avoid creating new files during make install

[p5sagit/p5-mst-13.2.git] / README.cygwin

If you read this file _as_is_, just ignore the funny characters you
see. It is written in the POD format (see pod/perlpod.pod) which is
specially designed to be readable as is.

=head1 NAME

README.cygwin - notes about porting Perl to Cygwin

=head1 SYNOPSIS

=over

=item Cygwin

The Cygwin tools are ports of the popular GNU development tools for
Windows NT, 95, and 98.  They run thanks to the Cygwin library which
provides the UNIX system calls and environment these programs expect.
More info about this project can be found at its home page
http://sourceware.cygnus.com/cygwin/

=item libperl.dll

These instructions and the default cygwin hints build a shared
libperl.dll Perl library and enables dynamically loaded extensions.

=back

=head1 BUILDING

=head2 Prerequisites

=over

=item Cygwin b20.1

The latest stable Cygwin suite is beta20.1.  It may be
downloaded from ftp://go.cygnus.com/pub/sourceware.cygnus.com/cygwin/latest/
or many mirror sites around the world.

=item egcs-1.1.2

This port was built with egcs-1.1.2 downloaded from
ftp://ftp.xraylith.wisc.edu/pub/khan/gnu-win32/cygwin/egcs-1.1.2/

=item install executable

To make life easier, you should download
ftp://ftp.franken.de/pub/win32/develop/gnuwin32/cygwin/porters/Humblet_Pierre_A/install-cygwin-b20.sh,
and use it as your install "executable." Just follow the instructions
that are embedded as comments in the .sh file.

=item Windows NT notes

You should execute a 'chmod -R +w *' on the entire perl source directory. 
The configuration process creates new files (and thus needs write access
to the directory) and sometimes, especially if you make repeated builds
in the same directory, overwrites old files. If you do not enable write
access, you're just asking for trouble. Reminder for B20.1: unless 'ntea' 
is included in the CYGWIN environment variable settings, chmod has no 
effect. See "environment," below.

It is best if you build, test, and install as a normal user, not as
Administrator or as any member of the Administrators group. There is
a well-known NT-ism that affects Cygwin: all files that are created
by any member of the Administrators B<group> are B<NOT> owned by 
that member. The ownership of those files is assigned to the 
Administrators group, instead. If the default access mode for new files
is -rw-r--r--, then the original creator of the file cannot overwrite 
it: he no longer owns the file, no B<user> does. It is owned by the 
group, but group members don't have write access to it. This causes 
any number of problems, including make test / perl harness failures, 
installation failures, etc.

In some cases, however, it is necessary to install as Administrator. For 
instance, if normal users are not allowed write access to the install 
directory. My solution, in this case, was to transfer ownership of the 
install directory tree (/usr/local) to a single, normal user, and 
set permissions to -rw-r--r-- (drwxr-xr-x for directories, of course). 
If you read the preceeding paragraph carefully, you might suspect that 
changing the permissions on the entire tree to -rw-rw-r--, but allowing 
the Administrators group to keep ownership should solve the problem. 
However, newly created directories (and the perl install creates a lot 
of them) will not allow group write access. Setting umask will not 
fix this problem, because umask is a B<negation> operator; it only 
specifies the types of accesses that will NOT be allowed on new files. 
For instance, umask u=rwx,g=rwx,o=rx means that world ('B<o>thers') will 
never be allowed write access, but owner ('B<u>ser') and B<g>roup B<might> 
be allowed write access. Everybody (u, g, and o) B<might>be allowed 
read access.

In any case, Corinna Vinschen's ntsec patches B<may> eventually 
alleviate this whole mess, and are included in the development 
snapshots as of 24 May 1999. You will need to include 'ntsec' as 
one of the items in the CYGWIN variable setting. However, initial 
tests indicate some incompatibility the 0524 snapshot and this perl
build.

=item environment

I (csw) found the following steps necessary for a successful build:

=over

=item path

I set my path so that none of the windows directories showed up -
otherwise Configure found the wrong executables (find, grep, etc).
It is, however, important that '.' be in the path, because otherwise
the build process can't execute the ld2 script that is created.

