万无一失的跨平台进程杀死守护进程

Question

我有一些python自动化,它产生telnet我用linux script命令记录的会话; script每个日志记录会话有两个进程ID(父级和子级).

我需要解决一个问题,如果python自动化脚本死了,script会话永远不会自己关闭; 由于某种原因,这比它应该更难.

到目前为止,我已经实现了watchdog.py(请参阅问题的底部)守护自身,并在循环中轮询python自动化脚本的PID.当它看到python自动化PID从服务器的进程表中消失时,它会尝试终止script会话.

我的问题是:

• script会话总是产生两个独立的进程,其中一个script会话是另一个script会话的父进程.
• watchdog.py script如果我script从自动化脚本启动会话,则不会终止子会话(请参阅下面的自动化示例)

自动化示例(reproduce_bug.py)

import pexpect as px
from subprocess import Popen
import code
import time
import sys
import os

def read_pid_and_telnet(_child, addr):
    time.sleep(0.1) # Give the OS time to write the PIDFILE
    # Read the PID in the PIDFILE
    fh = open('PIDFILE', 'r')
    pid = int(''.join(fh.readlines()))
    fh.close()
    time.sleep(0.1)
    # Clean up the PIDFILE
    os.remove('PIDFILE')
    _child.expect(['#', '\$'], timeout=3)
    _child.sendline('telnet %s' % addr)
    return str(pid)

pidlist = list()
child1 = px.spawn("""bash -c 'echo $$ > PIDFILE """
    """&& exec /usr/bin/script -f LOGFILE1.txt'""")
pidlist.append(read_pid_and_telnet(child1, '10.1.1.1'))

child2 = px.spawn("""bash -c 'echo $$ > PIDFILE """
    """&& exec /usr/bin/script -f LOGFILE2.txt'""")
pidlist.append(read_pid_and_telnet(child2, '10.1.1.2'))

cmd = "python watchdog.py -o %s -k %s" % (os.getpid(), ','.join(pidlist))
Popen(cmd.split(' '))
print "I started the watchdog with:\n   %s" % cmd

time.sleep(0.5)
raise RuntimeError, "Simulated script crash.  Note that script child sessions are hung"

现在举例说明当我运行上述自动化时会发生什么...请注意,PID 30017产生30018,PID 30020产生30021.所有上述PID都是script会话.

[mpenning@Hotcoffee Network]$ python reproduce_bug.py 
I started the watchdog with:
   python watchdog.py -o 30016 -k 30017,30020
Traceback (most recent call last):
  File "reproduce_bug.py", line 35, in <module>
    raise RuntimeError, "Simulated script crash.  Note that script child sessions are hung"
RuntimeError: Simulated script crash.  Note that script child sessions are hung
[mpenning@Hotcoffee Network]$

在我运行上面的自动化之后,所有子script会话仍在运行.

[mpenning@Hotcoffee Models]$ ps auxw | grep script
mpenning 30018  0.0  0.0  15832   508 ?        S    12:08   0:00 /usr/bin/script -f LOGFILE1.txt
mpenning 30021  0.0  0.0  15832   516 ?        S    12:08   0:00 /usr/bin/script -f LOGFILE2.txt
mpenning 30050  0.0  0.0   7548   880 pts/8    S+   12:08   0:00 grep script
[mpenning@Hotcoffee Models]$

我在Debian Squeeze linux系统(uname -a :)上运行Python 2.6.6下的自动化Linux Hotcoffee 2.6.32-5-amd64 #1 SMP Mon Jan 16 16:22:28 UTC 2012 x86_64 GNU/Linux.

题:

似乎守护进程在产生进程崩溃时无法生存.如果自动化程序死亡(如上例所示),如何修复watchdog.py以关闭所有脚本会话？

一个watchdog.py说明问题的日志(遗憾的是,PID与原始问题不一致)......

[mpenning@Hotcoffee ~]$ cat watchdog.log 
2012-02-22,15:17:20.356313 Start watchdog.watch_process
2012-02-22,15:17:20.356541     observe pid = 31339
2012-02-22,15:17:20.356643     kill pids = 31352,31356
2012-02-22,15:17:20.356730     seconds = 2
[mpenning@Hotcoffee ~]$

解析度

问题基本上是竞争条件.当我试图杀死"父" script进程时,他们已经死于与自动化事件巧合......

