Repository navigation
socket.readline() interface doesn't handle EINTR properly #44412
Description
| BPO | 1628205 |
|---|---|
| Nosy | @loewis, @gpshead |
| Files |
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
assignee = 'https://github.com/gpshead'
closed_at = <Date 2010-12-15.20:00:10.644>
created_at = <Date 2007-01-04.21:37:50.000>
labels = ['extension-modules', 'type-bug', 'library']
title = "socket.readline() interface doesn't handle EINTR properly"
updated_at = <Date 2010-12-15.20:00:10.642>
user = 'https://bugs.python.org/sobomax'bugs.python.org fields:
activity = <Date 2010-12-15.20:00:10.642>
actor = 'pitrou'
assignee = 'gregory.p.smith'
closed = True
closed_date = <Date 2010-12-15.20:00:10.644>
closer = 'pitrou'
components = ['Extension Modules', 'Library (Lib)']
creation = <Date 2007-01-04.21:37:50.000>
creator = 'sobomax'
dependencies = []
files = ['7694', '14634', '14667']
hgrepos = []
issue_num = 1628205
keywords = ['patch']
message_count = 11.0
messages = ['51655', '51656', '51657', '51658', '51659', '91209', '91359', '91409', '91414', '91532', '106334']
nosy_count = 8.0
nosy_names = ['loewis', 'gregory.p.smith', 'jorend', 'orenti', 'exarkun', 'sobomax', 'vila', 'rhettg']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'test needed'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue1628205'
versions = ['Python 2.6', 'Python 3.0', 'Python 3.1', 'Python 2.7']
Activity
sobomax commentedon Jan 4, 2007
The socket.readline() interface doesn't handle EINTR properly. Currently, when EINTR received exception is not handled and all data that has been in the buffer is lost. There is no way to recover that data from the code that uses the interface.
Correct behaviour would be to catch EINTR and restart recv(). Patch is attached.
Following is the real world example of how it affects httplib module:
File "/usr/local/lib/python2.4/xmlrpclib.py", line 1096, in __call__
return self.__send(self.__name, args)
File "/usr/local/lib/python2.4/xmlrpclib.py", line 1383, in __request
verbose=self.__verbose
File "/usr/local/lib/python2.4/xmlrpclib.py", line 1131, in request
errcode, errmsg, headers = h.getreply()
File "/usr/local/lib/python2.4/httplib.py", line 1137, in getreply
response = self._conn.getresponse()
File "/usr/local/lib/python2.4/httplib.py", line 866, in getresponse
response.begin()
File "/usr/local/lib/python2.4/httplib.py", line 336, in begin
version, status, reason = self._read_status()
File "/usr/local/lib/python2.4/httplib.py", line 294, in _read_status
line = self.fp.readline()
File "/usr/local/lib/python2.4/socket.py", line 325, in readline
data = recv(1)
error: (4, 'Interrupted system call')
-Maxim
orenti commentedon Jan 7, 2007
You may have encountered this on sockets but *all* Python I/O does not handle restart on EINTR.
The right place to fix this is probably in C, not the Python library. The places where an I/O operation could be interrupted are practically anywhere the GIL is released. This kind of change is likely to be controversial.
sobomax commentedon Jan 8, 2007
Well, it's not quite correct since for example httplib.py tries to handle EINTR. The fundamental problem with socket.readline() is that it does internal buffering so that getting EINTR results in data being lost.
I don't think it has to be fixed in C, since recv() is very low-level interface and it is expected to return EINTR on signal, so that "fixing" it there could possibly break software that relies on this behaviour. And I don't quite buy your reasoning - "since it's broken in few more places let's keep it consistently broken everywhere". To me it sounds like attempt to hide the head in the sand instead of facing the problem at hand. Fixing socket.readline() may be the first step in improvind the library to handle this condition properly.
loewis commentedon Feb 16, 2007
I agree that this should be fixed; I'm not sure I like the proposed fixed, though. It discards the exception and keeps running.
What it (IMO) should do instead is abort, then return the data on the next invocation. Of course, this may have problems in itself, since the file descriptor might not report read-ready when passed to select or poll, even though data are available.
Please discuss this on python-dev (and elsewhere), and report what recommendations people made.
jorend commentedon Mar 7, 2007
loewis: I think your idea is the right answer. I'm not worried about select/poll. Surely no one uses select/poll and socket._fileobject.readline() on the same socket. select/poll are for nonblocking sockets; this readline() method doesn't even catch EWOULDBLOCK.
...In fact even if you did use select/poll on the (blocking) socket after readline() threw EINTR--which no one should do--I think it would still work just as expected unless you were doing something truly weird.
rhettg commentedon Aug 2, 2009
I've hit this issue as well.
Attached is an updated patch to 2.6 branch and a test case.
I wrote up more details here:
http://nullhole.com/2009/08/02/anatomy-of-a-regression-test/
gpshead commentedon Aug 6, 2009
nice test case rhettg.
This is a correctness issue to prevent data loss on EINTR.
I've attached a patch that builds on rhettg's but allows the EINTR signal
to propagate upwards as desired by loweis and jorend for both read() and
readline() calls.
gpshead commentedon Aug 7, 2009
realistically, file objects (Objects/fileobject.c) never raise EINTR as
they use the C library fread/fwrite/fclose underneath. Why should a
socket based file object every be allowed to raise EINTR rather than
handling it internally?
IMHO people should only expect to handle EINTR when doing
socket.send/recv or os.read/write/close, not using a file-like object.
rhettg commentedon Aug 7, 2009
Good addition Gregory.
Totally agree on handling EINTR in even more cases.
You can't really expect the caller to know they need to deal with this
sort of thing when using these higher level call. The whole point is the
abstract away some of the complexity of dealing with the system call.
gpshead commentedon Aug 13, 2009
fixed in trunk r74426. socket.socket.sendall() and all
socket._fileobject methods (read/readline/write/flush) now properly
handle EINTR internally.
sendall will allow a python signal handler to raise an exception
aborting it in unknown state, otherwise it handles EINTR as it should
and continues sending.
I'm leaving this issue open until it is merged into py3k and considered
for 2.6 and/or 3.1 release backporting.
exarkun commentedon May 23, 2010
Backported to release26-maint in r81488.