#!/bin/bash
i=1
while : ; do
    n=$( 1230099.pl | wc -l )
    (( n == 15 )) || break
    printf '%d\r' $((i++))
done
echo "$i:$n"
[download]

It stopped after four another iterations with only 10 lines of output.

This is really weird…

I removed the sleep commands and printed the $pos in the start message.

Both processes tell the same startposition. Nevertheless I also get non-15 lines of output every now and then. 1536 loops were okay, then the parent didn't read any lines.

My modified 123099.pl code:

#!/usr/bin/perl
use IO::Handle;
use strict;
use warnings;

my $pos= tell *DATA;

my $pid= fork();
die unless defined $pid;
if ($pid == 0) {
  my $x= IO::Handle->new();
  $x->fdopen(fileno(DATA), "r");
  seek $x, $pos, 0;
  print "Start c $pos\n";
  while (<$x>) {
    print "c: $_";
  }
  print "stop c\n";
}
else {
  my $y= IO::Handle->new();
  $y->fdopen(fileno(DATA), "r");
  seek $y, $pos, 0;
  print "Start p $pos\n";
  while (<$y>) {
    print "p: $_";
  }
  print "waiting for child\n";
  wait;
  print "stop p\n";
}
__DATA__
a
b
c
d
e
[download]

And my modified version of your test-bash:

#!/bin/bash
i=1
while : ; do
    o=$( ./1230099.pl )
    n=$( echo "$o" | wc -l )
    (( n == 15 )) || break
    printf '%d\r' $((i++))
done
echo "$i:$n"
echo "$o"
[download]

Update I've run it several times now and this is the result:

2182:      10
1504:      10
560:      10
14:      10
919:      10
360:      10
908:      10
200:      10
65:      10
108:      10
43:      10
105:      10
296:      10
345:      10
100:      10
1084:      10
1278:      10
224:      10
6:      10
8:      10
125:      10
684:      10
96:      10
8:      10
579:      10
[download]

So after an unpredictable amount of iterations (6 to 2182) the scripts prints 10 instead of 15 lines…

---

s$$([},&%#}/&/]+}%&{})*;#$&&s&&$^X.($'^"%]=\&(|?*{%
+.+=%;.#_}\&"^"-+%*).}%:##%}={~=~:.")&e&&s""`$''`"e

Your race condition is: one side of fork gets to the (buffered) read before the other side gets to its seek().

I find it easier to see what is going on by using non buffered reads (sysread). Half-offsetting (timewise) the two forks leads to consistent results.

Removing the (unnecessary) seeks()'s should also lead to consistent output size (although it could be from either the parent or child).

#!/usr/bin/perl

# https://perlmonks.org/?node_id=1230099

use strict;
use warnings;
use IO::Handle;
$| = 1;

my $pos = tell DATA;
my $pid = fork();

my $x= IO::Handle->new();
$x->fdopen(fileno(DATA), "r");

$pid and select undef, undef, undef, 0.05;
seek $x, $pos, 0;

while( sysread $x, $_, 1 )
  {
  print $pid ? uc : lc;
  select undef, undef, undef, 0.1;
  }

1 while wait > 0;

__DATA__
aa
bb
cc
dd
ee
[download]

Outputs:

aAa
bB
Cc
dD
Ee
[download]

Note the extra 'a' as a result of the seek().

What' even weirder: If I run a sequential verison of the same script, the result is, at least in the last 12104 runs, always correct.

So it's not in the code it must have to do with perl itself.

This is the "serialized" code I used:

#!/usr/bin/perl
use IO::Handle;
use strict;
use warnings;

my $pos= tell *DATA;

  my $x= IO::Handle->new();
  $x->fdopen(fileno(DATA), "r");
  seek $x, $pos, 0;
  print "Start c $pos\n";
  #sleep 4;
  while (<$x>) {
    print "c: $_"; #sleep 1;
  }
  print "stop c\n";
  my $y= IO::Handle->new();
  $y->fdopen(fileno(DATA), "r");
  seek $y, $pos, 0;
  print "Start p $pos\n";
  while (<$y>) {
    print "p: $_"; #sleep 1;
  }
  print "stop p\n";
__DATA__
a
b
c
d
e
[download]

---

s$$([},&%#}/&/]+}%&{})*;#$&&s&&$^X.($'^"%]=\&(|?*{%
+.+=%;.#_}\&"^"-+%*).}%:##%}={~=~:.")&e&&s""`$''`"e

It's not perl, it's the OS. You'd have the same problem with an already open fd in C. DATA is just an open fd on the source file.

Try a seek DATA, 0, 0; and notice you can read the whole source file.