-------------------
Written by Ted T'so
-------------------

> https://www.gnu.org/software/libtool/manual/html_node/Updating-version-info.html
>
> I understood that, if there is no interface change but some implementation
> changes, I need to bump revision. If new interface is added, for example, I
> need to bump current while revision=0 and age++.

So part of the problem here is that libtool is doing something really
strange because they are trying to use some abstract concept that is
OS-independent.  I don't use libtool because I find it horribly
complex and doesn't add enough value to be worth the complexity.

So I'll tell you how things work with respect to Linux's ELF version
numbering system.  Translating this to libtool's wierd "current,
revision, age" terminology is left as an exercise to the reader.  I've
looked at the libtool documentation, and it confuses me horribly.
Reading it, I suspect it's wrong, but I don't have the time to
experiment to confirm that the documentation is wrong and how it
diverges from the libtool implementation.

So let me explain things using the ELF shared library terminology,
which is "major version, minor version, patchlevel".  This shows up in
the library name:

	libudev.so.1.6.11

So in this example, the major version number is 1, the minor version
is 6, and the patchlevel is 11.  The patchlevel is entirely optional,
and many packages don't use it at all.  The minor number is also
mostly useless on Linux, but it's still there for historical reasons.
The patchlevel and minor version numbers were useful back for SunOS
(and Linux a.out shared library), back when there weren't rpm and dpkg
as package managers.

So many modern Linux shared libraries will only use the major and
minor version numbers, e.g:

	libext2fs.so.2.4

The only thing you really need to worry about is the major version
number, really.  The minor version is *supposed* to change when new
interfaces has changed (but I and most other people don't do that any
more).  But the big deal is that the major number *must* get bumped if
an existing interface has *changed*.

So let's talk about the major version number, and then we'll talk
about why the minor version number isn't really a big deal for Linux.

So if you change any of the library's function signatures --- and this
includes changing a type from a 32-bit integer to a 64-bit integer,
that's an ABI breakage, and so you must bump the major version number
so that a program that was linked against libfoo.so.4 doesn't try to
use libfoo.so.5.  That's really the key --- will a program linked
against the previous version library break if it links against the
newer version.  If it does, then you need to bump the version number.

So for structures, if you change any of the existing fields, or if the
application program allocates the structure --- either by declaring it
on the stack, or via malloc() --- and you expand the structure,
obviously that will cause problem, and so that's an ABI break.

If however, you arrange to have structures allocated by the library,
and struct members are always added at the end, then an older program
won't have any problems.  You can guarantee this by simply only using
a pointer to the struct in your public header files, and defining the
struct in a private header file that is not available to userspace
programs.

Similarly, adding new functions never breaks the ABI.  That's because
older program won't try to use the newer interfaces.  So if I need to
change an interface to a function, what I'll generally do is to define
a new function, and then implement the older function in terms of the
newer one.  For example:

extern errcode_t ext2fs_open(const char *name, int flags, int superblock,
			     unsigned int block_size, io_manager manager,
			     ext2_filsys *ret_fs);

extern errcode_t ext2fs_open2(const char *name, const char *io_options,
			      int flags, int superblock,
			      unsigned int block_size, io_manager manager,
			      ext2_filsys *hret_fs);

As far as the minor version numbers are concerned, the dynamic linker
doesn't use it.  In SunOS 4, if you have a DT_NEEDED for libfoo.so.4,
and the dynamic linker finds in its search path:

    libfoo.so.4.8
    libfoo.so.4.9

It will preferentially use libfoo.so.4.9.

That's not how it works in Linux, though.  In Linux there will be a
symlink that points libfoo.so.4 to libfoo.so.4.9, and the linker just
looks for libfoo.so.4.  One could imagine a package manager which
adjusts the symlink to point at the library with the highest version,
but given that libfoo.so.4.9 is supposed to contain a superset of
libfoo.so.4.8, there's no point.  So we just in practice handle all of
this in the package manager, or via an ELF symbol map.  Or, we just
assume that since vast majority of software comes from the
distribution, the distro package manager will just update libraries to
the newer version as a matter of course, and nothing special needs to
be done.

