关于linux可安装模块的装载地址的研究
前言、写这篇文章的由来
最近在学习韦东山嵌入式培训视频(3期项目实战之USB摄像头监控)时,仿照视频教程,自己写了一个简化版的uvc摄像头驱动。在虚拟机上调试驱动模块时,入了一个大坑,折腾了很久才总算爬出来。而源头是由于自己对linux模块的装载地址认识不足,再加上一篇网文的误导。兹整理记录下来爬坑期间做的笔记,希望对自己和大家都有所助益。
一、实验环境
1.1 虚拟机环境
a) Vmware版本:Vmware Workstation 12.5.7
b) Ubuntu版本:9.10
c) 内核版本:2.6.31.14
d) gcc版本:4.4.1
e) gdb版本:7.0
1.2 开发板环境
1.2.1 硬件
开发板:百问网JZ2440开发板
摄像头:百问网自制uvc摄像头
1.2.2 软件
a) 内核版本: 2.6.31.14
b) toolchain版本:
arm-linux-gcc 3.4.5
arm-linux-gdb 6.8
二、 基础知识简介
2.1 linux可安装模块(简称LKM)
LKM是一种特殊的ELF文件。可以用objdmp –h 或者readelf -e来查看它的section信息。
LKM不能直接运行,而需要用insmod命令动态装载到内核中,才能运行。而内核装载LDM主要是在load_module()中实现的。它会为LKM分配内存空间,计算其各个section的装载地址,然后将它们复制到对应的装载地址上。
2.2 关于怎样用kgdb调试可安装模块,网上有很多文章。总结下来,基本套路都是类似的,大致如下:
i)配置目标机的内核,以支持kgdb调试
ii)配置目标机的引导程序的启动参数,以支持kgdb串口调试
iii) 目标机上insmod xxx.ko,然后用某种方法获取xxx.ko的text节的装载地址(假设为0xd099a000)(注:请见下文3.3)
iv) 目标机在shell中运行echo g > /proc/sysrq-trigger ,等待开发机的gdb来连接
v) 开发机在shell中进入linux源代码目录,然后执行:
gdb ./vmlinux
(gdb) set remotebaud 115200
(gdb) target remote /dev/ttyS0
Remote debugging using /dev/ttyS0
kgdb_breakpoint () at kernel/kgdb.c:1674
1674 wmb(); /*Sync point after breakpoint */
以上信息说明开发机的gdb和目标机的kgdb成功建立了连接。接下来,在开发机上运行:
(gdb) add-symbol-file /xxx_path/xxx.ko 0xd099a000
后续的调试方法,就和在单机上用gdb调试应用程序基本相同了。
2.3 关于怎样获取模块的代码段的装载地址,网上有两种方法:
方法1)经验证是可行的
cat /sys/module/xxx.ko/sections/.text
方法2)经验证是有问题的
“cat /proc/modules | grep test”(假设模块名称为test.ko,注意不要带".ko")找到模块的加载地址,如下图所示:
找到代码段(.text段)的偏移量,如下图所示:
偏移量是.text对应的行中第四个十六进制字段(或者说从左往右数第六个字 段)是.text段在文件中的偏移量。将这一偏移量与模块中的装载地址相加,就可以找到模块的代码在重定向之后的地址了。在我们的例子中,可以得到 0xffffffffa001b000 + 0x00000040 = 0xFFFFFFFFA001B040。
三、为什么第二种方法是有问题的
3.1 内核源代码的分析
(下文分析的内核版本是2.6.31.14)
模块的加载,是在kernel/module.c的load_module()里实现的:
static noinline struct module *load_module(void __user *umod, unsigned long len, const char __user *uargs) { . . . /* Determine total sizes, and put offsets in sh_entsize. For now this is done generically; there doesn't appear to be any special cases for the architectures. */ layout_sections(mod, hdr, sechdrs, secstrings);
. . .
