Re: [libvirt] gnulib tests in libvirt broken by newer glibc 2.26

[Bruno Haible]

Hi,

Christian Ehrhardt wrote:
> > Fedora 26 only has glibc 2.25 - you need to have Fedora rawhide to get
> > the broken behaviour, as that has glibc 2.26.90
> As Daniel said at least glibc 2.26 as in Fedora rawhide or Ubuntu Artful.

This tip is not helpful: I spent two hours trying Fedora Rawhide, and the
version of today is not functional: the live ISO shows a desktop in which
all you can do is to reboot, and the installation ISO does an installation
but the installed system does not boot.

So we are back to the support mode where we ask for pieces of information
for analysis by us (gnulib developers).

> ./configure gl_cv_func_getopt_posix=no ac_cv_header_getopt_h=no

This mode of configuration is not the usual one. (Why, if you are on a glibc
system, would you want to avoid the system's <getopt.h> header and getopt()
function?) But anyway, it should be supported, and the result should be
a 'test-getopt-posix' program that uses gnulib's replacement:

$ nm test-getopt-posix | grep getopt
000000000060a160 b getopt_data
0000000000405b90 T _getopt_internal_r
0000000000400f90 t getopt_loop.constprop.0
0000000000406170 T rpl_getopt
0000000000406110 T rpl_getopt_internal
0000000000401130 t test_getopt

> [1]: http://paste.ubuntu.com/25638847/

Please provide info as mail attachments (compressed if necessary) in the
future. This site provides a "download as text" button which requests the
user's Launchpad credentials and then attempts to transmit them without
encryption (at least that's what I understand from the Firefox warning).

> Here [1] a log of your commands on such a system showing the issue.
> I added a gdb to show the assert

Thanks. The essential lines are:

checking for getopt.h... (cached) no
checking whether getopt is POSIX compatible... (cached) no
...
{ echo '/* DO NOT EDIT! GENERATED AUTOMATICALLY! */'; \
  sed -e 's|@''GUARD_PREFIX''@|GL|g' \
      -e 's|@''HAVE_GETOPT_H''@|0|g' \
      -e 's|@''INCLUDE_NEXT''@|include_next|g' \
      -e 's|@''PRAGMA_SYSTEM_HEADER''@|#pragma GCC system_header|g' \
      -e 's|@''PRAGMA_COLUMNS''@||g' \
      -e 's|@''NEXT_GETOPT_H''@|<getopt.h>|g' \
      -e '/definition of _GL_ARG_NONNULL/r ./arg-nonnull.h' \
      < ./getopt.in.h; \
} > getopt.h-t && \
...
gcc -DHAVE_CONFIG_H -I. -I..  -DGNULIB_STRICT_CHECKING=1   -g -O2 -MT getopt.o 
-MD -MP -MF .deps/getopt.Tpo -c -o getopt.o getopt.c
...
gcc -DHAVE_CONFIG_H -I. -I..  -DGNULIB_STRICT_CHECKING=1   -g -O2 -MT getopt1.o 
-MD -MP -MF .deps/getopt1.Tpo -c -o getopt1.o getopt1.c

So, the build attempts to use the gnulib override of getopt.

More essential lines:

Program received signal SIGABRT, Aborted.
#0  __GI_raise (address@hidden) at ../sysdeps/unix/sysv/linux/raise.c:51
#1  0x00007ffff7a2df5d in __GI_abort () at abort.c:90
#2  0x0000555555558d23 in test_getopt () at test-getopt.h:754
#3  0x0000555555554b24 in main () at test-getopt-main.h:65

So, the test is executing test_getopt () with the variable POSIXLY_CORRECT
unset; but test_getopt () executes code that is conditional on 'posixly' being
true. Which means that __GETOPT_PREFIX must be defined (presumably to 'rpl_').

Can you show three things, please?

  1) The output of
     $ nm test-getopt-posix | grep getopt

  2) The output of
     $ gcc -DHAVE_CONFIG_H -DGNULIB_STRICT_CHECKING=1 -I. -I.. -I../gllib -g 
-O2 -E test-getopt-posix.c

  3) The output of
     $ gcc -DHAVE_CONFIG_H -DGNULIB_STRICT_CHECKING=1 -I. -I.. -I../gllib -g 
-O2 -E -dM test-getopt-posix.c

Thanks.

Bruno