So in practice I don't bump the minor version number for e2fsprogs
each time I add new interfaces, because in practice it really doesn't
matter for Linux.  We have a much better system that gets used for
Debian.

For example in Debian there is a file that contains when each symbol
was first introduced into a library, by its package version number.
See:

https://git.kernel.org/pub/scm/fs/ext2/e2fsprogs.git/tree/debian/libext2fs2.symbols

This file contains a version number for each symbol in libext2fs2, and
it tells us what version of libext2fs you need to guarantee that a
particular symbol is present in the library.  Then when *other*
packages are built that depend on libext2fs2, the minimum version of
libext2fs can be calculated based on which symbols they use.

So for example the libf2fs-format4 package has a Debian dependency of:

Depends: libblkid1 (>= 2.17.2), libc6 (>= 2.14), libf2fs5, libuuid1 (>= 2.16)

The minimum version numbers needed for libblkid1 and libuuid1 are
determined by figuring out all of the symbols used by the
libf2fs-format4 package, and determining the minimum version number of
libblkid1 that supports all of those blkid functions.

This gets done automatically, so I didn't have to figure this out.
All I have in the debian/control file is:

Depends: ${misc:Depends}, ${shlibs:Depends}

Sorry this got so long, but hopefully you'll find this useful.  How
you bend libtool to your will is something you'll have to figure out,
because I don't use libtool in my packages.[1]

Cheers,

					- Ted


[1] If you are interested in how I do things in e2fsprogs, take a look
at the Makefile.elf-lib, Makefile.solaris-lib, Makefile.darwin-lib,
etc. here:

https://git.kernel.org/pub/scm/fs/ext2/e2fsprogs.git/tree/lib

This these Makefile fragments are then pulled into the generated
makefile using autoconf's substitution rules, here:

https://git.kernel.org/pub/scm/fs/ext2/e2fsprogs.git/tree/lib/ext2fs/Makefile.in

(Search for "@MAKEFILE_ELF@" in the above Makefile.in).

So when someone runs "configure --enable-elf-shlibs", they get the ELF
shared libraries built.  On BSD and MacOS systems they just have to
run "configure --enable-bsd-shlibs", and so on.

Personally, since most people don't bother to write truly portable
programs, as their C code is full of Linux'isms, using libtool is just
overkill, because they probably can't build on any other OS *anyway*
so libtool's slow and complex abstraction layer is totally wasted.
Might as well not use autoconf, automake, and libtool at all.

On the other hand, if you really *do* worry about portability on other
OS's (e2fsprogs builds on MacOS, NetBSD, Hurd, Solaris, etc.) then
using autoconf makes sense --- but I *still* don't think the
complexity of libtool is worth it.

= Add-on =
If you are going to be making one less major update, this is the
perfect time to make sure that data structures are allocated by the
library, and are (ideally) opaque to the calling application (so they
only manipulate structure poitners).  That is, the structure
definition is not exposed in the public header file, and you use
accessor functions to set and get fields in the structure.

If you can't do that for all data structures, if you can do that with
your primary data structure that's going to make your life much easier
in the long term.  For ext2fs, that's the file systme handle.  It's
created by ext2fs_open(), and it's passed to all other library
functions as the first argument.

The other thing you might want to consider doing is adding a magic
number to the beginning of each structure.  That way you can tell if
the wrong structure gets passed to a library.  It's also helpful for
doing the equivalent of subclassing in C.

This is how we do it in libext2fs --- we use com_err to define the
magic numbers:

	error_table ext2

ec	EXT2_ET_BASE,
	"EXT2FS Library version @E2FSPROGS_VERSION@"

ec	EXT2_ET_MAGIC_EXT2FS_FILSYS,
	"Wrong magic number for ext2_filsys structure"

ec	EXT2_ET_MAGIC_BADBLOCKS_LIST,
	"Wrong magic number for badblocks_list structure"
	...

And then every single structure starts like so:

struct struct_ext2_filsys {
	errcode_t			magic;
	...

struct ext2_struct_inode_scan {
	errcode_t		magic;
	...

And then before we use any pointer we do this:

	if (file->magic != EXT2_ET_MAGIC_EXT2_FILE)
		return EXT2_ET_MAGIC_EXT2_FILE;