用GDB观察共享库函数的翻译过程
研究了一下共享库函数是怎样加载到当前进程中的.开始共享库函数地址放在GOT中,
第一次调用时,ld将其翻译成函数在程序空间的真实地址.用GDB跟踪了一下整个过程,
记录在下面.
PLT (Procedure Linkage Table) 和 GOT (Global Offset Table)背景,google.
-------------------------------------------------------------------------------
### 准备 ### ### 环境Ubuntu 11.04 amd64, ### 安装libc debug symbol.
sudo apt-get install libc-dbg
|
### 安装libc6 source,假设目录是~/codes/debsrc/eglibc-2.13
sudo apt-get install build-essential
sudo apt-get source libc6
|
### 使用实验源文件http://files.cnblogs.com/dyno/plt.zip,
mkdir whatever; cd whatever; make
main.c <---- 主程序
test.c <---- 共享库
test.h
Makefile
|
foo & foo2是两个共享库中的函数,
[dyno@ubuntu:plt]$ objdump --syms main.exe | grep -E "(foo|xyz)"
0000000000000000 F *UND* 0000000000000000 foo2 <---- 1
0000000000000000 F *UND* 0000000000000000 foo <---- 2
0000000000601028 g O .bss 0000000000000004 xyz
[dyno@ubuntu:plt]$ readelf --sections --wide main.exe | grep got
[22] .got PROGBITS 0000000000600fe0 000fe0 000008 08 WA 0 0 8
[23] .got.plt PROGBITS 0000000000600fe8 000fe8 000030 08 WA 0 0 8
|
-------------------------------------------------------------------------------
### 实验 ###
export LD_LIBRARY_PATH=$PWD
gdb main.exe
(gdb) break main
(gdb) run
Breakpoint 1, main () at main.c:4
4 xyz = 100;
|
### 加载ld的符号表,(/usr/lib/debug/lib/*是libc6-dbg安装的debug symbol。)
### 注意 add-symbol-file的第三个参数,地址是如何得到的。
(gdb) info sharedlibrary
From To Syms Read Shared Object Library
0x00007ffff7ddcaf0 0x00007ffff7df5a66 Yes (*) /lib64/ld-linux-x86-64.so.2
0x00007ffff7bda500 0x00007ffff7bda628 Yes /home/dyno/codes/plt/libtest.so
0x00007ffff7864c00 0x00007ffff79817ec Yes /lib/x86_64-linux-gnu/libc.so.6
(gdb) add-symbol-file /usr/lib/debug/lib/x86_64-linux-gnu/ld-2.13.so 0x00007ffff7ddcaf0
(gdb) directory ~/codes/debsrc/eglibc-2.13/elf
(gdb) set disassemble-next-line on
|
### foo() 现在是 <foo@plt>
(gdb) disassemble main
Dump of assembler code for function main:
0x0000000000400674 <+0>: push %rbp
0x0000000000400675 <+1>: mov %rsp,%rbp
0x0000000000400678 <+4>: movl $0x64,0x2009a6(%rip) # 0x601028 <xyz>
0x0000000000400682 <+14>: mov $0x0,%eax
0x0000000000400687 <+19>: callq 0x400578 <foo@plt> <----
0x000000000040068c <+24>: mov $0x0,%eax
0x0000000000400691 <+29>: callq 0x400578 <foo@plt>
...
(gdb) disassemble 0x400578
Dump of assembler code for function foo@plt:
0x0000000000400578 <+0>: jmpq *0x200a92(%rip) # 0x601010 <_GLOBAL_OFFSET_TABLE_+40>
0x000000000040057e <+6>: pushq $0x2 <----
0x0000000000400583 <+11>: jmpq 0x400548 <----
End of assembler dump.
</foo@plt></foo@plt></xyz>
|
### pushq是什么?翻译函数所需要的参数,这个是第一个参数reloc_index,是函数foo在GOT中的偏移量。
### $rip里存了下一条指令,所以实际上将要执行顺序下一条指令
(gdb) p/x 0x40057e + 0x200a92
$3 = 0x601010
|
### 这就是PLT的精妙之处,第一次执行,转到哪里去了呢?
(gdb) disassemble 0x400548
No function contains specified address.