=item mounts

I had to unmount my f: drive. I have cygwin installed under 
F:\cygnus\cygwin-b20\, which is mounted as \. I also ordinarily 
have F:\ mounted as /f (i.e. mounted onto the empty directory 
F:\cygnus\cygwin-b20\f\ ). However, this causes Configure to 
"locate" the awk, tr, sed, etc. programs at 
/f/cygnus/cygwin-b20/usr/bin instead of /usr/bin. 
This ended up causing problems.

I built and tested perl using all binary mounts. However, Eric Fifer 
has built and tested it using text mounts, but reported more failures
during make test and perl harness. Based on his findings, and experiments
performed by Sebastien Barre with the static build of perl, I can 
report that these test failures are B<not> due to any differences in
the perl executable. Most of the failures encountered during a make test
on text mounts can be eliminated by remounting as binary, and re-running
the tests using the same executable. These test failures are due to 
problems in the test scripts, not the executable. See Appendix.

One observation from experience with the static build of perl is that
it's a bad idea to a mix a perl executable that was compiled using binary
mounts with modules compiled using text mounts, and vice versa. Make
sure your mount environment matches. This observation has not been
confirmed with respect to the dynamically linked build of perl. 

=item environment variables

For NT users, the CYGWIN variable should include the 'ntea' setting. 
However, if you have FAT drives on your system, as opposed to NTFS, 
please read the Cygwin FAQ concerning ntea before including it in 
your system settings. If you do not use ntea, you will encounter a 
few extra make test and/or perl harness failures. These are not 
indicative of a faulty perl executable, but only that your system 
settings do not allow the types of file access and ownership checking 
that the test scripts are attempting to verify. See Appendix.

I unset INCLUDE and LIB (these two variables are set by MSVC5, and 
inherited from my Windows environment by cygwin). I'm not sure this made 
a difference, but it has caused problems in the past...

=back

=item crypt library

http://miracle.geol.msu.ru/sos/ points to two different crypt
libraries ported to cygwin.  This has been tested with the libcrypt.tgz
by Andy Piper. His home page can be found at
http://www.xemacs.freeserve.co.uk/

=item hacks that should be revisited after the next cygwin release

Some of the failures we encountered when running make test and/or perl harness
are due to bugs in the cygwin b20.1 distribution. We sometimes found it
necessary to use dirty little hacks to persuade make test and perl harness
to play nicely. Since cygwin is in active development, many of these hacks 
may not be necessary in the future. These include:

=over

=item fix for pragma/locale

the line '#undef MB_CUR_MAX' was added to ex/POSIX/POSIX.xs. This fixes
a failure in the pragma/locale.t test, which before this fix resulted in a 
coredump. It appears that MB_CUR_MAX is #defined __mb_cur_max, and __mb_cur_max
is declared 'extern' in Cygwin b20.1's stdlib.c, but is never defined. Thus,
the error.

=item fix for lib/io_sock

there is a rather extensive patch to t/lib/io_sock.t which works around
a failure related to fork() in the cygwin environment. Cygwin b20.1 does not 
properly remap manually loaded DLLs in the child after a fork.

=item fix for lib/filehand

during the make test/perl harness steps, a win32 popup complains about 
a "perl.exe Application Error - illegal memory access." This is due to to
a test in t/lib/filehand.t, and is related to the fork + dll problem.

=item fix for environ

there are a number of changes to miniperlmain.c, util.c, and mg.c that
are there to work around a Cygwin problem relating to environ.

=item fix for lib/posix

the following line was added to t/lib/posix.t to work around a Cygwin bug.

=begin text

kill 'CONT', $$ if($^O =~ /cygwin/); # XXX: Cygwin bug INT signal gets stuck

=end text

=back

=back

=head2 Configure

Check hints/cygwin.sh for any system specific settings.  In
particular change libpth to point to the correct location of
...../i586-cygwin32/lib.

run "sh Configure".  

When confronted with this prompt:

=begin text

        First time through, eh?  I have some defaults handy for the
        following systems:
           .
           .
           .
        Which of these apply, if any?

=end text

select "cygwin".

Do not use the malloc that comes with perl--using the perl malloc
collides with some cygwin startup routines. 

=head2 make

Run "make". If you're really feeling adventurous, type 
"make 2>&1 | tee make-log.txt".

