Philip Winston said:
We have a multiprocess Python program that uses Queue to communicate
between processes. Recently we've seen some errors while blocked
waiting on Queue.get:
IOError: [Errno 4] Interrupted system call
What causes the exception?
Unix divides system calls up into "slow" and "fast". The difference is
how the react to signals.
Fast calls are things which are expected to return quickly. A canonical
example would get getuid(), which just returns a number it looks up in a
kernel data structure. Fast syscalls cannot be interrupted by signals.
If a signal arrives while a fast syscall is running, delivery of the
signal is delayed until after the call returns.
Slow calls are things which may take an indeterminate amount of time to
return. An example would be a read on a network socket; it will block
until a message arrives, which may be forever. Slow syscalls get
interrupted by signals. If a signal arrives while a slow syscall is
blocking, the call returns EINTR. This lets your code have a chance to
do whatever is appropriate, which might be clean up in preparation for
process shutdown, or maybe just ignore the interrupt and re-issue the
system call.
Here's a short python program which shows how this works (tested on
MacOS-10.6, but should be portable to just about any posix box):
-----------------------------------------------------
#!/usr/bin/env python
import socket
import signal
import os
def handler(sig_num, stack_frame):
return
print "my pid is", os.getpid()
signal.signal(signal.SIGUSR1, handler)
s = socket.socket(type=socket.SOCK_DGRAM)
s.bind(("127.0.0.1", 0))
s.recv(1024)
-----------------------------------------------------
Run this in one window. It should print out its process number, then
block on the recv() call. In another window, send it a SIGUSR1. You
should get something like:
play$ ./intr.py
my pid is 6969
Traceback (most recent call last):
File "./intr.py", line 14, in <module>
s.recv(1024)
socket.error: [Errno 4] Interrupted system call
Is it necessary to catch this exception
and manually retry the Queue operation? Thanks.
That's a deeper question which I can't answer. My guess is the
interrupted system call is the Queue trying to acquire a lock, but
there's no predicting what the signal is. I'm tempted to say that it's
a bug in Queue that it doesn't catch this exception internally, but
people who know more about the Queue implementation than I do should
chime in.
We have some Python 2.5 and 2.6 machines that have run this program
for many 1,000 hours with no errors. But we have one 2.5 machine and
one 2.7 machine that seem to get the error very often.
Yup, that's the nature of signal delivery race conditions in
multithreaded programs. Every machine will behave a little bit
differently, with no rhyme or reason. Google "undefined behavior" for
more details
The whole posix signal delivery mechanism dates back
to the earliest Unix implementations, long before there were threads or
networks. At this point, it's got many layers of duct tape.
