GDB 调试入门

gdb 调试入门,大牛写的高质量指南 http://blog.jobbole.com/107759/

SysAdmim 必备:系统性能大牛 Brendan Gregg 分享的 Linux 性能工具 http://blog.jobbole.com/76103/

 

源码解读poll/select内核机制

http://gityuan.com/2019/01/05/linux-poll-select/

http://gityuan.com/2019/01/06/linux-epoll/

 

Brendan Gregg 目前是 Netflix 的高级性能架构师 ,他在那里做大规模计算机性能设计、分析和调优。他是《Systems Performance》等技术书的作者,因在系统管理员方面的成绩,获得过 2013年 USENIX LISA 大奖。他之前是 SUN 公司是性能领头人和内核工程师,他在 SUN 开发过 ZFS L2ARC,研究存储和网络性能。他也发明和开发过一大波性能分析工具,很多已集成到操作系统中了 。他的最近工作包括研究性能分析的方法论和可视化,其目标包括Linux内核。

上面这是 Gregg 的简介,正如其中说的,他个人站点上分享了很多 Linux 性能相关的资源,都是自己开发的:

 Linux observability tools | Linux 性能观测工具

 

 Linux benchmarking tools | Linux 性能测评工具

 

Linux tuning tools | Linux 性能调优工具

 

Linux observability sar

 

愈看更多图表和演讲,请移步:http://www.brendangregg.com/linuxperf.html

补充:去年在微博分享  Brendan Gregg 在 SCaLE 11x 大会上的演讲《Linux Performance Analysis and Tools | Linux 性能分析和工具》,这个链接(http://t.cn/zYHtxoL )中有演讲视频(Youtube)和幻灯片(SlideShare),所以请自带梯子访问。

 

先留着,还没来得及看:

没想到Brendan Gregg这样的大牛,会写出这样一篇gdb tutorials文章:gdb Debugging Full Example (Tutorial): ncurses 。但可能正如文章开头所说,大牛对网上的gdb文章都不太满意,所以才有了这篇高质量指南,gdb入门者的福音。—— 何登成

如果你是系统管理员,但还不认识 Brendan Gregg,那网上流传甚广的 3 张 Linux 性能工具图(链接),你应该看过的。—— 伯小乐。

8f8d0abbc7f7212d6f1c06c048c4433c-300x300.jpeguploading.4e448015.gif转存失败重新上传取消8f8d0abbc7f7212d6f1c06c048c4433c-300x300.jpeguploading.4e448015.gif转存失败重新上传取消brendan-gregg

( Brendan Gregg)

gdb 调试 ncurses 全过程:

发现网上的“gdb 示例”只有命令而没有对应的输出,我有点不满意。gdb 是 GNU 调试器,Linux 上的标配调试器。当我看 Greg Law 在 CppCon 2015 上的演讲《给我 15 分钟,我将改变你的对 GDB 的认知》的时候,我想起了示例输出的不足,幸运的是,这次有输出!这 15 分钟太值了。

它也启发我去分享一个完整的 gdb 调试实例,包含输出和每个步骤,甚至钻牛角尖的情况。这不是一个特别有趣或奇怪的问题,只是常规的 gdb 调试会话。但它包含了基础的东西可以勉强作为教程使用,记住 gdb 里还有很多东西我这里没用到。

我会以 root 权限运行下面的命令,因为我在调试一个工具,它需要 root 权限(目前)。需要的时候可用 sudo 获取 root 权限。你也没必要通读全篇︰ 我已列出每一步,你可以浏览它们找感兴趣的看。

1. 问题概述

BPF 工具箱里的 bcc 工具集有一个对cachetop.py 的 pull 请求,它通过程序使用 top-like display 显示 page cache 的统计。太好了 !然而,当我测试它时,遇到了段错误︰

 

1

2

# ./cachetop.py

Segmentation fault

注意它说的是“段错误”,不是“段错误(核心已转储)”。我想要一个核心转储文件用来调试。(核心转储文件是进程内存的拷贝 – 这个名字来源于磁芯存储器时代 – 可用调试器分析)

分析核心转储文件是一种方法,但不是调试这个问题的唯一方法。我可以在 gdb 中运行此程序,来检查这个问题。我也可以在段错误发生时,用外部追踪器去抓数据和栈帧。我们从核心转储文件入手。

2. 解决核心转储问题

我检查一下核心转储的设置:

 

1

2

3

4

# ulimit -c

0

# cat /proc/sys/kernel/core_pattern

core

ulimit -c 显示核心转储文件大小的最大值,这里是零:禁止核心转储(对于本进程和它的子进程)。

/proc/…/core_pattern 仅仅被设为 “core”,表示会在当前目录下生成一个文件名为 “core” 的 核心转储文件。目前这样就行了,但是我要演示如何把它设置为全局位置。

 

1

2

3

# ulimit -c unlimited

# mkdir /var/cores

# echo "/var/cores/core.%e.%p" > /proc/sys/kernel/core_pattern

你可以进一步定制 core_pattern;例如,%h 为主机名,%t 为转储的时间。这些选项被写在 Linux 内核源码 Documentation/sysctl/kernel.txt中。

要使 core_pattern 保持不变,重启之后仍然有效,你可以通过设置 /etc/sysctl.conf 里的 “kernel.core_pattern” 实现。

再来一次:

 

1

2

3

4

5

6

7

# ./cachetop.py

Segmentation fault (core dumped)

# ls -lh /var/cores

total 19M

-rw------- 1 root root 20M Aug  7 22:15 core.python.30520

# file /var/cores/core.python.30520

/var/cores/core.python.30520: ELF 64-bit LSB core file x86-64, version 1 (SYSV), SVR4-style, from 'python ./cachetop.py'

好多了:我们有了自己的核心转储文件。

3. 启动 GDB

现在我要用 gdb 启动目标程序(用 shell 替换符,”`”,不过在你确定能用的情况下,也可指定完整路径),和核心转储文件:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

# gdb `which python` /var/cores/core.python.30520

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

Copyright (C) 2016 Free Software Foundation, Inc.

License GPLv3+: GNU GPL version 3 or later

This is free software: you are free to change and redistribute it.

There is NO WARRANTY, to the extent permitted by law.  Type "show copying"

and "show warranty" for details.

This GDB was configured as "x86_64-linux-gnu".

Type "show configuration" for configuration details.

For bug reporting instructions, please see:

.

Find the GDB manual and other documentation resources online at:

.

For help, type "help".

Type "apropos word" to search for commands related to "word"...

Reading symbols from /usr/bin/python...(no debugging symbols found)...done.

warning: core file may not match specified executable file.

[New LWP 30520]

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

warning: JITed object file architecture unknown is not compatible with target architecture i386:x86-64.

Core was generated by `python ./cachetop.py'.

Program terminated with signal SIGSEGV, Segmentation fault.

#0  0x00007f0a37aac40d in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

最后两行很有趣:它告诉我们这个段错误发生在 libncursesw 库里 doupdate() 函数中。可以先在网上搜一下,以防这是个很常见的问题。我搜了一下,可是没发现一个常见的原因。

我已经猜到 libncursesw 是什么了,如果你对它很陌生,它在 “/lib” 目录下以 “.so.*” 结尾表明这是一个动态库文件,可能有 man 手册、网站、包描述等。

 

1

2

3

# dpkg -l | grep libncursesw

ii  libncursesw5:amd64                  6.0+20160213-1ubuntu1                    amd64

     shared libraries for terminal handling (wide character support)

我是碰巧在 Ubuntu 上调试,但用什么 Linux发行版对使用 gdb 并没有影响。

4. 回溯

栈回溯显示我们是如何到达失败点的,通常足够帮助我们确定常见的问题。bt (backtrace的简写)常常是我在 gdb 中使用的第一条命令:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

(gdb) bt

#0  0x00007f0a37aac40d in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

#1  0x00007f0a37aa07e6 in wrefresh () from /lib/x86_64-linux-gnu/libncursesw.so.5

#2  0x00007f0a37a99616 in ?? () from /lib/x86_64-linux-gnu/libncursesw.so.5

#3  0x00007f0a37a9a325 in wgetch () from /lib/x86_64-linux-gnu/libncursesw.so.5

#4  0x00007f0a37cc6ec3 in ?? () from /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

#5  0x00000000004c4d5a in PyEval_EvalFrameEx ()

#6  0x00000000004c2e05 in PyEval_EvalCodeEx ()

#7  0x00000000004def08 in ?? ()

#8  0x00000000004b1153 in PyObject_Call ()

#9  0x00000000004c73ec in PyEval_EvalFrameEx ()

#10 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#11 0x00000000004caf42 in PyEval_EvalFrameEx ()

#12 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#13 0x00000000004c2ba9 in PyEval_EvalCode ()

#14 0x00000000004f20ef in ?? ()

#15 0x00000000004eca72 in PyRun_FileExFlags ()

#16 0x00000000004eb1f1 in PyRun_SimpleFileExFlags ()

#17 0x000000000049e18a in Py_Main ()

#18 0x00007f0a3be10830 in __libc_start_main (main=0x49daf0 <main>, argc=2, argv=0x7ffd33d94838, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>,

    stack_end=0x7ffd33d94828) at ../csu/libc-start.c:291

#19 0x000000000049da19 in _start ()

从下往上,按照从父函数到子函数的顺序看。有 “??” 的地方是因为符号解析失败。遍历栈 – 用来生成栈帧 — 也会失败。在这种情况下你可能会看到一个正常的栈帧,跟着一个小数值的假地址。如果符号或栈破损很严重,导致无法理解栈回溯,这里有几个常用的办法来修复:安装 debug info 包(给 gdb 提供更多的符号,让它来做基于 DWARF 的栈遍历),或者重新用源码编译(-fno-omit-frame-pointer -g)一个带帧指针和调试信息的版本。以上大多数 “??”
可以通过安装 python-dbg 包来修复。

这些栈看起来不太有用:帧 5 到 17 (左边的索引)在 Python 内部,虽然还看不到 Python 方法。帧 4 是 _curses 库,然后就到了 libncursesw。看起来调用顺序是 wgetch()->wrefresh()->doupdate()。根据函数名来看,我猜是刷新窗口。为什么会导致核心转储 呢?

