NAME
IO::FD - Faster accept, socket, listen with file descriptors, not handles
SYNOPSIS
Create and bind a STREAM socket (server):
use IO::FD;
use Socket ":all";
die "Error creating socket"
unless defined IO::FD::socket(my $listen_fd, AF_INET, SOCK_STREAM, 0);
my ($err, @sockaddr)=addrinfo "0.0.0.0", 80, {
family=> AF_INET,
socktype=> SOCK_STREAM,
flags=> AI_NUMERICHOST|AI_PASSIVE
};
die "Error binding"
unless defined FD::IO::bind($listen_fd, $sockaddr[0]{addr});
die "Error accepting"
unless defined IO::FD::accept(my $client_fd, $listen_fd);
#read and write hereCreate and connect a STREAM socket(client):
use IO::FD;
use Socket ":all";
die "Error creating socket"
unless defined IO::FD::socket(my $fd, AF_INET,SOCK_STREAM,0);
my ($err,@sockaddr)=addrinfo "127.0.0.1", 80, {
family=> AF_INET,
socktype=> SOCK_STREAM,
flags=> AI_NUMERICHOST
};
die "Error connecting";
unless defined FD::IO::connect($fd, $sockaddr[0]{addr});
#read and write hereOpen a file
use IO::FD;
use Fcntl;
die "could not open file"
unless defined IO::FD::sysopen(my $fd, "path.txt", O_RDONLY);Read/Write/Close an fd
use IO::FD;
my $fd; #From IO::FD::socket, IO::FD::accept IO::FD::sysopen, POSIX::open
die "Error writing"
unless defined IO::FD::syswrite $fd, "This is some data"; #Length and optional offset
die "Error reading"
unless defined IO::FD::sysread $fd, my $buffer, $length);
die "Error closing"
unless defined IO::FD::close $fd;Advanced:
fcntl
sendfile
accept4
accept_multiple
#TODO:
ioctl...DESCRIPTION
IO::FD is an XS module implementing common core Perl system I/O functions to work with file descriptors instead of Perl file handles. Functions include but are not limited to accept, connect, socket, bind, sysopen, sysread, and syswrite.
Many non core system functions such as sendfile, dup and mkstemp, pread, pwrite, mkfifo which work with file descriptors are also implemented.
Additional support for streamlined connection accepting is included via accept_multiple.
This module can significantly lower memory usage per file descriptor and decrease file/socket opening and socket accepting times. accept performance is particularly improved with much higher connection handling rates for a given backlog.
Actual byte throughput (read/write) is basically unchanged compared to the core Perl sysread/syswrite. Please see the PERFORMANCE section later in this document
The supported interfaces mostly resemble the core Perl implementations of similarly named functions. The largest difference being you should check for defined values for error checks. For example:
#Perl:
sysopen(my $file_handle, ...) or die $!;
sysread($file_handle, ...) or die $!;
#IO::FD
defined IO::FD::sysopen(my $file_descriptor, ...) or die $!
defined IO::FD::sysread($file_descriptor, ...) or die $!;This modules IS NOT intended to be a drop in replacement for core IO subroutines in existing code. If you want a 'drop in replacement' please look at IO::FD::DWIM which is part of the same distribution.
Currently this module is focused on UNIX/Linux systems, as this is the natural habitat of a file descriptor.
IMPORTANT VERSION DIFFERENCES
v0.3.0 and later
New functions:
pread, pwrite, mkfifo, mkfifoat, open, openat
v0.2.0 and later
New functions:
accept4, accept_multiple, sendfile
Changes:
All functions creating a new fd now behave more perlish and apply O_CLOEXEC if larger than $^F to prevent fd leakage. This may result in an extra system call you didn't need if your program never calls exec. To disable this, increase the value of $^F as per normal.
Functions now throw exceptions when output variables (fds) are read only when they need to be writable. This matches Perl behaviour in the same scenario for sysopen etc.
When function input fd variables doesn't look like an fd (an IV), a warning 'IO::FD::xxxx called with something other than a file descriptor' is generated, return value is undef and the $! variable is set to EBADF (bad file descriptor>). This is analogous to Perl behaviour when checking for valid GLOB/refs with sysread and friends.
WHERE SHOULD I USE THIS MODULE?
Networking ... Oh Yes
Socket centric programs will benefit greatly from this module. The process of socket creation/opening/accepting/listening, where it is INET/INET6 or UNIX families is much improved.