/* Transfer each section which specifies SHF_ALLOC */ DEBUGP("final section addresses:\n"); for (i = 0; i < hdr->e_shnum; i++) { void *dest; if (!(sechdrs[i].sh_flags & SHF_ALLOC)) continue; if (sechdrs[i].sh_entsize & INIT_OFFSET_MASK) dest = mod->module_init + (sechdrs[i].sh_entsize & ~INIT_OFFSET_MASK); else dest = mod->module_core + sechdrs[i].sh_entsize;
if (sechdrs[i].sh_type != SHT_NOBITS)
memcpy(dest, (void *)sechdrs[i].sh_addr,
sechdrs[i].sh_size);
/* Update sh_addr to point to copy in image. */
sechdrs[i].sh_addr = (unsigned long)dest;
DEBUGP("\t0x%lx %s\n", sechdrs[i].sh_addr, secstrings + sechdrs[i].sh_name);
} }
模块里各个section的定位,主要就是在以上两段代码中实现的。
- 第一段代码layout_sections()的作用,正如注释所说:a)确定模块总大小,b) 计算模块里各个section的偏移量,并保存到sechdrs.sh_entsize里(内核会为模块分
配一个Elf_Shdr类型的数组sechdrs[],数组元素分别对应该模块的每个section的头信息)
- 第二段代码的作用,是对模块内带有SHF_ALLOC标志的section,将其内容复制到最终的目的地址上。
下面主要分析layout_sections():
/* Lay out the SHF_ALLOC sections in a way not dissimilar to how ld might -- code, read-only data, read-write data, small data. Tally sizes, and place the offsets into sh_entsize fields: high bit means it belongs in init. */ static void layout_sections(struct module *mod, const Elf_Ehdr *hdr, Elf_Shdr *sechdrs, const char *secstrings) { static unsigned long const masks[][2] = { //关于这些mask的含义,暂未深究,详见参考资料5.8(ELF格式探析之三:sections): /* NOTE: all executable code must be the first section * in this array; otherwise modify the text_size * finder in the two loops below */ { SHF_EXECINSTR | SHF_ALLOC, ARCH_SHF_SMALL }, { SHF_ALLOC, SHF_WRITE | ARCH_SHF_SMALL }, { SHF_WRITE | SHF_ALLOC, ARCH_SHF_SMALL }, { ARCH_SHF_SMALL | SHF_ALLOC, 0 } }; unsigned int m, i; for (i = 0; i < hdr->e_shnum; i++) sechdrs[i].sh_entsize = ~0UL; //先初始化sh_entsize 为-1 DEBUGP("Core section allocation order:\n"); for (m = 0; m < ARRAY_SIZE(masks); ++m) { for (i = 0; i < hdr->e_shnum; ++i) { Elf_Shdr *s = &sechdrs[i]; if ((s->sh_flags & masks[m][0]) != masks[m][0] || (s->sh_flags & masks[m][1]) || s->sh_entsize != ~0UL || strstarts(secstrings + s->sh_name, ".init")) //对每个section进行一些判断,细节暂未深究 continue; s->sh_entsize = get_offset(mod, &mod->core_size, s, i); //计算各个section的偏移量,并保存到sechdrs.sh_entsize里 DEBUGP("\t%s\n", secstrings + s->sh_name); } if (m == 0) mod->core_text_size = mod->core_size; } DEBUGP("Init section allocation order:\n");
. . . //省略和init节相关的处理代码 } }
可见,核心的代码就是get_offset()这个函数:
/* Update size with this section: return offset. */ static long get_offset(struct module *mod, unsigned int *size, Elf_Shdr *sechdr, unsigned int section) { long ret; *size += arch_mod_section_prepend(mod, section); //在非parisc平台下,arch_mod_section_prepend定义为return 0; ret = ALIGN(*size, sechdr->sh_addralign ?: 1);//对齐修正。注意:这里的三目表达式等价于sechdr->sh_addralign ? sechdr->sh_addralign : 1 *size = ret + sechdr->sh_size; return ret; }
从代码可以看出:内核在计算LKM装载后的各个section的偏移地址时,并没有用.ko文件中的offset信息,而是按照以下公式来计算的:
offset_align = ALIGN(*size, sechdr->sh_addralign ?: 1)
其中,offset_align是计算出的偏移地址,*size是该section之前所有section的总大小,ALIGN是进行对齐修正,对齐边界是该section的sh_addralign或者1。
对于.text节来说,由于它是模块的第一个section,所以*size即mod->core_size,且因mod->core_size初始值是0(这是我在单步时确定的,但我没找到源代码中对其显式初始化的地方?),所以text节的offset_align即等于0,换句话说,text节的装载地址就是模块的装载地址。
3.2 实验验证(在JZ2440开发板上实验)
step1) 在get_offset()里加printk后,编译安装内核
step2)insmod my_uvc.ko,然后cat /proc/modules|grep my_uvc获取模块的装载地址,然后通过sysfs获取各个section的装载地址
#cat /proc/modules |grep my_uvc
my_uvc 15028 0 - Live 0xbf000000
# ls -al /sys/module/my_uvc/sections/
drwxr-xr-x 2 0 0 0 Jan 1 00:04 .
drwxr-xr-x 6 0 0 0 Jan 1 00:01 ..
