arm linux kernel 从入口到start_kernel 的代码分析

参考资料：

《ARM体系结构与编程》

《嵌入式Linux应用开发完全手册》

Linux_Memory_Address_Mapping

http://www.chinaunix.net/old_jh/4/1021226.html

更多文档参见：http://pan.baidu.com/s/1mg3DbHQ

本文针对arm linux, 从kernel的第一条指令开始分析,一直分析到进入start_kernel()函数. 我们当前以linux-2.6.19内核版本作为范例来分析,本文中所有的代码,前面都会加上行号以便于和源码进行对照, 例: 在文件init/main.c中: 00478: asmlinkage void __init start_kernel(void) 前面的"00478:" 表示478行,冒号后面的内容就是源码了. 在分析代码的过程中,我们使用缩进来表示各个代码的调用层次. 由于启动部分有一些代码是平台特定的,虽然大部分的平台所实现的功能都比较类似,但是为了更好的对code进行说明,对于平台相关的代码,我们选择at91(ARM926EJS)平台进行分析. 另外,本文是以uncomPRessed kernel开始讲解的.对于内核解压缩部分的code,在 arch/arm/boot/compressed中,本文不做讨论.

通常从系统上电到执行到linux kenel这部分的任务是由boot loader来完成. 关于boot loader的内容,本文就不做过多介绍. 这里只讨论进入到linux kernel的时候的一些限制条件,这一般是boot loader在最后跳转到kernel之前要完成的: 1. CPU必须处于SVC(supervisor)模式,并且IRQ和FIQ中断都是禁止的; 2. MMU(内存管理单元)必须是关闭的, 此时虚拟地址对物理地址; 3. 数据cache(Data cache)必须是关闭的 4. 指令cache(Instruction cache)可以是打开的,也可以是关闭的,这个没有强制要求; 5. CPU 通用寄存器0 (r0)必须是 0; 6. CPU 通用寄存器1 (r1)必须是 ARM Linux machine type (关于machine type, 我们后面会有讲解) 7. CPU 通用寄存器2 (r2) 必须是 kernel parameter list 的物理地址(parameter list 是由boot loader传递给kernel,用来描述设备信息属性的列表,详细内容可参考"Booting ARM Linux"文档).

首先，我们先对几个重要的宏进行说明(我们针对有MMU的情况)：

宏 位置 默认值 说明 KERNEL_RAM_ADDR arch/arm/kernel/head.S +26 0xc0008000 kernel在RAM中的的虚拟地址 PAGE_OFFSET include/asm-arm/memeory.h +50 0xc0000000 内核空间的起始虚拟地址 TEXT_OFFSET arch/arm/Makefile +137 0x00008000 内核相对于存储空间的偏移 TEXTADDR arch/arm/kernel/head.S +49 0xc0008000 kernel的起始虚拟地址 PHYS_OFFSET include/asm-arm/arch-xxx/memory.h 平台相关 RAM的起始物理地址 内核的入口是stext,这是在arch/arm/kernel/vmlinux.lds.S中定义的: 00011: ENTRY(stext) 对于vmlinux.lds.S,这是ld script文件,此文件的格式和汇编及C程序都不同,本文不对ld script作过多的介绍,只对内核中用到的内容进行讲解,关于ld的详细内容可以参考ld.info 这里的ENTRY(stext) 表示程序的入口是在符号stext. 而符号stext是在arch/arm/kernel/head.S中定义的: 下面我们将arm linux boot的主要代码列出来进行一个概括的介绍,然后,我们会逐个的进行详细的讲解. 在arch/arm/kernel/head.S中 72 - 94 行,是arm linux boot的主代码: 00072: ENTRY(stext) 00073: msr cpsr_c, #PSR_F_BIT | PSR_I_BIT | SVC_MODE @ ensure svc mode 00074: @ and irqs disabled 00075: mrc p15, 0, r9, c0, c0 @ get processor id 00076: bl __lookup_processor_type @ r5=procinfo r9=cpuid 00077: movs r10, r5 @ invalid processor (r5=0)? 00078: beq __error_p @ yes, error 'p' 00079: bl __lookup_machine_type @ r5=machinfo 00080: movs r8, r5 @ invalid machine (r5=0)? 00081: beq __error_a @ yes, error 'a' 00082: bl __create_page_tables 00083: 00084: /* 00085: * The following calls CPU specific code in a position independent 00086: * manner. See arch/arm/mm/proc-*.S for details. r10 = base of 00087: * xxx_proc_info structure selected by __lookup_machine_type 00088: * above. On return, the CPU will be ready for the MMU to be 00089: * turned on, and r0 will hold the CPU control register value. 00090: */ 00091: ldr r13, __switch_data @ address to jump to after 00092: @ mmu has been enabled 00093: adr lr, __enable_mmu @ return (PIC) address 00094: add pc, r10, #PROCINFO_INITFUNC 其中,73行是确保kernel运行在SVC模式下,并且IRQ和FIRQ中断已经关闭,这样做是很谨慎的. arm linux boot的主线可以概括为以下几个步骤: 1. 确定 processor type (75 - 78行) 2. 确定 machine type (79 - 81行) 3. 创建页表 (82行) 4. 调用平台特定的__cpu_flush函数 (在struct proc_info_list中) (94 行) 5. 开启mmu (93行) 6. 切换数据 (91行) 最终跳转到start_kernel (在__switch_data的结束的时候,调用了 b start_kernel) 下面,我们按照这个主线,逐步的分析Code.

