rc-1.4beta

rc-1.4beta

[Byron Rakitzis]

Work on rc has been proceeding at a steady, but far reduced pace
ever since the last major release (1.2, Sept. 1991). I think it's
about time to put out another version into the various archives,
but quite a lot has happened since February (when 1.3beta was
announced to this list) so I would like to try your patience and
ask you to review what I am now calling "1.4beta".

rc unfortunately still needs some work before I can put my conscience
to rest and say that I've done everything that needs doing for
now, e.g., fixed every bug I knew about. In particular, there are
these issues in my mind:

1) editline. Last time I checked it, it was a little flaky, and if
it's going to be distributed with rc I want to make sure it's solid.
If you'll allow a little immodesty, having put hundreds of hours
of my own time into rc (not to speak of the generous help I have
received from many of you), if I am going to pair something else
with rc in an archive, I would like to make sure it is at least as
stable. This is something I just haven't gotten around to personally
overseeing yet. Maybe I shouldn't be so paranoid, but...

2) trip.rc. It needs to be improved, and work on trip.rc is just
tedious, boring and time-consuming. Here's a current list of
basic-block coverage I've got with a ~500 line rc script now:

	main.tcov:	100.00	Percent of the file executed
	list.tcov:	100.00	Percent of the file executed
	getopt.tcov:	100.00	Percent of the file executed
	glob.tcov:	98.91	Percent of the file executed
	hash.tcov:	96.52	Percent of the file executed
	match.tcov:	95.56	Percent of the file executed
	heredoc.tcov:	95.56	Percent of the file executed
	lex.tcov:	94.77	Percent of the file executed
	var.tcov:	94.67	Percent of the file executed
	fn.tcov:	94.44	Percent of the file executed
	which.tcov:	92.31	Percent of the file executed
	wait.tcov:	92.11	Percent of the file executed
	tree.tcov:	91.89	Percent of the file executed
	except.tcov:	90.32	Percent of the file executed
	builtins.tcov:	89.82	Percent of the file executed
	nalloc.tcov:	88.24	Percent of the file executed
	status.tcov:	87.50	Percent of the file executed
	exec.tcov:	87.50	Percent of the file executed
	input.tcov:	87.10	Percent of the file executed
	footobar.tcov:	83.21	Percent of the file executed
	walk.tcov:	82.64	Percent of the file executed
	glom.tcov:	80.77	Percent of the file executed
	utils.tcov:	77.78	Percent of the file executed
	signal.tcov:	76.00	Percent of the file executed
	redir.tcov:	71.88	Percent of the file executed
	print.tcov:	66.67	Percent of the file executed
	open.tcov:	66.67	Percent of the file executed

Some of these percentages will never reach 100, unless I simulate
failures to fork(), malloc(), pipe(), etc. Some of the code coverage
in trip.rc is completely superficial, and real tests need to be
devised. Other parts of the code are just completely untested (by
trip.rc). Of course, in principle such a test suite is useless
because it will never uncover all the bugs. If anyone has any
suggestions about how to make a test suite like this more useful,
I'm listening.

You could help out the trip.rc effort by running it on your Unix
and letting me know if 1) there are any bugs in the script, or 2) if
there are any bugs in rc.

Of course, if you're foolhardy enough to want to extend trip.rc
yourself, you are more than welcome!

3) race condition. There is a race condition in the fork-exec
sequence that rc uses to start a process and wait for it to
finish. Essentially, it is possible to send a SIGINT to rc
and to have it get lost. This is because rc uses the following
paradigm for spawning a process:

	fork/exec

	ignore SIGINT
	wait()
	restore SIGINT

If a SIGINT arrives after the wait() but before SIGINT's handler
is restored, the signal gets lost. The solution that /bin/sh
uses is to simply collect SIGINTs that arrive during a wait().
However, I find this unacceptable for the following reason:
a signal destined for a child is usually not destined for the
shell as well. For example, in sh:

	if true; then
		ed
		bar
	fi

If I type "^C" inside ed to abort some ed operation, let ed catch the
SIGINT, continue to edit and then exit ed normally, the command "bar"
will never get executed. Whereas,

	if (true) {
		ed
		bar
	}

does try to find bar.

So it's a dilemma: adopt poor semantics for SIGINT, or allow
a race condition into the code. I really don't know which one
is worse right now, so I'm leaving things as they are (though
I must note that 1.4beta signals are more reliable than in
previous versions---there were other problems with signals
that have been fixed in the current release).

4) FIFO's. As long as I'm picking nits, I should mention that
this (mis)feature of rc seems to cause more trouble than it's
worth. I've mostly ignored it up to now, and it's easily the
flakiest part of the shell. Part of the problem is that FIFO's
themselves are so flaky. e.g., on a Sun if you put one on a
tmpfs filesystem, the machine can crash(!). I want to do
something to clean up this mess, if possible.

Anyway, here's a list of changes since 1.3beta, and of course
rc is available in the usual way. I'm using a tar file format
because of the non-printable characters in trip.rc.

Changes since 1.3beta:

New Makefile/config.h setup.

builtin echo may now be conditionally included out, to use a Goldwynism.

builtin exit takes any legal exit status. If the status is not a single
digit, (e.g., (0 1 2), sigint, (sigiot+core 0), etc.) rc exits with 1.
rc's exit status is zero if and only if the status given to exit is all
zeros. (or $status is all zeros, of course) (BTW, having "exit sigiot"
produce a core dump would be going a little far, I think.)

limit does not append a unit after a zero limit; 0g was too confusing.

exec > /nonexistentfile does not cause rc to exit any more.

If a noninteractive rc is started with sigint ignored, rc does not install
its own signal handler.

error messages produced by rc in a subshell were cleaned up. (rc erroneously
reset the 'interactive' flag after a fork)

print.c was cleaned up a little; no functionality was changed, but should
be more portable now.

a bug in rc-1.3beta (not previous versions) was fixed: setting the first
element of $path to '' caused PATH to be exported as '':etc..

getopt's "illegal option" message was gratuitously changed to something
less abrupt.

some dead code was removed from input.c

%term was changed to %token in parse.y; apparently newer yacc's don't grok
%term any more.

a race condition in the signal handler was fixed.

the variable in for() was getting evaluated each time through the loop
(e.g., for (`{echo i;date>[1=2]} in 1 2 3)echo $i would print the date
three times). This was cleaned up.

a redundant fork() was removed from walk.c; this showed up when running
a braced command with a redirection in the background. e.g., {a;b}>c&

man pages for history and rc were cleaned up by david (thanks).

rc set SIGQUIT and SIGTERM to SIG_DFL on background jobs---even when
trying to do old-style backgrounding (i.e., don't use process groups,
just ignore SIGINT & SIGQUIT & SIGTERM).

$0 is now changed to the name of the signal when entering a signal
handler. Thus it's possible to write code like

	fn sigint sigterm sigquit {
		switch ($0) {
		case sigint
			...
		case sigterm
			...

wait with no arguments now prints the pid of any and all children
that died with a signal. e.g.,

	; wait
	25321: terminated
	25325: terminated

as opposed to

	; wait
	terminated

An error saving/restoring state in the input stream code would
cause rc to exit with the (erroneous) command:

	eval '|[a'