Same code but different problem as this issue: http://bugs.python.org/issue10897

The above-linked issue was closed as WONTFIX, but there is a secondary problem: the file descriptor that mmap.mmap() allocates is not set to close-on-exec.

This means that any process you spawn using os.system() and the like, its children, its children's children etc will inherit this file descriptor.

This is unexpected, unnecessary behavior, and potentially a huge pain if you mmap() a file and then spawn, say, inetd or something that does similar stuff, because now you have to kill every process that spawned from that in order to unmount filesystems or remove kernel modules that are being blocked as a result of the fds being open.

I should add a caveat: the fd that is created by mmap /will/ be set to close-on-exec if the fd you passed in was. But even if it's not, I see no reason why mmap should not be setting it anyway.

At the very least, the documentation should bring the user's attention to this potential stumbling block. Currently it makes no mention of the fact that mmap duplicates file descriptors at all, much less how the flags are set.

Can you suggest a documentation update?

"the file descriptor that mmap.mmap() allocates is not set to close-on-exec"

In Python 3.4, the file descriptor is now non-inheritable, as a side effect of the PEP 446.
http://www.python.org/dev/peps/pep-0446/

I'm adding Library again because I think the current behavior is a bug and should be fixed in the 2.7 tree. Perhaps the documentation in older versions should be updated 

mmap.mmap should always set the FD_CLOEXEC flag on the descriptor that it gets from dup(). I don't see why we wouldn't do this, because if we are calling exec(), we are blowing away all python state, and only the mmap object should be using this file descriptor (since its existence is hidden from the user).

Users should not be depending on the old behavior, because the interface which is presented to them gives them no indication that this fd even exists. Users are probably expecting the proposed behavior, because anyone who has worked with unix file descriptors in C would not expect:

fd = os.open(...)
mmap = mmap.mmap(fd, size)
# do some stuff
os.close(fd)

To result in an extra inheritable fd still being there. Nothing in the documentation indicates that this would happen, either.

Sorry, I correct my earlier statement: even if the fd you pass to mmap.mmap() is set to FD_CLOEXEC, the dup'd fd /will not be/

So this is a REALLY bad bug because users cannot workaround it except by just not using mmap

I agree this should be fixed.
Robert, want to submit a patch?

> Robert, want to submit a patch?

If someone wants to work on a patch, here is the code of Python 3.4 to duplicate a file descriptor and make it non-inheritable.
  http://hg.python.org/cpython/file/e97d9926da83/Python/fileutils.c#l978
and
  http://hg.python.org/cpython/file/e97d9926da83/Python/fileutils.c#l618

The code is complex because it has optimized versions for each operating system. Only supporting Windows and fcntl() is enough for Python 2.

Closing because Victor has proposed a patch in Issue 26769.