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .bss
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .data
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .gnu.linkonce.this_module
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .note.gnu.build-id
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .rodata
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .rodata.str1.1
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .strtab
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .symtab
-r--r--r-- 1 0 0 4096 Jan 1 00:04 .text
# cat /sys/module/my_uvc/sections/.text
0xbf000000
# cat /sys/module/my_uvc/sections/.note.gnu.build-id
0xbf0013e8
# cat /sys/module/my_uvc/sections/.rodata
0xbf00140c
# cat /sys/module/my_uvc/sections/.rodata.str1.1
0xbf0015c4
# cat /sys/module/my_uvc/sections/.symtab
0xbf001a08
# cat /sys/module/my_uvc/sections/.strtab
0xbf002438
# cat /sys/module/my_uvc/sections/.gnu.linkonce.this_module
0xbf002a70
# cat /sys/module/my_uvc/sections/.bss
0xbf002b84
对比printk的日志(注:红字是我手工加上的结算结果)
sh_name:50, name:.text, sh_type:1, sh_flags:6, sh_offset:88, sh_size:5096, sh_addralign:4, offset_align:0 0xbf000000+0=0xbf000000 sh_name:27, name:.note.gnu.build-id, sh_type:7, sh_flags:2, sh_offset:52, sh_size:36, sh_addralign:4, offset_align:0x13e8 0xbf000000+0x13e8=0xbf0013e8 sh_name:60, name:.rodata, sh_type:1, sh_flags:2, sh_offset:5184, sh_size:440, sh_addralign:4, offset_align:0x140c 0xbf000000+0x140c=0xbf00140c sh_name:77, name:.rodata.str1.1, sh_type:1, sh_flags:50, sh_offset:5682, sh_size:1092, sh_addralign:1, offset_align:0x15c4 0xbf000000+0x15c4=0xbf0015c4 sh_name:1, name:.symtab, sh_type:2, sh_flags:2, sh_offset:130584, sh_size:2608, sh_addralign:4, offset_align:0x1a08 0xbf000000+0x1a08=0xbf001a08 sh_name:9, name:.strtab, sh_type:3, sh_flags:2, sh_offset:133192, sh_size:1397, sh_addralign:1, offset_align:0x2438 0xbf000000+0x2438=0xbf002438 i:8, sh_name:96, name:.data, sh_type:1, sh_flags:3, sh_offset:6776, sh_size:192, sh_addralign:4, offset_align:0x29b0 0xbf000000+0x29b0=0xbf0029b0 i:10, sh_name:106, name:.gnu.linkonce.this_module, sh_type:1, sh_flags:3, sh_offset:6968, sh_size:276, sh_addralign:4, offset_align:0x2a70 0xbf000000+0x2a70=0xbf002a70 i:12, sh_name:132, name:.bss, sh_type:8, sh_flags:3, sh_offset:7244, sh_size:3888, sh_addralign:4, offset_align:0x2b84 0xbf000000+0x2b84=0xbf002b84
可见,实际结果和分析源代码得出的结论是一致的。
四、如果使用第二种方法获取LKM的text节的装载地址,实际调试时,会表现出什么症状,原因是什么
4.1 实践发现,在X86平台下,和在arm平台下,实际调试时表现出的症状,差别比较大
1) 在X86平台下面,会出现很多奇怪的现象,例如:
- 虽然加了断点,但continue让程序跑起来后,却直接无视断点;
- 有时虽然能进断点,但单步之后,程序就跑飞掉了;
- 甚至加了断点,continue让程序跑起来后,会报segmentation fault,或者instruction fault
2) 在arm平台下,可以正常的调试(包括加断点,单步,监视等),没有表现出显式的症状
4.2 为什么在两种平台下,会有这种差别呢?
