用GDB排查Python程序故障

某Team在用Python开发一些代码，涉及子进程以及设法消除僵尸进程的需求。实践中他们碰上Python程序非预期退出的现象。最初他们决定用GDB调试Python解释器，查看exit()的源头。我听了之后，觉得这个问题应该用别的调试思路。帮他们排查这次程序故障时，除去原始问题，还衍生了其他问题。

这次的问题相比西安研发中心曾经碰上的Python信号处理问题，有不少基础知识、先验知识是共用的，此处不做再普及，感兴趣的同学可以翻看我以前发过的文章。

下文是一次具体的调试、分析记录。为了简化现场、方便调试，已将原始问题、衍生问题浓缩成DebugPythonWithGDB_6.py、DebugPythonWithGDB_7.py。

$ vi DebugPythonWithGDB_6.py

#!/usr/bin/env python

# -*- encoding: utf-8 -*-

import sys,os,signal,subprocess,shlex,traceback

def on_SIGCHLD(signum,frame):

print"[on_SIGCHLD"

sys.stdout.write("signum = %u "%signum)

traceback.print_stack(frame)

print os.waitpid(-1,os.WNOHANG)

"""

try :

print os.waitpid( -1, os.WNOHANG )

except OSError :

sys.stdout.write( 'Line[%u]: OSError ' % sys.exc_info()[2].tb_lineno )

"""

print"on_SIGCHLD]"

def do_more(count):

print'[do_more() begin %u]'%count

os.system(r'printf "Child = %u " $$;/bin/sleep 1')

"""

# 这里存在竞争条件，可以增加触发OSError异常的概率

os.system( r'printf "Child =%un" $$;/bin/sleep 1' )

"""

print'[do_more() end %u]'%count

def main(prog,args):

if0==len(args):

print'Usage: %s '%prog

else:

sys.stdout.write("Parent = %u "%os.getpid())

# 本例中，即使有下列代码，Ctrl-C仍然无效。

signal.signal(signal.SIGINT,signal.SIG_DFL)

# signal.signal( signal.SIGCHLD, signal.SIG_IGN )

signal.signal(signal.SIGCHLD,on_SIGCHLD)

count =0

whileTrue:

# 本例中父进程只是一个调度框架，不需要与子进程进行通信，因此不

# 需要特别处理"stdin=None, stdout=None, stderr=None"。

child =subprocess.Popen

(

# 不要直接用args[0].split()，它在处理单、双引号时不是我们

# 期望的行为。考虑这种例子，ls -l "/tmp/non exist"

shlex.split(args[0]),

# all file descriptors except 0, 1 and 2 will be closed

# before the child process is executed

close_fds =True,

cwd ="/tmp"

)

sys.stdout.write("Child = %u "%child.pid)

# child.send_signal( signal.SIGTERM )

# child.terminate()

child.kill()

# child.wait()

do_more(count)

count +=1

if'__main__'==__name__:

try:

main(os.path.basename(sys.argv[0]),sys.argv[1:])

except KeyboardInterrupt:

pass

$python DebugPythonWithGDB_6.py'python -c "import time;time.sleep(3600)"'

Parent =10244

Child =10245

[do_more()begin0]

[on_SIGCHLD

signum =17

File"DebugPythonWithGDB_6.py",line81,in

main(os.path.basename(sys.argv[0]),sys.argv[1:])

File"DebugPythonWithGDB_6.py",line76,inmain

do_more(count)

File"DebugPythonWithGDB_6.py",line20,indo_more

print'[do_more() begin %u]'%count

(10245,9)

on_SIGCHLD]

Child =10246

[on_SIGCHLD

signum =17

File"DebugPythonWithGDB_6.py",line81,in

main(os.path.basename(sys.argv[0]),sys.argv[1:])

File"DebugPythonWithGDB_6.py",line76,inmain

do_more(count)

File"DebugPythonWithGDB_6.py",line21,indo_more

os.system(r'printf "Child = %u " $$;/bin/sleep 1')

Traceback(most recent call last):

File"DebugPythonWithGDB_6.py",line81,in

main(os.path.basename(sys.argv[0]),sys.argv[1:])

File"DebugPythonWithGDB_6.py",line76,inmain

do_more(count)

File"DebugPythonWithGDB_6.py",line21,indo_more

os.system(r'printf "Child = %u " $$;/bin/sleep 1')

File"DebugPythonWithGDB_6.py",line10,inon_SIGCHLD

print os.waitpid(-1,os.WNOHANG)

