Re: [dpdk-dev] [PATCH] examples/l3fwd: adding event queue support

[Jerin Jacob <jerin.jacob@caviumnetworks.com>]

-----Original Message-----
> Date: Thu, 12 Apr 2018 15:03:22 +0100
> From: Bruce Richardson <bruce.richardson@intel.com>
> To: Sunil Kumar Kori <sunil.kori@nxp.com>
> CC: "dev@dpdk.org" <dev@dpdk.org>, Hemant Agrawal <hemant.agrawal@nxp.com>
> Subject: Re: [dpdk-dev] [PATCH] examples/l3fwd: adding event queue support
> User-Agent: Mutt/1.9.4 (2018-02-28)
> 
> On Thu, Apr 12, 2018 at 06:09:04AM +0000, Sunil Kumar Kori wrote:
> > Gentle reminder to review the RFC.
> > 
> > Regards
> > Sunil Kumar
> > 
> 
> Hi,
> 
> sorry for the delay in review.
> 
> /Bruce
> 
> > -----Original Message-----
> > From: Sunil Kumar Kori [mailto:sunil.kori@nxp.com] 
> > Sent: Monday, March 19, 2018 7:15 PM
> > To: dev@dpdk.org
> > Cc: Sunil Kumar Kori <sunil.kori@nxp.com>; Hemant Agrawal <hemant.agrawal@nxp.com>
> > Subject: [PATCH] examples/l3fwd: adding event queue support
> > 
> > This patch set to add the support for eventdev based queue mode support to the l3fwd application.
> > 1. Eventdev support with parallel queue
> > 2. Eventdev support with atomic queue
> > 
> > This patch adds
> > - New command line parameter is added named as "dequeue-mode" which
> >   identifies dequeue method i.e. dequeue via eventdev or polling
> >   (default is polling)
> > . If dequeue mode is via:
> >          a. eventdev: New parameters are added -e, -a, -l  to cater
> > 	    eventdev config, adapter config and link configuration
> >             respectively. "--config" option will be invalid in this case.
> >          b. poll mode: It will work as of existing way and option for
> > 	    eventdev parameters(-e, -a, -l) will be invalid.
> > 
> > - Functions are added in l3fwd_em.c and l3fwd_lpm.c for packet I/O
> >   operation
> > 
> > The main purpose of this RFC is get comments on the approach.
> > This is a *not tested* code.
> > 
> > Signed-off-by: Sunil Kumar Kori <sunil.kori@nxp.com>
> > ---
> >  examples/l3fwd/Makefile         |   2 +-
> >  examples/l3fwd/l3fwd.h          |  21 ++
> >  examples/l3fwd/l3fwd_em.c       | 100 ++++++++
> >  examples/l3fwd/l3fwd_eventdev.c | 541 ++++++++++++++++++++++++++++++++++++++++
> >  examples/l3fwd/l3fwd_eventdev.h |  85 +++++++
> >  examples/l3fwd/l3fwd_lpm.c      | 100 ++++++++
> >  examples/l3fwd/main.c           | 318 +++++++++++++++++++----
> >  examples/l3fwd/meson.build      |   2 +-
> >  8 files changed, 1120 insertions(+), 49 deletions(-)  create mode 100644 examples/l3fwd/l3fwd_eventdev.c  create mode 100644 examples/l3fwd/l3fwd_eventdev.h
> > 
> 
> My initial impression is that this seems like an awful lot of new code just
> to support reading from an eventdev rather than from an ethdev. Looking at
> the datapath main function loop, is the only difference there that
> rte_eth_rx_burst has been changed to rte_eventdev_dequeue_burst or are
> there more significant changes than that?
> 
> If this is the case, is this scale of changes really needed to this app?
> What about the other examples, how many of them will need to be similarly
> updated?
> 
> I'm also wondering if it would help, or be useful, to have a vdev type
> which wraps an eventdev queue as an ethdev. That would eliminate the need
> for the datapath code, and may help abstract away some parts of the setup.
> It would also help with re-use if you anticipate wanting to make a similar
> change to other apps.

Exposing as an ethdev-vdev device will introduce cyclic build dependency(now
eventdev is depended on ethdev). I think, maybe a helper function
in eventdev area to setup Rx adapter and similar slow path logic may work.


> 
> /Bruce
>