为了找出原因,尝试在两个平台下分别做以下实验(为了简化,目标代码用了宋宝华的《Linux设备驱动开发详解:基于最新的Linux 4.0内核》第六章的globalmem):
step1)用两种方法(详见上文2.3小节),分别获取LKM的text节的装载地址;
step2)在gdb中,分别用这两种地址来add-symbol-file,然后b globalmem_read(第二种不要continue),第一种当continue命中断点后,打印反汇编代码,结果是:
a) X86平台(假设cat /proc/modules|grep globalmem得到LKM的装载地址是0xe0c26000,而objdump –h 得到text节的offset是40)
用第二种方法获取text节的装载地址,然后add-symbol-file,b globalmem_read:
(gdb) add-symbol-file /home/book/workspace/baohua/drivers/ch6/globalmem.ko 0xe0c26040 add symbol table from file "/home/book/workspace/baohua/drivers/ch6/globalmem.ko" at .text_addr = 0xe0c26040 (y or n) y Reading symbols from /home/book/workspace/baohua/drivers/ch6/globalmem.ko...done. (gdb) b globalmem_read Breakpoint 1 at 0xe0c261a0: file /home/book/workspace/baohua/drivers/ch6/globalmem.c, line 64.
用第一种方法获取text节的装载地址,然后add-symbol-file,b globalmem_read:
(gdb) add-symbol-file /home/book/workspace/baohua/drivers/ch6/globalmem.ko 0xe0c26000 add symbol table from file "/home/book/workspace/baohua/drivers/ch6/globalmem.ko" at .text_addr = 0xe0c26000 (y or n) y Reading symbols from /home/book/workspace/baohua/drivers/ch6/globalmem.ko...done. (gdb) b globalmem_read Breakpoint 1 at 0xe0c26174: file /home/book/workspace/baohua/drivers/ch6/globalmem.c, line 70. (gdb) c Continuing. [New Thread 3101] [Switching to Thread 3101] Breakpoint 1, globalmem_read (filp=0xda625b00, buf=0x9163000 <Address 0x9163000 out of bounds>, size=32768, ppos=0xd7ae9f98) at /home/book/workspace/baohua/drivers/ch6/globalmem.c:70 70 if (p >= GLOBALMEM_SIZE) (gdb)disassemble 0xe0c26160 <globalmem_read+0>: push %ebp 0xe0c26161 <globalmem_read+1>: mov %esp,%ebp 0xe0c26163 <globalmem_read+3>: sub $0x18,%esp 0xe0c26166 <globalmem_read+6>: mov %ebx,-0xc(%ebp) 0xe0c26169 <globalmem_read+9>: mov %esi,-0x8(%ebp) 0xe0c2616c <globalmem_read+12>: mov %edi,-0x4(%ebp) 0xe0c2616f <globalmem_read+15>: nopl 0x0(%eax,%eax,1) 0xe0c26174 <globalmem_read+20>: xor %esi,%esi ;可见,用第一种方法获取的地址做实验,断点的位置是正确的 0xe0c26176 <globalmem_read+22>: mov 0x70(%eax),%eax 0xe0c26179 <globalmem_read+25>: mov %edx,%edi 0xe0c2617b <globalmem_read+27>: mov 0x8(%ebp),%edx 0xe0c2617e <globalmem_read+30>: mov (%edx),%ebx 0xe0c26180 <globalmem_read+32>: cmp $0xfff,%ebx 0xe0c26186 <globalmem_read+38>: ja 0xe0c261a8 <globalmem_read+72> 0xe0c26188 <globalmem_read+40>: mov $0x1000,%si 0xe0c2618c <globalmem_read+44>: sub %ebx,%esi 0xe0c2618e <globalmem_read+46>: cmp %ecx,%esi 0xe0c26190 <globalmem_read+48>: ja 0xe0c261b8 <globalmem_read+88> 0xe0c26192 <globalmem_read+50>: lea 0x3c(%eax,%ebx,1),%edx 0xe0c26196 <globalmem_read+54>: mov %esi,%ecx 0xe0c26198 <globalmem_read+56>: mov %edi,%eax 0xe0c2619a <globalmem_read+58>: call 0xc0326d60 <copy_to_user> 0xe0c2619f <globalmem_read+63>: test %eax,%eax
;可见,用第二种方法获取的地址做实验,由于在代码段中,0xe0c261a0这个地址上并没有有效的指令,所以如果在其上加断点,就会产生各种奇怪的问题 0xe0c261a1 <globalmem_read+65>: je 0xe0c261bc <globalmem_read+92> 0xe0c261a3 <globalmem_read+67>: mov $0xfffffff2,%esi 0xe0c261a8 <globalmem_read+72>: mov %esi,%eax 0xe0c261aa <globalmem_read+74>: mov -0xc(%ebp),%ebx 0xe0c261ad <globalmem_read+77>: mov -0x8(%ebp),%esi 0xe0c261b0 <globalmem_read+80>: mov -0x4(%ebp),%edi 0xe0c261b3 <globalmem_read+83>: mov %ebp,%esp 0xe0c261b5 <globalmem_read+85>: pop %ebp 0xe0c261b6 <globalmem_read+86>: ret
b) arm平台(假设cat /proc/modules|grep globalmem得到LKM的装载地址是0xbf000000,而objdump –h 得到text节的offset是34)
用第二种方法获取text节的装载地址,然后add-symbol-file,b globalmem_read:
(gdb) add-symbol-file /home/book/workspace/baohua/drivers/ch6/globalmem.ko 0xbf000034 add symbol table from file "/home/book/workspace/baohua/drivers/ch6/globalmem.ko" at .text_addr = 0xbf000034 (y or n) y Reading symbols from /home/book/workspace/baohua/drivers/ch6/globalmem.ko...done. (gdb) b globalmem_read Breakpoint 1 at 0xbf000200: file /home/book/workspace/baohua/drivers/ch6/globalmem.c, line 65.
