Title: Popen.communicate() breaks when child closes its side of pipe but not exits
Type: behavior Stage: commit review
Components: Library (Lib) Versions: Python 3.9, Python 3.8, Python 3.7
Status: closed Resolution: fixed
Dependencies: Superseder:
Assigned To: gregory.p.smith Nosy List: and800, gregory.p.smith, josh.r, miss-islington
Priority: normal Keywords: patch

Created on 2018-11-07 10:23 by and800, last changed 2020-01-23 23:33 by gregory.p.smith. This issue is now closed.

Pull Requests
URL Status Linked Edit
PR 17020 closed and800, 2019-10-31 17:01
PR 17023 closed and800, 2019-10-31 23:40
PR 18117 merged apr, 2020-01-22 15:23
PR 18148 merged apr, 2020-01-23 14:56
PR 18151 merged apr, 2020-01-23 15:18
Messages (9)
msg329412 - (view) Author: Andriy Maletsky (and800) * Date: 2018-11-07 10:23
When communicate() is called in a loop, it crashes when the child process has already closed any piped standard stream, but still continues to be running.

How this happens:
1) the parent waits for the child events inside communicate() call
2) the child closes its side of any attached pipes long before exiting (in my case there is some complex c++ application which had messed with its termination)
3) communicate() receives an epoll event, tries to read/write, receives SIGPIPE (for stdin) or EOF (for stdout), decides to close corresponding file descriptors from its side
4) communicate() waits for the death of the child, but a timeout is fired
5) parent handles timeout exception and calls communicate() again
6) an exception is raised when communicate() tries to register closed file in epoll

I think there may be a simple solution: before registering file descriptors in epoll, we may check whether any of them is already closed, and don't register it in that case.

Here is a simple reproducible example, ran on Linux 4.15.0-1021-aws x86_64:

import subprocess

child = subprocess.Popen(
    ['/usr/local/bin/python3.7', '-c', 'import os, time; os.close(1), time.sleep(30)'],

while True:
    except subprocess.TimeoutExpired:
        # do something useful here

Here is a stacktrace:

Traceback (most recent call last):
  File "", line 10, in <module>
  File "/usr/local/lib/python3.7/", line 933, in communicate
    stdout, stderr = self._communicate(input, endtime, timeout)
  File "/usr/local/lib/python3.7/", line 1666, in _communicate
    selector.register(self.stdout, selectors.EVENT_READ)
  File "/usr/local/lib/python3.7/", line 352, in register
    key = super().register(fileobj, events, data)
  File "/usr/local/lib/python3.7/", line 238, in register
    key = SelectorKey(fileobj, self._fileobj_lookup(fileobj), events, data)
  File "/usr/local/lib/python3.7/", line 225, in _fileobj_lookup
    return _fileobj_to_fd(fileobj)
  File "/usr/local/lib/python3.7/", line 40, in _fileobj_to_fd
    "{!r}".format(fileobj)) from None
ValueError: Invalid file object: <_io.BufferedReader name=3>
msg329435 - (view) Author: Josh Rosenberg (josh.r) * (Python triager) Date: 2018-11-07 20:04
Sounds like the solution you'd want here is to just change each if check in _communicate, so instead of:

    if self.stdout:
        selector.register(self.stdout, selectors.EVENT_READ)
    if self.stderr:
        selector.register(self.stderr, selectors.EVENT_READ)

it does:

    if self.stdout and not self.stdout.closed:
        selector.register(self.stdout, selectors.EVENT_READ)
    if self.stderr and not self.stderr.closed:
        selector.register(self.stderr, selectors.EVENT_READ)

The `if self.stdin and input:` would also have to change. Right now it's buggy in a related, but far more complex way. Specifically if you call it with input the first time:

1. If some of the input is sent but not all, and the second time you call communicate you rely on the (undocumented, but necessary for consistency) input caching and don't pass input at all, it won't register the stdin handle for read (and in fact, will explicitly close the stdin handle), and the remaining cached data won't be sent. If you try to pass some other non-empty input, it just ignores it and sends whatever remains in the cache (and fails out as in the stdout/stderr case if the data in the cache was sent completely before the timeout).

2. If all of the input was sent on the first call, you *must* pass input=None, or you'll die trying to register self.stdin with the selector

The fix for this would be to either:

1. Follow the pattern for self.stdout/stderr (adding "and not self.stdin.closed"), and explicitly document that repeated calls to communicate must pass the exact same input each time (and optionally validate this in the _save_input function, which as of right now just ignores the input if a cache already exists); if input is passed the first time, incompletely transmitted, and not passed the second time, the code will error as in the OP's case, but it will have violated the documented requirements (ideally the error would be a little more clear though)


2. Change the code so populating the cache (if not already populated) is the first step, and replace all subsequent references to input with references to self._input (for setup tests, also checking if self._input_offset >= len(self._input), so it doesn't register for notifications on self.stdin if all the input has been sent), so it becomes legal to pass input=None on a second call and rely on the first call to communicate caching it. It would still ignore new input values on the subsequent calls, but at least it would behave in a sane way (not closing sys.stdin despite having unsent cached data, then producing a confusing error that is several steps removed from the actual problem)

Either way, the caching behavior for input should be properly documented; we clearly specify that output is preserved after a timeout and retrying communicate ("If the process does not terminate after timeout seconds, a TimeoutExpired exception will be raised. Catching this exception and retrying communication will not lose any output."), but we don't say anything about input, and right now, the behavior is the somewhat odd and hard to express:

"Retrying a call to communicate when the original call was passed non-None/non-empty input requires subsequent call(s) to pass non-None, non-empty input. The input on said subsequent calls is otherwise ignored; only the unsent remainder of the original input is sent. Also, it will just fail completely if you pass non-empty input and it turns out the original input was sent completely on the previous call, in which case you *must* call it with input=None."

