| msg51655 - (view) |
Author: Maxim Sobolev (sobomax) |
Date: 2007-01-04 21:37 |
| 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 |
|
|
| msg51656 - (view) |
Author: Oren Tirosh (orenti) |
Date: 2007-01-07 18:24 |
| 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. |
|
|
| msg51657 - (view) |
Author: Maxim Sobolev (sobomax) |
Date: 2007-01-08 10:51 |
| 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. |
|
|
| msg51658 - (view) |
Author: Martin v. Löwis (loewis) *  |
Date: 2007-02-16 13:05 |
| 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. |
|
|
| msg51659 - (view) |
Author: Jason Orendorff (jorend) |
Date: 2007-03-07 17:54 |
| 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. |
|
|
| msg91209 - (view) |
Author: Rhett Garber (rhettg) |
Date: 2009-08-02 22:05 |
| 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/ |
|
|
| msg91359 - (view) |
Author: Gregory P. Smith (gregory.p.smith) * |
Date: 2009-08-06 13:21 |
| 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. |
|
|
| msg91409 - (view) |
Author: Gregory P. Smith (gregory.p.smith) * |
Date: 2009-08-07 18:05 |
| 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. |
|
|
| msg91414 - (view) |
Author: Rhett Garber (rhettg) |
Date: 2009-08-07 23:19 |
| 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. |
|
|
| msg91532 - (view) |
Author: Gregory P. Smith (gregory.p.smith) * |
Date: 2009-08-13 18:57 |
| 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. |
|
|
| msg106334 - (view) |
Author: Jean-Paul Calderone (exarkun) * |
Date: 2010-05-23 15:22 |
| Backported to release26-maint in r81488. |
|
|