On 12/11/2017 12:33 PM, Shahaf Shuler wrote:
Hi Radu,
Monday, December 11, 2017 1:48 PM, Radu Nicolau :
Hi,
Comment inline
On 11/23/2017 12:19 PM, Shahaf Shuler wrote:
Ethdev offloads API has changed since:
commit ce17eddefc20 ("ethdev: introduce Rx queue offloads API") commit
cba7f53b717d ("ethdev: introduce Tx queue offloads API")
This commit support the new API.
Signed-off-by: Shahaf Shuler <shah...@mellanox.com>
---
examples/ipsec-secgw/ipsec-secgw.c | 27
+++++++++++++++++++++++++--
1 file changed, 25 insertions(+), 2 deletions(-)
diff --git a/examples/ipsec-secgw/ipsec-secgw.c
b/examples/ipsec-secgw/ipsec-secgw.c
index c98454a90..6e538a1ab 100644
--- a/examples/ipsec-secgw/ipsec-secgw.c
+++ b/examples/ipsec-secgw/ipsec-secgw.c
@@ -217,6 +217,8 @@ static struct rte_eth_conf port_conf = {
},
.txmode = {
.mq_mode = ETH_MQ_TX_NONE,
+ .offloads = (DEV_TX_OFFLOAD_IPV4_CKSUM |
+ DEV_TX_OFFLOAD_MULTI_SEGS),
},
};
@@ -1394,6 +1396,22 @@ port_init(uint16_t portid)
if (dev_info.tx_offload_capa & DEV_TX_OFFLOAD_SECURITY)
port_conf.txmode.offloads |= DEV_TX_OFFLOAD_SECURITY;
+ if ((dev_info.rx_offload_capa & port_conf.rxmode.offloads) !=
+ port_conf.rxmode.offloads) {
+ printf("Some Rx offloads are not supported "
+ "by port %d: requested 0x%lx supported 0x%lx\n",
+ portid, port_conf.rxmode.offloads,
+ dev_info.rx_offload_capa);
+ port_conf.rxmode.offloads &= dev_info.rx_offload_capa;
+ }
+ if ((dev_info.tx_offload_capa & port_conf.txmode.offloads) !=
+ port_conf.txmode.offloads) {
+ printf("Some Tx offloads are not supported "
+ "by port %d: requested 0x%lx supported 0x%lx\n",
+ portid, port_conf.txmode.offloads,
+ dev_info.tx_offload_capa);
+ port_conf.txmode.offloads &= dev_info.tx_offload_capa;
+ }
I don't think that clearing the offload flags that are not advertised in the
capabilities is a good approach, although it may be the right one.
From what I can see there are more PMDs that don't fully populate the
offload capabilities, but actually check for them in the configure/start
function. One of them is ixgbe, which needs CRC strip enabled when IPSec is
enabled, and will fail to start otherwise. So although it supports CRC strip it
does not set the flag in the capabilities, but checks it in the start function.
Why ixgbe don't expose the CRC cap then? It seems wrong behavior to expect the
application to set it without any cap reported.
It is bad behavior but from what I can see most, if not all, PMDs don't
expose CRC strip (or jumbo frames) while still supporting it.
I would propose to just print a warning if a requested offload is not set in the
capabilities, but let the pmd start fail if it is not really supported.
I think I agree, however not from the reason you mentioned.
It is bad to mask the un-supported offloads because the application relies on
them to be set successfully. The application will not run successfully if the
IPV4 checksum is not actually set (for example).
On v2 I will print just the warn.
ret = rte_eth_dev_configure(portid, nb_rx_queue, nb_tx_queue,
&port_conf);
if (ret < 0)
@@ -1420,7 +1438,8 @@ port_init(uint16_t portid)
printf("Setup txq=%u,%d,%d\n", lcore_id, tx_queueid,
socket_id);
txconf = &dev_info.default_txconf;
- txconf->txq_flags = 0;
+ txconf->txq_flags = ETH_TXQ_FLAGS_IGNORE;
+ txconf->offloads = port_conf.txmode.offloads;
ret = rte_eth_tx_queue_setup(portid, tx_queueid, nb_txd,
socket_id, txconf);
@@ -1434,6 +1453,8 @@ port_init(uint16_t portid)
/* init RX queues */
for (queue = 0; queue < qconf->nb_rx_queue; ++queue) {
+ struct rte_eth_rxconf rxq_conf;
+
if (portid != qconf->rx_queue_list[queue].port_id)
continue;
@@ -1442,8 +1463,10 @@ port_init(uint16_t portid)
printf("Setup rxq=%d,%d,%d\n", portid, rx_queueid,
socket_id);
+ rxq_conf = dev_info.default_rxconf;
+ rxq_conf.offloads = port_conf.rxmode.offloads;
ret = rte_eth_rx_queue_setup(portid, rx_queueid,
- nb_rxd, socket_id, NULL,
+ nb_rxd, socket_id,
&rxq_conf,
socket_ctx[socket_id].mbuf_pool);
if (ret < 0)
rte_exit(EXIT_FAILURE,
