[ovs-dev] Why not implement netdev datapath in a separated process

[Aaron Conole]

Huanle Han <hanxueluo at gmail.com> writes:

> Hi,

Hi Huanle,

> I'm thinking about this quesion, after experience the dpdk feature in ovs.
> Could you give me some answer?

I'll take a stab at it.

> Currently, netdev datapath(or called dpdk) is implemented inside daemon
> process "ovs-vswitchd".
> It makes ovs-vswitchd not that stable as before:
> 1. some ovsbd config for dpdk requires restart ovs-vswitchd.

That's a limitation of the library, and one that we want to fix.
Note also that Daniele has proposed a change to allow -enabling- dpdk
after startup.  I'm currently reviewing that series.

> 2. sometimes dpdk fails to init and aborts process because of inappropriate
> configuration.

Yep.  There's ongoing work in dpdk upstream to rectify this situation,
though (I think?)

> 3. dpdk lib crash bug.

I don't think bugs aren an argument in this case.  We find them and fix
them, that's that.  Otherwise, bubble-sort might be implemented in it's
own process ;)

> These suitation seriously effect the forwarding of other datapath(system).
>
> Is it possibale that netdev datapath is implemented in a new process, and
> connects witch ovs-vswitchd using IPC?
> If possibale, what the main difficulty might be?

If you think it's a worthwhile effort, go ahead and submit patches.
It's easy to dismiss an idea out-of-hand if it's just an idea.
If you have code that can provide this feature, and submit it, it will
be looked at.

I haven't thought of any such benefits.  I could be wrong.

-Aaron