arch/arm/kernel/head.S中: 00075: mrc p15, 0, r9, c0, c0 @ get processor id 00076: bl __lookup_processor_type @ r5=procinfo r9=cpuid 00077: movs r10, r5 @ invalid processor (r5=0)? 00078: beq __error_p @ yes, error 'p' 75行: 通过cp15协处理器的c0寄存器来获得processor id的指令. 关于cp15的详细内容可参考相关的arm手册 76行: 跳转到__lookup_processor_type.在__lookup_processor_type中,会把processor type 存储在r5中 77,78行: 判断r5中的processor type是否是0,如果是0,说明是无效的processor type,跳转到__error_p(出错) __lookup_processor_type 函数主要是根据从cpu中获得的processor id和系统中的proc_info进行匹配,将匹配到的proc_info_list的基地址存到r5中, 0表示没有找到对应的processor type. 下面我们分析__lookup_processor_type函数 arch/arm/kernel/head-common.S中: 00145: .type __lookup_processor_type, %function 00146: __lookup_processor_type: 00147: adr r3, 3f 00148: ldmda r3, {r5 - r7} 00149: sub r3, r3, r7 @ get offset between virt&phys 00150: add r5, r5, r3 @ convert virt addresses to 00151: add r6, r6, r3 @ physical address space 00152: 1: ldmia r5, {r3, r4} @ value, mask 00153: and r4, r4, r9 @ mask wanted bits 00154: teq r3, r4 00155: beq 2f 00156: add r5, r5, #PROC_INFO_SZ @ sizeof(proc_info_list) 00157: cmp r5, r6 00158: blo 1b 00159: mov r5, #0 @ unknown processor 00160: 2: mov pc, lr 00161: 00162: /* 00163: * This provides a C-API version of the above function. 00164: */ 00165: ENTRY(lookup_processor_type) 00166: stmfd sp!, {r4 - r7, r9, lr} 00167: mov r9, r0 00168: bl __lookup_processor_type 00169: mov r0, r5 00170: ldmfd sp!, {r4 - r7, r9, pc} 00171: 00172: /* 00173: * Look in include/asm-arm/procinfo.h and arch/arm/kernel/arch.[ch] for 00174: * more information about the __proc_info and __arch_info structures. 00175: */ 00176: .long __proc_info_begin 00177: .long __proc_info_end 00178: 3: .long . 00179: .long __arch_info_begin 00180: .long __arch_info_end 145, 146行是函数定义 147行: 取地址指令,这里的3f是向前symbol名称是3的位置,即第178行,将该地址存入r3. 这里需要注意的是,adr指令取址,获得的是基于pc的一个地址,要格外注意,这个地址是3f处的"运行时地址", 由于此时MMU还没有打开,也可以理解成物理地址(实地址).(详细内容可参考arm指令手册) 148行: 因为r3中的地址是178行的位置的地址,因而执行完后: (ldmda表示栈指针递减，即r3递减，内存的地址编号较大的对应寄存器编号较大的) r5存的是176行符号 __proc_info_begin的地址; r6存的是177行符号 __proc_info_end的地址; r7存的是3f处的地址. 这里需要注意链接地址和运行时地址的区别. r3存储的是运行时地址(物理地址),而r7中存储的是链接地址(虚拟地址). __proc_info_begin和__proc_info_end是在arch/arm/kernel/vmlinux.lds.S中: 00031: __proc_info_begin = .; 00032: *(.proc.info.init) 00033: __proc_info_end = .; 这里是声明了两个变量:__proc_info_begin 和 __proc_info_end,其中等号后面的"."是location counter(详细内容请参考ld.info) 这三行的意思是: __proc_info_begin 的位置上,放置所有文件中的 ".proc.info.init" 段的内容,然后紧接着是 __proc_info_end 的位置. kernel 使用struct proc_info_list来描述processor type. 