freebsd-dev/contrib/cvs/HACKING
2004-06-10 19:05:38 +00:00

230 lines
9.7 KiB
Plaintext

How to write code for CVS
* Source
Patches against the development version of CVS are most likely to be accepted:
$ cvs -d:pserver:anoncvs@cvs.cvshome.org/cvsroot co ccvs
* Compiler options
If you are using GCC, you'll want to configure with -Wall, which can
detect many programming errors. This is not the default because it
might cause spurious warnings, but at least on some machines, there
should be no spurious warnings. For example:
$ CFLAGS="-g -Wall" ./configure
Configure is not very good at remembering this setting; it will get
wiped out whenever you do a ./config.status --recheck, so you'll need
to use:
$ CFLAGS="-g -Wall" ./config.status --recheck
* Indentation style
CVS mostly uses a consistent indentation style which looks like this:
void
foo (arg)
char *arg;
{
if (arg != NULL)
{
bar (arg);
baz (arg);
}
switch (c)
{
case 'A':
aflag = 1;
break;
}
}
The file cvs-format.el contains settings for emacs and the NEWS file
contains a set of options for the indent program which I haven't tried
but which are correct as far as I know. You will find some code which
does not conform to this indentation style; the plan is to reindent it
as those sections of the code are changed (one function at a time,
perhaps).
In a submitted patch it is acceptable to refrain from changing the
indentation of large blocks of code to minimize the size of the patch;
the person checking in such a patch should reindent it.
* Portability
The general rule for portability is that it is only worth including
portability cruft for systems on which people are actually testing and
using new CVS releases. Without testing, CVS will fail to be portable
for any number of unanticipated reasons.
The current consequence of that general rule seems to be that if it
is in ANSI C and it is in SunOS4 (using /bin/cc), generally it is OK
to use it without ifdefs (for example, assert() and void * as long as
you add more casts to and from void * than ANSI requires. But not
function prototypes). Such constructs are generally portable enough,
including to NT, OS/2, VMS, etc.
* Run-time behaviors
Use assert() to check "can't happen" conditions internal to CVS. We
realize that there are functions in CVS which instead return NULL or
some such value (thus confusing the meaning of such a returned value),
but we want to fix that code. Of course, bad input data, a corrupt
repository, bad options, etc., should always print a real error
message instead.
Do not use arbitrary limits (such as PATH_MAX) except perhaps when the
operating system or some external interface requires it. We spent a
lot of time getting rid of them, and we don't want to put them back.
If you find any that we missed, please report it as with other bugs.
In most cases such code will create security holes (for example, for
anonymous readonly access via the CVS protocol, or if a WWW cgi script
passes client-supplied arguments to CVS).
Although this is a long-term goal, it also would be nice to move CVS
in the direction of reentrancy. This reduces the size of the data
segment and will allow a multi-threaded server if that is desirable.
It is also useful to write the code so that it can be easily be made
reentrant later. For example, if you need to pass data from a
Parse_Info caller to its callproc, you need a static variable. But
use a single pointer so that when Parse_Info is fixed to pass along a
void * argument, then the code can easily use that argument.
* Coding standards in general
Generally speaking the GNU coding standards are mostly used by CVS
(but see the exceptions mentioned above, such as indentation style,
and perhaps an exception or two we haven't mentioned). This is the
file standards.text at the GNU FTP sites.
Filenames for .c and .h files may contain _ but should not contain -
(the latter causes Visual C++ 2.1 to create makefiles which Visual C++
4.0 cannot use).
* Regenerating Build Files
On UNIX, if you wish to change the Build files, you will need Autoconf and
Automake.
Some combinations of Automake and Autoconf versions may break the
CVS build if file timestamps aren't set correctly and people don't
have the same versions the developers do, so the rules to run them
automatically aren't included in the generated Makefiles unless you run
configure with the --enable-maintainer-mode option.
The CVS Makefiles and configure script were built using Automake 1.7.9 and
Autoconf 2.58, respectively.
There is a known bug in Autoconf 2.57 that will prevent the configure
scripts it generates from working on some platforms. Other combinations of
autotool versions may or may not work. If you get other versions to work,
please send a report to <bug-cvs@gnu.org>.
