure(uint16_t port_id, uint16_t nb_rx_q, uint16_t nb_tx_q,
 		memcpy(&dev->data->dev_conf, dev_conf,
 		       sizeof(dev->data->dev_conf));
 
+	/* Backup mtu for rollback */
+	old_mtu = dev->data->mtu;
+
 	ret = rte_eth_dev_info_get(port_id, &dev_info);
 	if (ret != 0)
 		goto rollback;
 
+	/* Get the real Ethernet overhead length */
+	if (dev_info.max_mtu != UINT16_MAX &&
+	    dev_info.max_rx_pktlen > dev_info.max_mtu)
+		overhead_len = dev_info.max_rx_pktlen - dev_info.max_mtu;
+	else
+		overhead_len = RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN;
+
 	/* If number of queues specified by application for both Rx and Tx is
 	 * zero, use driver preferred values. This cannot be done individually
 	 * as it is valid for either Tx or Rx (but not both) to be zero.
@@ -1409,12 +1421,17 @@  rte_eth_dev_configure(uint16_t port_id, uint16_t nb_rx_q, uint16_t nb_tx_q,
 			ret = -EINVAL;
 			goto rollback;
 		}
+
+		/* Scale the MTU size to adapt max_rx_pkt_len */
+		dev->data->mtu = dev->data->dev_conf.rxmode.max_rx_pkt_len -
+				overhead_len;
 	} else {
-		if (dev_conf->rxmode.max_rx_pkt_len < RTE_ETHER_MIN_LEN ||
-			dev_conf->rxmode.max_rx_pkt_len > RTE_ETHER_MAX_LEN)
+		uint16_t pktlen = dev_conf->rxmode.max_rx_pkt_len;
+		if (pktlen < RTE_ETHER_MIN_MTU + overhead_len ||
+		    pktlen > RTE_ETHER_MTU + overhead_len)
 			/* Use default value */
 			dev->data->dev_conf.rxmode.max_rx_pkt_len =
-							RTE_ETHER_MAX_LEN;
+						RTE_ETHER_MTU + overhead_len;
 	}
 
 	/*
@@ -1549,6 +1566,8 @@  rte_eth_dev_configure(uint16_t port_id, uint16_t nb_rx_q, uint16_t nb_tx_q,
 	eth_dev_tx_queue_config(dev, 0);
 rollback:
 	memcpy(&dev->data->dev_conf, &orig_conf, sizeof(dev->data->dev_conf));
+	if (old_mtu != dev->data->mtu)
+		dev->data->mtu = old_mtu;
 
 	rte_ethdev_trace_configure(port_id, nb_rx_q, nb_tx_q, dev_conf, ret);
 	return ret;