Slurp entire file ... Yes
If a file can be loaded completely into memory for processing, this module will provide improved opening and closing times. Any decoding and line processing will need to be done manually
Line Processing ... Hmmm, No
General text file line processing is best left to Perl file handles. File handles do the heavy lifting of line splitting, EOL handling, encodings, which this modules does not implement.
You can do it, but it is not in the scope of this module.
LIMITATIONS
Perl does a lot of nice things when working with files and handles. When using file descriptors directly you will loose:
- Buffering for file small read/write performance (via print and <FH>)
- Automatic close when out of scope
- Special variables not supported (ie '_' in stat)
- <FH> 'readline' support
- IO::Handle inheritance
MOTIVATION
Perl makes working with text files easy, thanks to file handles. Line splitting, UTF-8, EOL processing etc. are awesome and make your life easier.
However, the benefits of file handles when working within a network or binary file context are not so clear cut. All the nice line ending and encoding support doesn't help in these scenarios.
In addition, the OS kernel does a lot of buffering for networking already. Do we really need to add more?
So if these features are not being fully utilised for binary/network programming, the hypothesis is that opening and accepting operations would be faster with file descriptors as less setup is required internally.
APIs
Each of the APIs mimic the Perl counterpart, if applicable, as much as possible. Unless explicitly mentioned, they should operate like built in routines. Consult perldoc -f FUNCTION for details. The general exception however is return values should be explicitly tested for definedness and not relying on a 'true' value
As none of these functions are exported, they must be called with full package name.
Socket Manipulation
IO::FD::socket
IO::FD::socketpair
IO::FD::bind
IO::FD::listen
IO::FD::accept
IO::FD::accept4
my $ok=defined IO::FD::accept4 $new, $listen, $flags;
Constants: IO::FD::SOCK_NONBLOCK, IO::FD::SOCK_CLOEXECImplements the linux accept4 syscall. On non linux systems this is emulated by calling fcntl to set the FD_CLOEXEC flag and O_NONBLOCK status. Returns undef on error, for the user to test $!.
The flags argument can be the bitwise or'ed value of SOCK_NONBLOCK and SOCK_CLOEXEC from the Socket module on linux and bsd. Darwin (macos) does not have these values. so please use IO::FD::SOCK_NONBLOCK, IO::FD::SOCK_CLOEXEC on that platform.
NOTE:Unlike other functions returning new file descriptors, this DOES NOT automatically apply the CLOEXEC flag. The $flags argument must be set accordingly to achieve this.
NOTE: On emulated systems, any errors reported are only from the accept call, not subsequent fcntl calls
IO::FD::accept_multiple
my @new_fds;
my @peers;
my $count=accept_multiple(@new_fds, @peers, $listen_sock);NOTE: DO NOT use this function on a blocking socket!!
Accepts as many new connection sockets as available. The new sockets are stored in new_fds, which is an array, not a array ref. The corresponding peers to the connections are stored in @peers, also an array not a reference.
$listen_sock is the file descriptor from which the sockets are accepted from. It MUST be configured for non blocking operation, otherwise your program will just loop forever in this function
Because this function will only works for non blocking listening sockets, the sockets/fds returned are configured for non blocking mode also. On BSD type systems the socket will already be non blocking. On linux systems the accept4 call is used to set the SOCK_NONBLOCK flag.
Returns the number of sockets accepted until an error condition occurred. Returns undef if no sockets where accepted. Check the $! for normal non blocking error codes.
IO::FD::connect
IO::FD::getsockopt
IO::FD::setsockopt
Note: Implements the integer shorthand as per perldoc -f setsockopt
IO::FD::getpeername
IO::FD::getsockname
IO::FD::sendfile
sendfile $socket, $source_fd, $length, $offsetCalls system sendfile. Returns undef on error or the number of bytes transferred otherwise. The error might be an EAGAIN for non blocking sockets. Please reference the manual page for sendfile on your system, but be mindful the position of the arguments might not match.
Currently advanced header/trailer features of BSD sendfile are not supported.
File Maniupulation
IO::FD::sysopen
IO::FD::sysopen4
Same as IO::FD::sysopen, but expects all four arguments
IO::FD::open
Binding to open. Please see your system manual. If no mode is specified, the 'perlish' 0666, is used.
IO::FD::openat
Binding to openat. Please see your system manual. If no mode is specified, the 'perlish' 0666, is used.
IO::FD::mktemp
Behaves similar to File::Temp::mktemp
Requires at least six 'X' characters at the end of the template
The template string used as input is modified and is the same as the return value on success
NOTE: This function does not return a file descriptor. It might be included in future versions of this module
IO::FD::mkstemp
Behaves like File::Temp::mkstemp
Requires at least six 'X' characters at the end of the template
In list context returns ($fd,$path), where $fd is the already open file descriptor, and $path is the unique path generated from the template.
