Bug#884662: reproducible-builds.org: regular files sometimes treated as directories

[Christoph Berg]

> On Mon, 18 Dec 2017 at 23:30:05 +0000, Simon McVittie wrote:
> > dpkg-deb: building package 'libglib2.0-data' in '../libglib2.0-data_2.54.2-2_all.deb'.
> > tar: ./usr/share/locale/en_CA/LC_MESSAGES/glib20.mo/: Cannot savedir: Not a directory
> 
> This seems to be a symptom of some more general problem on the
> reproducible-builds builders - I would guess it's either the
> (FUSE?) filesystem, or a LD_PRELOAD hack that intercepts stat(), like
> fakeroot does. Build logs indicate that regular files are sometimes
> treated as directories by install(1), resulting in the build being
> incomplete and unreproducible even in cases where it doesn't FTBFS:

This "Not a directory" problem has started popping up in PostgreSQL
build logs:

https://buildd.debian.org/status/fetch.php?pkg=pgagent&arch=alpha&ver=4.0.0-7&stamp=1602009768&raw=0

Dropping cluster 13/regress ...
### End 13 installcheck ###
make[1]: Leaving directory '/<<PKGBUILDDIR>>'
   dh_strip_nondeterminism -a
   dh_compress -a
   dh_fixperms -a
   dh_missing -a
   dh_dwz -a -a
Can't opendir(debian/pgagent/usr/share/doc/pgagent/changelog.Debian.gz): Not a directory
 at /usr/bin/dh_dwz line 119.

https://buildd.debian.org/status/fetch.php?pkg=pgagent&arch=sparc64&ver=4.0.0-7&stamp=1601991893&raw=0

It has also been seen on amd64.

PG extension packages have started to run tests at build time, which
is done via pg_virtualenv. Internally, for creating the temporary
PostgreSQL server instance, LD_PRELOAD is unset, and that seems to be
what confuses the "dh" and dpkg-buildpackage processes that share the
same fakeroot instance.

jwilk did some debugging (thanks!) and came up with this simple
recipe:

$ fakeroot sh -c 'mkdir foo; env -u LD_PRELOAD rmdir foo; touch bar; stat bar | grep directory' Size: 0               Blocks: 0          IO Block: 4096 directory

So, if "foo" is removed without fakeroot knowing, the "bar" file is
reported as a directory. (It doesn't get it wrong for me, it depends
on inode numbers being recycled and similar.)

jwilk also noted that glib-2.0's debian/rules unsets LD_PRELOAD for
the test suite too which strengthens the evidence that fakeroot is to
blame.

I'll see to move the PG extension packages to `R³: no`, that seems to
be the best option in my case.

Christoph