Re: [dpdk-dev] [PATCH] reserve 'make install' for future use

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2015-11-30 11:49, Bruce Richardson:
> On Mon, Nov 30, 2015 at 11:41:32AM +0000, Richardson, Bruce wrote:
> > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > Sent: Monday, November 30, 2015 11:27 AM
> > > To: Richardson, Bruce <bruce.richardson@intel.com>
> > > Cc: Panu Matilainen <pmatilai@redhat.com>; dev@dpdk.org;
> > > olivier.matz@6wind.com
> > > Subject: Re: [dpdk-dev] [PATCH] reserve 'make install' for future use
> > > 
> > > 2015-11-30 11:08, Richardson, Bruce:
> > > > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > > > Why is it a step in the right direction?
> > > > >
> > > > > We just need to install the files in a different hierarchy and adapt
> > > > > the makefiles to be able to compile an application while keeping the
> > > > > RTE_SDK variable to specify the root directory (previously built
> > > > > thanks to DESTDIR).
> > > > > As the hierarchy could be tuned, we need more variables, e.g.:
> > > > > 	DPDK_INC_DIR (default = RTE_SDK/include/dpdk)
> > > > > 	DPDK_LIB_DIR (default = RTE_SDK/lib)
> > > > >
> > > > > While doing it, we can have a specific handling of T= to keep
> > > > > compatibility with the current (old) syntax.
> > > > >
> > > > > What have I missed?
> > > >
> > > > I'm not sure our existing "make install" is suitable for use for this,
> > > without having it heavily overloaded. The existing T= behavior has support
> > > for wildcards and compiling multiple instances at the same time -
> > > something that won't work with a scheme to actually install DPDK
> > > throughout the filesystem hierarchy. Having it sometimes behave as now,
> > > and sometimes behave as a standard make install is a bad idea IMHO, as it
> > > confuses things. Having lots of extra environment variables is also not a
> > > great idea, to my mind.
> > > 
> > > Yes I agree.
> > > I forgot to mention it, but in my idea, we can drop the support for
> > > multiple targets. So the T= compatibility would be only a shortcut to do
> > > "make config" and name the build directory based on the template name.
> > > 
> > > About the environment variables:
> > > An application requires CFLAGS and LDFLAGS (at least). The standard way to
> > > provide them is pkgconfig (not implemented yet).
> > > For applications using the DPDK makefiles, the only input is RTE_SDK.
> > > When allowing more tuning in paths, we need more variables when using the
> > > DPDK makefiles to build an application.
> > > 
> > > > My opinion is that we should rename our existing "make install" to
> > > something more suitable - my patch suggestion was "make sdk" but it could
> > > be "make target" or something else if people prefer. Once that is done, we
> > > can then look to implement a proper "make install" command that works in a
> > > standard way, perhaps alongside a configure script of some description.
> > > 
> > > I think we don't need to rename or move some code.
> > > Just drop and replace some of them.
> > > 
> > > The configure script is a great idea but it is a totally different idea.
> > > I do not think that installation and configuration should be related.
> > > Please let's consider "make install" first.
> > > 
> > > > For an easy enough solution, I would look to apply this patch to create
> > > "make sdk" and also http://dpdk.org/dev/patchwork/patch/8076/ to have a
> > > "make install" command that works in the build dir. That way:
> > > > * you can have existing behavior using "make sdk T=<target>"
> > > > * you can have standard(ish) configure/make/make install behavior using:
> > > > 	make config T=<target>
> > > > 	cd build
> > > > 	make
> > > > 	make install
> > > >   and the "make config" step can subsequently be wrapped in a configure
> > > script to eliminate the need to know what the best target to use is, etc.
> > > 
> > > As Panu commented, I do not think it is a good idea to have different
> > > behaviours inside and outside of the build directory.
> > > I would even say that this embedded makefile is only confusing and should
> > > be dropped.
> > > We need to have *one* right building methods, not to bring more confusion.
> > 
> > I disagree. I don't think we can have *one* right building method, because to
> > do so means completely throwing away our existing methods of building DPDK
> > and using sample applications. That general method, using RTE_SDK and RTE_TARGET
> > needs to be supported for some time for those projects already familiar with it
> > and using it.

We can keep it for some time while allowing other tree hierarchies.

> > As well as this, we also need a sane way of building DPDK inside the "build" 
> > directory, and having a "make install" target that installs the libraries
> > and headers inside /usr/local (or whatever was specified as $prefix).
> > 
> > With regards to different behavior, since different targets are provided, I
> > don't see it as a problem. In the root directory, "make config" and "make sdk"
> > are provided for backward compatibility. Inside the build directory you have
> > your standard "make" and "make install" commands. Since the command set is
> > very limited, it's easy enough to print a suitable error when the wrong
> > command is used in the wrong place. 
> 
> By way of follow-up to my own email, I'd also state that I would indeed prefer
> not to have different targets in different places, and that ideally you would
> do configure/make/make-install from the root directory. The reason I suggested
> having "make install" work inside the build directory is because of our
> existing use of "make install" for something different in the root directory.
> This is also the reason I sent out this patch. By renaming the "make install"
> command in 2.2, we give ourselves the option in future releases of adding in
> a new "make install" command that behaves as we want, without having to worry
> about conflict with a legacy make install.
> 
> That is why I feel this one patch should go in - it opens up more options for
> us in future releases. It's not an end in itself. :-)

If we do not agree on something else (I'll try to submit some patches),
yes your patch to introduce "make sdk" will be integrated.
But I'd prefer avoiding to document a new command which will be deprecated
when the new-new "make install" will be implemented.
I think there is another solution (I may be wrong).

> > Yes, I would like the ideal state where we have one set of build commands that
> > are run from just one location. However, I don't think we can get to that objective
> > without going through a transition phase where we support both old and new options.
> > 
> > /Bruce