Host OpenVSwitch+DPDK process can segfault from a VM running DPDK-enabled application

[Mikhail Chernik]

Description:

Due to a bug in DPDK prior to 16.04, a dpdk-enabled process running in VM can crash host dpdk-enabled OpenVSwitch process. Detailed description can be found here: http://www.dpdk.org/ml/archives/dev/2016-May/039422.html

Steps to reproduce:

1. Enable experimental features on MOS 9.0 Fuel Master
2. Deploy a compute node with configured DPDK
3. Create a VM on the compute node
4. Download and compile DPDK ver <= 2.2 on the VM
5. Follow instructions from link above

Expected result:
Host OVS process is not affected

Actual result:
Host OVS crashes with segfault after second port initialization

OVS+DPDK pmd segfault on host also could be reproduces next way:

1. Enable experimental features on MOS 9.0 Fuel Master
2. Deploy a compute node with configured DPDK
3. Create a VM on the compute node. Add one extra interface.
4. Download and compile DPDK
5. Download DPDK-based application. 100% reproducebility achieved with pktgen-dpdk. http://dpdk.org/browse/apps/pktgen-dpdk
6. Start DPDk-based application with extra interface attached to DPDK on guest.
7. Enable high workload traffic on the interface that attached to DPDK on guest.
8. Kill DPDK-based application
9. Start DPDK-based application immediately.

Expected result:
Host OVS process is not affected

Actual result:
Host OVS crashes with segfault after DPDK-based app init.

[Sergey Matov (smatov)]

Described bug is partly fixed by
https://review.fuel-infra.org/#/c/25699/
Re-init of testpmd ports does not cause segfault anymore.

For the situation where DPDK-based application on guest stops incorrectly(e.g. killing pid) and re-starts after we are awaiting fix in new versions of DPDK. To avoid segfault of host vswitchd process segfault after re-starting guest DPDK-based app HugePages for guest application should be freed(or re-mounted).

Appropriate information will be added to the release notes.

[Sergey Matov (smatov)]

Release notes

 - Referring to bug https://bugs.launchpad.net/fuel/+bug/1615012 for DPDK-accelerated compute node if DPDK-based application on guest stops without freeing HugePages it is recommended to re-mount HugePages on guest VM to avoid OVS-DPDK segmentation fault on compute node.