Re: [dpdk-dev] [PATCH 0/3] null driver improvements for testability

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Feb 18, 2016 at 12:19:58PM +0000, Paul Atkins wrote:
> 
> 
> On 17/02/16 17:23, Bruce Richardson wrote:
> >On Fri, Jan 29, 2016 at 04:47:58PM +0000, Paul Atkins wrote:
> >>Hi Thomas,
> >>
> >>On 29/01/16 16:31, Thomas Monjalon wrote:
> >>>Hi Paul,
> >>>
> >>>2016-01-29 16:18, Paul Atkins:
> >>>>This patchset adds functionality to the null driver help when testing
> >>>>a dataplane that uses dpdk.  The idea is that the the dataplane can
> >>>>have multiple null interfaces attached, and each of theses can be
> >>>>assigned a mac address. Packets can then be injected into the null
> >>>>drivers by adding them to a ring, giving the application complete
> >>>>control of the packets that arrive.  Packets that are sent by a null
> >>>>driver can be stored on a ring, where the application can pick them up
> >>>>and verify it is what was expected.  To allow the application to know
> >>>>when packets have been pulled of the rx ring, counters of the number of
> >>>>times an rx poll has been made are kept, and these can be retrieved via
> >>>>the existing APIs.
> >>>I have not read your code, just read this description.
> >>>It sounds like being a ring PMD. Have you already checked it?
> >>>https://urldefense.proofpoint.com/v2/url?u=http-3A__dpdk.org_browse_dpdk_tree_drivers_net_ring_rte-5Feth-5Fring.c&d=CwICAg&c=IL_XqQWOjubgfqINi2jTzg&r=45ezphVDEm8OnEpH-fLWdXvR3RneLhhNZRDLQRgR6LY&m=wJLO24XFe_B0nZve6mkvocCt7fQWo3PULCTWxrC8rZk&s=bIWycJrY-PYgzkQsBeRfkl8JCHFcxRAHhHDrqRSzHYs&e=
> >>I hadn't seen the ring PMD. I will have a look at it and see if I can make
> >>it do what i need.
> >>
> >>thanks,
> >>Paul
> >Hi Paul,
> >
> >any update on this. Your patches are still showing as pending in patchwork, but
> >if ring pmd is more what need, we can set these patches aside as unneeded, and
> >remove them from the patchwork merge backlog.
> 
> Hi Bruce,
> 
> Sorry for the delay.  The patchset adds 3 things: assigning a mac addr to
> the null pmd, adding the rings to the null pmd and adding xstats for how
> many times the null pmd has been polled.  I could move to using the ring
> pmd, but I would still need the other 2 parts (mac addr and stats).  It
> seems like the ring pmd shouldn't really have these two extra things added,
> but i could do that if it that is preferred over what is in the current
> patchset.
> 
Adding a mac address to be reported by the ring PMD should not be a problem.
Having a stat that tracks polls might be depending on how it is done - if it
uses an atomic, as in this patchset, it would problematic as it would add a
severe performance hit for the SP/SC ring case. However, you could get around
that by copying what is already done in the PMD for tracking packet counts.

Overall, though, it seems that it might be worthwhile doing the work to extend
the ring pmd rather than turning the null pmd into a second ring one.

Regards,
/Bruce