No IPFW binary compat across versions ?

[Bruce Evans]

On Mon, 5 Sep 2011, Arnaud Lacombe wrote:

> On Mon, Sep 5, 2011 at 7:50 PM, Adrian Chadd <adrian at freebsd.org> wrote:
>> That's not what the COMPAT_* hooks are for.
>>
>> They're for backwards compatibility of normal userland binaries, not
>> binaries which use a FreeBSD-specific kernel ABI.
>>
> What do you define as "normal" and where do you draw the line ?

It's hard to say, and probably undocumented, but system-y programs
should be the only unnormal ones.  Until about 10 years ago, anything
using kinfo broke every other day with a binary change to struct
kinfo_proc, but this was mostly fixed so kinfo only breaks every
couple of major releases.  Using sysctls instead of kmem significantly
increased the portability of statistics utilities like systat and
netstat, so they only break across every 1 and a half major releases
(new features don't work of course).

I run kernels between FreeBSD-4 and FreeBSD-9-current-6-months
ago under a FreeBSD-5 userland with only the following changes:
- hack signal handling for FreeBSD-4 compatibility in otherwise
   FreeBSD-5 userland.
- static linkage throughout (except in ports, which are mostly not
   not even installed except under FreeBSD-5 (FreeBSD-4 ports)).
   No dll hell.
- scripts to select a working binary.  E.g., for ps:

% #!/bin/sh
% 
% case "$(uname -m)-$(uname -r)" in
% amd64-8.*-*)	p=/c/tmp/rela/;;
% i386-9.*-*)	p=/c/tmp/rel9/;;
% i386-8.*-*)	p=/c/tmp/relc/;;
% i386-[6-7].*-*)	p=/c/tmp/rel6/;;
% *)		p=/;;
% esac
% exec "$p"bin/ps "$@"

I only need a few binaries to work on all kernels, and add a script like
the above, or lines to an old script as necessary.  The above shows the
FreeBSD-6 ps working for FreeBSD-5 and FreeBSD-7 on i386 and ps only
needed under FreeBSD-8 for amd64.  My list of system-scripts is:

% atacontrol
% bsdtar	# not really system-y; minor portability problems
% fdisk
% fstat
% ifconfig
% kdump
% killall
% mdconfig
% mount		# mostly not needed, and also not used for mounting root
% mount_msdosfs	# all mount utils mostly not needed, and never should be
% mount_ntfs
% netstat	# this breaks too often
% nfsstat
% pmccontrol	# FreeBSD-5 doesn't even have pmc
% pmcstat	# but pmc changed within major releases and I only have old vers
% ps
% pstat
% rpc.lockd
% systat	# surprisingly binary-compatible (same binaries for 6-7 and 8-9)
% top		# as binary-compatible as systat
% vmstat	# even more binary-compatible (6 binary for i386 works on 6-9)
% w

>> From my point of view, I should be able to run a FreeBSD 9.0 kernel
> (when released) on top of a FreeBSD 5 userland without such issues.
> That's what backward compatibility is. _Every_ piece of the ABI
> exposed by the kernel[0] should be kept compatible, even funky
> behavior userland might ends up relying on.

Keeping the ABI compatible is not easy, and mostly wasn't attempted
except for simple syscalls in FreeBSD until about 10 years ago.

/etc/rc.d stuff tends to be a bit broken.  No new features anyway.  I
mostly use FreeBSD >5 in my amd64 laptop which could use some new
features, but I avoid some problems by not using any special power
saving kernel or userland features on it at all.  This laptop mostly
gets used for benchmarking new kernels and libm.  It nfs-mounts /c
from an i386 FreeBSD-5 system, and gets some amd64 binaries from there.
(I should use i386 binaries nfs/mounted from the FreeBSD-5 system more,
but haven't investigated this.  Only much the same system-y binaries
should have a machine-dependence that stops the i386 version from
working.)

> [0]: ... with very few exception, such are security issue when there
> is no other choices available.

My old userland must have many security problems.

Bruce