要解决这个问题......首先,监视程序守护程序需要在轮询观察到的PID之前识别要杀死的整个子项列表(我的原始脚本试图在观察到的PID崩溃后识别子项).接下来,我不得不修改我的看门狗守护进程,以允许某些script进程可能因观察到的PID而死亡.

---

watchdog.py:

#!/usr/bin/python
"""
Implement a cross-platform watchdog daemon, which observes a PID and kills 
other PIDs if the observed PID dies.

Example:
--------

watchdog.py -o 29322 -k 29345,29346,29348 -s 2

The command checks PID 29322 every 2 seconds and kills PIDs 29345, 29346, 29348 
and their children, if PID 29322 dies.

Requires:
----------

 * https://github.com/giampaolo/psutil
 * http://pypi.python.org/pypi/python-daemon
"""
from optparse import OptionParser
import datetime as dt
import signal
import daemon
import logging
import psutil
import time
import sys
import os

class MyFormatter(logging.Formatter):
    converter=dt.datetime.fromtimestamp
    def formatTime(self, record, datefmt=None):
        ct = self.converter(record.created)
        if datefmt:
            s = ct.strftime(datefmt)
        else:
            t = ct.strftime("%Y-%m-%d %H:%M:%S")
            s = "%s,%03d" % (t, record.msecs)
        return s

def check_pid(pid):        
    """ Check For the existence of a unix / windows pid."""
    try:
        os.kill(pid, 0)   # Kill 0 raises OSError, if pid isn't there...
    except OSError:
        return False
    else:
        return True

def kill_process(logger, pid):
    try:
        psu_proc = psutil.Process(pid)
    except Exception, e:
        logger.debug('Caught Exception ["%s"] while looking up PID %s' % (e, pid))
        return False

    logger.debug('Sending SIGTERM to %s' % repr(psu_proc))
    psu_proc.send_signal(signal.SIGTERM)
    psu_proc.wait(timeout=None)
    return True

def watch_process(observe, kill, seconds=2):
    """Kill the process IDs listed in 'kill', when 'observe' dies."""
    logger = logging.getLogger(__name__)
    logger.setLevel(logging.DEBUG)
    logfile = logging.FileHandler('%s/watchdog.log' % os.getcwd())
    logger.addHandler(logfile)
    formatter = MyFormatter(fmt='%(asctime)s %(message)s',datefmt='%Y-%m-%d,%H:%M:%S.%f')
    logfile.setFormatter(formatter)


    logger.debug('Start watchdog.watch_process')
    logger.debug('    observe pid = %s' % observe)
    logger.debug('    kill pids = %s' % kill)
    logger.debug('    seconds = %s' % seconds)
    children = list()

    # Get PIDs of all child processes...
    for childpid in kill.split(','):
        children.append(childpid)
        p = psutil.Process(int(childpid))
        for subpsu in p.get_children():
            children.append(str(subpsu.pid))

    # Poll observed PID...
    while check_pid(int(observe)):
        logger.debug('Poll PID: %s is alive.' % observe)
        time.sleep(seconds)
    logger.debug('Poll PID: %s is *dead*, starting kills of %s' % (observe, ', '.join(children)))

    for pid in children:
        # kill all child processes...
        kill_process(logger, int(pid))
    sys.exit(0) # Exit gracefully

def run(observe, kill, seconds):

    with daemon.DaemonContext(detach_process=True, 
        stdout=sys.stdout,
        working_directory=os.getcwd()):
        watch_process(observe=observe, kill=kill, seconds=seconds)

if __name__=='__main__':
    parser = OptionParser()
    parser.add_option("-o", "--observe", dest="observe", type="int",
                      help="PID to be observed", metavar="INT")
    parser.add_option("-k", "--kill", dest="kill",
                      help="Comma separated list of PIDs to be killed", 
                      metavar="TEXT")
    parser.add_option("-s", "--seconds", dest="seconds", default=2, type="int",
                      help="Seconds to wait between observations (default = 2)", 
                      metavar="INT")
    (options, args) = parser.parse_args()
    run(options.observe, options.kill, options.seconds)

Answer 1

问题本质上是竞争条件。当我试图杀死“父”脚本进程时，它们已经与自动化事件同时死亡......

