I have a Perl program that createse lots of forks (18 in parallel and about 2000 during its total lifetime). I've noticed the following exception:

Error in tempdir() using /tmp/XXXXXXXXXX: Have exceeded the maximum number of attempts (1000) to open temp file/dir at

The following short script can reproduced the problem:


use strict;
use warnings;
use File::Temp qw(tempdir);
use File::Path qw(rmtree);

rand();    # implicitly call srand()

my $f = 5;
my $n = 1000;
my $dir = "/tmp/" . time;
mkdir $dir;
print "$dir\n";

for (1 .. $f) {
    my $pid = fork();
    if (not $pid) { # in the child process
        # srand();  # explicitely set srand in the child process
        for (1 .. $n) {
            tempdir( CLEANUP => 1, DIR => $dir );

for (1 .. $f ) {
rmtree $dir;

The higher the fork-count (in $f) and the higher the repeat-count in $n the more time we encounter the error.

The problem apparently is that the "random" temporary directory (or temporary files) is just going over a list of pseudo-random numbers starting from the current srand. If the randomly created directory or filename already exists then it will try the next one, but only up to 1,000 attempts. After that it throws an exception.

In our example we called rand() in our main code which calls srand the first time it is used in a process. After that each fork continued to use the same random sequence. If we remove the call to rand() then we don't get the exception any more.

The reason is that in that case the rand() is only called in the already forked processess (by tempdir) and thus each forked process will set its own srand().

This might be a "solution", but if for any reason we later add some code to the main part of the application that calls rand then we get back to this problem.

The real solution is to call srand() explicitly in each one of the child processes. We have a commented out line doing just that.

We have actually already encountered this problem in a different setting and described in random numbers in forked processes.