The London Perl and Raku Workshop takes place on 26th Oct 2024. If your company depends on Perl, please consider sponsoring and/or attending.


Proc::FastSpawn - fork+exec, or spawn, a subprocess as quickly as possible


   use Proc::FastSpawn;

   # simple use
   my $pid = spawn "/bin/echo", ["echo", "hello, world"];
   waitpid $pid, 0;

   # with environment
   my $pid = spawn "/bin/echo", ["echo", "hello, world"], ["PATH=/bin", "HOME=/tmp"];

   # inheriting file descriptors
   pipe R, W or die;
   fd_inherit fileno W;
   my $pid = spawn "/bin/sh", ["sh", "-c", "echo a pipe >&" . fileno W];
   close W;
   print <R>;


The purpose of this small (in scope and footprint) module is simple: spawn a subprocess asynchronously as efficiently and/or fast as possible. Basically the same as calling fork+exec (on POSIX), but hopefully faster than those two syscalls.

Apart from fork overhead, this module also allows you to fork+exec programs when otherwise you couldn't - for example, when you use POSIX threads in your perl process then it generally isn't safe to call fork from perl, but it is safe to use this module to execute external processes.

If neither of these are problems for you, you can safely ignore this module.

So when is fork+exec not fast enough, how can you do it faster, and why would it matter?

Forking a process requires making a complete copy of a process. Even thought almost every implementation only copies page tables and not the memory itself, this is still not free. For example, on my 3.6GHz amd64 box, I can fork a 5GB process only twenty times a second. For a real-time process that must meet stricter deadlines, this is too slow. For a busy and big web server, starting CGI scripts might mean unacceptable overhead.

A workaround is to use vfork - this function isn't very portable, but it avoids the memory copy that fork has to do. Some systems have an optimised implementation of spawn, and some systems have nothing.

This module tries to abstract these differences away.

As for what improvements to expect - on the 3.6GHz amd64 box that this module was originally developed on, a 3MB perl process (basically just perl + Proc::FastSpawn) takes 3.6s to run /bin/true 10000 times using fork+exec, and only 2.6s when using vfork+exec. In a 22MB process, the difference is already 5.0s vs 2.6s, and so on.


All the following functions are currently exported by default.

$pid = spawn $path, \@argv[, \@envp]

Creates a new process and tries to make it execute $path, with the given arguments and optionally the given environment variables, similar to calling fork + execv, or execve.

Returns the PID of the new process if successful. On any error, undef is currently returned. Failure to execution might or might not be reported as undef, or via a subprocess exit status of 127.

$pid = spawnp $file, \@argv[, \@envp]

Like spawn, but searches $file in $ENV{PATH} like the shell would do.

fd_inherit $fileno[, $on]

File descriptors can be inherited by the spawned processes or not. This is decided on a per file descriptor basis. This module does nothing to any preexisting handles, but with this call, you can change the state of a single file descriptor to either be inherited ($on is true or missing) or not $on is false).

Free portability pro-tip: it seems native win32 perls ignore $^F and set all file handles to be inherited by default - but this function can switch it off.


On POSIX systems, this module currently calls vfork+exec, spawn, or fork+exec, depending on the platform. If your platform has a good vfork or spawn but is misdetected and falls back to slow fork+exec, drop me a note.

On win32, the _spawn family of functions is used, and the module tries hard to patch the new process into perl's internal pid table, so the pid returned should work with other Perl functions such as waitpid. Also, win32 doesn't have a meaningful way to quote arguments containing "special" characters, so this module tries it's best to quote those strings itself. Other typical platform limitations (such as being able to only have 64 or so subprocesses) are not worked around.


 Marc Lehmann <>