[dpdk-dev] secondary processes and private data

[Alejandro Lucero]

On Wed, Sep 26, 2018 at 3:33 PM Stephen Hemminger <
stephen at networkplumber.org> wrote:

> On Wed, 26 Sep 2018 15:21:52 +0200
> Thomas Monjalon <thomas at monjalon.net> wrote:
>
> > Hi Alejandro,
> >
> > 25/09/2018 16:10, Alejandro Lucero:
> > > I've a problem when part of device private data needs to be private per
> > > process.
> >
> > It appears we are facing the same issue to support multi-process in tap.
> >
> > > Current multiprocess support shares device private data between
> primary and
> > > secondaries but it is all dependent on a pointer initialized to the
> same
> > > memory address by the multiprocess support code. If there is a
> per-process
> > > data, If a secondary process changes it the primary gets affected, and
> the
> > > same for additional secondaries which will affect not just the primary
> but
> > > other previous secondaries.
> >
> > Yes, the field rte_eth_dev.data.dev_private is private to the device,
> > but shared between processes.
> >
> > > The solution is to add support for this inside struct rte_eth_dev,
> > > something like
> > >
> > > void *secondary_priv_data;
> > >
> > > so it is up to the secondaries to use this field if necessary.
> >
> > I would say it is not only for secondary process.
> > What about this name:
> >
> >       rte_eth_dev.process_private
> >
> > > NFP PMD creates the required rte_eth_devs specifically, similar to
> what is
> > > done inside rte_ethdev.c but adding initialization for an interface
> needed
> > > when calling device ethdev_init function. There are other PMDs doing
> this
> > > but none has this requirement for per-process private data.
> >
> > Actually tap has a per-process requirement for its file descriptors.
> >
> > > Please, let me know what you think about this change to struct
> rte_ethdev
> > > or if you have a better idea for solving this problem.
> >
> > I support the idea, but we need to agree on name bikeshedding :-)
>
> Good idea, as long as it stays contained to DPDK. Don't want additional
> user API
> pointers buried in internal structures (like ethdev).  If application
> needs device
> private data it should manage its own state.
>
>
AFAIK the per-device data is not used by apps but only PMDs, and the new
field should be the same.