在 include/asm-arm/procinfo.h 中: 00029: struct proc_info_list { 00030: unsigned int cpu_val; 00031: unsigned int cpu_mask; 00032: unsigned long __cpu_mm_mmu_flags; /* used by head.S */ 00033: unsigned long __cpu_io_mmu_flags; /* used by head.S */ 00034: unsigned long __cpu_flush; /* used by head.S */ 00035: const char *arch_name; 00036: const char *elf_name; 00037: unsigned int elf_hwcap; 00038: const char *cpu_name; 00039: struct processor *proc; 00040: struct cpu_tlb_fns *tlb; 00041: struct cpu_user_fns *user; 00042: struct cpu_cache_fns *cache; 00043: }; 我们当前以at91为例,其processor是926的. 在arch/arm/mm/proc-arm926.S 中: 00464: .section ".proc.info.init", #alloc, #execinstr 00465: 00466: .type __arm926_proc_info,#object 00467: __arm926_proc_info: 00468: .long 0x41069260 @ ARM926EJ-S (v5TEJ) 00469: .long 0xff0ffff0 00470: .long PMD_TYPE_SECT | / 00471: PMD_SECT_BUFFERABLE | / 00472: PMD_SECT_CACHEABLE | / 00473: PMD_BIT4 | / 00474: PMD_SECT_AP_WRITE | / 00475: PMD_SECT_AP_READ 00476: .long PMD_TYPE_SECT | / 00477: PMD_BIT4 | / 00478: PMD_SECT_AP_WRITE | / 00479: PMD_SECT_AP_READ 00480: b __arm926_setup 00481: .long cpu_arch_name 00482: .long cpu_elf_name 00483: .long HWCAP_SWP|HWCAP_HALF|HWCAP_THUMB|HWCAP_FAST_MULT|HWCAP_VFP|HWCAP_EDSP|HWCAP_java 00484: .long cpu_arm926_name 00485: .long arm926_processor_functions 00486: .long v4wbi_tlb_fns 00487: .long v4wb_user_fns 00488: .long arm926_cache_fns 00489: .size __arm926_proc_info, . - __arm926_proc_info 从464行,我们可以看到 __arm926_proc_info 被放到了".proc.info.init"段中. 对照struct proc_info_list,我们可以看到 __cpu_flush的定义是在480行,即__arm926_setup.(我们将在"4. 调用平台特定的__cpu_flush函数"一节中详细分析这部分的内容.) 从以上的内容我们可以看出: r5中的__proc_info_begin是proc_info_list的起始地址, r6中的__proc_info_end是proc_info_list的结束地址. 149行: 从上面的分析我们可以知道r3中存储的是3f处的物理地址,而r7存储的是3f处的虚拟地址,这一行是计算当前程序运行的物理地址和虚拟地址的差值,将其保存到r3中. 150行: 将r5存储的虚拟地址(__proc_info_begin)转换成物理地址 151行: 将r6存储的虚拟地址(__proc_info_end)转换成物理地址 152行: 对照struct proc_info_list,可以得知,这句是将当前proc_info的cpu_val和cpu_mask分别存r3, r4中 153行: r9中存储了processor id(arch/arm/kernel/head.S中的75行),与r4的cpu_mask进行逻辑与操作,得到我们需要的值 154行: 将153行中得到的值与r3中的cpu_val进行比较 155行: 如果相等,说明我们找到了对应的processor type,跳到160行,返回 156行: (如果不相等) , 将r5指向下一个proc_info, 157行: 和r6比较,检查是否到了__proc_info_end. 158行: 如果没有到__proc_info_end,表明还有proc_info配置,返回152行继续查找 159行: 执行到这里,说明所有的proc_info都匹配过了,但是没有找到匹配的,将r5设置成0(unknown processor) 160行: 返回