The template string used as input is modified and is the same as the $path return value on success
IO::FD::sysseek
Pipes
IO::FD::pipe
IO::FD::syspipe
A alias of IO::FD::pipe.
mkfifo
Binding to mkfifo. Please see your system manual. If no mode is specified, the 'perlish' 0666, is used.
mkfifoat
Binding to mkfifoat. Please see your system manual. If no mode is specified, the 'perlish' 0666, is used.
Common
IO::FD::dup
IO::FD::dup2
IO::FD::close
IO::FD::recv
IO::FD::send
IO::FD::sysread
NOTE: Versions prior to 0.1.4 would end up using fd = 0 (normally STDIN) when it was non numeric. This is fixed in 0.1.4. An fd which is not numeric will cause an immediate return of undefined.
IO::FD::sysread3
Same as IO::FD::sysread, but expects only 3 of 4 arguments
IO::FD::sysread4
Same as IO::FD::sysread, but expects all four arguments
IO::FD::syswrite
NOTE: Versions prior to 0.1.4 would end up using fd = 0 (normally STDIN) when it was non numeric. This is fixed in 0.1.4. An fd which is not numeric will cause an immediate return of undefined.
IO::FD::syswrite2
Same as IO::FD::syswrite, but expect 2 of 4 arguments.
IO::FD::syswrite3
Same as IO::FD::syswrite, but expect 3 of 4 arguments.
IO::FD::syswrite4
Same as IO::FD::syswrite, but expect 4 of 4 arguments.
IO::FD::fcntl
IO::FD::sysfcntl
Alias to IO::FD::fcntl
IO::FD::stat
Almost the same as CORE::stat. If your system uses a signed dev_t for st_dev and st_rdev, (ie macos), this module will preserve both. At the time of writing, Perl CORE::stat will only preserve the sign of st_dev, and assume st_rdev is signed in all cases.
Attempts to replicate string expansion of values for some stat values as CORE::stat does. Tests of equality should be done using the eq operator (just like CORE::stat results) if your Perl is not 64bit.
IO::FD::lstat
As above.
IO::FD::pread
Binding to pread. Please see your system manual.
IO::FD:pwrite
Binding to pwrite. Please see your system manual.
Experimental
These functions haven't really been tested, documented or finished. They exist none the less. You will need to Look at the code for documentation at the moment. Their behaviour and interface are LIKELY TO CHANGE without notice.
IO::FD::ioctl
Not complete
IO::FD::sysioctl
Alias to ioctl
IO::FD::clock_gettime_monotonic
IO::FD::select
Broken. Probably will be removed as core Perl has this already.
IO::FD::poll
Constants for use with poll are available via IO::FD:Constants
IO::FD::kqueue
IO::FD::kevent
This is broken ok 32 bit BSD at the moment. Constants for use with kevent are available via IO::FD:Constants
IO::FD::pack_kevent
IO::FD::sv_to_pointer
IO::FD::pointer_to_sv
IO::FD::SV
IO::FD::SV($size)Allocates a string SV with the given size preallocated. The current string length is set to 0. For short string this is not the fastest way to allocate. For 4k and above, it is much faster, and doesn't use extra memory in compilation
IO::FD::readline
#SLURP A FILE
local $/=undef;
my $slurp=IO::FD::readline;
#or
#SLURP ALL RECORDS OF KNOWN LENGTH
local $/=\1234;
my @records=IO::FD::readline;A read line function is available, but is only operates in file slurp or record slurp mode (see perldoc -f readline). As no buffering is used, It does not attempt to split lines or read a line at a time like the normal Perl readline or <> operator
PERFORMANCE
Part of this distribution are benchmarking scripts. The following are typical outputs from my Intel 2020 Macbook Pro.