(gdb) x/5i 0x400548
0x400548: pushq 0x200aa2(%rip) # 0x600ff0 <_GLOBAL_OFFSET_TABLE_+8> <----
0x40054e: jmpq *0x200aa4(%rip) # 0x600ff8 <_GLOBAL_OFFSET_TABLE_+16> <----
0x400554: nopl 0x0(%rax)
0x400558 <__libc_start_main@plt>: jmpq *0x200aa2(%rip) # 0x601000 <_GLOBAL_OFFSET_TABLE_+24>
0x40055e <__libc_start_main@plt+6>: pushq $0x0
|
### 又一个pushq, link_map .got.plt,是翻译需要的第二个参数。
### 再次jumpq,where?where?
(gdb) x/a 0x600ff8
0x600ff8 <_GLOBAL_OFFSET_TABLE_+16>: 0x7ffff7df0760
(gdb) info symbol 0x7ffff7df0760
_dl_runtime_resolve in section .text of /usr/lib/debug/lib/x86_64-linux-gnu/ld-2.13.so
|
### 看看_dl_runtime_resolve是怎么工作的...
(gdb) break _dl_runtime_resolve
(gdb) info breakpoints
Num Type Disp Enb Address What
1 breakpoint keep y 0x0000000000400678 in main at main.c:4
breakpoint already hit 1 time
2 breakpoint keep y 0x00007ffff7df0760 ../sysdeps/x86_64/dl-trampoline.S:30
(gdb) si
0x0000000000400548 in ?? ()
=> 0x0000000000400548: ff 35 a2 0a 20 00 pushq 0x200aa2(%rip) # 0x600ff0 <_GLOBAL_OFFSET_TABLE_+8>
|
### 上面提到的第二个参数
(gdb) x/x 0x600ff0
0x600ff0 <_GLOBAL_OFFSET_TABLE_+8>: 0x00007ffff7ffe2e8
(gdb) list _dl_runtime_resolve
...
29 _dl_runtime_resolve:
30 subq $56,%rsp
31 cfi_adjust_cfa_offset(72) # Incorporate PLT
32 movq %rax,(%rsp) # Preserve registers otherwise clobbered.
...
(gdb) list +
...
39 movq 64(%rsp), %rsi # Copy args pushed by PLT in register.
40 movq 56(%rsp), %rdi # %rdi: link_map, %rsi: reloc_index <----前面提到的两个参数
41 call _dl_fixup # Call resolver.
42 movq %rax, %r11 # Save return value <----真正的共享库里函数地址
43 movq 48(%rsp), %r9 # Get register content back.
...
|
### 设置断点,看地址在GOT表中的变化
(gdb) info line _dl_runtime_resolve
Line 30 of "../sysdeps/x86_64/dl-trampoline.S" starts at address 0x7ffff7df0760 <_dl_runtime_resolve>
and ends at 0x7ffff7df0764 <_dl_runtime_resolve+4>.
(gdb) break ../sysdeps/x86_64/dl-trampoline.S:40
Breakpoint 3 at 0x7ffff7df078b: file ../sysdeps/x86_64/dl-trampoline.S, line 40.
(gdb) c
(gdb) x/a 0x601010
0x601010 <_GLOBAL_OFFSET_TABLE_+40>: 0x40057e <foo@plt+6> <---- _dl_fixup 之前
(gdb) ni
42 movq %rax, %r11 # Save return value
=> 0x00007ffff7df0795 <_dl_runtime_resolve+53>: 49 89 c3 mov %rax,%r11
(gdb) x/a 0x601010
0x601010 <_GLOBAL_OFFSET_TABLE_+40>: 0x7ffff7bda5cc <foo> <---- _dl_fixup 之后
</foo></foo@plt+6>
|
### 以后再次调用foo就直接到这里了。
-------------------------------------------------------------------------------
### 延伸阅读 ###
[1] Reversing the ELF Stepping with GDB during PLT uses and .GOT fixup
http://packetstormsecurity.org/files/view/25642/elf-runtime-fixup.txt
[2] AMD64 Application Binary Interface (v 0.99)
http://www.x86-64.org/documentation/abi.pdf
[3] PLT and GOT - the key to code sharing and dynamic libraries
http://www.technovelty.org/linux/pltgot.html
[4] examining PLT/GOT structures
http://althing.cs.dartmouth.edu/secref/resources/plt-got.txt
[5] Debugging with GDB
http://sourceware.org/gdb/current/onlinedocs/gdb/
[6] 共享库函数调用原理
http://blog.csdn.net/absurd/article/details/3169860
[7] How main() is executed on Linux
http://linuxgazette.net/issue84/hawk.html
[8] Gentle Introduction to x86-64 Assembly
http://www.x86-64.org/documentation/assembly.html