arch/arm/kernel/head.S中: 00079: bl __lookup_machine_type @ r5=machinfo 00080: movs r8, r5 @ invalid machine (r5=0)? 00081: beq __error_a @ yes, error 'a' 79行: 跳转到__lookup_machine_type函数,在__lookup_machine_type中,会把struct machine_desc的基地址(machine type)存储在r5中 80,81行: 将r5中的 machine_desc的基地址存储到r8中,并判断r5是否是0,如果是0,说明是无效的machine type,跳转到__error_a(出错) __lookup_machine_type 函数 下面我们分析__lookup_machine_type 函数: arch/arm/kernel/head-common.S中: 00176: .long __proc_info_begin 00177: .long __proc_info_end 00178: 3: .long . 00179: .long __arch_info_begin 00180: .long __arch_info_end 00181: 00182: /* 00183: * Lookup machine architecture in the linker-build list of architectures. 00184: * Note that we can't use the absolute addresses for the __arch_info 00185: * lists since we aren't running with the MMU on (and therefore, we are 00186: * not in the correct address space). We have to calculate the offset. 00187: * 00188: * r1 = machine architecture number 00189: * Returns: 00190: * r3, r4, r6 corrupted 00191: * r5 = mach_info pointer in physical address space 00192: */ 00193: .type __lookup_machine_type, %function 00194: __lookup_machine_type: 00195: adr r3, 3b 00196: ldmia r3, {r4, r5, r6} 00197: sub r3, r3, r4 @ get offset between virt&phys 00198: add r5, r5, r3 @ convert virt addresses to 00199: add r6, r6, r3 @ physical address space 00200: 1: ldr r3, [r5, #MACHINFO_TYPE] @ get machine type 00201: teq r3, r1 @ matches loader number? 00202: beq 2f @ found 00203: add r5, r5, #SIZEOF_MACHINE_DESC @ next machine_desc 00204: cmp r5, r6 00205: blo 1b 00206: mov r5, #0 @ unknown machine 00207: 2: mov pc, lr 193, 194行: 函数声明 195行: 取地址指令,这里的3b是向后symbol名称是3的位置,即第178行,将该地址存入r3. 和上面我们对__lookup_processor_type 函数的分析相同,r3中存放的是3b处物理地址. 196行: r3是3b处的地址,因而执行完后:（ldmia 表示栈是递增的，即r3递增，低内存地址对应小号寄存器） r4存的是 3b处的地址 r5存的是__arch_info_begin 的地址 r6存的是__arch_info_end 的地址 __arch_info_begin 和 __arch_info_end是在 arch/arm/kernel/vmlinux.lds.S中: 00034: __arch_info_begin = .; 00035: *(.arch.info.init) 00036: __arch_info_end = .; 这里是声明了两个变量:__arch_info_begin 和 __arch_info_end,其中等号后面的"."是location counter(详细内容请参考ld.info) 这三行的意思是: __arch_info_begin 的位置上,放置所有文件中的 ".arch.info.init" 段的内容,然后紧接着是 __arch_info_end 的位置. kernel 使用struct machine_desc 来描述 machine type. 在 include/asm-arm/mach/arch.h 中: 00017: struct machine_desc { 00018: /* 00019: * Note! The first four elements are used 00020: * by assembler code in head-armv.S 00021: */ 00022: unsigned int nr; /* architecture number */ 00023: unsigned int phys_io; /* start of physical io */ 00024: unsigned int io_pg_offst; /* byte offset for io 00025: * page tabe entry */ 00026: 00027: const char *name; /* architecture name */ 00028: unsigned long boot_params; /* tagged list */ 00029: 00030: unsigned int video_start; /* start of video RAM */ 00031: unsigned int video_end; /* end of video RAM */ 00032: 00033: unsigned int reserve_lp0 :1; /* never has lp0 */ 00034: unsigned int reserve_lp1 :1; /* never has lp1 */ 00035: unsigned int reserve_lp2 :1; /* never has lp2 */ 00036: unsigned int soft_reboot :1; /* soft reboot */ 00037: void (*fixup)(struct machine_desc *, 00038: struct tag *, char **, 00039: struct meminfo *); 00040: void (*map_io)(void);/* IO mapping function */ 00041: void (*init_irq)(void); 00042: struct sys_timer *timer; /* system tick timer */ 00043: void (*init_machine)(void); 00044: }; 00045: 00046: /* 00047: * Set of macros to define architecture features. This is built into 00048: * a table by the linker. 00049: */ 00050: #define MACHINE_START(_type,_name) / 00051: static const struct machine_desc __mach_desc_##_type / 00052: __attribute_used__ / 00053: __attribute__((__section__(".arch.info.init")) = { / 00054: .nr = MACH_TYPE_##_type, / 00055: .name = _name, 00056: 00057: #define MACHINE_END / 00058: }; 内核中,一般使用宏MACHINE_START来定义machine type. 对于at91, 在 arch/arm/mach-at91rm9200/board-ek.c 中: 00137: MACHINE_START(AT91RM9200EK, "Atmel AT91RM9200-EK" 00138: /* Maintainer: SAN People/Atmel */ 00139: .phys_io = AT91_BASE_SYS, 00140: .io_pg_offst = (AT91_VA_BASE_SYS >> 1 & 0xfffc, 00141: .boot_params = AT91_SDRAM_BASE + 0x100, 00142: .timer = &at91rm9200_timer, 00143: .map_io = ek_map_io, 00144: .init_irq = ek_init_irq, 00145: .init_machine = ek_board_init, 00146: MACHINE_END 197行: r3中存储的是3b处的物理地址,而r4中存储的是3b处的虚拟地址,这里计算处物理地址和虚拟地址的差值,保存到r3中 198行: 将r5存储的虚拟地址(__arch_info_begin)转换成物理地址 199行: 将r6存储的虚拟地址(__arch_info_end)转换成物理地址 200行: MACHINFO_TYPE 在 arch/arm/kernel/asm-offset.c 101行定义, 这里是取 struct machine_desc中的nr(architecture number) 到r3中 201行: 将r3中取到的machine type 和 r1中的 machine type(见前面的"启动条件"进行比较 202行: 如果相同,说明找到了对应的machine type,跳转到207行的2f处,此时r5中存储了对应的struct machine_desc的基地址 203行: (不相同), 取下一个machine_desc的地址 204行: 和r6进行比较,检查是否到了__arch_info_end. 205行: 如果不相同,说明还有machine_desc,返回200行继续查找. 206行: 执行到这里,说明所有的machind_desc都查找完了,并且没有找到匹配的, 将r5设置成0(unknown machine). 207行: 返回