=over

=item ld2

The make script will install ld2 into your $installbin directory (i.e.
wherever you said to put the perl.exe) during the *make* process. It 
does not wait until the *make install* process to install the ld2 script.
This is because the remainder of the make refers to ld2 without fully
specifying its path, and does this from multiple subdirectories (so ./ld2
won't work.) The assumption here is that $installbin is in your current
$PATH. If this is not the case, or if you do not have an install 
executable, the make will fail at some point. Don't panic. Just manually
copy ld2 from the source directory to someplace in your path.

This cannot be done prior to make, because ld2 is created during the
make process.

=back

=head2 make test

Run "make test" to see how stable your system is. I (csw) got the 
following errors/warnings:

=over

=item op/taint

Got two "missing cygwin1.dll" warning popups. This is because
op/taint wants cygwin1.dll to be somewhere in the system path 
(\WINDOWS\SYSTEM, etc) or in the build directory. Can be ignored.

=item lib/filehand

Got an "Application Error - memory could not be read" popup. While 
this looks alarming, it can be ignored - just click OK. It is 
because Cygwin B20.1 doesn't properly remap manually loaded DLLs 
in the child after a fork.

=item lib/io_sock

Got an "Application Error - memory could not be read" popup. Again,
just click OK and ignore it.

=back

=head2 perl harness

Once you've run make test, then cd into the t/ subdirectory and
execute './perl harness'. I (csw) got the following results:

=over

=item op/taint

Got four "missing cygwin1.dll" warning popups. Click OK and
ignore.

=item lib/filehand

Got an "Application Error - memory could not be read" popup. Again,
click OK and ignore.

=item lib/io_sock

Got an "Application Error - memory could not be read" popup. Again,
clock OK and ignore.

=item final results

After the ./perl harness, I got the following results summary.

 Failed Test  Status Wstat Total Fail  Failed  List of failed
 -------------------------------------------------------------------------------
 op/taint.t                  149    3   2.01%  1, 3, 31
 9 tests skipped, plus 35 subtests skipped.
 Failed 1/190 test scripts, 99.47% okay. 3/6452 subtests failed, 99.95% okay.

=back

=head2 make install

Finally, run "make install". In my case, the install process was unable
to copy the files from <SRCDIR>/pod/* to /usr/local/lib/perl5/5.00503/pod/.
I (csw) just copied them by hand after the install finished. I believe
this is because I'm using Sergey Okhapkin's coolview version of the
cygwin1.dll, which provides case sensitivity. The directory is created
as "/usr/local/lib/perl5/5.00503/Pod" but the copy is done into
"/usr/local/lib/perl5/5.00503/pod". This fails -- but probably won't 
fail if you're using the default cygwin1.dll.

=head1 BUGS

A lot of warnings about incompatible pointer types and comparison
lacking a cast.  This is because of __declspec(dllimport).

Upon each start, make warns that a rule for perlmain.o is overrided.
Yes, it is.  In order to use libperl.dll, perlmain needs to import
symbols from there.  According to alex smishlajev, there seems to be 
no better solution than adding an explicit define to the rule.

make clean does not remove library .def and .exe.core files.

ld2 script is installed with reference to source directory.  You should
change this to /usr/local/bin (or whatever) after install.

.bat wrappers for installed utility scripts are not made during installation.

=head1 AUTHOR

alexander smishlajev <als@turnhere.com>

=head1 DISCLAIMER

I (alex) am not going to maintain this document or this port.  I only wanted
to make perl porting a bit easier.  If failed, I can't be helpful for you.
Contact one of the others listed in the history section.

