Re: [dpdk-dev] Heads up: Fedora packaging plans

[Neil Horman <nhorman@tuxdriver.com>]

On Mon, May 19, 2014 at 06:28:47PM +0200, Thomas Monjalon wrote:
> 2014-05-19 09:18, Neil Horman:
> > On Mon, May 19, 2014 at 12:11:35PM +0200, Thomas Monjalon wrote:
> > > My main concerns are about naming and extensions.
> > > We must keep "dpdk-core" naming in order to distinguish it from PMD
> > > extensions.
> > 
> > I don't see why.  We can name packages whatever we want, as long as the spec
> > and srpm share the same name. It seems to me that the core should be the
> > base name of the package while the extensions should have some extension on
> > their name.
> 
> OK
> 
> > > And then, packaging of memnic and non-uio paravirtualization PMDs
> > > (virtio/vmxnet3) are missing.
> > 
> > They're in separate repositories, I was planning on packaging them at a
> > later time separately, since their versioning and development is handled
> > separately.
> 
> OK
> 
> > > We should try to get .spec for Fedora and in-tree .spec as common as
> > > possible. There are probably some things to push.
> > 
> > Ok, sure, just keep in mind that different distributions have different
> > packaging requirements that may affect the contents of the spec file, and so
> > attaining parity may not be possible (or even worthwhile).
> 
> Yes, the in-tree file should be a common base.
> 
> > > I thought that we should put version in the name, in order to be able to
> > > install many versions together. How is it handled by yum?
> > 
> > So, I spent some time thinking about this, and I _really_ want to avoid the
> > inclusion of a version with the package name.  Doing so, while it allows yum
> > to install multiple versions side-by-side, is a real overhead for me, as it
> > requires that I go through a new pacakge review process for each released
> > version that we want to package.  I do not have time to do that.  If
> > someone from 6wind or intel wants to get involved in the Packaging process
> > we can look at that as a solution, but while I'm doing it, its really just
> > too much overhead.  This method will allow multiple version to be installed
> > side by side as well.  The tradeoff is that yum doesn't directly allow
> > that, as it will just preform an upgrade.  The multiple version solution
> > will require that you download older versions and install them directly
> > using rpm commands.  I think thats a fair tradeoff.
> 
> OK
> 
> > > > * Added config files to match desired configs for Fedora (i.e. disabled
> > > > PMD's that require out of tree kernel modules
> > > 
> > > It would be clearer to make your configuration changes with "sed -i".
> > > In a near future we would probably need a "configure" script to do it.
> > 
> > I really disagree.  Its not clearer in my mind at all - in that the final
> > config file is a product of two pieces of information (the base config
> > file, and the sed scripts that you run on it), as opposed to one piece (the
> > canonical modified config specified in the source line).  Using sed also
> > implies that you need to list sed as a BuildRequires (minimal buildroots
> > may not include sed when they are spun up).
> 
> I don't understand the logic.
> When packaging a library using autotools, you are setting some options to 
> "configure" script which are not saved elsewhere than in the spec file, right?
> Why it wouldn't be the same here (options in spec file)? 
> 
When you need to override options that can't be auto-detected, then yes.  In
that case, you can override options that can't be automatically determined from
the build environment.  Thats a common use case and its well understood by
everyone, which is fine.

But thats not what you have here.  DPDK has a build environment that more akin
to the linux kernel, in that, from an rpm build perspective a flat file is
inserted as the configuration.  the spec file just drops it in the proper place
and uses it, unmodified.  Theres no expectation that, when a configuration is
completely supplied, that it will be modified prior to build time.

Its the difference between static and dynamic generation of configuration data.
If you're using autoconf, then you know that you don't have a set configuration
until you run ./configure, which means you know to look to the spec file to
determine any options for building, and you know that your configuration will be
generated after the spec file runs configure, whatever it is.  If conversely you
have a static configuration file (as DPDK does), the assumption is that the
selected configuration is included in the source package and is unaltered at
build time.

> > > So you don't package igb_uio but you build it because there is no option
> > > to disable it currently. We should add such option.
> > 
> > Not sure what you mean here.  The only uio code I see in the package is the
> > uio unbind script for igb, which should still work just fine (save for the
> > fact that we don't have a user space PMD to attach the hardware to).  I can
> > certainly remove the script though so it doesn't appear in the package
> > until such time as the LAD group integrates the uio code in the upstream
> > driver.
> 
> I'm speaking about lib/librte_eal/linuxapp/igb_uio.
> I think you are building it and it would fail if you no kernel header 
> installed.
> 
Thats true, but you'll always have those header files because, at least for
fedora uio_driver.h is packaged as part of the kernel-devel package, which is a
BuildRequires: depedency for dpdk.  Note that may vary from distro to distro, so
my assertion isn't always true (see previous note about pushing changes to
upstream spec files).  It is safe for Fedora however.


That all said, I noted a patch series to convert DPDK to VFIO, which I am highly
supportive of, which I think would also affect this in that it would modify this
script (or arguably obviate it), as I think there is a common vfio script that
can detach any pci device.

Neil

> -- 
> Thomas
>