Hi,
# I'm not sure about netiquette here:
# I decided to continue posting to the python-list without CCing to
everyone.
First of all, here's the prototype.
It's a prototype and I know it's far from perfect, but it works for me
(in production code) - however, I did not yet test it on Non-Windows.
-
#!/bin/env python
# -*- coding: iso-8859-1 -*-
"""
File ftools.py: Useful tools for working with files.
"""
import os
import os.path
import time
import shutil
rmtree = shutil.rmtree
move = shutil.move
builtin_open = open
if os.name != "nt":
import fcntl
def open(fname, mode="r", bufsize=None):
"""
Like the "open" built-in, but does not inherit to child processes.
The code is using os.open and os.fdopen.
On Windows, to avoid inheritance, os.O_NOINHERIT is used
directly in the open call, thus it should be thread-safe.
On other operating systems, fcntl with FD_CLOEXEC is used
right after opening the file; however in a mutli-threaded program
it may happen that another thread starts a child process in the
fraction of a second between os.open and fcntl.
Note: The bufsize argument is ignored (not yet implemented).
"""
flags = 0
if "r" in mode:
flags += os.O_RDONLY
elif "w" in mode:
flags += os.O_RDWR + os.O_CREAT + os.O_TRUNC
elif "a" in mode:
flags += os.O_RDWR + os.O_CREAT + os.O_APPEND
else:
raise NotImplementedError ("mode=" + mode)
if os.name == "nt":
if "b" in mode:
flags += os.O_BINARY
else:
flags += os.O_TEXT
flags += os.O_NOINHERIT
try:
fd = os.open (fname, flags)
if os.name != "nt":
old = fcntl.fcntl(fd, fcntl.F_GETFD)
fcntl.fcntl(fd, fcntl.F_SETFD, old | fcntl.FD_CLOEXEC)
return os.fdopen (fd, mode)
except OSError, x:
raise IOError(x.errno, x.strerror, x.filename)
def copyfile(src, dst):
"""
Copies a file - like shutil.copyfile, but the files are opened
non-inheritable.
Note: This prototype does not test _samefile like shutil.copyfile.
"""
fsrc = None
fdst = None
try:
fsrc = open(src, "rb")
fdst = open(dst, "wb")
shutil.copyfileobj(fsrc, fdst)
finally:
if fdst:
fdst.close()
if fsrc:
fsrc.close()
"""
blah blah:
I googled around a bit, and it more and more seems to me
that the Posix system has a serious design flaw, since it seems
to be SO hard to write multi-threaded programs that also
start child-processes.
It's funny that right now Linus Torvalds himself seems to be
aware of this problem and that the Linux kernel developers
are discussing ways to solve it.
Let's hope they find a way to get around it on the OS level.
To me, the design of MS Windows looks better in the aspect
of process-creation, handle inheritance and multi-threading...
Anyway, it has its drawbacks, too.
For example, I still cannot specify in a thread-safe way that
a handle should be inherited to one child process but not
to another - I would have to use a lock to synchronize it,
which has its own problems, as Ross Cohen noted.
The best solution at the OS level would be to explitly
specify the handles/fds I want to be inherited in a
"create child process" system call.
BTW the Linux kernel developers face the same situation
as we do: They could somehow implement a new system
function like "open_noinherit", but there's a whole bunch
of existing "standard code" that uses open and similar
functions like socket(), accept() etc., and they don't want
to change all these calls.
So perhaps, for Python development, we just have to accept
that the problem persists and that at this time a 100% solution
just does not exist - and we should watch the discussion
on http://lwn.net/Articles/237722/ to see how they solve it for
Linux.
"""
That being said, I still think it's necessary for Python to provide
a solution as good as possible.
For example, in my production application, by consequently
using the ftools module above, I could reduce the error rate dramatically:
* With built-in open and shutil.copyfile:
Several "Permission denied" and other errors a day
* With ftools.open and ftools.copyfile:
program running for a week or more without errors.
There are still errors sometimes, and I suspect it has to do with the
unintenional inheritance of socket handles (I did not dig into
SocketServer.py, socket.py and socket.c to solve it).
(However, the errors are so rare now that our clients think
it's just errors in their network :-).
Martin, you mentioned that for sockets, inheritance is not a problem
unless accept(), recv() or select() is called in the child process
(as far as I understood it).
Though I am not an expert in socket programming at the C level,
I doubt that you are right here. Apart from by own experiences,
I've found some evidence in the W