通过前面的两步,我们已经确定了processor type 和 machine type. 此时,一些特定寄存器的值如下所示: r8 = machine info (struct machine_desc的基地址) r9 = cpu id (通过cp15协处理器获得的cpu id) r10 = procinfo (struct proc_info_list的基地址) 创建页表是通过函数 __create_page_tables 来实现的. 这里,我们使用的是arm的L1主页表,L1主页表也称为段页表(section page table) L1 主页表将4 GB 的地址空间分成若干个1 MB的段(section),因此L1页表包含4096个页表项(section entry). 每个页表项是32 bits(4 bytes) 因而L1主页表占用 4096 *4 = 16k的内存空间. 对于ARM926,其L1 section entry的格式为可参考arm926EJS TRM):

(一级描述符的格式 可以参考《ARM体系结构与编程》P180)

下面我们来分析 __create_page_tables 函数: 在 arch/arm/kernel/head.S 中: 00206: .type __create_page_tables, %function 00207: __create_page_tables: 00208: pgtbl r4 @ page table address 00209: 00210: /* 00211: * Clear the 16K level 1 swapper page table 00212: */ 00213: mov r0, r4 00214: mov r3, #0 00215: add r6, r0, #0x4000 00216: 1: str r3, [r0], #4 00217: str r3, [r0], #4 00218: str r3, [r0], #4 00219: str r3, [r0], #4 00220: teq r0, r6 00221: bne 1b 00222: 00223: ldr r7, [r10, #PROCINFO_MM_MMUFLAGS] @ mm_mmuflags 00224: 00225: /* 00226: * Create identity mapping for first MB of kernel to 00227: * cater for the MMU enable. This identity mapping 00228: * will be removed by paging_init(). We use our current program 00229: * counter to determine corresponding section base address. 00230: */ 00231: mov r6, pc, lsr #20 @ start of kernel section 00232: orr r3, r7, r6, lsl #20 @ flags + kernel base 00233: str r3, [r4, r6, lsl #2] @ identity mapping 00234: 00235: /* 00236: * Now setup the pagetables for our kernel direct 00237: * mapped region. 00238: */ 00239: add r0, r4, #(TEXTADDR & 0xff000000) >> 18 @ start of kernel 00240: str r3, [r0, #(TEXTADDR & 0x00f00000) >> 18]! 00241: 00242: ldr r6, =(_end - PAGE_OFFSET - 1) @ r6 = number of sections 00243: mov r6, r6, lsr #20 @ needed for kernel minus 1 00244: 00245: 1: add r3, r3, #1 << 20 00246: str r3, [r0, #4]! 00247: subs r6, r6, #1 00248: bgt 1b 00249: 00250: /* 00251: * Then map first 1MB of ram in case it contains our boot params. 00252: */ 00253: add r0, r4, #PAGE_OFFSET >> 18 00254: orr r6, r7, #PHYS_OFFSET 00255: str r6, [r0] ... 00314: mov pc, lr 00315: .ltorg 206, 207行: 函数声明 208行: 通过宏 pgtbl 将r4设置成页表的基地址(物理地址) 宏pgtbl 在 arch/arm/kernel/head.S 中: 00042: .macro pgtbl, rd 00043: ldr /rd, =(__virt_to_phys(KERNEL_RAM_ADDR - 0x4000)) 00044: .endm 可以看到,页表是位于 KERNEL_RAM_ADDR 下面 16k 的位置 宏 __virt_to_phys 是在incude/asm-arm/memory.h 中: 00125: #ifndef __virt_to_phys 00126: #define __virt_to_phys(x) ((x) - PAGE_OFFSET + PHYS_OFFSET) 00127: #define __phys_to_virt(x) ((x) - PHYS_OFFSET + PAGE_OFFSET) 00128: #endif 下面从213行 - 221行, 是将这16k 的页表清0. 213行: r0 = r4, 将页表基地址存在r0中 214行: 将 r3 置成0 215行: r6 = 页表基地址 + 16k, 可以看到这是页表的尾地址 216 - 221 行: 循环,从 r0 到 r6 将这16k页表用0填充. 223行: 获得proc_info_list的__cpu_mm_mmu_flags的值,并存储到 r7中. (宏PROCINFO_MM_MMUFLAGS是在arch/arm/kernel/asm-offset.c中定义，值为8)（可以参考《嵌入式Linux应用完全开发手册》P118）（r7的值就是设置这个段描述符的权限、域字段，）

