(solved, I think) Re: openjdk build failures on "nm"

Van: Harald Eilertsen <[email protected]>
Datum: donderdag, 11 september 2025 10:15
Aan: Ronald Klop <[email protected]>
CC: [email protected]
Onderwerp: Re: openjdk build failures on "nm"
> 
> On Thu, Sep 11, 2025 at 09:32:08AM +0200, Ronald Klop wrote:
> > As it seems hard to debug, what do the java@ maintainers think of this experiment?
> >
> > diff --git a/java/openjdk17/Makefile b/java/openjdk17/Makefile
> > index 5cb25e6aee72..ba7001ff2bd0 100644
> > --- a/java/openjdk17/Makefile
> > +++ b/java/openjdk17/Makefile
> > @@ -38,7 +38,7 @@ GH_PROJECT=   jdk17u
> >
> > NO_CCACHE=     yes
> >
> > -_MAKE_JOBS=    #
> > +MAKE_JOBS_UNSAFE=      yes
> > MAKE_ENV=      LANG="C" \
> >                LC_ALL="C" \
> >                CLASSPATH="" \
> 
> That's well spotted! I think this is definitely worth trying. I'm no
> expert on the ports Makefiles myself, but was told to use
> MAKE_JOBS_UNSAFE instead in the review for openjdk24. Haven't gone
> through all the ancient version yet, as my focus is mainly on mainline,
> and these days on openjdk25.
> 
> > I can commit this today to openjdk17 and if builds start to succeed to other openjdk* ports.
> 
> I'm in favour!
> 
> H!
> 
> 
> 

Hi Harald,

I found out what happens in openjdk23 and openjdk24 around _MAKE_JOBS.

_MAKE_JOBS is an internal variable in Mk/bsd.port.mk used to add "-j<MAKE_JOBS_NUMBER>" to the make command in do-build. It was set empty to suppress passing -j, because the port otherwise gives this error:
===>  Building for openjdk17-17.0.16+8.1_2
Error: 'make -jN' is not supported, use 'make JOBS=N'

NB: this error comes from work/jdk17u-jdk-17.0.16-8-1/make/InitSupport.gmk:119
In /usr/ports/Mk/bsd.port.mk we have these lines of code:
.    if defined(DISABLE_MAKE_JOBS) || defined(MAKE_JOBS_UNSAFE)
_MAKE_JOBS?=        #
MAKE_JOBS_NUMBER=   1
.    else
So setting _MAKE_JOBS is a side effect of MAKE_JOBS_UNSAFE. The suggestion of using MAKE_JOBS_UNSAFE was just another way of setting _MAKE_JOBS and MAKE_JOBS_NUMBER.
But the MAKE_JOBS_NUMBER is never passed to the (implicit) do-build target of the ports Makefile.

If you look in https://pkg-status.freebsd.org/beefy17/data/15stable-amd64-default/56c5cbbc8d99/logs/openjdk23-23.0.2+7.1_2.log and search for "jobs" you will find this:
checking for number of cores... 48
checking for memory size... 130685 MB
checking for appropriate number of jobs to run in parallel... 48
...
Build performance summary:
* Build jobs:     48
So, in summary
* MAKE_JOBS_UNSAFE sets _MAKE_JOBS empty and gives the suggestion of running without parallel jobs. But the port further ignores MAKE_JOBS_NUMBER and ...
* The port decides to run with 48 parallel jobs as it looks at the number of CPUs.
* A bit more documentation/comments around overriding the _MAKE_JOBS variable in the Makefile could have saved us quite some research. 😉

The current construction in openjdk17 handles MAKE_JOBS_NUMBER properly. And with that also the MAKE_JOBS_UNSAFE variable. After more testing we can use this blueprint in other openjdk ports.
Also in openjdk23 and -24. They don't have flapping builds but can improve handling of MAKE_JOBS_NUMBER.

Regards,
Ronald.