About

sendfile(2) is a system call which provides a "zero-copy" way of copying data from one file descriptor to another (a socket). The phrase "zero-copy" refers to the fact that all of the copying of data between the two descriptors is done entirely by the kernel, with no copying of data into userspace buffers. This is particularly useful when sending a file over a socket (e.g. FTP).

The normal way of sending a file over a socket involves reading data from the file into a userspace buffer, then write that buffer to the socket via send() or sendall():

# how a file is tipically sent

import socket

file = open("somefile", "rb")
sock = socket.socket()
sock.connect(("127.0.0.1", 8021))

while True:
    chunk = file.read(65536)
    if not chunk: 
        break  # done
    sock.sendall(chunk)

This copying of the data twice (once into the userland buffer, and once out from that userland buffer) imposes some performance and resource penalties. sendfile(2) syscall avoids these penalties by avoiding any use of userland buffers; it also results in a single system call (and thus only one context switch), rather than the series of read(2) / write(2) system calls (each system call requiring a context switch) used internally for the data copying.

import socket
from sendfile import sendfile

file = open("somefile", "rb")
sock = socket.socket()
sock.connect(("127.0.0.1", 8021))
offset = 0

while True:
    sent = sendfile(sock.fileno(), file.fileno(), offset, 65536)
    if sent == 0:
        break  # done
    offset += sent

A simple benchmark

This benckmark script implements the two examples above and compares plain socket.send() and sendfile() performances in terms of CPU time spent and bytes transmitted per second resulting in sendfile() being about 3x faster.
These are the results I get on my Linux 2.6.38 box, AMD dual-core 1.6 GHz:

send()

sendfile()

When do you want to use it?

Basically any application sending files over the network can take advantage of sendfile(2). HTTP and FTP servers are a typical example. proftpd and vsftpd are known to use it, so is pyftpdlib.

API documentation

sendfile module provides a single function: sendfile().

sendfile.sendfile(out, in, offset, nbytes, header="", trailer="", flags=0)

Copy nbytes bytes from file descriptor in (a regular file) to file descriptor out (a socket) starting at offset.
Return the number of bytes just being sent. When the end of file is reached return 0.
On Linux, if offset is given as None, the bytes are read from the current position of in and the position of in is updated.
headers and trailers are strings that are written before and after the data from in is written. In cross platform applications their usage is discouraged (socket.send() or socket.sendall() can be used instead).
On Solaris, out may be the file descriptor of a regular file or the file descriptor of a socket. On all other platforms, out must be the file descriptor of an open socket.
flags argument is only supported on FreeBSD.

sendfile.SF_NODISKIO
sendfile.SF_MNOWAIT
sendfile.SF_SYNC

Parameters for the flags argument, if the implementation supports it. They are available on FreeBSD platforms. See FreeBSD's man sendfile(2).

Differences with send()

• sendfile(2) works with regular (mmap-like) files only (e.g. you can't use it with a StringIO object).
• Also, it must be clear that the file can only be sent "as is" (e.g. you can't modify the content while transmitting).
• There might be problems with non regular filesystems such as NFS, SMBFS/Samba and CIFS. For this please refer to proftpd documentation.
• OSError is raised instead of socket.error. The accompaining error codes have the same meaning though: EAGAIN, EWOULDBLOCK, EBUSY meaning you are supposed to retry, ECONNRESET, ENOTCONN, ESHUTDOWN, ECONNABORTED in case of disconnection. Some examples:
• benchmark script
• test suite
• pyftpdlib wrapper

Supported platforms

This module works with Python versions from 2.5 to 3.3. The supported platforms are:

• Linux
• Mac OSX
• FreeBSD
• Dragon Fly BSD
• Sun OS
• AIX (not properly tested)

Support

Feel free to mail me at g.rodola[AT]gmail[DOT]com or post on the the mailing list: http://groups.google.com/group/py-sendfile

Status

As of now the code includes a solid test suite and it should be ready for production use. Recently it's been included in pyftpdlib project; hopefully that's going to tell whether it's actually stable or not.

Authors

pysendfile was originally written by Ben Woolley including Linux, FreeBSD and DragonFly BSD support.
Later on Niklas Edmundsson took over maintenance and added AIX support.
After a couple of years of project stagnation Giampaolo Rodola' took over maintenance and rewrote it from scratch adding support for:

• Python 3
• non-blocking sockets
• large files support
• Mac OSX
• Sun OS
• FreeBSD flag argument
• multiple threads (release GIL)
• a simple benchmark suite
• unit tests
• documentation