OSError:[Errno10]No childprocesses

流程进入on_SIGCHLD()，但os.waitpid()抛出OSError异常。帮助里写的是，如果系统调用waitpid()返回-1，就抛出异常: An OSError is raised with the value of errno when the syscall returns -1. 10245号子进程在on_SIGCHLD()里waitpid()成功，(10245, 9)中的9表示该进程是被SIGKILL干掉的，符合预期。 10246号子进程是do_more()里的os.system()产生的shell进程，它结束时向10244号父进程投递了SIGCHLD信号。on_SIGCHLD()里waitpid()时，已经在别处wait*()过，10246号子进程已经彻底消失，系统调用waitpid()返回-1，Python函数os.waitpid()抛出异常。整个过程非常复杂，用伪代码描述如下:

do_more()

os.system()

posix_system() // posixmodule.c

__libc_system() // weak_alias (__libc_system, system)

do_system() // sysdeps/posix/system.c

* SIG_IGN

* Ctrl-C暂时失效

sigaction(SIGINT,&sa,&intr)

* 屏蔽(阻塞)SIGCHLD信号

sigaddset(&sa.sa_mask,SIGCHLD)

sigprocmask(SIG_BLOCK,&sa.sa_mask,&omask)

fork()

子进程(10246号子进程)

* 恢复原有SIGINT信号处理方式

sigaction(SIGINT,&intr,(structsigaction *)NULL)

* 调用"sh -c ..."

execve()

[shell子进程结束，向DebugPythonWithGDB_6.py投递SIGCHLD]

[由于SIGCHLD信号已被屏蔽(阻塞)，其保持在内核态的未决信号链上]

父进程(10244号父进程)

* 同步调用，会阻塞。不是在信号句柄中异步调用。

* 10246号子进程在此被wait*()回收后彻底消失

waitpid(pid,&status,0)

* 恢复原有SIGINT信号处理方式

sigaction(SIGINT,&intr,(structsigaction *)NULL)

* 取消对SIGCHLD的屏蔽(阻塞)

sigprocmask(SIG_SETMASK,&omask,(sigset_t *)NULL)

[SIGCHLD信号的屏蔽(阻塞)被取消]

[DebugPythonWithGDB_6.py的C级信号句柄signal_handler()安排"延迟调用"后返回]

[DebugPythonWithGDB_6.py的on_SIGCHLD()此时并未得到执行，因为built-in函数os.system()尚未返回]

* built-in函数os.system()返回后，10244号父进程开始处理"延迟调用"，调用

* Python级信号句柄。这个SIGCHLD信号是10246号子进程投递过来的。

* DebugPythonWithGDB_6.py的on_SIGCHLD()得到执行

on_SIGCHLD()

* 调用waitpid( -1, &status, WNOHANG )，试图处理10246号子进程。

* 10246号子进程已为前述waitpid( pid, &status, 0 )所处理，此处系统调用

* 返回-1，导致os.waitpid()抛出OSError异常。

os.waitpid(-1,os.WNOHANG)

整个过程之所以如此复杂，主要是因为Python的信号处理机制比较复杂，让已经非常复杂的Linux信号机制再添变数。参看:

1 2	《2.50对Python解释器进行调试》《22.0Linux信号机制》

就本例而言，为了确保DebugPythonWithGDB_6.py不因OSError异常而终止，只需在on_SIGCHLD()中调用os.waitpid()时捕捉OSError异常:

def on_SIGCHLD(signum,frame):

try:

print os.waitpid(-1,os.WNOHANG)

except OSError:

sys.stdout.write('Line[%u]: OSError '%sys.exc_info()[2].tb_lineno)

前述观点有些是动态调试得到，有些是静态分析得到。有人可能问了，为什么不拦截Python进程的C级信号句柄，查看SIGCHLD信号源，以此确认10246号子进程可能被回收两次？其实我最初也想这么干来着，但这是行不通的，因为Python的C级信号句柄signal_handler()是那种最原始的单形参信号句柄，不是高大上的三形参信号句柄。用GDB调试Python解释器:

# gdb -q -ex "b *signal_handler" -ex r --args /usr/bin/python2.7-dbg DebugPythonWithGDB_6.py '/usr/bin/python2.7-dbg -c "import time;time.sleep(3600)"'

...

Breakpoint1at0x8216f2d:file../Modules/signalmodule.c,line185.

Starting program:/usr/bin/python2.7-dbg DebugPythonWithGDB_6.py/usr/bin/python2.7-dbg-c"importtime;time.sleep(3600)"

[Thread debugging using libthread_db enabled]

Using host libthread_db library"/lib/i386-linux-gnu/i686/cmov/libthread_db.so.1".

