On 29/04/2025 13:00, Daniel Ding wrote:
> The broadcom nic needs allow extended mbufs for rx-ag ring.
>
> Output warn log like:
> dpdk|WARN|bnxt_init_one_rx_ring(): init'ed AG ring 8 with 512/8192 mbufs
> only
> dpdk|WARN|bnxt_init_one_rx_ring(): init'ed rx ring 9 with 0/2048 mbufs only
> ---
> lib/netdev-dpdk.c | 22 +++++++++++++++++++---
> 1 file changed, 19 insertions(+), 3 deletions(-)
>
+cc: Ajit (DPDK bnxt maintainer)
H Daniel,
It seems like you are using an EoL DPDK/OVS. i.e.
bnxt_init_one_rx_ring(): init'ed AG ring 8 with 512/8192 mbufs only
bnxt_init_one_rx_ring(): init'ed rx ring 9 with 0/2048 mbufs only
these logs were replaced in:
commit f9f0b5121f62d8701165d20bfdd14638bc72654d
Author: Ajit Khaparde <ajit.khapa...@broadcom.com>
Date: Wed Nov 17 13:10:09 2021 -0800
net/bnxt: fail init when mbuf allocation fails
...
and that commit is merged since DPDK 21.11:
$ git tag --contains f9f0b5121f6 | head -1
v21.11
The oldest maintained DPDK LTS is 22.11, but the oldest DPDK LTS
actually used in a maintained OVS branch is DPDK 23.11 (for OVS branch-3.3).
Even if it still exists for the current code, I don't see an issue with
the OVS code, so it looks like it should be fixed in the DPDK driver.
In general, we don't want to take code in OVS for specific NICs as they
should work through the DPDK API.
Also, in this case, it is impacting OVS user functionality by changing
the amount of requested memory and overriding some of the user memory
config options. See
https://docs.openvswitch.org/en/latest/topics/dpdk/memory/ .
Is this still an issue with maintained versions ? I suggest to discuss
more with Ajit on DPDK mailing list or submit a bug at https://bugs.dpdk.org
thanks,
Kevin.
> diff --git a/lib/netdev-dpdk.c b/lib/netdev-dpdk.c
> index 29a6bf032..b1b37b75b 100644
> --- a/lib/netdev-dpdk.c
> +++ b/lib/netdev-dpdk.c
> @@ -610,6 +610,15 @@ netdev_dpdk_get_ingress_policer(const struct netdev_dpdk
> *dev);
> static void netdev_dpdk_mbuf_dump(const char *prefix, const char *message,
> const struct rte_mbuf *);
>
> +static bool
> +is_dpdk_driver(struct netdev_dpdk *dev, const char *name)
> +{
> + struct rte_eth_dev_info info;
> +
> + rte_eth_dev_info_get(dev->port_id, &info);
> + return !strcmp(info.driver_name, name);
> +}
> +
> static bool
> is_dpdk_class(const struct netdev_class *class)
> {
> @@ -770,6 +779,10 @@ dpdk_calculate_mbufs(struct netdev_dpdk *dev, int mtu)
> + MIN_NB_MBUF;
> }
>
> + if (is_dpdk_driver(dev, "net_bnxt")) { // Alloc mbufs for bnxt rx-ag
> ring.
> + n_mbufs += dev->requested_n_rxq * dev->requested_rxq_size * 4;
> + }
> +
> return n_mbufs;
> }
>
> @@ -891,9 +904,10 @@ dpdk_mp_get(struct netdev_dpdk *dev, int mtu)
> bool reuse = false;
>
> ovs_mutex_lock(&dpdk_mp_mutex);
> +
> /* Check if shared memory is being used, if so check existing mempools
> * to see if reuse is possible. */
> - if (!per_port_memory) {
> + if (!per_port_memory && !is_dpdk_driver(dev, "net_bnxt")) {
> /* If user has provided defined mempools, check if one is suitable
> * and get new buffer size.*/
> mtu = dpdk_get_user_adjusted_mtu(mtu, dev->requested_mtu,
> @@ -922,7 +936,8 @@ dpdk_mp_get(struct netdev_dpdk *dev, int mtu)
> * dmp to point to the existing entry and increment the refcount
> * to avoid being freed at a later stage.
> */
> - if (per_port_memory && rte_errno == EEXIST) {
> + if ((per_port_memory || is_dpdk_driver(dev, "net_bnxt"))
> + && rte_errno == EEXIST) {
> LIST_FOR_EACH (next, list_node, &dpdk_mp_list) {
> if (dmp->mp == next->mp) {
> rte_free(dmp);
> @@ -971,7 +986,8 @@ netdev_dpdk_mempool_configure(struct netdev_dpdk *dev)
> /* With shared memory we do not need to configure a mempool if the MTU
> * and socket ID have not changed, the previous configuration is still
> * valid so return 0 */
> - if (!per_port_memory && dev->mtu == dev->requested_mtu
> + if (!per_port_memory && !is_dpdk_driver(dev, "net_bnxt")
> + && dev->mtu == dev->requested_mtu
> && dev->socket_id == dev->requested_socket_id) {
> return ret;
> }
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