Listen Backlog
Results from benchmark/server-perl.pl benchmark/server.pl and benchmark/client.pl
Listen Backlog: 10
Perl server:
Connections before client refused: 18
IO::FD server
Connections before client refuse: 9285
Listen Backlog: 100
Perl server:
Connections before client refused: 190
IO::FD server
Connections before client refuse: (none refused)
Listen Backlog: 1000
Perl server:
Connections before client refused: 245
IO::FD server
Connections before client refuse: (none refused)Accept
Results from benchmark/server-perl.pl benchmark/server.pl and benchmark/client.pl
Listen Backlog: 100
Perl accept rate: 73568.4857256754/s
IO::FD Accept rate: 150984.798776367/sMemory Usage
Results from benchmark/file-memory.pl
Creating 2000 file handles/descriptors
Start maxrss (kB): 4500
Perl file handles
Bytes: 905216, per handle: 452.608
IO::FD
Bytes: 4096, per fd: 2.048
End maxrss (kB): 5692Socket creation
Results from benchmark/socket-create.pl
Rate perl_socket_INET iofd_socket_INET
perl_socket_INET 81919/s -- -56%
iofd_socket_INET 185679/s 127% --
Rate perl_socket_INET6 iofd_socket_INET6
perl_socket_INET6 81498/s -- -57%
iofd_socket_INET6 189253/s 132% --
Rate perl_socket_UNIX iofd_socket_UNIX
perl_socket_UNIX 113778/s -- -78%
iofd_socket_UNIX 508970/s 347% --File open and close
Results from benchmark/file-open-close.pl
Rate file_handle file_desc_posix io_fd
file_handle 91897/s -- -35% -37%
file_desc_posix 140549/s 53% -- -4%
io_fd 146161/s 59% 4% --Read Performance
Result from benchmark/file-read-write.pl
Read performance:
Read (bytes): 1024 x 2^0
Rate file_desc_posix file_handle io_fd
file_desc_posix 1803743/s -- -5% -5%
file_handle 1889325/s 5% -- -0%
io_fd 1890461/s 5% 0% --
Read (bytes): 1024 x 2^1
Rate file_desc_posix io_fd file_handle
file_desc_posix 1799026/s -- -1% -2%
io_fd 1823610/s 1% -- -1%
file_handle 1837458/s 2% 1% --
Read (bytes): 1024 x 2^2
Rate file_desc_posix io_fd file_handle
file_desc_posix 1731140/s -- -1% -1%
io_fd 1747626/s 1% -- -0%
file_handle 1747627/s 1% 0% --
Read (bytes): 1024 x 2^3
Rate io_fd file_desc_posix file_handle
io_fd 1458670/s -- -1% -3%
file_desc_posix 1470359/s 1% -- -2%
file_handle 1499189/s 3% 2% --
Read (bytes): 1024 x 2^4
Rate file_desc_posix file_handle io_fd
file_desc_posix 1146879/s -- -3% -6%
file_handle 1180322/s 3% -- -3%
io_fd 1214700/s 6% 3% --Write Performance
Result from benchmark/file-read-write.pl
Write performance:
Write (bytes): 1024 x 2^0
Rate file_desc_posix io_fd file_handle
file_desc_posix 1978800/s -- -7% -12%
io_fd 2117316/s 7% -- -6%
file_handle 2244774/s 13% 6% --
Write (bytes): 1024 x 2^1
Rate file_desc_posix io_fd file_handle
file_desc_posix 2007408/s -- -6% -9%
io_fd 2143700/s 7% -- -3%
file_handle 2205537/s 10% 3% --
Write (bytes): 1024 x 2^2
Rate file_desc_posix io_fd file_handle
file_desc_posix 1978800/s -- -7% -12%
io_fd 2123851/s 7% -- -5%
file_handle 2244774/s 13% 6% --
Write (bytes): 1024 x 2^3
Rate file_desc_posix io_fd file_handle
file_desc_posix 1960478/s -- -7% -9%
io_fd 2117316/s 8% -- -2%
file_handle 2163924/s 10% 2% --
Write (bytes): 1024 x 2^4
Rate file_desc_posix io_fd file_handle
file_desc_posix 1997468/s -- -5% -8%
io_fd 2104367/s 5% -- -3%
file_handle 2163924/s 8% 3% --SEE ALSO
The POSIX module provides an open, close, read and write routines which return/work with file descriptors. If you are only concerned with working with files, this is a better option as it is a core module, and will give you the purported benefits of this module. However it does not provide any networking/socket support.
FUTURE WORK (IDEAS/TODO)
Further emulate linux/bsd SOCK_NONBLOCK and SOCK_CLOEXEC on darwin
Add more tests for stat and DWIM module
Wider compatability for older Perls
Add More system functions which work with fds
Work with win32 sockets
Maybe make an IO::Handle sub classAUTHOR
Ruben Westerberg, <drclaw@mac.com>
REPOSITORTY and BUGS
Please report any bugs via git hub: http://github.com/drclaw1394/perl-io-fd
COPYRIGHT AND LICENSE
Copyright (C) 2023 by Ruben Westerberg
This library is free software; you can redistribute it and/or modify it under the same terms as Perl or the MIT license.
DISCLAIMER OF WARRANTIES
THIS PACKAGE IS PROVIDED "AS IS" AND WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.