Re: [dpdk-users] Mellanox Unexpected CQE error syndrome

["Krauz, Pavel" <Pavel.Krauz@anritsu.com>]

Hi Slava,

Your suggestion made me realize that we do not set the flag for jumbo frames correctly. After setting the DEV_RX_OFFLOAD_JUMBO_FRAME the problem disappears.

However when we set the
port_conf.rxmode.max_rx_pkt_len = RTE_ETHER_MAX_LEN;

and do not enable jumbo frames offload, then we see the Mellanox debug files again.
(mbuf packet pool is configured with data_size = RTE_MBUF_DEFAULT_DATAROOM + RTE_PKTMBUF_HEADROOM)

It seems like if there are jumbo frames in the monitored network but not enabled in the DPDK then the card goes to error state.
 
> It would be nice to create an issue in DPDK Bugzilla and have the dedicated thread to handle.

Yes, will do it for the case of jumbo frames not enabled in DPDK and pkt_len set low.

> What DPDK version do you use?

We are using dpdk 19.11.4

b.r.
Pavel



-----Original Message-----
From: Slava Ovsiienko <viacheslavo@nvidia.com> 
Sent: pátek 11. prosince 2020 11:01
To: users@dpdk.org; Krauz, Pavel <Pavel.Krauz@anritsu.com>
Subject: RE: Mellanox Unexpected CQE error syndrome

>-----Original Message-----
>From: users <users-bounces@dpdk.org> On Behalf Of Krauz, Pavel
>Sent: Thursday, October 29, 2020 10:18 AM
>To: users@dpdk.org
>Subject: [dpdk-users] Mellanox Unexpected CQE error syndrome
>
>Hello,
>I am having problem with HPE Ethernet 100Gb 2-port 841QSFP28 Adapter 
>which is a Mellanox adapter for 100G network.
>
>The DPDK driver reports and generates lot of error files like
>dpdk_mlx5_port_0_rxq_0_2459159054 and loses traffic (because>IMHO it 
>must reset the card):
>
>the first line of the error report files is as follows:
>
>Unexpected CQE error syndrome 0x22 CQN = 1030 RQN = 12582977 
>wqe_counter = 10040 rq_ci = 494774062 cq_ci = 3586794130 Unexpected CQE 
>error syndrome 0x22 CQN = 1030 RQN = 12582977 wqe_counter = 27509 
>rq_ci= 808774458 cq_ci = 1527072213 Unexpected CQE error syndrome 0x0e 
>CQN = 1030 RQN = 12582977 wqe_counter = 0 rq_ci = 32768 cq_ci = 
>2413356687
>
..snip..
>
>I have tried latest card HP firmware and enable/disable CQE compression 
>in the mlx5 DPDK driver using rxq_cqe_comp_en=0/1, but no
improvement.
>
>Does anybody know what can be the problem and how to mitigate it?
>
>Thanks
>Pavel Krauz

Hi, Pavel.

Sorry, I missed this mail from users@ mailing list.
It would be nice to create an issue in DPDK Bugzilla and have the dedicated thread to handle.

What DPDK version do you use? Syndromes in this report are different ones, at my first glance it looks like as not correct MTU or max packet length is configured. The NIC receives the packet with the length exceeding for what queue was configured and packet data overwrite the descriptors, resulting into syndromes.

With best regards,
Slava