p5pRT
commented
19 years ago From @nwc10
Created by @nwc10
This is definitely platform dependent - I can reproduce this on x86 Linux\, but not x86 FreeBSD or OS X.
fork and __DATA__ don't mix. I don't think that we document this anywhere. I don't know if we should\, given the amount of documentation\, but it is a subtle gotcha\, and it had never occurred to me.
The basic problem is with modules that lazily read from the DATA file handle\, particularly if they read from another package's DATA file handle lazily\, and on demand. DATA is implemented by the Perl 5 compiler leaving the program's file handle open if it encounters the __DATA__ token. All is fine and dandy\, until you fork. At which point both processes now have a (buffered) DATA file handle pointing to the same kernel file descriptor. When one reads from DATA\, the other's DATA handle moves. Underneath it.
SelfLoader breaks.
$ cat Demo.pm #!perl -w
package Demo; sub import {}; use SelfLoader; @ISA = 'SelfLoader';
1; __DATA__ sub hash { print "pig-pen: $_[0]\n"; }
$ cat demo.pl #!perl -w use strict;
use Demo;
my $pid = fork(); sleep 2 if $pid; Demo::hash ($pid);
$ perl demo.pl pig-pen: 0 Undefined subroutine Demo::hash at demo.pl line 8
I'm not sure if/how we can fix SelfLoader. I'm not sure where we should document this gotcha with DATA.
Nicholas Clark
Migrated from rt.perl.org#37119 (status was 'open')
Searchable as RT37119$