I believe there are defects in the code supporting manual configuration of
port link speed and link duplex to values other than auto-negotiate. The
TESTPMD application from DPDK version 1.8.0 was executed on 2 different
systems having 4x1G NICs with the following Ethernet controllers:
* Intel Corporation 82576 Gigabit Network Connection (rev 01)
* Intel Corporation I350 Gigabit Network Connection (rev 01)
There appears to be two issues in the code:
* "hw->mac.autoneg? is always set to true (1). The force speed and duplex
code path is not followed.
* "hw->mac.forced_speed_duplex" is never set. A forced link speed
configuration will always default to 10Mb regardless of whether configured
for 100Mb or 10Mb. For example, e1000_phy_force_speed_duplex_setup() will
always configure link speed to 10mb since it checks for the following
condition "if (mac->forced_speed_duplex & E1000_ALL_100_SPEED)?.
Changes are needed within ?igb_ethdev.c? and ?em_ethdev.c? within
?lib/librte_pmd_e1000?. The switch statements that setup link speed and
link duplex within these files need to manually set "hw->mac.autoneg" and
"hw->mac.forced_speed_duplex" as shown below:
[root at box librte_pmd_e1000]# pwd
/home/marc/dpdk/dpdk-1.8.0/lib/librte_pmd_e1000
[root at box librte_pmd_e1000]# diff -p igb_ethdev.c.orig igb_ethdev.c
*** igb_ethdev.c.orig 2015-01-08 09:59:52.937215791 -0500
--- igb_ethdev.c 2015-01-08 10:01:44.073730592 -0500
*************** eth_igb_start(struct rte_eth_dev *dev)
*** 871,876 ****
--- 871,878 ----
hw->phy.autoneg_advertised = ADVERTISE_10_FULL;
else
goto error_invalid_config;
+ hw->mac.autoneg = 0;
+ hw->mac.forced_speed_duplex |= hw->phy.autoneg_advertised;
break;
case ETH_LINK_SPEED_100:
if (dev->data->dev_conf.link_duplex == ETH_LINK_AUTONEG_DUPLEX)
*************** eth_igb_start(struct rte_eth_dev *dev)
*** 881,886 ****
--- 883,890 ----
hw->phy.autoneg_advertised = ADVERTISE_100_FULL;
else
goto error_invalid_config;
+ hw->mac.autoneg = 0;
+ hw->mac.forced_speed_duplex |= hw->phy.autoneg_advertised;
break;
case ETH_LINK_SPEED_1000:
if ((dev->data->dev_conf.link_duplex ==
ETH_LINK_AUTONEG_DUPLEX) ||
*************** eth_igb_start(struct rte_eth_dev *dev)
*** 888,893 ****
--- 892,899 ----
hw->phy.autoneg_advertised = ADVERTISE_1000_FULL;
else
goto error_invalid_config;
+ hw->mac.autoneg = 0;
+ hw->mac.forced_speed_duplex |= hw->phy.autoneg_advertised;
break;
case ETH_LINK_SPEED_10000:
default:
[root at box librte_pmd_e1000]#
After only setting hw->mac.autoneg = 0 in the switch statement cases
described above I experimented with configuring the peer (switch) ports
for 100Mb full duplex. Within TESTPMD the ports were also configured for
100Mb full duplex using ?port config all speed 100 duplex full?. The
links failed to establish. I enabled debug statements within
?e1000_phy_force_speed_duplex_setup()? found in
?lib/librte_pmd_e1000/e1000/e1000_phy.c? to display whether 100mb or 10mb
was being forced. It was observed that 10Mb was being forced when the
link speed was manually configured for 100Mb. Setting
?hw->mac.force_speed_duplex? as shown above seemed to resolve this issue
and the links came up.
Are these known issues?
-Marc
