subprocess.popen（）使用管道/失败的stderr重定向

Question

I want to run a process which may produce a lot of output for up to timeout seconds, capturing the stdout / stderr . Using capture() and PIPE as stdout / stderr is prone to deadlocking according to the documentation for subprocess .

Now, I'm using poll() anyways -- because I want to be able to kill the process after the timeout -- but I still don't know how to avoid the deadlock using PIPE. How do I do that?

Currently I'm just working around by creating tempfiles:

#because of the shitty api, this has to be a file, because std.PIPE is prone to deadlocking with a lot of output, and I can't figure out what to do about it
out, outfile = tempfile.mkstemp()
err, errfile = tempfile.mkstemp()

now = datetime.datetime.now().strftime('%H:%M, %Ss')
print "Running '" + exe + "' with a timeout of ", timeout , "s., starting at ", now
p = subprocess.Popen(args = exe,
                     stdout = out,
                     #for some reason, err isn't working if the process is killed by the kernel for, say, using too much memory.
                     stderr = err,
                     cwd = dir)

start = time.time()

# take care of infinite loops
sleepDuration = 0.25
time.sleep(0.1)
lastPrintedDuration = 0
duration = 0
while p.poll() is None:
    duration = time.time() - start
    if duration > lastPrintedDuration + 1:
        lastPrintedDuration += 1
        #print '.',
        sys.stdout.flush()
    if duration >= timeout:
        p.kill()
        raise Exception("Killed after " + str(duration) + "s.")
    time.sleep(sleepDuration)

if p.returncode is not 0:
    with open(errfile, 'r') as f:
        e = f.read()
        #fix empty error messages
        if e == '':
            e = 'Program crashed, or was killed by kernel.'
        f.close()

    os.close(out)
    os.close(err)
    os.unlink(outfile)
    os.unlink(errfile)
    print "Error after " + str(duration) + 's: ',
    print "'" + e + "'"
    raw_input('test')
    raise Exception(e)
else:
    print "completed in " + str(duration) + 's.'

os.close(out)
os.close(err)
os.unlink(outfile)
os.unlink(errfile)

But even this fails to capture errors if the process is killed by, say, the kernel (out of memory, etc.).

What's the ideal solution to this problem?

Answer 1

Instead of using files for the output, go back to using pipes but use the fcntl module to put p.stdout and p.stderr into non-blocking mode. This will cause p.stdout.read() and p.stderr.read() to return whatever data is available or raise an IOError if there is no data, instead of blocking:

import fcntl, os

p = subprocess.Popen(args = exe,
                     stdout = subprocess.PIPE,
                     stderr = subprocess.PIPE,
                     cwd = dir)
fcntl.fcntl(p.stdout.fileno(), fcntl.F_SETFL, os.O_NONBLOCK)
fcntl.fcntl(p.stderr.fileno(), fcntl.F_SETFL, os.O_NONBLOCK)

outdata, errdata = '', ''
while p.poll() is None:
    try:
        outdata += p.stdout.read()
    except IOError:
        pass
    try:
        errdata += p.stderr.read()
    except IOError:
        pass
    time.sleep(sleepDuration)

As glglgl pointed out in comments, you should do some additional checking in the except IOError clause to make sure that it is not actually a real error.

Answer 2

The trouble with non-blocking mode is that you end up busy-waiting for I/O. The more conventional approach is to use one of the select calls. Even if you have only one file descriptor to read/write, you can stick your desired timeout on it, so you regain control after the specified interval with no further I/O.