
WARNING
=======

The fsopen() function in this module currently has a bug which causes it to
waste a filehandle every time it is called. Until this issue is resolved, the
sopen() function should generally be used instead.

The bug can be seen by running the following program:

	use strict;
	use warnings;

	use Win32::SharedFileOpen;

	my $file = 'test.txt';

	for (my $i = 1; $i <= 512; $i++) {
		local *FH;
		unless (fsopen(FH, $file, 'w', SH_DENYNO)) {
			print STDERR "Can't open '$file' for writing at i = $i: $!\n";
			last;
		}
		unless (close FH) {
			print STDERR "Can't close '$file' after writing at i = $i: $!\n";
			last;
		}
	}

	unless (unlink $file) {
		print STDERR "Can't delete '$file' after writing: $!\n";
	}

This program fails with the error:

	Can't open 'test.txt' for writing at i = 509:

That the error is related to a wasting of filehandles can be seen by inserting
the line:

	open FH2, '>test2.txt';

above the "for" loop. This causes the program to fail with the error:

	Can't open 'test.txt' for writing at i = 508:

The program fails one iteration sooner for every extra filehandle that is
opened before the "for" loop.

Furthermore, it is only the fsopen() function in this module that exhibits this
behaviour. If the call to fsopen() is replaced with any of the following:

	open FH, ">$file"

	sysopen FH, $file, O_WRONLY | O_CREAT | O_TRUNC

	sopen(FH, $file, O_WRONLY | O_CREAT | O_TRUNC, SH_DENYNO, S_IWRITE)

then the program completes and exits normally.

The reason why fsopen() currently wastes a filehandle every time it is called is
that in "C" programming terms it opens two filehandles at the "stdio" level
(i.e. the level dealt with by the fopen()/fclose() family of functions, which
includes _fsopen()), only one of which is properly closed later when the caller
has finished with the filehandle setup by fsopen().

Specifically, the C extension part of this module calls _fsopen(), which returns
a file stream (a "FILE *") - the first "stdio" level filehandle to be opened.
Because of the difficulty of passing this back to the Perl code (since it may be
being used for reading, writing or both), the C code instead simply returns the
corresponding file descriptor (an "int"), obtained by a call to fileno(). The
Perl code then effectively does an fdopen() on that to get a Perl filehandle -
the second "stdio" level filehandle to be opened.

The caller then reads from and/or writes to this Perl filehandle as required and
finally calls the Perl built-in function close() on it. This closes the second
"stdio" level filehandle that was opened and presumably the file descriptor with
which it was associated. However, the first "stdio" level filehandle that was
opened in the C code is never properly closed, so although it is no longer
accessible by any means, and presumably useless anyway since the underlying file
descriptor has been closed, it is now effectively wasted.

The default maximum number of filehandles that a process can have open at the
"stdio" level is 512. We can now see why the program above fails at i = 509.
The program starts with three "stdio" level filehandles - STDIN, STDOUT and
STDERR - and calls fsopen() followed by close() 508 times, each of which wastes
one "stdio" level filehandle. There are now 511 "stdio" level filehandles used.
On the 509th iteration the C code opens the 512th and last available "stdio"
level filehandle, returning the corresponding file descriptor to the Perl code.
The attempt to (effectively) fdopen() this file descriptor then fails because
there are already 512 "stdio" level filehandles in use.

Incidentally, the underlying file descriptor is not wasted whether the fsopen()
call succeeds or not, because it is always closed. If the fsopen() call succeeds
then the caller will subsequently close the Perl filehandle returned by
fsopen(), which has the effect of closing the underlying file descriptor. If the
fsopen() call fails then it is either because the _fsopen() call in the C code
failed, in which case nothing was opened in the first place, or because the
open() call subsequently done in the Perl code to (effectively) fdopen() the
same file descriptor failed. In the latter case, the Perl code explicitly closes
the file descriptor that the C code had opened (using the POSIX::close()
function) before returning failure.

The following C program does the equivalent of what the Perl program above does,
and exhibits the same behaviour for the same reason:

	#include <io.h>
	#include <stdio.h>
	#include <share.h>

	void main(void) {
		int i;
		char err[64];
		const char *file = "test.txt";

		for (i = 1; i <= 512; i++) {
			int fd;
			FILE *fp;

			if ((fd = my_fsopen(file)) == -1) {
				sprintf(err, "Can't open '%s' for writing at i = %d",
						file, i);
				perror(err);
				break;
			}

			if ((fp = fdopen(fd, "w")) == NULL) {
				sprintf(err, "Can't open '%s' for writing at i = %d",
						file, i);
				perror(err);
				if (close(fd) != 0) {
					sprintf(err, "Can't close file descriptor %d at i = %d",
							fd, i);
					perror(err);
				}
				break;
			}

			if (fclose(fp) != 0) {
				sprintf(err, "Can't close '%s' after writing at i = %d",
						file, i);
				perror(err);
				break;
			}
		}

		if (unlink(file) != 0) {
			sprintf(err, "Can't delete '%s' after writing", file);
			perror(err);
		}
	}

	int my_fsopen(const char *file) {
		FILE *fp;
		return((fp = _fsopen(file, "w", SH_DENYNO)) != NULL ? fileno(fp) : -1);
	}

The sopen() function does not suffer the same problem because it doesn't open a
"stdio" level filehandle in the C code. Instead, it opens a "lowio" level
filehandle (the level dealt with by the open()/close() family of functions) by
calling _sopen(). This is simply an "int" file descriptor, which is returned to
the Perl code to be (effectively) fdopen()'d as before. When the corresponding
Perl filehandle obtained thus is closed by the caller this "lowio" level
filehandle is closed, so these "lowio" level filehandles, which a process may
have upto 2048 of, are not wasted either.