It might also be worth changing the selectors module to raise a more obvious exception when register is passed a closed file-like object, but given it only requires non-integer fileobjs to have a .fileno() method, adding a requirement for a "closed" attribute/property could break other code.
msg329436 - (view) Author: Josh Rosenberg (josh.r) * (Python triager) Date: 2018-11-07 20:20
Hmm... Correction to my previous post. communicate itself has a test for:

"if self._communication_started and input:"

that raises an error if it passes, so the second call to communicate can only be passed None/empty input. And _communicate only explicitly closes self.stdin when input is falsy and _communication_started is False, so the required behavior right now is:

1. First call *may* pass input
2. Second call must not pass (non-empty) input under any circumstance

So I think we're actually okay on the code for stdin, but it would be a good idea to document that input *must* be None on all but the first call, and that the input passed to the first call is cached such that as long as at least one call to communicate completes without a TimeoutError (and the stdin isn't explicitly closed), it will all be sent.

Sorry for the noise; I should have rechecked communicate itself, not just _communicate.
msg355760 - (view) Author: Andriy Maletsky (and800) * Date: 2019-10-31 16:59
@josh.r but you’re correct regarding cached data that isn’t sent on subsequent communicate() calls. If the child consumes the input too slowly, and timeout occurs before sending all input, the remaining part will be lost.

Maybe it is not a bug, but it’s quite a confusing behavior, and I think it should be mentioned in the doc.
msg360520 - (view) Author: Gregory P. Smith (gregory.p.smith) * (Python committer) Date: 2020-01-22 23:28
New changeset d3ae95e1e945ed20297e1c38ba43a18b7a868ab6 by Gregory P. Smith (Alex Rebert) in branch 'master':
bpo-35182: fix communicate() crash after child closes its pipes (GH-17020) (GH-18117)
msg360521 - (view) Author: Gregory P. Smith (gregory.p.smith) * (Python committer) Date: 2020-01-22 23:31
backport automation appears unhappy at the moment.  I'm keeping this open and assigned to me to manually run cherry_picker on this for 3.8 and 3.7 (if still open for non-security fixes).
msg360584 - (view) Author: miss-islington (miss-islington) Date: 2020-01-23 23:31
New changeset 5654f83b9706af88040f515791f1cdc5d81cd9d6 by Miss Islington (bot) (Alex Rebert) in branch '3.8':
[3.8] bpo-35182: fix communicate() crash after child closes its pipes (GH-18117) (GH-18148)
msg360585 - (view) Author: miss-islington (miss-islington) Date: 2020-01-23 23:32
New changeset 61b3484cdf27ceca1c1069a351487d2db4b2b48c by Miss Islington (bot) (Alex Rebert) in branch '3.7':
[3.7] bpo-35182: fix communicate() crash after child closes its pipes (GH-18117) (GH-18151)
msg360586 - (view) Author: Gregory P. Smith (gregory.p.smith) * (Python committer) Date: 2020-01-23 23:33
thanks everyone!
Date User Action Args
2020-01-23 23:33:29gregory.p.smithsetstatus: open -> closed
versions: + Python 3.9
messages: + msg360586

resolution: fixed
stage: patch review -> commit review
2020-01-23 23:32:35miss-islingtonsetmessages: + msg360585
2020-01-23 23:31:42miss-islingtonsetnosy: + miss-islington
messages: + msg360584
2020-01-23 15:18:43aprsetpull_requests: + pull_request17537
2020-01-23 14:56:41aprsetstage: backport needed -> patch review
pull_requests: + pull_request17534
2020-01-22 23:31:30gregory.p.smithsetassignee: gregory.p.smith
stage: patch review -> backport needed
messages: + msg360521
versions: - Python 2.7, Python 3.4, Python 3.5, Python 3.6
2020-01-22 23:28:39gregory.p.smithsetnosy: + gregory.p.smith
messages: + msg360520
2020-01-22 15:23:06aprsetpull_requests: + pull_request17504
2019-10-31 23:40:22and800setpull_requests: + pull_request16540
2019-10-31 17:01:01and800setkeywords: + patch
stage: needs patch -> patch review
pull_requests: + pull_request16538
2019-10-31 16:59:47and800setmessages: + msg355760
2018-11-07 20:20:51josh.rsetmessages: + msg329436
2018-11-07 20:04:49josh.rsetnosy: + josh.r

messages: + msg329435
stage: needs patch
2018-11-07 12:13:58and800setcomponents: + Library (Lib)
2018-11-07 10:35:45and800settype: crash -> behavior
2018-11-07 10:23:30and800create