svn commit: r254273 - in head: . include lib lib/libc/iconv lib/libiconv_compat lib/libkiconv share/mk sys/sys tools/build/mk

[Joel Dahl]

On Sun, Aug 18, 2013 at 03:51:25PM -0700, Peter Wemm wrote:
> On 8/18/13 3:42 PM, Jilles Tjoelker wrote:
> > On Sun, Aug 18, 2013 at 09:53:04PM +0200, Joel Dahl wrote:
> >> On Sun, Aug 18, 2013 at 12:34:30AM +0200, Dimitry Andric wrote:
> >>> On Aug 13, 2013, at 09:15, Peter Wemm <peter at FreeBSD.org> wrote:
> >>>> Author: peter
> >>>> Date: Tue Aug 13 07:15:01 2013
> >>>> New Revision: 254273
> >>>> URL: http://svnweb.freebsd.org/changeset/base/254273
> > 
> >>>> Log:
> >>>>  The iconv in libc did two things - implement the standard APIs, the GNU
> >>>>  extensions and also tried to be link time compatible with ports libiconv.
> >>>>  This splits that functionality and enables the parts that shouldn't
> >>>>  interfere with the port by default.
> > 
> >>>>  WITH_ICONV (now on by default) - adds iconv.h, iconv_open(3) etc.
> >>>>  WITH_LIBICONV_COMPAT (off by default) adds the libiconv_open etc API, linker
> >>>>  symbols and even a stub libiconv.so.3 that are good enough to be able
> >>>>  to 'pkg delete -f libiconv' on a running system and reasonably expect it
> >>>>  to work.
> > 
> >>>>  I have tortured many machines over the last few days to try and reduce
> >>>>  the possibilities of foot-shooting as much as I can.  I've successfully
> >>>>  recompiled to enable and disable the libiconv_compat modes, ports that use
> >>>>  libiconv alongside system iconv etc.  If you don't enable the
> >>>>  WITH_LIBICONV_COMPAT switch, they don't share symbol space.
> > 
> >>>>  This is an extension of behavior on other system.  iconv(3) is a standard
> >>>>  libc interface and libiconv port expects to be able to run alongside it on
> >>>>  systems that have it.
> > 
> >>> Unfortunately I expect this will break many ports, when the libiconv
> >>> port is installed.  A simple example is the following:
> >> <SNIP>
> > 
> >> It also breaks installworld when /usr/src and /usr/obj are NFS exported
> >> read-only.
> > 
> > I think it has to do with share/i18n/csmapper and share/i18n/esdb using
> > directories as make targets. This apparently causes these files to be
> > rebuilt at 'make installworld' time, which is always bad but is only
> > detected when /usr/obj is read-only.
> > 
> > A hack that works is to enclose the four targets depending on ${SUBDIR}
> > in  .if !make(install)  .
> > 
> > Unfortunately, the Makefiles were written to depend on the directories
> > as make targets fairly deeply, so a real fix is harder.
> 
> I was looking at this yesterday, but was tied up with other things.  I'll
> take a look at it today after getting a few other things done.  It should be
> easy enough to replicate by changing /usr/obj to readonly on test systems.

FWIW, this is still broken.

-- 
Joel