5. 反汇编

我从反汇编发生段错误的函数 doupdate() 开始:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

(gdb) disas doupdate

Dump of assembler code for function doupdate:

   0x00007f0a37aac2e0 <+0>:   push   %r15

   0x00007f0a37aac2e2 <+2>:   push   %r14

   0x00007f0a37aac2e4 <+4>:   push   %r13

   0x00007f0a37aac2e6 <+6>:   push   %r12

   0x00007f0a37aac2e8 <+8>:   push   %rbp

   0x00007f0a37aac2e9 <+9>:   push   %rbx

   0x00007f0a37aac2ea <+10>:  sub    $0xc8,%rsp

[...]

---Type <return> to continue, or q <return> to quit---

[...]

   0x00007f0a37aac3f7 <+279>: cmpb   $0x0,0x21(%rcx)

   0x00007f0a37aac3fb <+283>: je     0x7f0a37aacc3b <doupdate+2395>

   0x00007f0a37aac401 <+289>: mov    0x20cb68(%rip),%rax        # 0x7f0a37cb8f70

   0x00007f0a37aac408 <+296>: mov    (%rax),%rsi

   0x00007f0a37aac40b <+299>: xor    %eax,%eax

=> 0x00007f0a37aac40d <+301>: mov    0x10(%rsi),%rdi

   0x00007f0a37aac411 <+305>: cmpb   $0x0,0x1c(%rdi)

   0x00007f0a37aac415 <+309>: jne    0x7f0a37aac6f7 <doupdate+1047>

   0x00007f0a37aac41b <+315>: movswl 0x4(%rcx),%ecx

   0x00007f0a37aac41f <+319>: movswl 0x74(%rdx),%edi

   0x00007f0a37aac423 <+323>: mov    %rax,0x40(%rsp)

[...]

部分输出。(我也可以只输入 “disas” 它会默认反汇编 doupdate )

“=>” 指向段错误地址,此处是一条 mov 指令 mov 0x10(%rsi),%rdi:从%rsi中指向内存地址的值加偏移量 0x10 处取值,送到 %rdi 寄存器中。接下来我会检查寄存器的状态。

6. 查看寄存器

使用 i r(info registers 的简写)打印寄存器值:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

(gdb) i r

rax            0x0  0

rbx            0x1993060    26816608

rcx            0x19902a0    26804896

rdx            0x19ce7d0    27060176

rsi            0x0  0

rdi            0x19ce7d0    27060176

rbp            0x7f0a3848eb10   0x7f0a3848eb10 <SP>

rsp            0x7ffd33d93c00   0x7ffd33d93c00

r8             0x7f0a37cb93e0   139681862489056

r9             0x0  0

r10            0x8  8

r11            0x202    514

r12            0x0  0

r13            0x0  0

r14            0x7f0a3848eb10   139681870703376

r15            0x19ce7d0    27060176

rip            0x7f0a37aac40d   0x7f0a37aac40d <doupdate+301>

eflags         0x10246  [ PF ZF IF RF ]

cs             0x33 51

ss             0x2b 43

ds             0x0  0

es             0x0  0

fs             0x0  0

gs             0x0  0

哦,%rsi是零,这就是我们的问题所在!零不太可能是有效地址,并且解引用一个未初始化的指针或空指针引起的段错误是常见的软件 bug。

7. 内存映射

你可以使用 i proc m (info proc mappings 的简写)核查零是不是有效地址:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

(gdb) i proc m

Mapped address spaces:

 

      Start Addr           End Addr       Size     Offset objfile

        0x400000           0x6e7000   0x2e7000        0x0 /usr/bin/python2.7

        0x8e6000           0x8e8000     0x2000   0x2e6000 /usr/bin/python2.7

        0x8e8000           0x95f000    0x77000   0x2e8000 /usr/bin/python2.7

  0x7f0a37a8b000     0x7f0a37ab8000    0x2d000        0x0 /lib/x86_64-linux-gnu/libncursesw.so.5.9

  0x7f0a37ab8000     0x7f0a37cb8000   0x200000    0x2d000 /lib/x86_64-linux-gnu/libncursesw.so.5.9

  0x7f0a37cb8000     0x7f0a37cb9000     0x1000    0x2d000 /lib/x86_64-linux-gnu/libncursesw.so.5.9

  0x7f0a37cb9000     0x7f0a37cba000     0x1000    0x2e000 /lib/x86_64-linux-gnu/libncursesw.so.5.9

  0x7f0a37cba000     0x7f0a37ccd000    0x13000        0x0 /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

  0x7f0a37ccd000     0x7f0a37ecc000   0x1ff000    0x13000 /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

  0x7f0a37ecc000     0x7f0a37ecd000     0x1000    0x12000 /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

  0x7f0a37ecd000     0x7f0a37ecf000     0x2000    0x13000 /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

  0x7f0a38050000     0x7f0a38066000    0x16000        0x0 /lib/x86_64-linux-gnu/libgcc_s.so.1

  0x7f0a38066000     0x7f0a38265000   0x1ff000    0x16000 /lib/x86_64-linux-gnu/libgcc_s.so.1

  0x7f0a38265000     0x7f0a38266000     0x1000    0x15000 /lib/x86_64-linux-gnu/libgcc_s.so.1

  0x7f0a38266000     0x7f0a3828b000    0x25000        0x0 /lib/x86_64-linux-gnu/libtinfo.so.5.9

  0x7f0a3828b000     0x7f0a3848a000   0x1ff000    0x25000 /lib/x86_64-linux-gnu/libtinfo.so.5.9

[...]

第一个有效的虚拟地址是 0x400000。任何小于它的地址都是非法的,这些地址如果被引用,就会引起段错误。

目前有几种不同的方式可做进一步分析。我先一步一步的看指令。

8. 断点

先回到反汇编:

 

1

2

3

4

   0x00007f0a37aac401 <+289>:   mov    0x20cb68(%rip),%rax        # 0x7f0a37cb8f70

   0x00007f0a37aac408 <+296>:   mov    (%rax),%rsi

   0x00007f0a37aac40b <+299>:   xor    %eax,%eax

=> 0x00007f0a37aac40d <+301>:   mov    0x10(%rsi),%rdi

看这四条指令:好像是从栈中取东西放到 %rax,然后解引用 %rax 到 %rsi,再将 %eax 置零( xor 是一个优化,替换掉移动 0 的动作),最后将 %rsi 解引用再加一个偏移,不过我们知道 %rsi 是零。这几条指令用来访问数据结构。 可能 %rax 会很有趣,但是它已经被前面的指令置零,所以我们在核心转储文件的寄存器里看不到它的值。

我可以在 doupdate+289 下个断点,然后逐条指令查看寄存器的值如何变化。首先,我需要启动 gdb 把程序跑起来:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

# gdb `which python`

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

Copyright (C) 2016 Free Software Foundation, Inc.

License GPLv3+: GNU GPL version 3 or later

This is free software: you are free to change and redistribute it.

There is NO WARRANTY, to the extent permitted by law.  Type "show copying"

and "show warranty" for details.

This GDB was configured as "x86_64-linux-gnu".

Type "show configuration" for configuration details.

For bug reporting instructions, please see:

.

Find the GDB manual and other documentation resources online at:

.

For help, type "help".

Type "apropos word" to search for commands related to "word"...

Reading symbols from /usr/bin/python...(no debugging symbols found)...done.

现在用 b (break 的简写)来下断点:

 

1

2

(gdb) b *doupdate + 289

No symbol table is loaded.  Use the "file" command.

哦。我想演示这个错误来解释为什么我们经常以在主函数设置断点作为开始,因为这时候符号可能被加载,可以设置感兴趣的断点。我直接在 doupdate 函数设断点,避开这个问题,一旦断点被触发就设置加了偏移的断点。

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

(gdb) b doupdate

Function "doupdate" not defined.

