bug#7892: test failures induced by stale NFS files on Solaris

[Ralf Wildenhues]

http://autobuild.josefsson.org/automake/log-201101181921337961000.txt
http://autobuild.josefsson.org/automake/log-201101181921319021000.txt
I've looked a bit now, and two of the leftover NFS files I could find
were configure and install-sh.

These failures are not new, see this older report:
http://thread.gmane.org/gmane.comp.sysutils.automake.patches/5038/focus=5092
The failures seem to happen only on some Solaris systems, more
specifically only on i386-pc-solaris2.10 and sparc-sun-solaris2.10.
The failures are 100% reliable AFAICS (i.e., happen every time).

A git bisect shows that the failures already happen with Automake 1.9.6,
so at least there seems to be no recent Automake regression.  Notably I
cannot find failure instances in my historic testsuite logs from 1.9a
in 2006 on this particular host however, and aclocal13.test passed as
late as 2009-04-19 in Automake 1.10c.  I did not try older Autoconf
or M4 versions yet, however.

Running lsof on the $(distdir) directory recursively right before the
final `rm -rf "$(distdir)"' does not show any processes holding files
open.

I suspect a Solaris NFS implementation update regression or a hardware
bug.  The fact that Dave saw similar issues makes the former more
likely.

If anyone has further data, please add them here.  Who can report this
upstream?

Thanks,
Ralf