=head1 HISTORY

=over

=item Release 1.4.1: 28-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

   Configure               minor fix for spaces in $PATH
   documentation updates

=item Release 1.4: 26-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

 From Eric Fifer:
   hints/cygwin32.sh       -L. and --export-dynamic not needed
   cygwin32/Makefile.SHs   no value needed for -DUSEIMPORTLIB
   t/lib/io_sock.t         -I../lib so "make test" works
   t/lib/posix.t           workaround a Cygwin bug so test works
   doio.c/perl.h           cleanup gcc warning "doio.c:789: warning:
                                pointer/integer type mismatch in 
                                conditional expression"
 From Charles Wilson:
   Configure               changes to findhdr script
   documentation updates
   built binary kit for release

=item Release 1.3: 26-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

 Changes to Cwd.pm to correct lib/findbin.t test failure from Eric Fifer
 Changes to t/op/magic.t to correct a test failure from Eric Fifer
 Changes to miniperlmain.c, util.c, and mg.c to correct t/op/magic.t #29
    test failure, from Eric Fifer
 more documentatino updates, patch merging, and a change to 
    cygwin/Makefile.SHs -- cw.
 99.95% okay!!!

=item Release 1.2: 25-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

 fixes for lib/io_sock and pragma/locale from Eric Fifer
 fixes for Configure, Makefile.SH, and cygwin32/Makefile.SHs from
    alex smishlajev
 documentation updates, and other fixes to the fixes from cw.
 99.91% okay!!!

=item Release 1.1: 21-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

minor change to Configure script, reversed a few changes made by
alexander's patch (made DOSISH #undefined again) and moved code
by alexander from dosish.h to perl.h. Reversed a change in 
pp_hot.c

=item Release 1.0: 16-May-1999

Charles Wilson  -  cwilson@ece.gatech.edu

Merged alexander's patch and Eric's patch into a single
monolithic patch. Minor cleanup. Built binary for distribution.
perl5.005_03-dynamic-patch-v1.0

=item Pre-release 3: 12-May-1999 

Eric Fifer  -  efifer@sanwaint.com

Removed all references to the impure_ptr hack since it is no longer
needed.  Some minor cleanup of alexander smishlajev's work and a few
bug fixes.

=item Pre-release 2 (initial dynamic build): 17..25-apr-1999

alexander smishlajev  -  als@turnher.com

perl 5.005_03.  cygwin b20.1  egcs 1.1.2.  far 1.60.  nescafe classic.

=item Pre-release 1 (static build): 5-Mar-1999

Charles Wilson  -  cwilson@ece.gatech.edu

Collected various patches that had been floating around the net, along
with build instructions.  Original authorship credit for those patches
goes to:

 Steven Morlock  -  newspost@morlock.net
 Sebastien Barre -  Sebastien.Barre@utc.fr
 Teun Burgers    -  burgers@ecn.nl

Created a monolithic patchkit (perl5.005_03-static-patch) and build
instructions for cygwin (beta 20.1). Also created a binary distribution
of the resulting static perl build.

=back

=head1 APPENDIX

Perl harness results from Eric Fifer, under various environments. The same
executable was used in all cases. The last item is a different executable
on a different machine, built by Charles Wilson.

There are a number of very good questions one could ask about anomalies 
in the test results presented below. "Why do the last two show different
results?" "Why did op/stat.t #18 pass in the first two tests and fail in
the third?" Short answer: I don't know. Long answer: I really have no
idea.

=over

=item text mounts, no 'ntea'

 Failed Test  Status Wstat Total Fail  Failed  List of failed
 ------------------------------------------------------------
 lib/anydbm.t      2   512    12    8  66.67%  5-12
 lib/sdbm.t        2   512    18   15  83.33%  2, 5-18
 op/split.t                   25    1   4.00%  11
 op/stat.t                    58    3   5.17%  9, 19, 26
 op/taint.t                  149    3   2.01%  1, 3, 31

=item binary mounts, no 'ntea'

 Failed Test  Status Wstat Total Fail  Failed  List of failed
 ------------------------------------------------------------
 lib/sdbm.t                   18    1   5.56%  2
 op/stat.t                    58    3   5.17%  9, 19, 26
 op/taint.t                  149    3   2.01%  1, 3, 31

=item binary mounts, 'ntea'

 Failed Test  Status Wstat Total Fail  Failed  List of failed
 ------------------------------------------------------------
 op/stat.t                    58    3   5.17%  18-19, 26
 op/taint.t                  149    3   2.01%  1, 3, 31

=item binary mounts, ntea (csw build)

 Failed Test  Status Wstat Total Fail  Failed  List of failed
 -------------------------------------------------------------------------------
 op/taint.t                  149    3   2.01%  1, 3, 31

=back

=cut