Make breakpoint pending on future shared library load? (y or [n]) y

Breakpoint 1 (doupdate) pending.

(gdb) r cachetop.py

Starting program: /usr/bin/python cachetop.py

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

warning: JITed object file architecture unknown is not compatible with target architecture i386:x86-64.

 

Breakpoint 1, 0x00007ffff34ad2e0 in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

(gdb) b *doupdate + 289

Breakpoint 2 at 0x7ffff34ad401

(gdb) c

Continuing.

 

 

Breakpoint 2, 0x00007ffff34ad401 in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

我们到了断点处。

如果你之前没有做这些,r (run) 命令会把参数传给我们早先在命令行指定的 gdb 目标(python)。这样的话程序会以执行 “python cachetop.py” 结束。

9. 单步调试

我跳到下一条指令(si,stepi的简写),然后检查寄存器:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

(gdb) si

0x00007ffff34ad408 in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

(gdb) i r

rax            0x7ffff3e8f948   140737285519688

rbx            0xaea060 11444320

rcx            0xae72a0 11432608

rdx            0xa403d0 10748880

rsi            0x7ffff7ea8e10   140737352732176

rdi            0xa403d0 10748880

rbp            0x7ffff3e8fb10   0x7ffff3e8fb10 <SP>

rsp            0x7fffffffd390   0x7fffffffd390

r8             0x7ffff36ba3e0   140737277305824

r9             0x0  0

r10            0x8  8

r11            0x202    514

r12            0x0  0

r13            0x0  0

r14            0x7ffff3e8fb10   140737285520144

r15            0xa403d0 10748880

rip            0x7ffff34ad408   0x7ffff34ad408 <doupdate+296>

eflags         0x202    [ IF ]

cs             0x33 51

ss             0x2b 43

ds             0x0  0

es             0x0  0

fs             0x0  0

gs             0x0  0

(gdb) p/a 0x7ffff3e8f948

$1 = 0x7ffff3e8f948 <cur_term>

又一条线索。所以我们解引用的空指针好像是一个叫 “cur_term” 的符号(p/a 是 print/a 的简写,这里 “/a” 指以地址的形式)。考虑到这是 ncurses, 是我们的环境变量 TERM 设置有问题吗?

 

1

2

# echo $TERM

xterm-256color

我试过将其设置为 vt100 并运行程序,还是遇到了同样的段错误。

注意我只是在 doupdate() 第一次被调用的时候查看了寄存器,但是它可以被多次调用,所以问题可能出在后边的调用中。我可以通过执行 c( continue 的简写)一步步到达出问题的地方。如果它被调用几次的话这样做是可行的,如果它被调用几千次的话我得用别的办法。(我会在 15 节的里介绍。)

10. 回退

gdb 有一个超棒的功能叫回退,Greg Law 在他的演讲中提到过。这里有一个例子。

我再启动一个 python 会话,从头演示:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

# gdb `which python`

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

Copyright (C) 2016 Free Software Foundation, Inc.

License GPLv3+: GNU GPL version 3 or later

This is free software: you are free to change and redistribute it.

There is NO WARRANTY, to the extent permitted by law.  Type "show copying"

and "show warranty" for details.

This GDB was configured as "x86_64-linux-gnu".

Type "show configuration" for configuration details.

For bug reporting instructions, please see:

<http://www.gnu.org/software/gdb/bugs/>.

Find the GDB manual and other documentation resources online at:

<http://www.gnu.org/software/gdb/documentation/>.

For help, type "help".

Type "apropos word" to search for commands related to "word"...

Reading symbols from /usr/bin/python...(no debugging symbols found)...done.

和之前一样我在 doupdate 下断点,一旦触发,我就启动 recording,然后继续运行程序直到崩溃。Recording 会增加相当大的开销,所以我不想在主函数里就将它打开。

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

(gdb) b doupdate

Function "doupdate" not defined.

Make breakpoint pending on future shared library load? (y or [n]) y

Breakpoint 1 (doupdate) pending.

(gdb) r cachetop.py

Starting program: /usr/bin/python cachetop.py

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

warning: JITed object file architecture unknown is not compatible with target architecture i386:x86-64.

 

Breakpoint 1, 0x00007ffff34ad2e0 in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

(gdb) record

(gdb) c

Continuing.

 

Program received signal SIGSEGV, Segmentation fault.

0x00007ffff34ad40d in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

这里我可以逐行或逐条指令的回退。它通过播放我们记录的寄存器状态来工作。我回退两条指令,然后打印寄存器值:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

(gdb) reverse-stepi

0x00007ffff34ad40d in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

(gdb) reverse-stepi

0x00007ffff34ad40b in doupdate () from /lib/x86_64-linux-gnu/libncursesw.so.5

(gdb) i r

rax            0x7ffff3e8f948   140737285519688

rbx            0xaea060 11444320

rcx            0xae72a0 11432608

rdx            0xa403d0 10748880

rsi            0x0  0

rdi            0xa403d0 10748880

rbp            0x7ffff3e8fb10   0x7ffff3e8fb10 <SP>

rsp            0x7fffffffd390   0x7fffffffd390

r8             0x7ffff36ba3e0   140737277305824

r9             0x0  0

r10            0x8  8

r11            0x302    770

r12            0x0  0

r13            0x0  0

r14            0x7ffff3e8fb10   140737285520144

r15            0xa403d0 10748880

rip            0x7ffff34ad40b   0x7ffff34ad40b <doupdate+299>

eflags         0x202    [ IF ]

cs             0x33 51

ss             0x2b 43

ds             0x0  0

es             0x0  0

fs             0x0  0

gs             0x0  0

(gdb) p/a 0x7ffff3e8f948

$1 = 0x7ffff3e8f948 <cur_term>

所以,又找到了 “cur_term” 的线索。我很想看这里的源代码,但我将从调试信息入手。

11. 调试信息

这是 libncursesw,我没有安装调试信息(Ubuntu):

 

1

2

3

4

5

6

7

8

# apt-cache search libncursesw

libncursesw5 - shared libraries for terminal handling (wide character support)

libncursesw5-dbg - debugging/profiling libraries for ncursesw

libncursesw5-dev - developer's libraries for ncursesw

# dpkg -l | grep libncursesw

 

 

ii  libncursesw5:amd64                  6.0+20160213-1ubuntu1                    amd64        shared libraries for terminal handling (wide character support)

我把它装上:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

# apt-get install -y libncursesw5-dbg

Reading package lists... Done

Building dependency tree      

Reading state information... Done

[...]

After this operation, 2,488 kB of additional disk space will be used.

Get:1 http://us-west-1.ec2.archive.ubuntu.com/ubuntu xenial/main amd64 libncursesw5-dbg amd64 6.0+20160213-1ubuntu1 [729 kB]

Fetched 729 kB in 0s (865 kB/s)          

Selecting previously unselected package libncursesw5-dbg.

(Reading database ... 200094 files and directories currently installed.)

Preparing to unpack .../libncursesw5-dbg_6.0+20160213-1ubuntu1_amd64.deb ...

Unpacking libncursesw5-dbg (6.0+20160213-1ubuntu1) ...

Setting up libncursesw5-dbg (6.0+20160213-1ubuntu1) ...

# dpkg -l | grep libncursesw

ii  libncursesw5:amd64                  6.0+20160213-1ubuntu1                    amd64        shared libraries for terminal handling (wide character support)

ii  libncursesw5-dbg                    6.0+20160213-1ubuntu1                    amd64        debugging/profiling libraries for ncursesw

太好了,版本匹配。那么现在我们的段错误是什么样子呢?

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

# gdb `which python` /var/cores/core.python.30520

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

[...]

warning: JITed object file architecture unknown is not compatible with target architecture i386:x86-64.

Core was generated by `python ./cachetop.py'.

Program terminated with signal SIGSEGV, Segmentation fault.

#0  ClrBlank (win=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

1129        if (back_color_erase)

(gdb) bt

#0  ClrBlank (win=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

#1  ClrUpdate () at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1147

#2  doupdate () at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1010

#3  0x00007f0a37aa07e6 in wrefresh (win=win@entry=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_refresh.c:65

#4  0x00007f0a37a99499 in recur_wrefresh (win=win@entry=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:384

#5  0x00007f0a37a99616 in _nc_wgetch (win=win@entry=0x1993060, result=result@entry=0x7ffd33d93e24, use_meta=1)

    at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:491

#6  0x00007f0a37a9a325 in wgetch (win=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:672

#7  0x00007f0a37cc6ec3 in ?? () from /usr/lib/python2.7/lib-dynload/_curses.x86_64-linux-gnu.so

#8  0x00000000004c4d5a in PyEval_EvalFrameEx ()

#9  0x00000000004c2e05 in PyEval_EvalCodeEx ()

#10 0x00000000004def08 in ?? ()

#11 0x00000000004b1153 in PyObject_Call ()

#12 0x00000000004c73ec in PyEval_EvalFrameEx ()

#13 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#14 0x00000000004caf42 in PyEval_EvalFrameEx ()

#15 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#16 0x00000000004c2ba9 in PyEval_EvalCode ()

#17 0x00000000004f20ef in ?? ()

#18 0x00000000004eca72 in PyRun_FileExFlags ()

#19 0x00000000004eb1f1 in PyRun_SimpleFileExFlags ()

#20 0x000000000049e18a in Py_Main ()

#21 0x00007f0a3be10830 in __libc_start_main (main=0x49daf0 <main>, argc=2, argv=0x7ffd33d94838, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>,

    stack_end=0x7ffd33d94828) at ../csu/libc-start.c:291

#22 0x000000000049da19 in _start ()

栈回溯看起来不太一样:我们确实不在 doupdate() 里边,而是在 ClrBlank() 中,它内联在 ClrUpdate() 里,ClrUpdate() 又内联在 doupdate() 中。

现在我真的要看源码了。

12. 源代码

安装了调试信息之后,gdb 可以同时列出源码和汇编:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

(gdb) disas/s

Dump of assembler code for function doupdate:

/build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:

759 {

   0x00007f0a37aac2e0 <+0>:   push   %r15

   0x00007f0a37aac2e2 <+2>:   push   %r14

   0x00007f0a37aac2e4 <+4>:   push   %r13

   0x00007f0a37aac2e6 <+6>:   push   %r12

[...]

