Bug 1366 - mlx5 secondary testpmd process crash when calling rte_eth_stats_get
Summary: mlx5 secondary testpmd process crash when calling rte_eth_stats_get
Status: UNCONFIRMED
Alias: None
Product: DPDK
Classification: Unclassified
Component: other (show other bugs)
Version: 22.11
Hardware: x86 Linux
: Normal minor
Target Milestone: ---
Assignee: dev
URL:
Depends on:
Blocks:
 
Reported: 2024-01-15 15:24 CET by Aidan van Wyk
Modified: 2024-01-16 17:43 CET (History)
1 user (show)



Attachments
Secondary lldb process output (3.87 KB, text/plain)
2024-01-15 15:24 CET, Aidan van Wyk
Details

Description Aidan van Wyk 2024-01-15 15:24:53 CET
Created attachment 271 [details]
Secondary lldb process output

Linux:
AlmaLinux 9.3

Kernel:
5.14.0-362.8.1.el9_3.x86_64

Software version:
rdma-core-48.0-1.el9.x86_64
dpdk-v22.11(tested on git branch v23.11-rc4 as well)

Hardware:
Mellanox Technologies MT27800 Family [ConnectX-5]

Running two testpmd processes at the same time causes the secondary process to crash with a segmentation fault from the libmlx5.so library. Two ports on a Mellanox MT27800 card are configured to be used by both testpmd apps.
The two separate instances are launched:
1. dpdk-testpmd --proc-type=primary
2. dpdk-testpmd --proc-type=secondary

I am not sure if this is a rdma-core issue which contains the libmlx5.so library where segmentation fault is thrown from. The primary dpdk process runs without any issues and there are no segfaults. The rte_eth_stats_get function only throws an error on the secondary process, not on the primary process. 
Attached is the lldb output when running a debug build secondary process with lldb.

Running dmesg also shows that there is a segfault:
dpdk-testpmd[67695]: segfault at b077f90 ip 00007f1ae42069e4 sp 00007ffcd49903b8 error 4 in libmlx5.so.1.24.48.0[7f1ae41b5000+5a000] likely on CPU 0 (core 0, socket 0)
Code: 77 90 0f b6 d2 48 8d 05 7a a3 00 00 44 0f be 24 10 e9 7c ff ff ff e8 fb f3 fa ff 66 66 2e 0f 1f 84 00 00 00 00 00 f3 0f 1e fa <48> 
8b 07 49 89 d2 4c 8d 0d 4f 88 01 00 48 8b 10 48 8b 92 98 02 00

Note You need to log in before you can comment on or make changes to this bug.