在arch/arm/mm/proc-arm926.S 中:

        00464:         .section ".proc.info.init", #alloc, #execinstr

        00465: 

        00466:         .type        __arm926_proc_info,#object

        00467: __arm926_proc_info:

        00468:         .long        0x41069260                        @ ARM926EJ-S (v5TEJ)

        00469:         .long        0xff0ffff0

        00470:         .long   PMD_TYPE_SECT | /

        00471:                 PMD_SECT_BUFFERABLE | /

        00472:                 PMD_SECT_CACHEABLE | /

        00473:                 PMD_BIT4 | /

        00474:                 PMD_SECT_AP_WRITE | /

        00475:                 PMD_SECT_AP_READ

        00476:         .long   PMD_TYPE_SECT | /

        00477:                 PMD_BIT4 | /

        00478:                 PMD_SECT_AP_WRITE | /

        00479:                 PMD_SECT_AP_READ

        00480:         b        __arm926_setup

        00481:         .long        cpu_arch_name

        00482:         .long        cpu_elf_name

        00483:         .long        HWCAP_SWP|HWCAP_HALF|HWCAP_THUMB|HWCAP_FAST_MULT|HWCAP_VFP|HWCAP_EDSP|HWCAP_JAVA

        00484:         .long        cpu_arm926_name

        00485:         .long        arm926_processor_functions

        00486:         .long        v4wbi_tlb_fns

        00487:         .long        v4wb_user_fns

        00488:         .long        arm926_cache_fns

        00489:         .size        __arm926_proc_info, . - __arm926_proc_info

231行: 通过pc值的高12位(右移20位),得到kernel的section,并存储到r6中.因为当前是通过运行时地址得到的kernel的section,因而是物理地址. 232行: r3 = r7 | (r6 << 20); flags + kernel base,得到页表中需要设置的值. 233行: 设置页表: mem[r4 + r6 * 4] = r3 这里,因为页表的每一项是32 bits(4 bytes),所以要乘以4(<<2). 上面这三行,设置了kernel的第一个section(物理地址所在的page entry)的页表项 239, 240行: TEXTADDR是内核的起始虚拟地址(0xc0008000), 这两行是设置kernel起始虚拟地址的页表项(注意,这里设置的页表项和上面的231 - 233行设置的页表项是不同的 ) 执行完后,r0指向kernel的第2个section的虚拟地址所在的页表项. /* TODO: 这两行的code很奇怪,为什么要先取TEXTADDR的高8位(Bit[31:24])0xff000000,然后再取后面的8位 (Bit[23:20])0x00f00000*/ 242行: 这一行计算kernel镜像的大小(bytes). _end 是在vmlinux.lds.S中162行定义的,标记kernel的结束位置(虚拟地址): 00158 .bss : { 00159 __bss_start = .; /* BSS */ 00160 *(.bss) 00161 *(COMMON) 00162 _end = .; 00163 } kernel的size = _end - PAGE_OFFSET -1, 这里 减1的原因是因为 _end 是 location counter,它的地址是kernel镜像后面的一个byte的地址. 243行: 地址右移20位,计算出kernel有多少sections（也就是有多少兆，因为段描述符每个可以映射1MiB的虚拟地址）,并将结果存到r6中 245 - 248行: 这几行用来填充kernel所有section虚拟地址对应的页表项. 253行: 将r0设置为RAM第一兆虚拟地址的页表项地址(page entry) 254行: r7中存储的是mmu flags, 逻辑或上RAM的起始物理地址,得到RAM第一个MB页表项的值. 255行: 设置RAM的第一个MB虚拟地址的页表. 上面这三行是用来设置RAM中第一兆虚拟地址的页表. 之所以要设置这个页表项的原因是RAM的第一兆内存中可能存储着boot params. 这样,kernel所需要的基本的页表我们都设置完了, 如下图所示:

