Demonstration:$ cat test.pyimport osimport subprocessimport sysfd = os.dup(sys.stdout.fileno())subprocess.call([sys.executable, '-c',                 'import sys;'                 'print("Hello stdout");'                 'print("Hello fd", file=open(int(sys.argv[1]), "w"))',                 str(fd)],                stdout=fd,                pass_fds=[fd])$ python3 test.pyHello stdoutTraceback (most recent call last):  File "<string>", line 1, in <module>OSError: [Errno 9] Bad file descriptorI see two issues here:1. The fact that file descriptors specified for redirection are closed (even with close_fds=False) unless in range [0, 2] is not documented and not tested. If I change the corresponding code in _posixsubprocess.c to close them only if they are ends of pipes created by subprocess itself, all tests still pass. Also, shells behave differently: e.g., in command 'echo 3>&1' fd 3 is duplicated but not closed in the child, so subprocess behavior may be not what people expect.2. pass_fds interaction with (1) should be fixed and documented. We may either raise ValueError if pass_fds contains a redirected fd or don't close such a redirected fd.Please provide your thoughts.

Thanks for the report and the investigation!> 1. The fact that file descriptors specified for redirection are closed (even with close_fds=False) unless in range [0, 2] is not documented and not tested.That sounds like a bug to me, so we should fix it if decently possible.> 2. pass_fds interaction with (1) should be fixed and documented. We may either raise ValueError if pass_fds contains a redirected fd or don't close such a redirected fd.Same here, it seems like a bug which deserves fixing.Do you want to submit a PR for this?

Regarding fixing (1), I'm worrying about backward compatibility a bit. Some people who discovered that behavior might rely on such "move" semantics and expect that the redirected descriptor is not leaked into the child. OTOH, since descriptors are non-inheritable by default since 3.4, it might be less of an issue now. What do you think?Otherwise, yes, I'd like to fix this, but fixing is a bit trickier than it may seem because sometimes descriptors are duplicated in _posixsubprocess.c (in case of low fds), so we need to track the ownership properly. I've already performed some work and discovered another issue: in a corner case involving closed standard fds, pipes created by subprocess may leak into the child and/or an incorrect redirection may occur. Since I can't completely fix this issue without affecting a discovered one, I think I'll do the opposite: file a new issue, fix it, and then go back to this one.

> Regarding fixing (1), I'm worrying about backward compatibility a bit.Well, people shouldn't rely on bugs.  Otherwise we would never be able to fix bugs, lest someone relies on it.

This bug stems from:https://github.com/python/cpython/blob/master/Modules/_posixsubprocess.c#L454When stdout=fd is passed, that results in c2pwrite=fd.  The stdin/stdout/stderr pipe fds are closed when > 2 without checking to see if they are listed in pass_fds.https://github.com/python/cpython/blob/master/Lib/subprocess.py#L1306 maps the following: stdin -> p2cread stdout -> c2pwrite stderr -> errwrite

Actually I've started to work on this with#32844, but got no feedback. This issue may of course be fixed independently, but the problems with descriptor ownership for fds <= 2 will remain unless all ownership problems are fixed at once.

An alternative fix is here:https://github.com/izbyshev/cpython/commit/b89b52f28490b69142d5c061604b3a3989cec66cFeel free to use the test if you don't like the approach :)

New changesetce34410b8b67f49d8275c05d51b3ead50cf97f48 by Gregory P. Smith in branch 'master':bpo-32270: Don't close stdin/out/err in pass_fds (GH-6242)https://github.com/python/cpython/commit/ce34410b8b67f49d8275c05d51b3ead50cf97f48

New changeset9c4a63fc17efea31ad41f90d28825be37469e0e2 by Gregory P. Smith (Miss Islington (bot)) in branch '3.7':bpo-32270: Don't close stdin/out/err in pass_fds (GH-6242) (GH-9148)https://github.com/python/cpython/commit/9c4a63fc17efea31ad41f90d28825be37469e0e2

New changeset2173bb818c6c726d831b106ed0d3fad7825905dc by Gregory P. Smith (Miss Islington (bot)) in branch '3.6':bpo-32270: Don't close stdin/out/err in pass_fds (GH-6242) (GH-9149)https://github.com/python/cpython/commit/2173bb818c6c726d831b106ed0d3fad7825905dc

Thank you, Gregory!