subprocess.Popen stdin read file

Question

I'm trying to call a process on a file after part of it has been read. For example:

with open('in.txt', 'r') as a, open('out.txt', 'w') as b:
  header = a.readline()
  subprocess.call(['sort'], stdin=a, stdout=b)

This works fine if I don't read anything from a before doing the subprocess.call, but if I read anything from it, the subprocess doesn't see anything. This is using python 2.7.3. I can't find anything in the documentation that explains this behaviour, and a (very) brief glance at the subprocess source didn't reveal a cause.

Answer 1

If you open the file unbuffered then it works:

import subprocess

with open('in.txt', 'rb', 0) as a, open('out.txt', 'w') as b:
    header = a.readline()
    rc = subprocess.call(['sort'], stdin=a, stdout=b)

subprocess module works at a file descriptor level (low-level unbuffered I/O of the operating system). It may work with os.pipe(), socket.socket(), pty.openpty(), anything with a valid .fileno() method if OS supports it.

It is not recommended to mix the buffered and unbuffered I/O on the same file.

On Python 2, file.flush() causes the output to appear e.g.:

import subprocess
# 2nd
with open(__file__) as file:
    header = file.readline()
    file.seek(file.tell()) # synchronize (for io.open and Python 3)
    file.flush()           # synchronize (for C stdio-based file on Python 2)
    rc = subprocess.call(['cat'], stdin=file)

The issue can be reproduced without subprocess module with os.read():

#!/usr/bin/env python
# 2nd
import os

with open(__file__) as file: #XXX fully buffered text file EATS INPUT
    file.readline() # ignore header line
    os.write(1, os.read(file.fileno(), 1<<20))

If the buffer size is small then the rest of the file is printed:

#!/usr/bin/env python
# 2nd
import os

bufsize = 2 #XXX MAY EAT INPUT
with open(__file__, 'rb', bufsize) as file:
    file.readline() # ignore header line
    os.write(2, os.read(file.fileno(), 1<<20))

It eats more input if the first line size is not evenly divisible by bufsize.

The default bufsize and bufsize=1 (line-buffered) behave similar on my machine: the beginning of the file vanishes -- around 4KB.

file.tell() reports for all buffer sizes the position at the beginning of the 2nd line. Using next(file) instead of file.readline() leads to file.tell() around 5K on my machine on Python 2 due to the read-ahead buffer bug (io.open() gives the expected 2nd line position).

Trying file.seek(file.tell()) before the subprocess call doesn't help on Python 2 with default stdio-based file objects. It works with open() functions from io, _pyio modules on Python 2 and with the default open (also io-based) on Python 3.

Trying io, _pyio modules on Python 2 and Python 3 with and without file.flush() produces various results. It confirms that mixing buffered and unbuffered I/O on the same file descriptor is not a good idea.

Answer 2

I solved it in python 2.7 by aligning the file descriptor position.

os.lseek(_file.fileno(), _file.tell(), os.SEEK_SET) truncate_null_cmd = ['tr','-d', '\\000'] subprocess.Popen(truncate_null_cmd, stdin=_file, stdout=subprocess.PIPE)

Answer 3

It happens because the subprocess module extracts the File handle from the File Object.

http://hg.python.org/releasing/2.7.6/file/ba31940588b6/Lib/subprocess.py

In line 1126, coming from 701.

The File Object uses buffers and has already read a lot from the file handle when the subprocess extracts it.