[dpdk-dev] [PATCH v4 08/12] net/failsafe: support offload capabilities

[Olga Shern]

L;w

> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Gaëtan Rivet
> Sent: Thursday, June 01, 2017 5:38 PM
> To: Stephen Hemminger <stephen at networkplumber.org>
> Cc: dev at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v4 08/12] net/failsafe: support offload
> capabilities
> 
> On Wed, May 31, 2017 at 08:23:09AM -0700, Stephen Hemminger wrote:
> > On Mon, 29 May 2017 15:42:20 +0200
> > Gaetan Rivet <gaetan.rivet at 6wind.com> wrote:
> >
> > > Signed-off-by: Gaetan Rivet <gaetan.rivet at 6wind.com>
> > > Acked-by: Olga Shern <olgas at mellanox.com>
> > > ---
> > >  doc/guides/nics/features/failsafe.ini |   6 ++
> > >  drivers/net/failsafe/failsafe_ops.c   | 131
> +++++++++++++++++++++++++++++++++-
> > >  2 files changed, 135 insertions(+), 2 deletions(-)
> >
> > Once again what about case of dumb synthetic NIC combined with SR-IOV
> VF?
> > The VF has offloads the virtual NIC does not.
> >
> 
> The rules for capabilities are a little complicated.
> In the case both VF and the virtual NIC are present at launch, then the logical
> AND is done both their capabilities sets.
> If one has additional capabilities that the user is requesting, and the fail-safe
> recognize them (currently, all RX offloads, as TX offloads were not yet
> expressed by flags), and this capability is not supported by one slave, then
> this offload is disabled in the configuration.
> 
> > What about late plugin. how do you program the offloads of the later
> > arriving VF device.
> 
> If the VF is not present at launch, then the fail-safe reads only the set of
> capabilities from the fallback device. It does not have to do any AND-ing of
> the flags.
> 
> The consequence is that upon plugin of the VF, the latter has to respect the
> current running configuration. Probing will actually fail if some capability is not
> supported (depending on PMDs), and the running configuration is not
> updated as it is considered "live".
> 
> There are only two solutions to this, either:
> 
> * Complicate a lot the fail-safe design and the rules applied in
>   the decision made on NIC configuration. The user then has bad
>   surprises upon seeing that his performance have been degraded
>   for arcane reasons.
> 
> * Emulate in software the offloads and try to advertize as many as
>   possible. This is done for example in the TAP PMD for some flags,
>   allowing those offloads to be used with hardware NICs.
>   The user then has a clear view of the available offloads by comparing
>   both sets of capabilities.
> 
> --
> Gaëtan Rivet
> 6WIND