Parent =10284

Child =10288

[do_more()begin0]

Child =10289

Breakpoint1,signal_handler(sig_num=17)at../Modules/signalmodule.c:185

185 {

(gdb)py-bt

#10 Frame 0xb7c20034, for file DebugPythonWithGDB_6.py, line 21, in do_more (count=0)

os.system(r'printf "Child = %u " $$;/bin/sleep 1')

#13 Frame 0xb7cb37dc, for file DebugPythonWithGDB_6.py, line 76, in main (prog='DebugPythonWithGDB_6.py', args=['/usr/bin/python2.7-dbg -c "import time;time.sleep(3600)"'], count=0, child=)

do_more(count)

#16 Frame 0xb7cbe49c, for file DebugPythonWithGDB_6.py, line 81, in ()

main(os.path.basename(sys.argv[0]),sys.argv[1:])

(gdb)bt7

#0 signal_handler (sig_num=17) at ../Modules/signalmodule.c:185

#2 0xb7fdcd3c in __kernel_vsyscall ()

#3 0xb7db25eb in __sigprocmask (how=how@entry=2, set=0x0, set@entry=0xbffff0d4, oset=oset@entry=0x0) at ../sysdeps/unix/sysv/linux/sigprocmask.c:57

#4 0xb7dc2084 in do_system (line=line@entry=0xb7cbf9e4 "printf "Child = %u\n" $$;/bin/sleep 1") at ../sysdeps/posix/system.c:161

#5 0xb7dc2380 in __libc_system (line=line@entry=0xb7cbf9e4 "printf "Child = %u\n" $$;/bin/sleep 1") at ../sysdeps/posix/system.c:184

#6 0xb7fa9bfb in system (line=0xb7cbf9e4 "printf "Child = %u\n" $$;/bin/sleep 1") at pt-system.c:28

(More stack frames follow...)

查看#4的system.c:161，这个位置已经在waitpid( pid, &status, 0 )之后: sigprocmask( SIG_SETMASK, &omask, (sigset_t *)NULL ) 其作用是取消对SIGCHLD的屏蔽(阻塞)。此时内存布局如下:

内存高址方向

fpstate // ESP+0x2DC output/x *(struct _fpstate *)($esp+0x2dc)

retcode // ESP+0x2D4 x/3i $esp+0x2d4

extramask // ESP+0x2D0 x/1wx $esp+0x2d0

fpstate_unused // ESP+0x60 output/x *(struct _fpstate *)($esp+0x60)

sigcontext_ia32// ESP+8 output/x *(struct sigcontext *)($esp+8)

sig // ESP+4 信号值，信号句柄第一形参

pretcode // ESP RetAddr=__kernel_sigreturn

// hexdump $esp 0x2dc

内存低址方向

(gdb)x/2wa$esp

0xbfffea6c: 0xb7fdcd18 0x11

(gdb)x/3i$esp+0x2d4

0xbfffed40: pop eax

0xbfffed41: mov eax,0x77

0xbfffed46: int 0x80

(gdb)output/x *(structsigcontext *)($esp+8)

{

gs=0x33,

__gsh=0x0,

fs=0x0,

__fsh=0x0,

es=0x7b,

__esh=0x0,

ds=0x7b,

__dsh=0x0,

edi=0xb7f2a000,

esi=0x8,

ebp=0x1,

esp=0xbfffeff0,

ebx=0x2,

edx=0x0,

ecx=0xbffff0d4,

eax=0x0,

trapno=0x1,

err=0x0,

eip=0xb7fdcd3c,

cs=0x73,

__csh=0x0,

eflags=0x246,

esp_at_signal=0xbfffeff0,

ss=0x7b,

__ssh=0x0,

fpstate=0xbfffed50,

oldmask=0x0,

cr2=0x0

}

因为是单形参信号句柄，没有siginfo，无法在用户态获知信号源。但我分析此时的信号源不是10289号子进程，而是10288号子进程。10288产生SIGCHLD时，SIGCHLD信号已被屏蔽(阻塞)，只能保持在内核态的未决信号链上。之后待10289产生SIGCHLD时，sigpending.signal中相应位已经置位，10289产生的SIGCHLD被丢弃，不会进入内核态的未决信号链。SIGCHLD信号的屏蔽(阻塞)被取消后，从内核态的未决信号链上取出10288产生的SIGCHLD进行处理。于是断点命中。如果完全理解了前述实验结果及分析，就会发现DebugPythonWithGDB_6.py存在竞争条件。subprocess.Popen()对应的子进程投递SIGCHLD信号时，父进程有两种可能:

