RFC: FreeBSD DB Port Reform - rev. 0.4

[Matthias Andree]

                      The FreeBSD BerkeleyDB Port Reform


    Matthias Andree

Revision: $Id: index.html,v 0.4 2003/11/06 14:30:55 emma Exp emma $

                                 Introduction

   The current databases/db* ports are a mess. They use nonstandard
   library names that require porters to ship a lot of FreeBSD
   files/patch-* patches to make the port compile with the non-canonical
   layout.

   I believe the layout has historical reasons in a.out. ELF, with its
   SONAME feature, allows for a more elegant and standard-conforming
   layout that allows us to drop FreeBSD specific patches from the ports.

                            Background information

     * Any ELF shared object (library), seen as .so file, bears an
       "SONAME" tag that can be made visible with "readelf -d FILENAME"
       and that needs not match the file name (it often does, see below).
     * Any ELF executable that links against shared object bears a
       corresponding "NEEDED" tag. A link time, here is recorded the
       SONAME of the .so file linked against, not the file name. Example:
       Linking a program with -L/usr/local/BerkeleyDB.4.1 -ldb will
       insert a "NEEDED: libdb-4.1.so.1" trait into the program. The
       program will, at run time, automatically choose
       /usr/local/lib/libdb-4.1.so.1. This NEEDED information is also
       visible in "readelf -d FILENAME".
     * An ELF executable MAY have one or more "RPATH" tag that specifies
       exactly one shared object search path element. It allows to
       specify the link path in an executable, and saves the end user the
       hassle of editing ld.so.conf or LD_LIBRARY_PATH for non-default
       library paths. This is, you guessed it, visible in "readelf -d
       FILENAME".

                                  Action Plan

    1. bsd.ports.mk should be extended to support USE_BDB_VER or
       WANT_BDB_VER or something that presets a set of variables, for
       instance BDB_VERSION, BDB_INCLUDES, BDB_LIBNAME, BDB_LIBPATH and
       LIB_DEPENDS, to lift shared code out of the individual ports into
       bsd.ports.mk.
    2. for db3, db4 and db41, the following changes are made:
         1. all programs move into $PREFIX/BerkeleyDB.N.M/bin/db_*. We
            will need to accomodate different versions with the same
            name. All other systems use $PREFIX/BerkeleyDB.N.M/bin.
         2. programs will no longer be renamed to db41_* db4_* or db3_*.
         3. all libraries maintain their official SleepyCat names,
            libdb-3.3, libdb-4.0, libdb-4.1, and will not be renamed to
            libdb3, libdb4 or libdb41 - this removes the need for a good
            deal of REINPLACE_CMD or patches.
         4. all libraries move into $PREFIX/BerkeleyDB.N.M/lib, because
            BerkeleyDB installs not only libdb-N.M* but also libdb-N* and
            libdb* (where * is .so.N, .so or .a), which cannot work in
            the same path. Sleepycat recommends linking against -ldb or
            libdb.so, see
            http://www.sleepycat.com/docs/ref/build_unix/shlib.html.
            These libaries are relevant at link time.
         5. a relative symbolic link from $PREFIX/lib/libdb-N.M.so.M to
            ../BerkeleyDB.N.M/lib/libdb-N.M.so.M (the SONAME, see above)
            is maintained. These are relevant at run time.
         6. All BerkeleyDB distributions will be compiled with
            --with-uniquename, so that the same application can link two
            different DB versions without conflicts. It's not too common,
            but can happen indirectly through other libraries. See
            http://www.sleepycat.com/docs/ref/install/multiple.html for
            rationale.
            NOTE: This changes the ABI (but not the API), which means all
            applications will have to be recompiled and relinked.
         7. documentation remains in $PREFIX/share/doc/dbNM (M omitted
            when 0).
    3. Dependent ports will have to be updated and recompiled, this will
       need the help of original maintainers.
    4. DB-2.7.7 ("databases/db2") will be subject to further discussion.
       Few ports depend on it (some of them are marked BROKEN on FreeBSD
       5-CURRENT), some offer it as an alternative to db3, and it's
       generally viewed as "vintage" (Sleepycat wording) version and
       slow.

                                  Advantages

     * no more makefile file patching for the arcane layout. -ldb with
       proper LDFLAGS will "just work".
     * ports that use the official library names such as -ldb-3.3,
       ldb-4.0, ldb-4.1, will also "just work".
     * ports that have --with-db switches and expect a standard
       BerkeleyDB.N.M directory under /usr/local will also "just work".
     * Fewer FreeBSD-specific patches in the ports tree will be needed.

                                 Disadvantages

   All ports that depend on DB will have to be updated once.

                                    Preview

   Port tarballs for preview are here. Comments are solicited. These
   currently correspond to Revision 0.1 of this draft.

                     History of this suggestion (from RCS)

$Log: index.html,v $
Revision 0.4  2003/11/06 14:30:55  emma
Fix up RCS tags in history :-(

Revision 0.3  2003/11/06 14:28:16  emma
Fix up RCS error of last version. No longer suggest DB-2 removal as some
ports DO depend on it, but leave it for further discussion.
Mention port tarballs correspond to version 0.1.

Revision 0.2  2003/11/06 14:10:40  emma
Remove WITH_BDB_VER, that is user <-> port communication namespace.
Clarify why we need $PREFIX/BerkeleyDB.N.M/bin/db_*, link to
http://www.sleepycat.com/docs/ref/build_unix/shlib.html.
Mention $PREFIX/lib/libdb-N.M* are relevant at run time.
Turn http://www.sleepycat.com/docs/ref/install/multiple.html into a link.
Mention that documentation remains in $PREFIX/share/doc/dbNM.
Add History section as <pre>$log</pre>. Turn RCS Id tag into
<PRE>to avoid tidy from word wrapping it.