为了解决这个问题...首先，看门狗守护进程需要在轮询观察到的 PID 之前识别要杀死的子进程的整个列表（我的原始脚本尝试在观察到的 PID 崩溃后识别子进程）。接下来，我必须修改我的看门狗守护程序，以允许某些脚本进程可能因观察到的 PID 而终止。

看门狗.py

#!/usr/bin/python
"""
Implement a cross-platform watchdog daemon, which observes a PID and kills 
other PIDs if the observed PID dies.

Example:
--------

watchdog.py -o 29322 -k 29345,29346,29348 -s 2

The command checks PID 29322 every 2 seconds and kills PIDs 29345, 29346, 29348 
and their children, if PID 29322 dies.

Requires:
----------

 * https://github.com/giampaolo/psutil
 * http://pypi.python.org/pypi/python-daemon
"""
from optparse import OptionParser
import datetime as dt
import signal
import daemon
import logging
import psutil
import time
import sys
import os

class MyFormatter(logging.Formatter):
    converter=dt.datetime.fromtimestamp
    def formatTime(self, record, datefmt=None):
        ct = self.converter(record.created)
        if datefmt:
            s = ct.strftime(datefmt)
        else:
            t = ct.strftime("%Y-%m-%d %H:%M:%S")
            s = "%s,%03d" % (t, record.msecs)
        return s

def check_pid(pid):        
    """ Check For the existence of a unix / windows pid."""
    try:
        os.kill(pid, 0)   # Kill 0 raises OSError, if pid isn't there...
    except OSError:
        return False
    else:
        return True

def kill_process(logger, pid):
    try:
        psu_proc = psutil.Process(pid)
    except Exception, e:
        logger.debug('Caught Exception ["%s"] while looking up PID %s' % (e, pid))
        return False

    logger.debug('Sending SIGTERM to %s' % repr(psu_proc))
    psu_proc.send_signal(signal.SIGTERM)
    psu_proc.wait(timeout=None)
    return True

def watch_process(observe, kill, seconds=2):
    """Kill the process IDs listed in 'kill', when 'observe' dies."""
    logger = logging.getLogger(__name__)
    logger.setLevel(logging.DEBUG)
    logfile = logging.FileHandler('%s/watchdog.log' % os.getcwd())
    logger.addHandler(logfile)
    formatter = MyFormatter(fmt='%(asctime)s %(message)s',datefmt='%Y-%m-%d,%H:%M:%S.%f')
    logfile.setFormatter(formatter)


    logger.debug('Start watchdog.watch_process')
    logger.debug('    observe pid = %s' % observe)
    logger.debug('    kill pids = %s' % kill)
    logger.debug('    seconds = %s' % seconds)
    children = list()

    # Get PIDs of all child processes...
    for childpid in kill.split(','):
        children.append(childpid)
        p = psutil.Process(int(childpid))
        for subpsu in p.get_children():
            children.append(str(subpsu.pid))

    # Poll observed PID...
    while check_pid(int(observe)):
        logger.debug('Poll PID: %s is alive.' % observe)
        time.sleep(seconds)
    logger.debug('Poll PID: %s is *dead*, starting kills of %s' % (observe, ', '.join(children)))

    for pid in children:
        # kill all child processes...
        kill_process(logger, int(pid))
    sys.exit(0) # Exit gracefully

def run(observe, kill, seconds):

    with daemon.DaemonContext(detach_process=True, 
        stdout=sys.stdout,
        working_directory=os.getcwd()):
        watch_process(observe=observe, kill=kill, seconds=seconds)

if __name__=='__main__':
    parser = OptionParser()
    parser.add_option("-o", "--observe", dest="observe", type="int",
                      help="PID to be observed", metavar="INT")
    parser.add_option("-k", "--kill", dest="kill",
                      help="Comma separated list of PIDs to be killed", 
                      metavar="TEXT")
    parser.add_option("-s", "--seconds", dest="seconds", default=2, type="int",
                      help="Seconds to wait between observations (default = 2)", 
                      metavar="INT")
    (options, args) = parser.parse_args()
    run(options.observe, options.kill, options.seconds)