1 2	1)os.system()调用sigprocmask(SIG_BLOCK,&sa.sa_mask,&omask)之前 2)os.system()调用sigprocmask(SIG_BLOCK,&sa.sa_mask,&omask)之后

情况1)会触发OSError异常，情况2)不会触发OSError异常。执行: $ python DebugPythonWithGDB_6.py ‘python -c “import time;time.sleep(3600)”‘ 有时会因OSError异常而终止，有时就一直循环执行下去。出现这种差异，正是竞争环境的表征。小结一下: 假设针对SIGCHLD安装了Python级信号句柄，其调用os.waitpid( -1, os.WNOHANG )回收子进程。如果别处会调用os.system()，则必须在os.waitpid()外侧捕捉OSError异常。不建议这种方式的混用。对waitpid()的分析到此就结束了，说点调试过程中出现的其他问题。意外地发现Ctrl-C无法终止情况2)，而我已经调用: signal.signal( signal.SIGINT, signal.SIG_DFL ) 这是因为do_system()中一上来就调用了:

1 2	sa.sa_handler =SIG_IGN; sigaction(SIGINT,&sa,&intr);

导致Ctrl-C暂时失效，直至do_system()结束。假设DebugPythonWithGDB_6.py已经出现情况2)，查看它的信号处理方式:

# ps auwx | grep python

root 10355 0.0 0.5 8116 5812pts/0 S+ 15:57 0:00pythonDebugPythonWithGDB_6.py python-c"import time;time.sleep(3600)"

root 10389 0.0 0.0 0 0pts/0 Z+ 15:57 0:00[python]

root 10393 0.0 0.0 2936 852pts/1 R+ 15:57 0:00grep python

# stap -DMAXACTION=10000 -g /usr/share/doc/systemtap-doc/examples/process/psig.stp -x 10355

10355: python

HUP default

INT ignored // 不是预期的default

QUIT ignored

ILL default

TRAP default

ABRT default

BUS default

FPE default

KILL default

USR1 default

SEGV default

USR2 default

PIPE ignored

ALRM default

TERM default

STKFLT default

CHLD blocked,caught0x818a480 0

...

上面显示SIGINT的处理方式是ignored，其实是ignored、default交叉出现，但我们基本上不可能看到default。 $ vi DebugPythonWithGDB_7.py

#!/usr/bin/env python

# -*- encoding: utf-8 -*-

import sys,os,subprocess,shlex,traceback

def do_more(count):

print'[do_more() begin %u]'%count

os.system(r'printf "Child = %u " $$;/bin/sleep 1')

print'[do_more() end %u]'%count

def main(prog,args):

if0==len(args):

print'Usage: %s '%prog

else:

sys.stdout.write("Parent = %u "%os.getpid())

count =0

whileTrue:

child =subprocess.Popen

(

shlex.split(args[0]),

close_fds =True,

cwd ="/tmp"

)

sys.stdout.write("Child = %u "%child.pid)

child.kill()

do_more(count)

count +=1

if'__main__'==__name__:

try:

main(os.path.basename(sys.argv[0]),sys.argv[1:])

except KeyboardInterrupt:

pass

$ python DebugPythonWithGDB_7.py ‘python -c “import time;time.sleep(3600)”‘ DebugPythonWithGDB_7.py没有显式调用wait()，它一直循环执行下去。我以为subprocess.Popen()会生成一堆僵尸进程。从另一个终端查看相关进程，发现始终只有一个僵尸进程，很快就被回收了。这个现象挺奇特，只能假设有隐式wait()存在。

# gdb -q -ex "b *waitpid" -ex r --args /usr/bin/python2.7-dbg DebugPythonWithGDB_7.py '/usr/bin/python2.7-dbg -c "import time;time.sleep(3600)"'

...

Parent =14508

Child =14512

[do_more()begin0]

Child =14513

[do_more()end0]

Breakpoint1,waitpid()at../sysdeps/unix/syscall-template.S:81

81 T_PSEUDO(SYSCALL_SYMBOL,SYSCALL_NAME,SYSCALL_NARGS)

(gdb)py-bt

#4 Frame 0xb7c21034, for file /usr/lib/python2.7/subprocess.py, line 1363, in _internal_poll (self=, _deadstate=2147483647, _waitpid=, _WNOHANG=1, _os_error=, _ECHILD=10)

pid,sts=_waitpid(self.pid,_WNOHANG)

#8 Frame 0xb7c6549c, for file /usr/lib/python2.7/subprocess.py, line 762, in __del__ (self=, _maxint=2147483647)

self._internal_poll(_deadstate=_maxint)