   0x00007f0a37aac3dd <+253>: jne    0x7f0a37aac6ca <doupdate+1002>

 

1009        if (CurScreen(SP_PARM)->_clear || NewScreen(SP_PARM)->_clear) {   /* force refresh ? */

   0x00007f0a37aac3e3 <+259>: mov    0x80(%rdx),%rax

   0x00007f0a37aac3ea <+266>: mov    0x88(%rdx),%rcx

   0x00007f0a37aac3f1 <+273>: cmpb   $0x0,0x21(%rax)

   0x00007f0a37aac3f5 <+277>: jne    0x7f0a37aac401 <doupdate+289>

   0x00007f0a37aac3f7 <+279>: cmpb   $0x0,0x21(%rcx)

   0x00007f0a37aac3fb <+283>: je     0x7f0a37aacc3b <doupdate+2395>

 

1129        if (back_color_erase)

   0x00007f0a37aac401 <+289>: mov    0x20cb68(%rip),%rax        # 0x7f0a37cb8f70

   0x00007f0a37aac408 <+296>: mov    (%rax),%rsi

 

1128        NCURSES_CH_T blank = blankchar;

   0x00007f0a37aac40b <+299>: xor    %eax,%eax

 

1129        if (back_color_erase)

=> 0x00007f0a37aac40d <+301>: mov    0x10(%rsi),%rdi

   0x00007f0a37aac411 <+305>: cmpb   $0x0,0x1c(%rdi)

   0x00007f0a37aac415 <+309>: jne    0x7f0a37aac6f7 <doupdate+1047>

好极了!看 “=>” 和它上边的代码。所以我们的段错误发生在 “if (back_color_erase)” ?看起来不可能。

这里我检查了一下,我的调试信息版本是对的,重新在 gdb 里边运行程序直到发生段错误。错误相同。

back_color_erase 有什么特殊吗?我们现在在 ClrBlank() 中,我先列出源码:

 

1

2

3

4

5

6

7

8

9

10

11

(gdb) list ClrBlank

1124    

1125    static NCURSES_INLINE NCURSES_CH_T

1126    ClrBlank(NCURSES_SP_DCLx WINDOW *win)

1127    {

1128        NCURSES_CH_T blank = blankchar;

1129        if (back_color_erase)

1130        AddAttr(blank, (AttrOf(BCE_BKGD(SP_PARM, win)) & BCE_ATTRS));

1131        return blank;

1132    }

1133

啊,在这个函数里边没定义,难道是全局变量?

13. TUI

有必要看看这些代码在 gdb 的文本用户界面(TUI)里是什么样的,我用的不多,是看了 Greg 的演讲之后受到的启发。

你可以用 –tui 来启动:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

# gdb --tui `which python` /var/cores/core.python.30520

   ┌───────────────────────────────────────────────────────────────────────────┐

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │             [ No Source Available ]                                       │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   │                                                                           │

   └───────────────────────────────────────────────────────────────────────────┘

None No process In:                                                L??   PC: ??

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

Copyright (C) 2016 Free Software Foundation, Inc.

License GPLv3+: GNU GPL version 3 or later

This is free software: you are free to change and redistribute it.

There is NO WARRANTY, to the extent permitted by law.  Type "show copying"

and "show warranty" for details.

This GDB was configured as "x86_64-linux-gnu".

---Type  to continue, or q  to quit---

它在抱怨没有 Python 源码。我可以搞定,但是我们是在 libncursesw 里边崩溃的。所以不管它敲回车让它完成加载,在发生错误的地方加载了 libncursesw 调试信息里的源码:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

   ┌──/build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c──────┐

   │1124                                                                       │

   │1125    static NCURSES_INLINE NCURSES_CH_T                                 │

   │1126    ClrBlank(NCURSES_SP_DCLx WINDOW *win)                              │

   │1127    {                                                                  │

   │1128        NCURSES_CH_T blank = blankchar;                                │

  >│1129        if (back_color_erase)                                          │

   │1130            AddAttr(blank, (AttrOf(BCE_BKGD(SP_PARM, win)) & BCE_ATTRS)│

   │1131        return blank;                                                  │

   │1132    }                                                                  │

   │1133                                                                       │

   │1134    /*                                                                 │

   │1135    **      ClrUpdate()                                                │

   │1136    **                                                                 │

   └───────────────────────────────────────────────────────────────────────────┘

multi-thre Thread 0x7f0a3c5e87 In: doupdate            L1129 PC: 0x7f0a37aac40d

warning: JITed object file architecture unknown is not compatible with target ar

chitecture i386:x86-64.

---Type <return> to continue, or q <return> to quit---

Core was generated by `python ./cachetop.py'.

Program terminated with signal SIGSEGV, Segmentation fault.

#0  ClrBlank (win=0x1993060)

    at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

(gdb)

棒极了!

“>” 指向发生崩溃的那行代码。更棒的是:用 layout split 命令,我们可以在不同的窗口查看源代码和汇编代码。

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

   ┌──/build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c──────┐

  >│1129        if (back_color_erase)                                          │

   │1130            AddAttr(blank, (AttrOf(BCE_BKGD(SP_PARM, win)) & BCE_ATTRS)│

   │1131        return blank;                                                  │

   │1132    }                                                                  │

   │1133                                                                       │

   │1134    /*                                                                 │

   │1135    **      ClrUpdate()                                                │

   └───────────────────────────────────────────────────────────────────────────┘

  >│0x7f0a37aac40d <doupdate+301>   mov    0x10(%rsi),%rdi                     │

   │0x7f0a37aac411 <doupdate+305>   cmpb   $0x0,0x1c(%rdi)                     │

   │0x7f0a37aac415 <doupdate+309>   jne    0x7f0a37aac6f7 <doupdate+1047>      │

   │0x7f0a37aac41b <doupdate+315>   movswl 0x4(%rcx),%ecx                      │

   │0x7f0a37aac41f <doupdate+319>   movswl 0x74(%rdx),%edi                     │

   │0x7f0a37aac423 <doupdate+323>   mov    %rax,0x40(%rsp)                     │

   │0x7f0a37aac428 <doupdate+328>   movl   $0x20,0x48(%rsp)                    │

   │0x7f0a37aac430 <doupdate+336>   movl   $0x0,0x4c(%rsp)                     │

   └───────────────────────────────────────────────────────────────────────────┘

multi-thre Thread 0x7f0a3c5e87 In: doupdate            L1129 PC: 0x7f0a37aac40d

chitecture i386:x86-64.

Core was generated by `python ./cachetop.py'.

Program terminated with signal SIGSEGV, Segmentation fault.

---Type <return> to continue, or q <return> to quit---

#0  ClrBlank (win=0x1993060)

    at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

(gdb) layout split

Greg 演示这个的时候,和这里的顺序相反,因此你可想像同时查看源代码和汇编的情景(这里我需要一个视频来演示)。

14. 外部工具:cscope

我需要对 back_color_erase 有更多了解,我可以试试 gdb 的 搜索命令,但是我发现用一个外部工具:cscope 更快。 cscope 是一个基于文本的代码浏览器 ,诞生于80年代的贝尔实验室。如果你有喜欢的现代 IDE,可以不用它。

安装 cscope:

 

1

2

3

4

5

6

# apt-get install -y cscope

# wget http://archive.ubuntu.com/ubuntu/pool/main/n/ncurses/ncurses_6.0+20160213.orig.tar.gz

# tar xvf ncurses_6.0+20160213.orig.tar.gz

# cd ncurses-6.0-20160213

# cscope -bqR

# cscope -dq

cscope -bqR 用来建立查找数据库。cscope -dq 用来启动 cscope。

查找 back_color_erase 的定义:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

Cscope version 15.8b                                   Press the ? key for help

 

 

 

 

 

 

 

 

 

 

 

 

Find this C symbol:

Find this global definition: back_color_erase

Find functions called by this function:

Find functions calling this function:

Find this text string:

Change this text string:

Find this egrep pattern:

Find this file:

Find files #including this file:

Find assignments to this symbol:

敲回车:

 

1

2

3

4

5

6

7

8

9

10

[...]

