STINNER Victor added the comment: Yeah, when stdin, stdout or stderr is a pipe, subprocess.Popen() calls CreateProcess() with the STARTF_USESTDHANDLES flag and bInheritHandles=TRUE.
This issue was fixed on UNIX for file descriptor inheritance. The fix is a major change: Python 3.4 now only creates non-inheritable file descriptors by default. See the PEP 446. For your issue, see the "Only Inherit Some Handles on Windows" section: https://www.python.org/dev/peps/pep-0446/#only-inherit-some-handles-on-windows But this section is not fixed by the PEP. I was decided to fix inheritance of file descriptors and inheritance of handles separatly. I guess that this issue is a duplicate of the issue #19575. I'm against the idea of adding a lock inside the subprocess module. I may introduce deadlock issues. You can already put such lock in your application, to call subprocess.Popen (directly or indirectly). > Currently, the Popen constructor will duplicate any stdout, stdin, and/or > stderr handle passed in and make them inheritable, by calling > DuplicateHandle. If two threads call Popen at the same time, the newly > created inheritable handles will leak into the subprocess that's running > being created in the opposite thread. This has consequences when two or more > subprocesses are piped together and executed at the time time. This is a race condition really specific to the subprocess module. Usually, handles are created non-inhertable on Windows, so calling CreateProcess() with bInheritHandles=TRUE is not an issue in the common case. Here the problem is that subprocess makes duplicated handles inheritable. There is a short window where a second thread can spawn a process and inherit these handles too. The real fix is to never make duplicated handles inheritable, but instead to use the new PROC_THREAD_ATTRIBUTE_HANDLE_LIST structure. Another option to fix the issue is to use a wrapper application and send handles from the parent to the child, but this option introduces a lot of complexity since the parent has to handle two processes, not only one! Again, see the issue #19575. ---------- _______________________________________ Python tracker <rep...@bugs.python.org> <http://bugs.python.org/issue24909> _______________________________________ _______________________________________________ Python-bugs-list mailing list Unsubscribe: https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
