Re: [dpdk-dev] Issues with openvswitch2.5+dpdk2.2+kvm/virtio-pci

[Christian Ehrhardt <christian.ehrhardt@canonical.com>]

On Tue, Mar 22, 2016 at 9:47 PM, Daniele Di Proietto <diproiettod@vmware.com
> wrote:

> Hi Christian,
>
> thanks for the report.  I've managed to reproduce the problem and I
> observed two separate issues:
>
> 1) short version: it appears to be a problem in DPDK 2.2 and it should be
> fixed by 9a0615af7746("virtio: fix restart"), now on master.
>
[...]

> It appears that simply backporting 9a0615af7746("virtio: fix restart")
> fixes the issue.
>

Yeah debugging it I came to the double configuration as cause as well
yesterday.
Eventually I saw that virtqueue->vq_ring structs got reallocated but stayed
zeroed out by the second call.

I didn't realize there was a fix for it already - thanks so much for
pointing it out.
I'll give a backport a try and let you know if it worked.



> 2) When ovs-vswitchd is started with --monitor, there will be a parent
> process (the monitor) that simply restarts the child if it crashes, while
> the child does the actual ovs-vswitchd job.
>
> It appears that the monitor feature is broken with DPDK, because the DPDK
> library is wrongly initialized in the parent process.  If the child
> crashes, all the DPDK memzones are preserved, meaning that new allocations
> will likely fail.
>
> The fix for this is calling rte_eal_init() in the child process, after
> parsing other ovs-vswitchd command line options. This is done (among other
> things) in Aaron's series currently under review:
>
> http://openvswitch.org/pipermail/dev/2016-March/067422.html
>
> I think we need a separate fix for branch-2.5.
>

Yeah I was on discussing and testing that series already as the series was
also related to a socket ownership/permission issue (back in time).
http://openvswitch.org/pipermail/dev/2015-December/063567.html
I think eventually we need separate fixes for both on the branch-2.5

[...]