#define non_dest_scroll_region         CUR Booleans[26]

#define can_change                     CUR Booleans[27]

#define back_color_erase               CUR Booleans[28]

#define hue_lightness_saturation       CUR Booleans[29]

#define col_addr_glitch                CUR Booleans[30]

#define cr_cancels_micro_mode          CUR Booleans[31]

 

 

[...]

哦,一个宏定义。(作为宏定义的常见的形式,它们至少应该大写)

好吧,那么 CUR 是什么呢? 用 cscope 查找定义易如反掌。

 

1

#define CUR cur_term->type.

起码这个宏定义是大写的!

我们通过逐条查看指令和寄存器找更早定义的 cur_term 。它是什么呢?

 

1

2

3

4

5

6

7

8

#if 0 && !0

extern NCURSES_EXPORT_VAR(TERMINAL *) cur_term;

#elif 0

NCURSES_WRAPPED_VAR(TERMINAL *, cur_term);

#define cur_term   NCURSES_PUBLIC_VAR(cur_term())

#else

extern NCURSES_EXPORT_VAR(TERMINAL *) cur_term;

#endif

cscope 读取了 /usr/include/term.h 。好吧,更多的宏。我用加粗来突出这行代码, 我认为它产生了影响。为什么这里会有 “if 0 && !0 … elif 0” ?我不清楚(需要再读些代码)。有时程序员会在他们想要在产品中失效的调试代码附近使用 “#if 0”,可是,这个好像是自动生成的。

查找 NCURSES_EXPORT_VAR 发现:

 

1

#  define NCURSES_EXPORT_VAR(type) NCURSES_IMPEXP type

… 和 NCURSES_IMPEXP:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

/* Take care of non-cygwin platforms */

#if !defined(NCURSES_IMPEXP)          

#  define NCURSES_IMPEXP /* nothing */

#endif                                

#if !defined(NCURSES_API)            

#  define NCURSES_API /* nothing */  

#endif                                

#if !defined(NCURSES_EXPORT)          

#  define NCURSES_EXPORT(type) NCURSES_IMPEXP type NCURSES_API

#endif                                

#if !defined(NCURSES_EXPORT_VAR)      

#  define NCURSES_EXPORT_VAR(type) NCURSES_IMPEXP type

#endif

… 还有 TERMINAL:

 

1

2

3

4

5

6

7

8

9

10

typedef struct term {       /* describe an actual terminal */

    TERMTYPE    type;       /* terminal type description */

    short   Filedes;    /* file description being written to */

    TTY     Ottyb,      /* original state of the terminal */

        Nttyb;      /* current state of the terminal */

    int     _baudrate;  /* used to compute padding */

    char *      _termname;      /* used for termname() */

 

 

} TERMINAL;

嗨!TERMINAL 是大写的。和宏混在一起,这个代码不太好跟踪 …

好吧,到底是谁给 cur_term 赋的值呢?记住我们的问题是它被赋值为零,也许因为它未被初始化或显式赋值。浏览给它赋值的代码路径可能会找到更多的线索,来回答为什么没被初始化,或为什么被赋值为零。使用 cscope 的第一个选项:

 

1

2

3

4

5

Find this C symbol: cur_term

Find this global definition:

Find functions called by this function:

Find functions calling this function:

[...]

快速浏览项发现:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

NCURSES_EXPORT(TERMINAL *)

NCURSES_SP_NAME(set_curterm) (NCURSES_SP_DCLx TERMINAL * termp)

{

    TERMINAL *oldterm;

 

    T((T_CALLED("set_curterm(%p)"), (void *) termp));

 

    _nc_lock_global(curses);

    oldterm = cur_term;

    if (SP_PARM)

    SP_PARM->_term = termp;

#if USE_REENTRANT

    CurTerm = termp;

#else

    cur_term = termp;

#endif

我加了高亮。甚至函数名称都被封装在宏里。但至少我们发现了 cur_term 如何被赋值的:通过 set_curterm()。也许它没被调用?

15. 外部工具:perf-tools/ftrace/uprobes

我稍后将介绍如何用 gdb 解决这个问题,可是我忍不住尝试我 perf-tools 工具箱里的 uprobe 工具,它使用 Linux 下的 ftrace 和 uprobes。用 tracers 的一个好处是它不会终止目标进程,像 gdb 一样(尽管对于这里的 cachetop.py 没什么用)。另一个好处是追踪几个和几千个进程一样容易。

我应该能追踪 libncursesw 对 set_curterm() 的调用,甚至打印出它的第一个参数:

 

1

2

# /apps/perf-tools/bin/uprobe 'p:/lib/x86_64-linux-gnu/libncursesw.so.5:set_curterm %di'

ERROR: missing symbol "set_curterm" in /lib/x86_64-linux-gnu/libncursesw.so.5

咦,没起作用。set_curterm() 在哪?有很多方法可以找到它,比如 gdb 或 objdump:

 

1

2

3

4

5

6

(gdb) info symbol set_curterm

set_curterm in section .text of /lib/x86_64-linux-gnu/libtinfo.so.5

# objdump -tT /lib/x86_64-linux-gnu/libncursesw.so.5 | grep cur_term

0000000000000000      DO *UND*  0000000000000000  NCURSES_TINFO_5.0.19991023 cur_term

# objdump -tT /lib/x86_64-linux-gnu/libtinfo.so.5 | grep cur_term

0000000000228948 g    DO .bss   0000000000000008  NCURSES_TINFO_5.0.19991023 cur_term

gdb 表现的好些。此外如果仔细看源代码,我注意到它是为 libtinfo 构建的。

试着在 libtinfo 里边查找 set_curterm() :

 

1

2

3

4

5

6

7

# /apps/perf-tools/bin/uprobe 'p:/lib/x86_64-linux-gnu/libtinfo.so.5:set_curterm %di'

Tracing uprobe set_curterm (p:set_curterm /lib/x86_64-linux-gnu/libtinfo.so.5:0xfa80 %di). Ctrl-C to end.

          python-31617 [007] d... 24236402.719959: set_curterm: (0x7f116fcc2a80) arg1=0x1345d70

          python-31617 [007] d... 24236402.720033: set_curterm: (0x7f116fcc2a80) arg1=0x13a22e0

          python-31617 [007] d... 24236402.723804: set_curterm: (0x7f116fcc2a80) arg1=0x14cdfa0

          python-31617 [007] d... 24236402.723838: set_curterm: (0x7f116fcc2a80) arg1=0x0

^C

找到了。所以 set_curterm() 被调用了,并且被调用了四次。最后一次被传了一个零,看起来这就是问题所在。

如果你觉得疑惑,我怎么就知道 %di 寄存器就是第一个参数呢,因为 AMD64/x86_64 ABI 写着呢(假设这个库和 ABI 兼容)。这里有提示:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

# man syscall

[...]

       arch/ABI      arg1  arg2  arg3  arg4  arg5  arg6  arg7  Notes

       ──────────────────────────────────────────────────────────────────

       arm/OABI      a1    a2    a3    a4    v1    v2    v3

       arm/EABI      r0    r1    r2    r3    r4    r5    r6

       arm64         x0    x1    x2    x3    x4    x5    -

       blackfin      R0    R1    R2    R3    R4    R5    -

       i386          ebx   ecx   edx   esi   edi   ebp   -

       ia64          out0  out1  out2  out3  out4  out5  -

       mips/o32      a0    a1    a2    a3    -     -     -     See below

       mips/n32,64   a0    a1    a2    a3    a4    a5    -

       parisc        r26   r25   r24   r23   r22   r21   -

       s390          r2    r3    r4    r5    r6    r7    -

       s390x         r2    r3    r4    r5    r6    r7    -

       sparc/32      o0    o1    o2    o3    o4    o5    -

       sparc/64      o0    o1    o2    o3    o4    o5    -

       x86_64        rdi   rsi   rdx   r10   r8    r9    -

[...]

我还想知道调用 arg1=0x0 的堆栈信息,但是 ftrace 还不支持栈追踪。

16. 外部工具:bcc/BPF

由于我们在调试 bcc 工具 cachetop.py,值得注意的是 bcc 里的 trace.py 有和我的老工具 uprobe 类似的功能:

 

1

2

3

4

5

6

# ./trace.py 'p:tinfo:set_curterm "%d", arg1'

TIME     PID    COMM         FUNC             -

01:00:20 31698  python       set_curterm      38018416

01:00:20 31698  python       set_curterm      38396640

01:00:20 31698  python       set_curterm      39624608

01:00:20 31698  python       set_curterm      0

是的,我们在用 bcc 调试 bcc !

