[DPDK Bug 1369] net/mlx5: RX packet metadata altered/incorrect after reception

[DPDK Bug 1369] net/mlx5: RX packet metadata altered/incorrect after reception

[bugzilla]

[-- Attachment #1: Type: text/plain, Size: 2539 bytes --]

https://bugs.dpdk.org/show_bug.cgi?id=1369

            Bug ID: 1369
           Summary: net/mlx5: RX packet metadata altered/incorrect after
                    reception
           Product: DPDK
           Version: 22.11
          Hardware: x86
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: Normal
         Component: ethdev
          Assignee: dev@dpdk.org
          Reporter: pt4hwmy@clavister.com
  Target Milestone: ---

Our platform operates with Mellanox ConnectX-5 (MT_0000000183) network
adapters, equipped with the latest firmware (version 16.35.3006). It utilizes
DPDK 22.11.3 and runs on Ubuntu 20.

Under heavy load, we occasionally observe the reception of packets where the
metadata is incorrect or altered at a later time. The incorrect data seem to
revolve around the "port" and "RSS" fields.

For packet reception, we use: 
rte_eth_rx_burst(port ..)

To detect the problem, we loop through the packets that have been received and
check if the "port" specified in the "mbuf" aligns with the interface
designated for the RX burst execution.

Occasionally, they will not match. Sometimes, the port appears to match
initially, only to change later during processing.

Attempts to modify the packet metadata after reception, such as changing the
port, result in its reverting to its previous state shortly after. 

The packet data for these anomalies seems accurate. For instance, if RX is
performed on port 1 and the packet mbuf indicates port 2, the IP data within
the packet still correlates with the IP ranges assigned to port 1.

Initially, we suspected a potential double free of packet buffers but found no
evidence supporting this. Also, our implementation functions correctly with
various other network adapters and drivers.

This implementation also worked with an older version of DPDK. We previously
used DPDK 20.02 without issue, but the problem emerged after upgrading to DPDK
22.11. Upgrading further to DPDK 23.11 did not resolve the issue.

We have tried different Mellanox firmware versions and even switched to
ConnectX-6 adapters, but the problem persisted. 

The only effective workaround so far has been to disable "hardware RX vector"
(rx_vec_en=0), which seem to make the problem go away.

Do you have any insights into what might be causing this problem, or
suggestions on how to proceed in resolving it?

-- 
You are receiving this mail because:
You are the assignee for the bug.

[-- Attachment #2: Type: text/html, Size: 4458 bytes --]