Re: [dpdk-dev] [PATCH v3 0/2] slow data path communication between DPDK port and Linux

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

On Wed, Mar 16, 2016 at 10:22:05AM +0200, Panu Matilainen wrote:
> On 03/16/2016 10:19 AM, Ferruh Yigit wrote:
> >On 3/16/2016 7:26 AM, Panu Matilainen wrote:
> >>On 03/14/2016 05:32 PM, Ferruh Yigit wrote:
> >>>On 3/9/2016 11:17 AM, Ferruh Yigit wrote:
> >>>>This patch sent to keep record of latest status of the work.
> >>>>
> >>>>
> >>>>This is slow data path communication implementation based on existing KNI.
> >>>>
> >>>>Difference is: librte_kni converted into a PMD, kdp kernel module is almost
> >>>>same except all control path functionality removed and some simplification done.
> >>>>
> >>>>Motivation is to simplify slow path data communication.
> >>>>Now any application can use this new PMD to send/get data to Linux kernel.
> >>>>
> >>>>PMD supports two communication methods:
> >>>>
> >>>>1) KDP kernel module
> >>>>PMD initialization functions handles creating virtual interfaces (with help of
> >>>>kdp kernel module) and created FIFO. FIFO is used to share data between
> >>>>userspace and kernelspace. This is default method.
> >>>>
> >>>>2) tun/tap module
> >>>>When KDP module is not inserted, PMD creates tap interface and transfers
> >>>>packets using tap interface.
> >>>>
> >>>>In long term this patch intends to replace the KNI and KNI will be
> >>>>depreciated.
> >>>>
> >>>
> >>>Self-NACK: Will work on another option that does not introduce new
> >>>kernel module.
> >>>
> >>
> >>Hmm, care to elaborate a bit? The second mode of this PMD already was
> >>free of external kernel modules. Do you mean you'll be just removing
> >>mode 1) from the PMD or looking at something completely different?
> >>
> >>Just thinking that tun/tap PMD sounds like a useful thing to have, I
> >>hope you're not abandoning that.
> >>
> >
> >It will be KNI PMD.
> >Plan is to have something like KDP, but with existing KNI kernel module.
> >There will be tun/tap support as fallback.
> 
> Hum, now I'm confused. I was under the impression everybody hated KNI and
> wanted to get rid of it, and certainly not build future solutions on top of
> it?
> 
KNI has it's issues - mainly: a) not being upstream and b) having large
amounts of code to do port management in it, that is best handled by other
means - but the code for transferring packets between kernel space and userspace
is more performant and scalable than TUN/TAP, so we need to keep that around
unless/until we can get TUN/TAP to reach the same performance levels.

Now, we are thinking of some ways in which that can be achieved, but any such
solution is going to be a bit out, so making any driver for transferring packets
from user->kernel and vice versa might as well take advantage of KNI as well as
TUN/TAP so as to allow those who want the extra performance to have it.

Regards,
/Bruce