如果你对 bcc 不熟悉,它值得一看。它为 Linux4.x 系列里的 BPF 新特性提供了 Python 和 lua 接口 。总之,它能让很多以前不可能或昂贵以致无法运行的性能工具运行起来。我以前发过贴介绍如何在 Ubuntu Xenial 上运行它。

bcc 的 trace.py 工具应该有一个开关来决定是否打印用户堆栈,因为内核从 Linux4.6 开始具备 BPF 堆栈功能,不过到写这篇文章的时候我们还没有加上这个开关。

17. 更多的断点

我真的应该从在 set_curterm() 下了断点的 gdb 入手,可是我觉得我们走的弯路,使用ftrace和BPF的还是蛮有趣的。

回到实时运行模式:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

# gdb `which python`

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

[...]

(gdb) b set_curterm

Function "set_curterm" not defined.

Make breakpoint pending on future shared library load? (y or [n]) y

Breakpoint 1 (set_curterm) pending.

(gdb) r cachetop.py

Starting program: /usr/bin/python cachetop.py

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

Breakpoint 1, set_curterm (termp=termp@entry=0xa43150) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=termp@entry=0xab5870) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=termp@entry=0xbecb90) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

好的,在这个断点我们可以看到 set_curterm() 被调用了,被传了一个 termp = 0x0 的参数, 多亏了 debuginfo 提供的信息。如果没有 debuginfo ,我只能在每个断点处打印寄存器值。

我打印栈帧出来,这样我们可以看到是谁将 curterm 设为零的。

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

(gdb) bt

#0  set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

#1  0x00007ffff5a44e75 in llvm::sys::Process::FileDescriptorHasColors(int) () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#2  0x00007ffff45cabb8 in clang::driver::tools::Clang::ConstructJob(clang::driver::Compilation&, clang::driver::JobAction const&, clang::driver::InputInfo const&, llvm::SmallVector<clang::driver::InputInfo, 4u> const&, llvm::opt::ArgList const&, char const*) const () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#3  0x00007ffff456ffa5 in clang::driver::Driver::BuildJobsForAction(clang::driver::Compilation&, clang::driver::Action const*, clang::driver::ToolChain const*, char const*, bool, bool, char const*, clang::driver::InputInfo&) const () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#4  0x00007ffff4570501 in clang::driver::Driver::BuildJobs(clang::driver::Compilation&) const () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#5  0x00007ffff457224a in clang::driver::Driver::BuildCompilation(llvm::ArrayRef<char const*>) () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#6  0x00007ffff4396cda in ebpf::ClangLoader::parse(std::unique_ptr<llvm::Module, std::default_delete<llvm::Module> >*, std::unique_ptr<std::vector<ebpf::TableDesc, std::allocator<ebpf::TableDesc> >, std::default_delete<std::vector<ebpf::TableDesc, std::allocator<ebpf::TableDesc> > > >*, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, bool, char const**, int) () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#7  0x00007ffff4344314 in ebpf::BPFModule::load_cfile(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, bool, char const**, int) ()

   from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#8  0x00007ffff4349e5e in ebpf::BPFModule::load_string(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, char const**, int) ()

   from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#9  0x00007ffff43430c8 in bpf_module_create_c_from_string () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

#10 0x00007ffff690ae40 in ffi_call_unix64 () from /usr/lib/x86_64-linux-gnu/libffi.so.6

#11 0x00007ffff690a8ab in ffi_call () from /usr/lib/x86_64-linux-gnu/libffi.so.6

#12 0x00007ffff6b1a68c in _ctypes_callproc () from /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so

#13 0x00007ffff6b1ed82 in ?? () from /usr/lib/python2.7/lib-dynload/_ctypes.x86_64-linux-gnu.so

#14 0x00000000004b1153 in PyObject_Call ()

#15 0x00000000004ca5ca in PyEval_EvalFrameEx ()

#16 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#17 0x00000000004def08 in ?? ()

#18 0x00000000004b1153 in PyObject_Call ()

#19 0x00000000004f4c3e in ?? ()

#20 0x00000000004b1153 in PyObject_Call ()

#21 0x00000000004f49b7 in ?? ()

#22 0x00000000004b6e2c in ?? ()

#23 0x00000000004b1153 in PyObject_Call ()

#24 0x00000000004ca5ca in PyEval_EvalFrameEx ()

#25 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#26 0x00000000004def08 in ?? ()

#27 0x00000000004b1153 in PyObject_Call ()

#28 0x00000000004c73ec in PyEval_EvalFrameEx ()

#29 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#30 0x00000000004caf42 in PyEval_EvalFrameEx ()

#31 0x00000000004c2e05 in PyEval_EvalCodeEx ()

#32 0x00000000004c2ba9 in PyEval_EvalCode ()

#33 0x00000000004f20ef in ?? ()

#34 0x00000000004eca72 in PyRun_FileExFlags ()

#35 0x00000000004eb1f1 in PyRun_SimpleFileExFlags ()

#36 0x000000000049e18a in Py_Main ()

#37 0x00007ffff7811830 in __libc_start_main (main=0x49daf0 <main>, argc=2, argv=0x7fffffffdfb8, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>,

    stack_end=0x7fffffffdfa8) at ../csu/libc-start.c:291

#38 0x000000000049da19 in _start ()

好了,有了更多的线索…我认为。我们在 llvm::sys::Process::FileDescriptorHasColors()里边。llvm 编译器有问题?

18. 外部工具:cscope,再来一次

代码较多的时候使用 cscope 查看,这次是 llvm。FileDescriptorHasColors() 函数:

 

1

2

3

4

5

6

static bool terminalHasColors(int fd) {

[...]

  // Now extract the structure allocated by setupterm and free its memory

  // through a really silly dance.

  struct term *termp = set_curterm((struct term *)nullptr);

  (void)del_curterm(termp); // Drop any errors here.

这是较早版本中使用的代码:

 

1

2

3

4

5

6

7

8

9

10

11

static bool terminalHasColors() {

  if (const char *term = std::getenv("TERM")) {

    // Most modern terminals support ANSI escape sequences for colors.

    // We could check terminfo, or have a list of known terms that support

    // colors, but that would be overkill.

    // The user can always ask for no colors by setting TERM to dumb, or

    // using a commandline flag.

    return strcmp(term, "dumb") != 0;

  }

  return false;

}

用空指针调用 set_curterm() 变成了 “愚蠢的舞蹈” 。

19. 写内存

作为实验,我要修改程序内存来避免 set_curterm() 被置零,用来探索可能的解决方法。

运行 gdb ,在 set_curterm() 下断点,跑到零调用的地方:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

# gdb `which python`

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1                                  

[...]

(gdb) b set_curterm

Function "set_curterm" not defined.

Make breakpoint pending on future shared library load? (y or [n]) y

Breakpoint 1 (set_curterm) pending.

(gdb) r cachetop.py

Starting program: /usr/bin/python cachetop.py

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

 

Breakpoint 1, set_curterm (termp=termp@entry=0xa43150) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80      {

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=termp@entry=0xab5870) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80      {

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=termp@entry=0xbecb90) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80      {

(gdb) c

Continuing.                                                                    

 

Breakpoint 1, set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80      {

这里我用 set 命令来改写内存,把零换成在前面看到的 set_curterm() 参数 0xbecb90 ,希望它仍是合法的。

警告:写内存不安全!gdb 不会问你 “你确定?”。如果你写错了或者敲错了,会搞坏程序。最好的情况是你的程序立即奔溃,你意识到自己做错了。最糟的情况,程序使用坏的数据继续运行几年之后被发现是错的。

这里,我在不用于生产的实验室机器上做试验,所以我继续。
我以16进制(p/x)的形式打印 %rdi 的值,然后将其设为之前的地址,再打印一次,最后打印所有寄存器的值:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

(gdb) p/x $rdi

$1 = 0x0

(gdb) set $rdi=0xbecb90

(gdb) p/x $rdi

$2 = 0xbecb90

(gdb) i r

rax            0x100    256

rbx            0x1  1

rcx            0xe71    3697

rdx            0x0  0

rsi            0x7ffff5dd45d3   140737318307283

rdi            0xbecb90 12503952

rbp            0x100    0x100

rsp            0x7fffffffa5b8   0x7fffffffa5b8

r8             0xbf0050 12517456

r9             0x1999999999999999   1844674407370955161

r10            0xbf0040 12517440

r11            0x7ffff7bb4b78   140737349634936

r12            0xbecb70 12503920

r13            0xbeaea0 12496544

r14            0x7fffffffa9a0   140737488333216

r15            0x7fffffffa8a0   140737488332960

rip            0x7ffff3c76a80   0x7ffff3c76a80 <set_curterm>

eflags         0x246    [ PF ZF IF ]

cs             0x33 51

ss             0x2b 43

ds             0x0  0

es             0x0  0

fs             0x0  0

gs             0x0  0

(因为这里我已经安装了调试信息,因此不必使用寄存器,我可以设置传给 set_curterm() 的参数参数 “termp”,而不是 $rdi。)

现在 %rdi 被用到了,所以那些寄存器看起来还能继续用。

 

1

2

3

4

5

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=termp@entry=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

好的,在调用 set_curterm() 时程序没崩!但遇到另一个参数也是零的问题。我们故技重施:

 

1

2

3

4

5

6

7

8

(gdb) set $rdi=0xbecb90

(gdb) c

Continuing.

warning: JITed object file architecture unknown is not compatible with target architecture i386:x86-64.