用第一种方法获取text节的装载地址,然后add-symbol-file,b globalmem_read:
(gdb) add-symbol-file /home/book/workspace/baohua/drivers/ch6/globalmem.ko 0xbf000000 add symbol table from file "/home/book/workspace/baohua/drivers/ch6/globalmem.ko" at .text_addr = 0xbf000000 (y or n) y Reading symbols from /home/book/workspace/baohua/drivers/ch6/globalmem.ko...done. (gdb) b globalmem_read Breakpoint 1 at 0xbf0001cc: file /home/book/workspace/baohua/drivers/ch6/globalmem.c, line 65. (gdb) c Continuing. [New Thread 787] [Switching to Thread 787] Breakpoint 1, globalmem_read (filp=0xc3df44e0, buf=0xbe962c80 "", size=8192, ppos=0xc06edf80) at /home/book/workspace/baohua/drivers/ch6/globalmem.c:65 65 unsigned long p = *ppos; (gdb)disassemble 0xbf0001c0 <globalmem_read+0>: mov r12, sp 0xbf0001c4 <globalmem_read+4>: push {r4, r5, r6, r7, r8, r11, r12, lr, pc} 0xbf0001c8 <globalmem_read+8>: sub r11, r12, #4 ; 0x4 0xbf0001cc <globalmem_read+12>: ldr r7, [r3] ;可见,用第一种方法获取的地址做实验,断点的位置是正确的 0xbf0001d0 <globalmem_read+16>: mov r8, r3 0xbf0001d4 <globalmem_read+20>: cmp r7, #4096 ; 0x1000 0xbf0001d8 <globalmem_read+24>: mov r12, r1 0xbf0001dc <globalmem_read+28>: ldr r0, [r0, #116] 0xbf0001e0 <globalmem_read+32>: movcs r6, #0 ; 0x0 0xbf0001e4 <globalmem_read+36>: bcs 0xbf00025c <globalmem_read+156> 0xbf0001e8 <globalmem_read+40>: mov r1, sp 0xbf0001ec <globalmem_read+44>: bic r3, r1, #8128 ; 0x1fc0 0xbf0001f0 <globalmem_read+48>: bic r3, r3, #63 ; 0x3f 0xbf0001f4 <globalmem_read+52>: rsb r5, r7, #4096 ; 0x1000 0xbf0001f8 <globalmem_read+56>: cmp r5, r2 0xbf0001fc <globalmem_read+60>: movcs r5, r2 0xbf000200 <globalmem_read+64>: ldr r3, [r3, #8] ;可见用第二种方法获取的地址做实验,断点位置已经偏离了正确值0x34个字节,但由于该地址是有效的地址,所以实际上还是可以正常调试的,并不会表现出显式的症状 0xbf000204 <globalmem_read+68>: adds r2, r12, r5 0xbf000208 <globalmem_read+72>: sbcscc r2, r2, r3 0xbf00020c <globalmem_read+76>: movcc r3, #0 ; 0x0 0xbf000210 <globalmem_read+80>: cmp r3, #0 ; 0x0 0xbf000214 <globalmem_read+84>: movne r0, r5