[DTrace-devel] [PATCH 0/9] relocatable DTrace

[Kris Van Hees]

On Wed, Jul 31, 2024 at 08:54:44PM +0100, Nick Alcock wrote:
> On 31 Jul 2024, Kris Van Hees said:
> 
> > General comments because my review covers the entire series...
> >
> > Why are there two pkg_config files being introduced when it seems like a
> > single dtrace.pc one would be sufficient.  The dtrace utility is used both
> > for tracing and for the building of libraries and executables with USDT
> > probes,  I see no reason why a single dtrace.pc couldn't cover both.
> 
> Simply because there are two distinct use cases here that require a
> different set of libraries and headers from different places: "we want
> to use USDT using <sdt.h>", and "we are a DTrace consumer". Almost
> nothing wants both, only one wants a library, etc etc.
> 
> Bear in mind that more-or-less universal usage of pkg-config (as in,
> with CMake, Autoconf and Meson it is hard to impossible to do anything
> else) has the caller appending $(pkg-config --cflags $pkg) to CFLAGS or
> CPPFLAGS and $(pkg-config --libs $pkg) to LIBS. If we tried to use the
> same pkg-config file, this would mean that *raw USDT probe users* would
> find themselves linking with libdtrace! This is *surely* not what we
> want.

If people do that without actually checking whether it makes any sense, then
that would be an argument (to me) to not use pkg-config at all.  Using
pkg-config to get --libs when there is no library to link with is pointless
and merely shows that the person using this has no diea what they are doing.

The cflags argument I can accept, but then perhaps we need *3* pkg-config
files?  What exactly do we need this for?  Testsuite needs to know where to
find dtrace, so that could need a pkg-config file.  Anyone wanting to use
libdtrace would need to know where to find the library and what cflags to
use (finding the include files), which would need a pkg-config file.  And
compiling code with SDT would need another pkg-config file.

This brings it back to my original thought that if we are going to provide
pkg-config, we need to carefully think about what we need to provide and do
it consistently.

> pkg-config is not meant to be "one pkg-config file per package", despite
> the name: a huge number of packages have multiple. (The worst offender
> on my system appears to be Qt 6, with 136, but even a simple thing like
> xcb-util has four.)
> 
> > On that note, why can't pkg-config then also be used as a means to get the
> > location of the dtrace executable?  Why go through the trouble of performing
> > text substitution on runtest.sh to insert the location of dtrace when you
> > already use pkg-config in it to get the appropriate include directory for
> > the installed case.
> 
> Because I didn't think of it. That's an excellent idea. Adjusting
> accordingly. Will send a v2 of the whole series (because it's been
> ages).
> 
> > I think this series needs to be reworked to be consistent.  If we are going
> > to use pkg-config, then let's use it to its fullest extent rather than using
> > it for some stuff but then still hardcoding other things.  That makes no sense
> > to me.
> 
> Agreed. Anything else you can think of that's hardcoded? I can't
> immediately think of anything (because we can't use pkg-config in
> dtprobed.service, it's not a program) but you looked at this much more
> recently than me...