 

Program received signal SIGSEGV, Segmentation fault.

0x00007ffff34ad411 in ClrBlank (win=0xaea060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

1129        if (back_color_erase)

啊。这就是我写内存的后果。所以这次试验以另一个段错误结束。

20. 条件断点

在前面一节,我用了 3 个 continues 到达断点的正确调用处。如果有几百次调用的话,就得用条件断点了。这里有个例子。

和之前一样我运行程序,在 set_curterm() 下断点:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

# gdb `which python`

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1                                  

[...]

(gdb) b set_curterm

Function "set_curterm" not defined.

Make breakpoint pending on future shared library load? (y or [n]) y

Breakpoint 1 (set_curterm) pending.

(gdb) r cachetop.py

Starting program: /usr/bin/python cachetop.py

[Thread debugging using libthread_db enabled]

Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

 

Breakpoint 1, set_curterm (termp=termp@entry=0xa43150) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

现在我要将 1 号断点变成条件断点,这样它只会在 %rdi 的值为零是被触发:

 

1

2

3

4

5

6

7

8

9

10

11

(gdb) cond 1 $rdi==0x0

(gdb) i b

Num     Type           Disp Enb Address            What

1       breakpoint     keep y   0x00007ffff3c76a80 in set_curterm at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

    stop only if $rdi==0x0

    breakpoint already hit 1 time

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

(gdb)

漂亮!cond 是 conditional 的简写。为什么当我第一次创建 “pending” 断点的时候没有立即运行它呢?因为我发现在 pending 断点上条件不管用,至少在这个版本的 gdb 上是这样。(要么是我哪里做错了。)我也用 i b (info breakpoints)列出了断点信息。

21. 返回命令

我曾经试过另一个改值的方法,但是这次我要改指令而不是数据。

警告:看前边的警告,这里也适用。

和之前一样我们来到 set_curterm 零断点处,然后敲入 ret (return 的简写),就会立即从此函数返回并且不执行这个函数。我想用不执行函数的方式让全局变量 curterm 不被置零。

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

[...]

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

 

(gdb) ret

Make set_curterm return now? (y or n) y

#0  0x00007ffff5a44e75 in llvm::sys::Process::FileDescriptorHasColors(int) () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

(gdb) c

Continuing.

 

Program received signal SIGSEGV, Segmentation fault.

                                                    _nc_free_termtype (ptr=ptr@entry=0x100) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/free_ttype.c:52

 

52      FreeIfNeeded(ptr->str_table);

又崩了。这是我搞砸的现场。

再试一次。在多看了一点代码之后,我想第二次尝试 ret,以防父函数被卷进来。再来一次,这只是一次非常规试验:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

[...]

(gdb) c

Continuing.

 

Breakpoint 1, set_curterm (termp=0x0) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tinfo/lib_cur_term.c:80

80  {

(gdb) ret

Make set_curterm return now? (y or n) y

#0  0x00007ffff5a44e75 in llvm::sys::Process::FileDescriptorHasColors(int) () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

(gdb) ret

Make selected stack frame return now? (y or n) y

#0  0x00007ffff45cabb8 in clang::driver::tools::Clang::ConstructJob(clang::driver::Compilation&, clang::driver::JobAction const&, clang::driver::InputInfo const&, llvm::SmallVector const&, llvm::opt::ArgList const&, char const*) const () from /usr/lib/x86_64-linux-gnu/libbcc.so.0

 

 

(gdb) c

屏幕清空暂停…然后刷新:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

07:44:22 Buffers MB: 61 / Cached MB: 1246

PID      UID      CMD              HITS     MISSES   DIRTIES  READ_HIT%  WRITE_HIT%

    2742 root     systemd-logind          3       66        2       1.4%      95.7%

   15836 root     kworker/u30:1           7        0        1      85.7%       0.0%

    2736 messageb dbus-daemon             8       66        2       8.1%      89.2%

       1 root     systemd                15        0        0     100.0%       0.0%

    2812 syslog   rs:main Q:Reg          16       66        8       9.8%      80.5%

     435 root     systemd-journal        32       66        8      24.5%      67.3%

    2740 root     accounts-daemon       113       66        2      62.0%      36.9%

   15847 root     bash                  160        0        1      99.4%       0.0%

   15864 root     lesspipe              306        0        2      99.3%       0.0%

   15854 root     bash                  309        0        2      99.4%       0.0%

   15856 root     bash                  309        0        2      99.4%       0.0%

   15866 root     bash                  309        0        2      99.4%       0.0%

   15867 root     bash                  309        0        2      99.4%       0.0%

   15860 root     bash                  313        0        2      99.4%       0.0%

   15868 root     bash                  341        0        2      99.4%       0.0%

   15858 root     uname                 452        0        2      99.6%       0.0%

   15858 root     bash                  453        0        2      99.6%       0.0%

   15866 root     dircolors             464        0        2      99.6%       0.0%

   15861 root     basename              465        0        2      99.6%       0.0%

   15864 root     dirname               468        0        2      99.6%       0.0%

   15856 root     ls                    476        0        2      99.6%       0.0%

[...]

哇!成功了!

22. 更好的方案

我已经把调试输出发布到 github,因为 BPF 首席工程师,Alexei Starovoitov 对 llvm 也很精通,问题的根源好像是 llvm 的一个 bug。当我在用写内存和返回命令瞎搞的时候,他建议我在 bcc 加上 llvm 选项 -fno-color-diagnostics,来避免这个问题。成功了!把它加到 bcc 里是一个解决办法。(我还是希望 llvm 的 bug 能被修复)

23. Python 环境

至此问题已经解决了,但是你可能会好奇想看修复好的堆栈回溯。

安装 python-dbg:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

# apt-get install -y python-dbg

Reading package lists... Done

[...]

The following additional packages will be installed:

  libpython-dbg libpython2.7-dbg python2.7-dbg

Suggested packages:

  python2.7-gdbm-dbg python2.7-tk-dbg python-gdbm-dbg python-tk-dbg

The following NEW packages will be installed:

  libpython-dbg libpython2.7-dbg python-dbg python2.7-dbg

0 upgraded, 4 newly installed, 0 to remove and 20 not upgraded.

Need to get 11.9 MB of archives.

After this operation, 36.4 MB of additional disk space will be used.

[...]

现在我回到 gdb 来看堆栈回溯:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

# gdb `which python` /var/cores/core.python.30520

GNU gdb (Ubuntu 7.11.1-0ubuntu1~16.04) 7.11.1

[...]

Reading symbols from /usr/bin/python...Reading symbols from /usr/lib/debug/.build-id/4e/a0539215b2a9e32602f81c90240874132c1a54.debug...done.

[...]

(gdb) bt

#0  ClrBlank (win=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1129

#1  ClrUpdate () at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1147

#2  doupdate () at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/tty/tty_update.c:1010

#3  0x00007f0a37aa07e6 in wrefresh (win=win@entry=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_refresh.c:65

#4  0x00007f0a37a99499 in recur_wrefresh (win=win@entry=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:384

#5  0x00007f0a37a99616 in _nc_wgetch (win=win@entry=0x1993060, result=result@entry=0x7ffd33d93e24, use_meta=1)

    at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:491

#6  0x00007f0a37a9a325 in wgetch (win=0x1993060) at /build/ncurses-pKZ1BN/ncurses-6.0+20160213/ncurses/base/lib_getch.c:672

#7  0x00007f0a37cc6ec3 in PyCursesWindow_GetCh.lto_priv.109 (self=0x7f0a3c57b198, args=()) at /build/python2.7-HpIZBG/python2.7-2.7.11/Modules/_cursesmodule.c:853

#8  0x00000000004c4d5a in call_function (oparg=<optimized out>, pp_stack=0x7ffd33d93f30) at ../Python/ceval.c:4350

#9  PyEval_EvalFrameEx () at ../Python/ceval.c:2987

#10 0x00000000004c2e05 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

#11 0x00000000004def08 in function_call.lto_priv () at ../Objects/funcobject.c:523

#12 0x00000000004b1153 in PyObject_Call () at ../Objects/abstract.c:2546

#13 0x00000000004c73ec in ext_do_call (nk=0, na=<optimized out>, flags=<optimized out>, pp_stack=0x7ffd33d941e8, func=<function at remote 0x7f0a37edcc80>)

    at ../Python/ceval.c:4662

#14 PyEval_EvalFrameEx () at ../Python/ceval.c:3026

#15 0x00000000004c2e05 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

#16 0x00000000004caf42 in fast_function (nk=0, na=<optimized out>, n=<optimized out>, pp_stack=0x7ffd33d943f0, func=<function at remote 0x7f0a38039140>)

    at ../Python/ceval.c:4445

#17 call_function (oparg=<optimized out>, pp_stack=0x7ffd33d943f0) at ../Python/ceval.c:4370

