A single close(fd) syscall is cheap, but when MAXFDS (maximum file
descriptor number) is high, the loop calling close(fd) on each file
descriptor can take several milliseconds.
The default value of subprocess.Popen "close_fds" parameter changed to True
in Python 3. Compared to Python 2, close_fds=True can make Popen 10x
slower: see bpo-37790 [1]
The present workaround on FreeBSD to improve performance is to load and
mount the fdescfs kernel module, but this is not enabled by default.
This change adds minimum viable (and upstreamable) closefrom(2) syscall
support to Python's subprocess and posix modules, improving performance
significantly for loads that involve working with many processes, such as
diffoscope, ansible, and many others.
For additional optimizations, upstream recently (3.8) landed posix_spawn(2)
support [3] and has stated that they will adopt close_range(2) after Linux
merges it [4]. Linux/FreeBSD developers are already collaborating on
ensuring compatible implementations, with FreeBSD's implementation pending
in D21627. [5]
Thank you emaste, cem, kevans for providing analysis, input,
clarifications, comms/upstream support and patches.
[1] https://bugs.python.org/issue37790
[2] https://bugs.python.org/issue38061
[3] https://bugs.python.org/issue35537
[4] https://lwn.net/Articles/789023/
[5] https://reviews.freebsd.org/D21627
Additional References:
https://bugs.python.org/issue8052https://bugs.python.org/issue11284https://bugs.python.org/issue13788https://bugs.python.org/issue1663329https://www.python.org/dev/peps/pep-0446/
PR: 242274, 221700
Submitted by: kevans (emaste, cem)
Approved by: koobs (python (maintainer), santa)
e4c2b30ce8