Scott Walters


Sys::Mmap - uses mmap to map in a file as a Perl variable


    use Sys::Mmap;

    Sys::Mmap->new( my $str, 8192, '' ) or die $!;
    Sys::Mmap->new( $var, 8192 ) or die $!;

    mmap( $foo, 0, PROT_READ, MAP_SHARED, FILEHANDLE ) or die "mmap: $!";
    @tags = $foo =~ /<(.*?)>/g;
    munmap($foo) or die "munmap: $!";

    substr( $bar, 1024, 11 ) = "Hello world";

    mmap( $baz, 8192, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, STDOUT );

    $addr = mmap( $baz, 8192, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, STDOUT );
    Sys::Mmap::hardwire( $qux, $addr, 8192 );


The Sys::Mmap module uses the POSIX mmap call to map in a file as a Perl variable. Memory access by mmap may be shared between threads or forked processes, and may be a disc file that has been mapped into memory. Sys::Mmap depends on your operating system supporting UNIX or POSIX.1b mmap, of course.

Note that PerlIO now defines a :mmap tag and presents mmap'd files as regular files, if that is your cup of joe.

Several processes may share one copy of the file or string, saving memory, and concurrently making changes to portions of the file or string. When not used with a file, it is an alternative to SysV shared memory. Unlike SysV shared memory, there are no arbitrary size limits on the shared memory area, and sparse memory usage is handled optimally on most modern UNIX implementations.

Using the new() method provides a tie()'d interface to mmap() that allows you to use the variable as a normal variable. If a filename is provided, the file is opened and mapped in. If the file is smaller than the length provided, the file is grown to that length. If no filename is provided, anonymous shared inheritable memory is used. Assigning to the variable will replace a section in the file corresponding to the length of the variable, leaving the remainder of the file intact and unmodified. Using substr() allows you to access the file at an offset, and does not place any requirements on the length argument to substr() or the length of the variable being inserted, provided it does not exceed the length of the memory region. This protects you from the pathological cases involved in using mmap() directly, documented below.

When calling mmap() or hardwire() directly, you need to be careful how you use the variable. Some programming constructs may create copies of a string which, while unimportant for smallish strings, are far less welcome if you're mapping in a file which is a few gigabytes big. If you use PROT_WRITE and attempt to write to the file via the variable you need to be even more careful. One of the few ways in which you can safely write to the string in-place is by using substr() as an lvalue and ensuring that the part of the string that you replace is exactly the same length. Other functions will allocate other storage for the variable, and it will no longer overlay the mapped in file.


Maps LENGTH bytes of (the contents of) OPTIONALFILENAME if OPTIONALFILENAME is provided, otherwise uses anonymous, shared inheritable memory. This memory region is inherited by any fork()ed children. VARIABLE will now refer to the contents of that file. Any change to VARIABLE will make an identical change to the file. If LENGTH is zero and a file is specified, the current length of the file will be used. If LENGTH is larger then the file, and OPTIONALFILENAME is provided, the file is grown to that length before being mapped. This is the preferred interface, as it requires much less caution in handling the variable. VARIABLE will be tied into the "Sys::Mmap" package, and mmap() will be called for you.

Assigning to VARIABLE will overwrite the beginning of the file for a length of the value being assigned in. The rest of the file or memory region after that point will be left intact. You may use substr() to assign at a given position:


Maps LENGTH bytes of (the underlying contents of) FILEHANDLE into your address space, starting at offset OFFSET and makes VARIABLE refer to that memory. The OFFSET argument can be omitted in which case it defaults to zero. The LENGTH argument can be zero in which case a stat is done on FILEHANDLE and the size of the underlying file is used instead.

The PROTECTION argument should be some ORed combination of the constants PROT_READ, PROT_WRITE and PROT_EXEC, or else PROT_NONE. The constants PROT_EXEC and PROT_NONE are unlikely to be useful here but are included for completeness.

The FLAGS argument must include either MAP_SHARED or MAP_PRIVATE (the latter is unlikely to be useful here). If your platform supports it, you may also use MAP_ANON or MAP_ANONYMOUS. If your platform supplies MAP_FILE as a non-zero constant (necessarily non-POSIX) then you should also include that in FLAGS. POSIX.1b does not specify MAP_FILE as a FLAG argument and most if not all versions of Unix have MAP_FILE as zero.

mmap returns undef on failure, and the address in memory where the variable was mapped to on success.


Unmaps the part of your address space which was previously mapped in with a call to mmap(VARIABLE, ...) and makes VARIABLE become undefined.

munmap returns 1 on success and undef on failure.


Specifies the address in memory of a variable, possibly within a region you've mmap()ed another variable to. You must use the same precautions to keep the variable from being reallocated, and use substr() with an exact length. If you munmap() a region that a hardwire()ed variable lives in, the hardwire()ed variable will not automatically be undefed. You must do this manually.


The Sys::Mmap module exports the following constants into your namespace:


Of the constants beginning with MAP_, only MAP_SHARED and MAP_PRIVATE are defined in POSIX.1b and only MAP_SHARED is likely to be useful.


Scott Walters doesn't know XS, and is just winging it. There must be a better way to tell Perl not to reallocate a variable in memory...

The tie() interface makes writing to a substring of the variable much less efficient. One user cited his application running 10-20 times slower when Sys::Mmap->new() is used than when mmap() is called directly.

Malcolm Beattie has not reviewed Scott's work and is not responsible for any bugs, errors, omissions, stylistic failings, importabilities, or design flaws in this version of the code.

There should be a tied interface to hardwire() as well.

Scott Walter's spelling is awful.

hardwire() will segfault Perl if the mmap() area it was refering to is munmap()'d out from under it.

munmap() will segfault Perl if the variable was not successfully mmap()'d previously, or if it has since been reallocated by Perl.


CowboyTim added support for MAP_NORESERVE, MAP_HUGETLB, MAP_HUGE_2MB, and MAP_HUGE_1GB. Thanks CowboyTim!

Todd Rinaldo cleaned up code, modernized again, and merged in many fixes, 2010-2011.

Scott Walters updated for Perl 5.6.x, additions, 2002.

Malcolm Beattie, 21 June 1996.