I hit a variant of issue #12251, namely when you redirect both stdout and stderr of a child process and one of them uses a low fd.
Testcase:

import subprocess, sys
subprocess.call(["ls", "asda"], stderr = sys.stdout, stdout = open("/dev/null", "w"))

strace output:

open("/dev/null", O_WRONLY|O_CREAT|O_TRUNC, 0666) = 3
[child process]
dup2(3, 1)   = 1  // overwrites the stdout filedescriptor..
dup2(1, 2)   = 2  // that was supposed to be duped here.
close(3)    = 0   // okay

The testcase is supposed to print "ls: file not found" on stdout, but is silent.

> stdout = open("/dev/null", "w"), stderr = sys.stdout

You ask to write all outputs to /dev/null. Why do you expect anything on stdout?

I expect that 'ls' print the error message on its stderr, which would be redirected to stdout of the test Python program.

If I had been using stderr = subprocess.STDOUT, I can agree with you that both output streams would go to /dev/null.

> stdout = open("/dev/null", "w"), stderr = sys.stdout

Oh, I read subprocess.STDOUT instead of sys.stdout.

It is indeed a problem.

It seems like the problem comes about due to the "swapping" of fds.
i.e. using stdout as stderr. The reverse appears to work due to the order in which the dup() calls are performed.

Attached is a patch which fixes the issue.

Thanks for the patch, I haven't tried it (I'm still on Python 2.7) but it looks very special-cased to my case. I can think about exotic cases like

stdin = sys.stderr, stdout = sys.stdin, stderr = sys.stdout

It can happen in reality if a daemon have closed stdio (see bug 10806) and reusing fd's 0,1,2 for other purposes.

Attached is a patch which tests all combinations and includes a testcase.

save_fds should be moved outside the try block:

+            try:
+                # save a copy of the standard file descriptors
+                saved_fds = [os.dup(fd) for fd in range(3)]
+                ...
+            finally:
+                for std, saved in enumerate(saved_fds):
+                    os.dup2(saved, std)
+                    os.close(saved)


"temp_fds = [fd for fd, fname in temps]" should also be moved outside its try block.

Why not using .communicate() to write into stdin and read stdout/stderr?

communicate() requires setting stdin, stdout and stderr to subprocess.PIPE which defeats the purpose of the test: setting them to 0, 1 and 2 (in various orders) so that they need to be swapped...

i12607_v2.patch: the test fails if Python is compiled with pydebug. Add "err = support.strip_python_stderr(err)" before self.assertEqual(err, b"err") in test_subprocess.py to fix the failure.

New changeset 1140b32747c9 by Ross Lagerwall in branch '3.2':
Issue #12607: In subprocess, fix issue where if stdin, stdout or stderr is
http://hg.python.org/cpython/rev/1140b32747c9

New changeset 943d323cb4b8 by Ross Lagerwall in branch 'default':
Issue #12607: Merge with 3.2.
http://hg.python.org/cpython/rev/943d323cb4b8

New changeset 6119440bae30 by Ross Lagerwall in branch '2.7':
Issue #12607: In subprocess, fix issue where if stdin, stdout or stderr is
http://hg.python.org/cpython/rev/6119440bae30