下面是linux-2.6.30.4中的arch/arm/kernel/head.S，代码有一些不同，但是效果一样：

   1: /*

   2:  *  linux/arch/arm/kernel/head.S

   3:  *

   4:  *  Copyright (C) 1994-2002 Russell King

   5:  *  Copyright (c) 2003 ARM Limited

   6:  *  All Rights Reserved

   7:  *

   8:  * This program is free software; you can redistribute it and/or modify

   9:  * it under the terms of the GNU General Public License version 2 as

  10:  * published by the Free Software Foundation.

  11:  *

  12:  *  Kernel startup code for all 32-bit CPUs

  13:  */

  14: #include <linux/linkage.h>

  15: #include <linux/init.h>

  16: 

  17: #include <asm/assembler.h>

  18: #include <asm/domain.h>

  19: #include <asm/ptrace.h>

  20: #include <asm/asm-offsets.h>

  21: #include <asm/memory.h>

  22: #include <asm/thread_info.h>

  23: #include <asm/system.h>

  24: 

  25: #if (PHYS_OFFSET & 0x001fffff)

  26: #error "PHYS_OFFSET must be at an even 2MiB boundary!"

  27: #endif

  28: 

  29: #define KERNEL_RAM_VADDR    (PAGE_OFFSET + TEXT_OFFSET)

  30: #define KERNEL_RAM_PADDR    (PHYS_OFFSET + TEXT_OFFSET)

  31: 

  32: 

  33: /*

  34:  * swapper_pg_dir is the virtual address of the initial page table.

  35:  * We place the page tables 16K below KERNEL_RAM_VADDR.  Therefore, we must

  36:  * make sure that KERNEL_RAM_VADDR is correctly set.  Currently, we expect

  37:  * the least significant 16 bits to be 0x8000, but we could probably

  38:  * relax this restriction to KERNEL_RAM_VADDR >= PAGE_OFFSET + 0x4000.

  39:  */

  40: #if (KERNEL_RAM_VADDR & 0xffff) != 0x8000

  41: #error KERNEL_RAM_VADDR must start at 0xXXXX8000

  42: #endif

  43: 

  44:     .globl    swapper_pg_dir

  45:     .equ    swapper_pg_dir, KERNEL_RAM_VADDR - 0x4000

  46: 

  47:     .macro    pgtbl, rd

  48:     ldr    /rd, =(KERNEL_RAM_PADDR - 0x4000)

  49:     .endm

  50: 

  51: #ifdef CONFIG_XIP_KERNEL

  52: #define KERNEL_START    XIP_VIRT_ADDR(CONFIG_XIP_PHYS_ADDR)

  53: #define KERNEL_END    _edata_loc

  54: #else

  55: #define KERNEL_START    KERNEL_RAM_VADDR

  56: #define KERNEL_END    _end

  57: #endif

  58: 

  59: /*

  60:  * Kernel startup entry point.

  61:  * ---------------------------

  62:  *

  63:  * This is normally called from the decompressor code.  The requirements

  64:  * are: MMU = off, D-cache = off, I-cache = dont care, r0 = 0,

  65:  * r1 = machine nr, r2 = atags pointer.

  66:  *

  67:  * This code is mostly position independent, so if you link the kernel at

  68:  * 0xc0008000, you call this at __pa(0xc0008000).

  69:  *

  70:  * See linux/arch/arm/tools/mach-types for the complete list of machine

  71:  * numbers for r1.

  72:  *

  73:  * We're trying to keep crap to a minimum; DO NOT add any machine specific

  74:  * crap here - that's what the boot loader (or in extreme, well justified

  75:  * circumstances, zImage) is for.

  76:  */

  77:     .section ".text.head", "ax"

  78: ENTRY(stext)

  79:     msr    cpsr_c, #PSR_F_BIT | PSR_I_BIT | SVC_MODE @ ensure svc mode

  80:                         @ and irqs disabled

  81:     mrc    p15, 0, r9, c0, c0        @ get processor id

  82:     bl    __lookup_processor_type        @ r5=procinfo r9=cpuid

  83:     movs    r10, r5                @ invalid processor (r5=0)?