#18 PyEval_EvalFrameEx () at ../Python/ceval.c:2987

#19 0x00000000004c2e05 in PyEval_EvalCodeEx () at ../Python/ceval.c:3582

#20 0x00000000004c2ba9 in PyEval_EvalCode (co=<optimized out>, globals=<optimized out>, locals=<optimized out>) at ../Python/ceval.c:669

#21 0x00000000004f20ef in run_mod.lto_priv () at ../Python/pythonrun.c:1376

#22 0x00000000004eca72 in PyRun_FileExFlags () at ../Python/pythonrun.c:1362

#23 0x00000000004eb1f1 in PyRun_SimpleFileExFlags () at ../Python/pythonrun.c:948

#24 0x000000000049e18a in Py_Main () at ../Modules/main.c:640

#25 0x00007f0a3be10830 in __libc_start_main (main=0x49daf0 <main>, argc=2, argv=0x7ffd33d94838, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>,

    stack_end=0x7ffd33d94828) at ../csu/libc-start.c:291

#26 0x000000000049da19 in _start ()

没有 “??” 了,但也没什么大用。

python 调试包给 gdb 加入了别的功能。现在我们可以看 python 的回溯:

 

1

2

3

4

5

6

7

8

(gdb) py-bt

Traceback (most recent call first):

  File "./cachetop.py", line 188, in handle_loop

    s = stdscr.getch()

  File "/usr/lib/python2.7/curses/wrapper.py", line 43, in wrapper

    return func(stdscr, *args, **kwds)

  File "./cachetop.py", line 260, in

    curses.wrapper(handle_loop, args)

… 和Python 源码:

 

1

2

3

4

5

6

7

8

9

10

11

12

(gdb) py-list

183        b.attach_kprobe(event="mark_buffer_dirty", fn_name="do_count")

184    

185        exiting = 0

186    

187        while 1:

>188            s = stdscr.getch()

> 189            if s == ord('q'):

> 190                exiting = 1

> 191            elif s == ord('r'):

> 192                sort_reverse = not sort_reverse

> 193            elif s == ord('<'):

它识别出了我们之前执行的 python 代码中的段错误。真是太棒了!

原先堆栈回溯的问题是我们看到了 python 内部在执行方法,却看不到方法本身。如果你调试别的语言,要取决于它的编译选项和运行环境,还有怎么结束执行代码。如果你在网上搜索 “语言名” 和 “gdb” 你可能会找到像 Python 一样的 gdb 扩展。如果没有的话,坏消息是你需要自己写,好消息是这样做是可行的!当它们可以用 Python 来写的时候,请搜索 “adding new GDB commands in Python” 的资料。

24. 更多命令

看起来好像我写了一个 gdb 的全面介绍,但我真的没有:gdb 里还有很多命令我没提到。help 命令列出了主要部分:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

(gdb) help

List of classes of commands:

 

aliases -- Aliases of other commands

breakpoints -- Making program stop at certain points

data -- Examining data

files -- Specifying and examining files

internals -- Maintenance commands

obscure -- Obscure features

running -- Running the program

stack -- Examining the stack

status -- Status inquiries

support -- Support facilities

tracepoints -- Tracing of program execution without stopping the program

user-defined -- User-defined commands

 

Type "help" followed by a class name for a list of commands in that class.

Type "help all" for the list of all commands.

Type "help" followed by command name for full documentation.

Type "apropos word" to search for commands related to "word".

Command name abbreviations are allowed if unambiguous.

你可以对每一类命令执行 help。例如,这是 breakpoints 类的全部清单:

 

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

96

97

98

(gdb) help breakpoints

Making program stop at certain points.

 

List of commands:

 

awatch -- Set a watchpoint for an expression

break -- Set breakpoint at specified location

break-range -- Set a breakpoint for an address range

catch -- Set catchpoints to catch events

catch assert -- Catch failed Ada assertions

catch catch -- Catch an exception

catch exception -- Catch Ada exceptions

catch exec -- Catch calls to exec

catch fork -- Catch calls to fork

catch load -- Catch loads of shared libraries

catch rethrow -- Catch an exception

catch signal -- Catch signals by their names and/or numbers

catch syscall -- Catch system calls by their names and/or numbers

catch throw -- Catch an exception

catch unload -- Catch unloads of shared libraries

catch vfork -- Catch calls to vfork

clear -- Clear breakpoint at specified location

commands -- Set commands to be executed when a breakpoint is hit

condition -- Specify breakpoint number N to break only if COND is true

delete -- Delete some breakpoints or auto-display expressions

delete bookmark -- Delete a bookmark from the bookmark list

delete breakpoints -- Delete some breakpoints or auto-display expressions

delete checkpoint -- Delete a checkpoint (experimental)

delete display -- Cancel some expressions to be displayed when program stops

delete mem -- Delete memory region

delete tracepoints -- Delete specified tracepoints

delete tvariable -- Delete one or more trace state variables

disable -- Disable some breakpoints

disable breakpoints -- Disable some breakpoints

disable display -- Disable some expressions to be displayed when program stops

disable frame-filter -- GDB command to disable the specified frame-filter

disable mem -- Disable memory region

disable pretty-printer -- GDB command to disable the specified pretty-printer

disable probes -- Disable probes

disable tracepoints -- Disable specified tracepoints

disable type-printer -- GDB command to disable the specified type-printer

disable unwinder -- GDB command to disable the specified unwinder

disable xmethod -- GDB command to disable a specified (group of) xmethod(s)

dprintf -- Set a dynamic printf at specified location

enable -- Enable some breakpoints

enable breakpoints -- Enable some breakpoints

enable breakpoints count -- Enable breakpoints for COUNT hits

enable breakpoints delete -- Enable breakpoints and delete when hit

enable breakpoints once -- Enable breakpoints for one hit

enable count -- Enable breakpoints for COUNT hits

enable delete -- Enable breakpoints and delete when hit

enable display -- Enable some expressions to be displayed when program stops

enable frame-filter -- GDB command to disable the specified frame-filter

enable mem -- Enable memory region

enable once -- Enable breakpoints for one hit

enable pretty-printer -- GDB command to enable the specified pretty-printer

enable probes -- Enable probes

enable tracepoints -- Enable specified tracepoints

enable type-printer -- GDB command to enable the specified type printer

enable unwinder -- GDB command to enable unwinders

enable xmethod -- GDB command to enable a specified (group of) xmethod(s)

ftrace -- Set a fast tracepoint at specified location

hbreak -- Set a hardware assisted breakpoint

ignore -- Set ignore-count of breakpoint number N to COUNT

rbreak -- Set a breakpoint for all functions matching REGEXP

rwatch -- Set a read watchpoint for an expression

save -- Save breakpoint definitions as a script

save breakpoints -- Save current breakpoint definitions as a script

save gdb-index -- Save a gdb-index file

save tracepoints -- Save current tracepoint definitions as a script

skip -- Ignore a function while stepping

skip delete -- Delete skip entries

skip disable -- Disable skip entries

skip enable -- Enable skip entries

skip file -- Ignore a file while stepping

skip function -- Ignore a function while stepping

strace -- Set a static tracepoint at location or marker

tbreak -- Set a temporary breakpoint

tcatch -- Set temporary catchpoints to catch events

tcatch assert -- Catch failed Ada assertions

tcatch catch -- Catch an exception

tcatch exception -- Catch Ada exceptions

tcatch exec -- Catch calls to exec

tcatch fork -- Catch calls to fork

tcatch load -- Catch loads of shared libraries

tcatch rethrow -- Catch an exception

tcatch signal -- Catch signals by their names and/or numbers

tcatch syscall -- Catch system calls by their names and/or numbers

tcatch throw -- Catch an exception

tcatch unload -- Catch unloads of shared libraries

tcatch vfork -- Catch calls to vfork

thbreak -- Set a temporary hardware assisted breakpoint

trace -- Set a tracepoint at specified location

watch -- Set a watchpoint for an expression

 

Type "help" followed by command name for full documentation.

Type "apropos word" to search for commands related to "word".

Command name abbreviations are allowed if unambiguous.

这些帮助表明了 gdb 有很多功能,也说明了我在示例中用到的只是一小部分。

25. 结语

好吧,这个问题有点恶心:一个 LLVM bug 破坏了 ncurses 并引起了 Python 程序的段错误。但是我用来调试的命令和步骤很常见:看堆栈,检查寄存器,下断点,逐步排查,看源码。

当我第一次使用 gdb 的时候(多年前),我真的不喜欢它。觉得它不灵活而且功能有限。从那之后 gdb 进步了很多,我也掌握了 gdb 的技巧,我现在认为它是一个强大的现代调试器。不同的调试器特性可能不同,但是 gdb 可能是目前基于文本的最强大的调试器,lldb 正奋起直追。

我希望我分享的有完整输出的 gdb 示例和我提到的不同的警告,会对搜到它的人有帮助。有机会的话我会发布更多的 gdb 示例,特别是其他运行环境比如Java。

用 q 退出 gdb。

 

评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值