[dpdk-dev] [PATCH 0/3] generic channel for multi-process communication

[Tan, Jianfeng]

Hi Anatoly,

> -----Original Message-----
> From: Burakov, Anatoly
> Sent: Monday, December 11, 2017 6:00 PM
> To: Tan, Jianfeng; dev at dpdk.org
> Cc: Richardson, Bruce; Ananyev, Konstantin; thomas at monjalon.net
> Subject: Re: [PATCH 0/3] generic channel for multi-process communication
> 
> On 30-Nov-17 6:44 PM, Jianfeng Tan wrote:
> > This patchset adds a generic channel for multi-process (primary/secondary)
> > communication.
> >
> > Patch 1: addess the purpose and howto;
> > Patch 2: add a syncrhonous way for those messages which need a response
> immediately.
> > Patch 3: Rework vfio to use this generic communication channel.
> >
> 
> Hi Jianfeng,
> 
> Just a general comment: I am assuming this has the limitation of
> "everything happens through primary process's involvement". This will
> work for VFIO, as secondary always needs to ask the primary before doing
> anything, but it doesn't address other issues that could have been
> addressed with IPC.
> 
> For example, if a primary process would've hotplugged a device, it can't
> notify all secondary processes about this; rather, it has to wait until
> secondary processes ask for this info.

No need to wait the secondary to pull such info.

It can work like this:
(1) Register a hotplug callback for each process at initialization;
(2) Whenever a process wants to hotplug a device, it will broadcast the info, by broadcast, I mean:
  - if plugin happens at the primary, the primary will tell all of the secondary processes;
  - if plugin happens at one secondary, it will firstly tell the primary, and the primary will broadcast it to all the secondary process.

> Neither can it do anything if
> secondary requests a primary to do something, and notify other secondary
> processes about it (i.e. if secondary wants to hotplug a device, but
> there are other secondaries also running). It would be great to have a
> standard way of doing things like this in future revisions of our IPC.

Please review above thing; If you are OK with that, I'll include this in the next version.

Thanks,
Jianfeng

> 
> --
> Thanks,
> Anatoly