net/mlx5: wrong Rx/Tx descriptor limits when DevX is off

[Edwin Brossette <edwin.brossette@6wind.com>]

[-- Attachment #1: Type: text/plain, Size: 4127 bytes --]

Hello,

I have run into a regression following an update to stable dpdk-24.11 with
a number of my Mellanox cx4/5/6 nics. This regression occurs with all nics
in my lab which have DevX disabled: using mstconfig utility, I can see the
flag UCTX_EN is not set.

Mainly, the issue is that the ports cannot be started, with the following
error logs in the journal:

Set nb_rxd=1 (asked=512) for port=0
Set nb_txd=1 (asked=512) for port=0
starting port 0
Initializing port 0 [7c:fe:90:65:e6:54]
port 0: ntfp1 (mlx5_pci)
nb_rxq=2 nb_txq=2
rxq0=c9 rxq1=c25
txq0=c9 txq1=c25
port 0: rx_scatter=0 tx_scatter=0 max_rx_frame=1526
mlx5_net: port 0 number of descriptors requested for Tx queue 0 must be
higher than MLX5_TX_COMP_THRESH, using 33 instead of 1
mlx5_net: port 0 increased number of descriptors in Tx queue 0 to the next
power of two (64)
mlx5_net: port 0 number of descriptors requested for Tx queue 1 must be
higher than MLX5_TX_COMP_THRESH, using 33 instead of 1
mlx5_net: port 0 increased number of descriptors in Tx queue 1 to the next
power of two (64)
mlx5_net: Port 0 Rx queue 0 CQ creation failure.
mlx5_net: port 0 Rx queue allocation failed: Cannot allocate memory
rte_eth_dev_start(port 0) failed, error=-12
Failed to start port 0, set link down
Failed to start port 0

Looking more precisely into the problem, it appears that the number of Rx
and Tx descriptors configured for my queues is 1. This happens because
mlx5_dev_infos_get() return a limit of 1 for both Rx and Tx, which is
unexpected. I identified this patch to be responsible for the regression:

4c3d7961d9002: net/mlx5: fix reported Rx/Tx descriptor limits
https://git.dpdk.org/dpdk/commit/?id=4c3d7961d9002bb715a8ee76bcf464d633316d4c

After doing some debugging, I noticed that hca_attr.log_max_wq_sz is never
configured. This should be done in mlx5_devx_cmd_query_hca_attr() which is
called in this bit of code:

https://git.dpdk.org/dpdk/tree/drivers/common/mlx5/mlx5_common.c#n681

/*
* When CTX is created by Verbs, query HCA attribute is unsupported.
* When CTX is imported, we cannot know if it is created by DevX or
* Verbs. So, we use query HCA attribute function to check it.
*/
if (cdev->config.devx || cdev->config.device_fd != MLX5_ARG_UNSET) {
/* Query HCA attributes. */
ret = mlx5_devx_cmd_query_hca_attr(cdev->ctx, &cdev->config.hca_attr);
if (ret) {
DRV_LOG(ERR, "Unable to read HCA caps in DevX mode.");
rte_errno = ENOTSUP;
goto error;
}
cdev->config.devx = 1;
}
DRV_LOG(DEBUG, "DevX is %ssupported.", cdev->config.devx ? "" : "NOT ");

I deduced that following the above patch, the correct value for maximum Rx
and Tx descriptors will only be set if DevX is enabled (see the if
condition on cdev->config.devx). If it is disabled, then maximum Rx and Tx
descriptors will be 1, which will make the ports fail to start. Perhaps we
should keep the previous default value (65535) if config.devx == 0 (DevX
off)? This could be done like this, for example:

diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
index 7708a0b80883..8ba3eb4a32de 100644
--- a/drivers/net/mlx5/mlx5_ethdev.c
+++ b/drivers/net/mlx5/mlx5_ethdev.c
@@ -359,10 +359,12 @@ mlx5_dev_infos_get(struct rte_eth_dev *dev, struct
rte_eth_dev_info *info)
        info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
        mlx5_set_default_params(dev, info);
        mlx5_set_txlimit_params(dev, info);
-       info->rx_desc_lim.nb_max =
-               1 << priv->sh->cdev->config.hca_attr.log_max_wq_sz;
-       info->tx_desc_lim.nb_max =
-               1 << priv->sh->cdev->config.hca_attr.log_max_wq_sz;
+       if (priv->sh->cdev->config.devx) {
+               info->rx_desc_lim.nb_max =
+                       1 << priv->sh->cdev->config.hca_attr.log_max_wq_sz;
+               info->tx_desc_lim.nb_max =
+                       1 << priv->sh->cdev->config.hca_attr.log_max_wq_sz;
+       }
        if (priv->sh->cdev->config.hca_attr.mem_rq_rmp &&
            priv->obj_ops.rxq_obj_new == devx_obj_ops.rxq_obj_new)
                info->dev_capa |= RTE_ETH_DEV_CAPA_RXQ_SHARE;

Thanks in advance for your help.

Regards,
Edwin Brossette.

[-- Attachment #2: Type: text/html, Size: 5377 bytes --]