RPMs for ULN do not build in MySQL 5.6: Patches + libmysqld.so

Bug #15972480

This is the change for 5.5:
    a cleanup in the way "libmysqld.so" is created.
(Patches were adapted for 5.5 previously.)

Originally, the ".so" was created by taking all modules in
"libmysqld.a", after removing some few which caused unresolved
references.
This is no good idea, rather "ld" should be used to follow all
references from some few start modules.

At the same time, the ".so" version needed to be corrected:
The original "0.0.1" is both wrong and risky.
Rather, the server version is used to identify the ".so" file,
but for linkage the first two levels are sufficient (so upgrades
are possible without re-building the embedded application).

1 files changed, 26 insertions, 11 deletions

diff --git a/packaging/rpm-uln/mysql.spec.sh b/packaging/rpm-uln/mysql.spec.sh
index 8b25b160be1..3e2c5c3f11c 100644
--- a/packaging/rpm-uln/mysql.spec.sh
+++ b/packaging/rpm-uln/mysql.spec.sh
@@ -508,21 +508,28 @@ mkdir release
 # TODO / FIXME: Do we need "scriptstub"?
 gcc $CFLAGS $LDFLAGS -o scriptstub "-DLIBDIR=\"%{_libdir}/mysql\"" %{SOURCE4}
 
-# TODO / FIXME: "libmysqld.so" should have been produced above  - WORK in PROGRESS
+# TODO / FIXME: "libmysqld.so" should have been produced above
 # regular build will make libmysqld.a but not libmysqld.so :-(
 cd release
 mkdir libmysqld/work
 cd libmysqld/work
-ar -x ../libmysqld.a
-rm rpl_utility.cc.o sql_binlog.cc.o  # Try-and-Error: These modules cause unresolved references
-gcc $CFLAGS $LDFLAGS -shared -Wl,-soname,libmysqld.so.0 -o libmysqld.so.0.0.1 \
-	*.o \
+# "libmysqld" provides the same ABI as "libmysqlclient", but it implements the server:
+# The shared object is identified by the full version,
+# for linkage selection the first two levels are sufficient so that upgrades are possible
+# (see "man ld", option "-soname").
+SO_FULL='%{mysql_version}'
+SO_USE=`echo $SO_FULL | sed -e 's/\([0-9]\.[0-9]\)\.[0-9]*/\1/'`
+# These two modules should pull everything else which is needed:
+ar -x ../libmysqld.a client.c.o signal_handler.cc.o
+gcc $CFLAGS $LDFLAGS -shared -Wl,-soname,libmysqld.so.$SO_USE -o libmysqld.so.$SO_FULL \
+	*.o ../libmysqld.a \
 	-lpthread -lcrypt -laio -lnsl -lssl -lcrypto -lz -lrt -lstdc++ -lm -lc
 # this is to check that we built a complete library
 cp %{SOURCE9} .
-ln -s libmysqld.so.0.0.1 libmysqld.so.0
-gcc -I../../include -I../../../%{src_dir}/include $CFLAGS mysql-embedded-check.c libmysqld.so.0
-LD_LIBRARY_PATH=. ldd ./a.out
+PROGNAME=`basename %{SOURCE9} .c`
+ln -s libmysqld.so.$SO_FULL libmysqld.so.$SO_USE
+gcc -I../../include -I../../../%{src_dir}/include $CFLAGS -o $PROGNAME %{SOURCE9} libmysqld.so.$SO_USE
+LD_LIBRARY_PATH=. ldd $PROGNAME
 cd ../..
 cd ..
 
@@ -618,9 +625,11 @@ mv ${RPM_BUILD_ROOT}%{_bindir}/mysql_config ${RPM_BUILD_ROOT}%{_libdir}/mysql/my
 install -m 0755 scriptstub ${RPM_BUILD_ROOT}%{_bindir}/mysql_config
 
 rm -f ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.a
-install -m 0755 release/libmysqld/work/libmysqld.so.0.0.1 ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so.0.0.1
-ln -s libmysqld.so.0.0.1 ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so.0
-ln -s libmysqld.so.0 ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so
+SO_FULL='%{mysql_version}'
+SO_USE=`echo $SO_FULL | sed -e 's/\([0-9]\.[0-9]\)\.[0-9]*/\1/'`
+install -m 0755 release/libmysqld/work/libmysqld.so.$SO_FULL ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so.$SO_FULL
+ln -s libmysqld.so.$SO_FULL ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so.$SO_USE
+ln -s libmysqld.so.$SO_USE  ${RPM_BUILD_ROOT}%{_libdir}/mysql/libmysqld.so
 
 rm -f ${RPM_BUILD_ROOT}%{_bindir}/comp_err
 rm -f ${RPM_BUILD_ROOT}%{_mandir}/man1/comp_err.1*
@@ -964,6 +973,12 @@ fi
 %{_mandir}/man1/mysql_client_test.1*
 
 %changelog
+* Thu Dec  7 2012 Joerg Bruehe <joerg.bruehe@oracle.com>
+- Change the way in which "libmysqld.so" is created: Using all object modules
+  was wrong, gcc / ld can resolve the dependencies from "libmysqld.a".
+  Also, identify the ".so" version from the MySQL version, "0.0.1" was wrong.
+  Bug#15972480
+
 * Tue Sep 18 2012 Joerg Bruehe <joerg.bruehe@oracle.com>
 - Restrict the vendor check to Oracle: There is no history here
   which we have to allow for.