freebsd/contrib/cvs/TESTS

To run the tests:

	$ make check

Note that if your /bin/sh doesn't support shell functions, you'll
have to try something like this, where "/bin/sh5" is replaced by the
pathname of a shell which handles normal shell functions:

	$ make SHELL=/bin/sh5 check

WARNING:  This test can take quite a while to run, esp. if your
disks are slow or over-loaded.

The tests work in /tmp/cvs-sanity (which the tests create) by default.
If for some reason you want them to work in a different directory, you
can set the TESTDIR environment variable to the desired location
before running them.  In particular, using SGI's Irix 6, the tests
will fail if TESTDIR is an XFS filesystem (which /tmp often is);
you'll want to set TESTDIR to a non-XFS filesystem.

You will probably need GNU expr, which is part of the GNU sh-utils
package (this is just for running the tests; CVS itself doesn't use
expr).

If there is some unexpected output, that is a failure which can be
somewhat hard to track down.  Finding out which test is producing the
output is not always easy.  The newer tests (that is, ones using
dotest*) will not have this problem, but there are many old tests
which have not been converted.

If running the tests produces the output "FAIL:" followed by the name
of the test that failed, then the details on the failure are in the
file check.log.  If it says "exit status is " followed by a number,
then the exit status of the command under test was not what the test
expected.  If it says "** expected:" followed by a regular expression
followed by "** got:" followed by some text, then the regular
expression is the output which the test expected, and the text is the
output which the command under test actually produced.  In some cases
you'll have to look closely to see how they differ.

If output from "make remotecheck" is out of order compared to what is
expected (for example,

   a
   b
   cvs foo: this is a demo

is expected and

   a
   cvs foo: this is a demo
   b

is output), this is probably a well-known bug in the CVS server
(search for "out-of-order" in src/server.c for a comment explaining
the cause).  It is a real pain in running the testsuite, but if you
are lucky and/or your machine is fast and/or lightly loaded, you won't
run into it.  Running the tests again might succeed if the first run
failed in this manner.

For more information on what goes in check.log, and how the tests are
run in general, you'll have to read sanity.sh.  Depending on just what
you are looking for, and how familiar you are with the Bourne shell
and regular expressions, it will range from relatively straightforward
to obscure.

If you choose to submit a bug report based on tests failing, be
aware that, as with all bug reports, you may or may not get a
response, and your odds might be better if you include enough
information to reproduce the bug, an analysis of what is going
wrong (if you have the time to provide one), etc.  The check.log
file is the first place to look.

ABOUT STDOUT AND STDERR
***********************

The sanity.sh test framework combines stdout and stderr and for tests
to pass requires that output appear in the given order.  Some people
suggest that ordering between stdout and stderr should not be
required, or to put it another way, that the out-of-order bug referred
to above, and similar behaviors, should be considered features, or at
least tolerable.  The reasoning behind the current behavior is that
having the output appear in a certain order is the correct behavior
for users using CVS interactively--that users get confused if the
order is unpredictable.

ABOUT TEST FRAMEWORKS
*********************

People periodically suggest using dejagnu or some other test
framework.  A quick look at sanity.sh should make it clear that there
are indeed reasons to be dissatisfied with the status quo.  Ideally a
replacement framework would achieve the following:

1.  Widely portable, including to a wide variety of unices, NT, Win95,
OS/2, VMS, probably DOS and Win3, etc.

2.  Nicely match extended regular expressions of unlimited length.

3.  Be freely redistributable, and if possible already the kind of
thing people might have already installed.  The harder it is to get
and install the framework, the less people will run the tests.

The various contenders are:

* Bourne shell and GNU expr (the status quo).  Falls short on #1
(we've only tried unix and NT, although MKS might help with other DOS
mutants).  #3 is pretty good (the main dependency is GNU expr which is
fairly widely available).

* Bourne shell with a new regexp matcher we would distribute with
CVS.  This means maintaining a regexp matcher and the makefiles which
go with it.  Not clearly a win over Bourne shell and GNU expr.

* Bourne shell, and use sed to remove variable portions of output, and
thus produce a form that can be compared with cmp or diff (this
sidesteps the need for a full regular expression matcher as mentioned
in #2 above).  The C News tests are said to work this way.  This would
appear to rely on variable portions of output having a certain syntax
and might spuriously recognize them out of context (this issue needs
more investigation; it isn't clear how big a problem it is in
practice).  Same portability issues as the other choices based on the
Bourne shell.

* Dejagnu.  This is overkill; most of dejagnu is either unnecessary
(e.g. libraries for communicating with target boards) or undesirable
(e.g. the code which stats every file in sight to find the tests).  On
the plus side, dejagnu is probably closer than any of the other
choices to having everything which is needed already there.

* Write our own small framework directly in tcl and distribute with
CVS.  The tests would look much like dejagnu tests, but we'd avoid the
unnecessary baggage.  The only dependency would be on tcl (that is,
wish).

* perl or python or <any other serious contenders here?>