* Writing patches (strategy)
Only some kinds of changes are suitable for inclusion in the
"official" CVS. Bugfixes, where CVS's behavior contradicts the
documentation and/or expectations that everyone agrees on, should be
OK (strategically). For features, the desirable attributes are that
the need is clear and that they fit nicely into the architecture of
CVS. Is it worth the cost (in terms of complexity or any other
tradeoffs involved)? Are there better solutions?
If the design is not yet clear (which is true of most features), then
the design is likely to benefit from more work and community input.
Make a list of issues, or write documentation including rationales for
how one would use the feature. Discuss it with coworkers, a
newsgroup, or a mailing list, and see what other people think.
Distribute some experimental patches and see what people think. The
intention is arrive at some kind of rough community consensus before
changing the "official" CVS. Features like zlib, encryption, and
the RCS library have benefitted from this process in the past.
If longstanding CVS behavior, that people may be relying on, is
clearly deficient, it can be changed, but only slowly and carefully.
For example, the global -q option was introduced in CVS 1.3 but the
command -q options, which the global -q replaced, were not removed
until CVS 1.6.
* Writing patches (tactics)
When you first distribute a patch it may be suitable to just put forth
a rough patch, or even just an idea. But before the end of the
process the following should exist:
- ChangeLog entry (see the GNU coding standards for details).
- Changes to the NEWS file and cvs.texinfo, if the change is a
user-visible change worth mentioning.
- Somewhere, a description of what the patch fixes (often in
comments in the code, or maybe the ChangeLog or documentation).
- Most of the time, a test case (see TESTS). It can be quite
frustrating to fix a bug only to see it reappear later, and adding
the case to the testsuite, where feasible, solves this and other
problems. See the TESTS file for notes on writing new tests.
If you solve several unrelated problems, it is generally easier to
consider the desirability of the changes if there is a separate patch
for each issue. Use context diffs or unidiffs for patches.
Include words like "I grant permission to distribute this patch under
the terms of the GNU Public License" with your patch. By sending a
patch to bug-cvs@gnu.org, you implicitly grant this permission.
Submitting a patch to bug-cvs is the way to reach the people who have
signed up to receive such submissions (including CVS developers), but
there may or may not be much (or any) response. If you want to pursue
the matter further, you are probably best off working with the larger
CVS community. Distribute your patch as widely as desired (mailing
lists, newsgroups, web sites, whatever). Write a web page or other
information describing what the patch is for. It is neither practical
nor desirable for all/most contributions to be distributed through the
"official" (whatever that means) mechanisms of CVS releases and CVS
developers. Now, the "official" mechanisms do try to incorporate
those patches which seem most suitable for widespread usage, together
with test cases and documentation. So if a patch becomes sufficiently
popular in the CVS community, it is likely that one of the CVS
developers will eventually try to do something with it. But dealing
with the CVS developers may be the last step of the process rather
than the first.
* What is the schedule for the next release?
There isn't one. That is, upcoming releases are not announced (or
even hinted at, really) until the feature freeze which is
approximately 2 weeks before the final release (at this time test
releases start appearing and are announced on info-cvs). This is
intentional, to avoid a last minute rush to get new features in.
* Mailing lists
Anyone can add themselves to the following mailing lists:
dev. Unless you are accepted as a CVS developer as
described in the DEVEL-CVS file, you will only be able to
read this list, not send to it. The charter of the list is
also in DEVEL-CVS.
cvs. The only messages sent to this list are sent
automatically, via the CVS `loginfo' mechanism, when someone
checks something in to the master CVS repository.
test-results. The only messages sent to this list are sent
automatically, daily, by a script which runs "make check"
and "make remotecheck" on the master CVS sources.
To subscribe to dev, cvs, or test-results, send
a message to "<list>-subscribe@ccvs.cvshome.org" or visit
http://ccvs.cvshome.org/servlets/ProjectMailingListList and follow the
instructions there.
One other list related to CVS development is bug-cvs. This is the
list which users are requested to send bug reports to. Anyone can
subscribe; to do so send mail to bug-cvs-request@gnu.org.
Other CVS discussions take place on the info-cvs mailing list
(send mail to info-cvs-request@gnu.org to subscribe) or on
the newsgroup comp.software.config-mgmt.