#18 Frame 0xb7cb37dc, for file DebugPythonWithGDB_7.py, line 22, in main (prog='DebugPythonWithGDB_7.py', args=['/usr/bin/python2.7-dbg -c "import time;time.sleep(3600)"'], count=1, child=)

cwd ="/tmp"

#21 Frame 0xb7cbe49c, for file DebugPythonWithGDB_7.py, line 31, in ()

main(os.path.basename(sys.argv[0]),sys.argv[1:])

(gdb)bt9

#0 waitpid () at ../sysdeps/unix/syscall-template.S:81

#1 0x081f80a3 in posix_waitpid (self=0x0, args=(14512, 1)) at ../Modules/posixmodule.c:6207

#2 0x080bc300 in PyCFunction_Call (func=, arg=(14512, 1), kw=0x0) at ../Objects/methodobject.c:81

#3 0x08149d0b in call_function (pp_stack=0xbfffebd4, oparg=2) at ../Python/ceval.c:4033

#4 0x081454ec in PyEval_EvalFrameEx (f=Frame 0xb7c21034, for file /usr/lib/python2.7/subprocess.py, line 1363, in _internal_poll (self=, _deadstate=2147483647, _waitpid=, _WNOHANG=1, _os_error=, _ECHILD=10), throwflag=0) at ../Python/ceval.c:2679

#5 0x08147a77 in PyEval_EvalCodeEx (co=0xb7c60448, globals={'STDOUT': -2, '_has_poll': True, 'gc': , 'check_call': , 'mswindows': False, 'select': , 'list2cmdline': , '__all__': ['Popen', 'PIPE', 'STDOUT', 'call', 'check_call', 'check_output', 'CalledProcessError'], 'errno': , '_demo_posix': , '__package__': None, 'PIPE': -1, '_cleanup': , '_eintr_retry_call': , 'call': , '__doc__': 'subprocess - Subprocesses with accessible I/O streams This module allows you to spawn processes, connect to their input/output/error pipes, and obtain their return codes. This module intends to replace several older modules and functions: os.system os.spawn* os.popen* popen2.* commands.* Information about how the subprocess module can be used to replace these modules and functions can be found below. ...(truncated), locals=0x0, args=0xb7c655e4, argcount=1, kws=0xb7c655e8, kwcount=1, defs=0xb7c73a20, defcount=5, closure=0x0) at ../Python/ceval.c:3265

#6 0x0814a1e5 in fast_function (func=, pp_stack=0xbfffeef4, n=3, na=1, nk=1) at ../Python/ceval.c:4129

#7 0x08149e93 in call_function (pp_stack=0xbfffeef4, oparg=256) at ../Python/ceval.c:4054

#8 0x081454ec in PyEval_EvalFrameEx (f=Frame 0xb7c6549c, for file /usr/lib/python2.7/subprocess.py, line 762, in __del__ (self=, _maxint=2147483647), throwflag=0) at ../Python/ceval.c:2679

(More stack frames follow...)

调用栈回溯表明:

查看: /usr/lib/python2.7/subprocess.py:1363

try:

_waitpid(self.pid,_WNOHANG)

...

except _os_error ase:

...

ife.errno==_ECHILD:

# This happens if SIGCLD is set to be ignored or waiting for child

# processes has otherwise been disabled for our process. This

# child is dead, we can't get the status.

# http://bugs.python.org/issue15756

...

考虑这种情形，调用subprocess.Popen()之前做过如下动作之一:

1 2	signal.signal(signal.SIGCHLD,signal.SIG_IGN) signal.signal(signal.SIGCHLD,on_SIGCHLD)

_internal_poll()调用_waitpid()时，在外侧捕捉了_os_error异常，就是对付上述可能。subprocess.Popen()不是built-in函数，对应很多条PVM指令，不像os.system()是built-in函数，对应单条PVM指令。前者执行过程中，Python级信号句柄on_SIGCHLD()有很大机会得到执行，_internal_poll()调用_waitpid()，遭遇_os_error异常的可能性不低。 Popen()对象的析构函数自动调用wait*()，于是Popen()对象离开作用域时自动回收子进程。 Python能够改变变量作用域的是def、class、lamda、global。下面这些并不涉及作用域变更:

if/elif/else

try/except/finally

for/while

DebugPythonWithGDB_7.py中Popen()对象的作用域是main()。尽管没有离开作用域，但对child变量的重新赋值会触发对前一个Popen()对象的析构。这就解释了为什么始终只有一个僵尸进程。回顾原始问题与衍生问题，再次表明，掌握对Python解释器的调试技术可以快速排查众多看似神秘实则基础的程序故障。