  84:     beq    __error_p            @ yes, error 'p'

  85:     bl    __lookup_machine_type        @ r5=machinfo

  86:     movs    r8, r5                @ invalid machine (r5=0)?

  87:     beq    __error_a            @ yes, error 'a'

  88:     bl    __vet_atags

  89:     bl    __create_page_tables

  90: 

  91:     /*

  92:      * The following calls CPU specific code in a position independent

  93:      * manner.  See arch/arm/mm/proc-*.S for details.  r10 = base of

  94:      * xxx_proc_info structure selected by __lookup_machine_type

  95:      * above.  On return, the CPU will be ready for the MMU to be

  96:      * turned on, and r0 will hold the CPU control register value.

  97:      */

  98:     ldr    r13, __switch_data        @ address to jump to after

  99:                         @ mmu has been enabled

 100:     adr    lr, __enable_mmu        @ return (PIC) address

 101:     add    pc, r10, #PROCINFO_INITFUNC

 102: ENDPROC(stext)

 103: 

 104: #if defined(CONFIG_SMP)

 105: ENTRY(secondary_startup)

 106:     /*

 107:      * Common entry point for secondary CPUs.

 108:      *

 109:      * Ensure that we're in SVC mode, and IRQs are disabled.  Lookup

 110:      * the processor type - there is no need to check the machine type

 111:      * as it has already been validated by the primary processor.

 112:      */

 113:     msr    cpsr_c, #PSR_F_BIT | PSR_I_BIT | SVC_MODE

 114:     mrc    p15, 0, r9, c0, c0        @ get processor id

 115:     bl    __lookup_processor_type

 116:     movs    r10, r5                @ invalid processor?

 117:     moveq    r0, #'p'            @ yes, error 'p'

 118:     beq    __error

 119: 

 120:     /*

 121:      * Use the page tables supplied from  __cpu_up.

 122:      */

 123:     adr    r4, __secondary_data

 124:     ldmia    r4, {r5, r7, r13}        @ address to jump to after

 125:     sub    r4, r4, r5            @ mmu has been enabled

 126:     ldr    r4, [r7, r4]            @ get secondary_data.pgdir

 127:     adr    lr, __enable_mmu        @ return address

 128:     add    pc, r10, #PROCINFO_INITFUNC    @ initialise processor

 129:                         @ (return control reg)

 130: ENDPROC(secondary_startup)

 131: 

 132:     /*

 133:      * r6  = &secondary_data

 134:      */

 135: ENTRY(__secondary_switched)

 136:     ldr    sp, [r7, #4]            @ get secondary_data.stack

 137:     mov    fp, #0

 138:     b    secondary_start_kernel

 139: ENDPROC(__secondary_switched)

 140: 

 141:     .type    __secondary_data, %object

 142: __secondary_data:

 143:     .long    .

 144:     .long    secondary_data

 145:     .long    __secondary_switched

 146: #endif /* defined(CONFIG_SMP) */

 147: 

 148: 

 149: 

 150: /*

 151:  * Setup common bits before finally enabling the MMU.  Essentially

 152:  * this is just loading the page table pointer and domain access

 153:  * registers.

 154:  */

 155: __enable_mmu:

 156: #ifdef CONFIG_ALIGNMENT_TRAP

 157:     orr    r0, r0, #CR_A

 158: #else

 159:     bic    r0, r0, #CR_A

 160: #endif

 161: #ifdef CONFIG_CPU_DCACHE_DISABLE

 162:     bic    r0, r0, #CR_C

 163: #endif

 164: #ifdef CONFIG_CPU_BPREDICT_DISABLE

 165:     bic    r0, r0, #CR_Z

 166: #endif

 167: #ifdef CONFIG_CPU_ICACHE_DISABLE

 168:     bic    r0, r0, #CR_I

 169: #endif

 170:     mov    r5, #(domain_val(DOMAIN_USER, DOMAIN_MANAGER) | /

 171:               domain_val(DOMAIN_KERNEL, DOMAIN_MANAGER) | /

 172:               domain_val(DOMAIN_TABLE, DOMAIN_MANAGER) | /

 173:               domain_val(DOMAIN_IO, DOMAIN_CLIENT))

 174:     mcr    p15, 0, r5, c3, c0, 0        @ load domain access register

 175:     mcr    p15, 0, r4, c2, c0, 0        @ load page table pointer

 176:     b    __turn_mmu_on

 177: ENDPROC(__enable_mmu)

 178: 

 179: /*

 180:  * Enable the MMU.  This completely changes the structure of the visible

 181:  * memory space.  You will not be able to trace execution through this.

 182:  * If you have an enquiry about this, *please* check the linux-arm-kernel