[dpdk-dev] [dpdk-techboard] Napatech pmd

[Bruce Richardson]

On Tue, Jan 09, 2018 at 07:24:45PM -0500, Neil Horman wrote:
> On Tue, Jan 09, 2018 at 01:21:29PM -0800, Stephen Hemminger wrote:
> > On Tue, 09 Jan 2018 21:36:03 +0100 Thomas Monjalon
> > <thomas at monjalon.net> wrote:
> > 
> > > 09/01/2018 21:20, Neil Horman:
> > > > On Tue, Jan 09, 2018 at 07:57:50PM +0000, Michael Lilja wrote:  
> > > > > > On Mon, Jan 08, 2018 at 04:15:47PM +0100, Thomas Monjalon
> > > > > > wrote:  
> > > > > > > Hi,
> > > > > > >
> > > > > > > 08/01/2018 14:08, Finn Christensen:  
> > > > > > > > Hi Thomas,
> > > > > > > >
> > > > > > > > Thanks for bringing this discussion up again.
> > > > > > > >
> > > > > > > > The Napatech PMD is build on top of our proprietary
> > > > > > > > driver. The  
> > > > > > reason is basically that we utilize many years of driver
> > > > > > development and thus reuses the FPGA controlling code in the
> > > > > > DPDK PMD. The Napatech driver suite is still closed source.  
> > > > > > > > The current NTNIC PMD dynamically links a Napatech
> > > > > > > > proprietary  
> > > > > > NTAPI library to control the FPGA on our NICs.  
> > > > > > > >
> > > > > > > > We did think of the PMD as being our responsibility to
> > > > > > > > keep updated  
> > > > > > towards the Napatech NIC communication, and that we would be
> > > > > > engaged and asked to modify accordingly if changes in DPDK
> > > > > > required that (maintainer). Furthermore, the PMD compiles
> > > > > > with no issues, when NTNIC is enabled.  
> > > > > > > > We have plans to write a stand-alone PMD, but this is
> > > > > > > > not a small task  
> > > > > > to do, therefore we haven't got to that yet.  
> > > > > > >
> > > > > > > This standalone PMD would be open and BSD licensed?
> > > > > > >  
> > > > > > > > If the DPDK community would accept the dynamic linking
> > > > > > > > to a  
> > > > > > proprietary library, from inside our PMD, then it would be
> > > > > > great.  
> > > > > > >
> > > > > > > Dynamic linking is OK.  I think we can accept such PMD at
> > > > > > > the condition that we can build it, meaning we can easily
> > > > > > > download the build dependencies for free.
> > > > > > >  
> > > > > > > > Let me know what you think. Or maybe you have ideas to
> > > > > > > > what else  
> > > > > > we could do to make it upstream.  
> > > > > > >
> > > > > > > My thinking is to allow every hardware to have a good DPDK
> > > > > > > support.  Every step in this direction is a progress.
> > > > > > >  
> > > > > >
> > > > > > I have to ask the question:  Why not open source your FPGA
> > > > > > code?  That would make all of this a non issue.
> > > > > >
> > > > > > While I knows it to various degrees been done in the past, I
> > > > > > really don't like the idea of including drivers (even open
> > > > > > source drivers), if they have dependencies on closed source
> > > > > > software.  It means that we as a community assume some level
> > > > > > of responsibility for that pmd, but have no ability to make
> > > > > > fixes to that pmd without accepting your license terms.  I
> > > > > > understand that you are saying you currently have
> > > > > > responsibility for it as the license owner, but if that
> > > > > > chages in the future, the PMD has no use to the community.
> > > > > > It would be preferable if access to controlling the hardware
> > > > > > was just free of a proprietary license.  Then you wouldn't
> > > > > > have to write a stand alone pmd.
> > > > > >
> > > > > > Neil  
> > > > > I understand your concern, but it is quite normal that BSP
> > > > > (Board Support Package) is binary and has an API that is BSD
> > > > > licensed. The Napatech Suite is basically a BSP. The API that
> > > > > will be used in the PMD is a BSD licensed API and so will the
> > > > > PMD be. If you understand the API you will be able to control
> > > > > the hardware and thereby also be able to change the DPDK
> > > > > driver. The API is public available and so is the BSP binary
> > > > > package. A good analogy is how Android does open source
> > > > > develop for Quallcomm based HW boards, where the Quallcomm
> > > > > firmware is proprietary. Any user can download full Android
> > > > > stack and BSP packages free of charge to do Android
> > > > > development.
> > > > > 
> > > > > /Michael
> > > > >   
> > > > You can couch it any way you like, but regardless of how you
> > > > want to view the proprietary part of this system, the hardware
> > > > is being used as a NIC in the DPDK, and for that purpose you
> > > > need a driver.  As you currently have it architected, the driver
> > > > (open source or not), is useless without the binary portion
> > > > underneath it, and so is useless to any user without agreeing to
> > > > your license terms.  Pert of the benefit of open source software
> > > > is that users can continue to modify/enchance/maintain the
> > > > software powering your hardware after you decide to stop
> > > > supporting it.  And so I'm not comfortable with accepting code
> > > > that doesn't allow this community to do that.
> > > > 
> > > > Neil  
> > > 
> > > How different is it of having a proprietary firmware?
> > > 
> > > I think it is better for the Napatech users to have this PMD
> > > supported upstream.  If it is too complicate to maintain and not
> > > supported by Napatech, we are free to drop it.
> > > 
> > > Adding the technical board Cc.
> > 
> > 
> > Agree with Thomas.  "Bad breath is better than no breath at all"
> > 
> I understand your intent here, but in fairness, theres a position
> between bad and no breath, metaphorically - that is to say,
> maintaining the driver out of tree.
> 
> > Though I suspect that DPDK linked to proprietary code is going to
> > unmaintainable across releases, and very hard to debug problems. 
> > 
> My point exactly, this PMD will receive exactly no testing from anyone
> in the community as they won't accept the license terms of the binary
> blob underneath it, and so this driver will likely see almost exactly
> the same level of support as it would were it maintained out of tree.
> 
I wonder is this really an issue? For many NIC drivers, now many
community users actively test (not just use in their app, but test a
full range of functions) the driver, apart from the maintainers from the
NIC vendors? [Thinking here especially of the less commonly used ones].

So long as there is active maintenance of the driver by the vendor, I
don't think there is a problem. Once code is unmaintained, it *is* a
problem, whether the underlying requirements are proprietary or not - 
just look at Xen support in DPDK! In my view, by accepting a driver
with a proprietary dependency, the only thing we are lacking is the
ability for someone else to step up to maintain the driver should the
original vendor drop out. Given our difficulty in finding maintainers
for things, I wouldn't view that as a major loss.

Regards,
/Bruce