这是ARM-Linux运行的第一个文件，这些代码是一个比较独立的代码包裹器。其作用就是解压Linux内核，并将PC指针跳到内核（vmlinux）的第一条指令。
Bootloader中传入到Linux中的参数总共有三个，Linux中用到的是第二个和第三个。第二个参数是architecture id，第三个是taglist的地址。Architecture id的arm芯片在Linux中一定要唯一。Taglist是bootload向Linux传入的参数列表（详细的解释请参考《booting arm linux.pdf》）。
//程序的入口点
.section ".start", #alloc, #execinstr
/*
* sort out different calling conventions
*/
.align
start:
.type start,#function
.rept 8//重复8次下面的指令，也就是空出中断向量表的位置
mov r0, r0//就是nop指令
.endr
b 1f
.word 0x016f2818 @ Magic numbers to help the loader
.word start @ absolute load/run zImage address
.word _edata @ zImage end address
1: mov r7, r1 @ save architecture ID
mov r8, r2 @ save atags pointer
#ifndef __ARM_ARCH_2__
/*
* Booting from Angel - need to enter SVC mode and disable
* FIQs/IRQs (numeric definitions from angel arm.h source).
* We only do this if we were in user mode on entry.
*/
mrs r2, cpsr @ get current mode
tst r2, #3 @ not user?
bne not_angel
mov r0, #0x17 @ angel_SWIreason_EnterSVC
swi 0x123456 @ angel_SWI_ARM
not_angel:
mrs r2, cpsr @ turn off interrupts to
orr r2, r2, #0xc0 @ prevent angel from running
msr cpsr_c, r2
#else
teqp pc, #0x0c000003 @ turn off interrupts
#endif
一定要保证当前运行在SVC模式下，否则会跳到swi里面去（为什么？我不清楚，而且我没有处理过这个swi）。然后再关闭irq和fiq。
/*
* Note that some cache flushing and other stuff may
* be needed here - is there an Angel SWI call for this?
*/
/*
* some architecture specific code can be inserted
* by the linker here, but it should preserve r7, r8, and r9.
*/
读入地址表。因为我们的代码可以在任何地址执行，也就是位置无关代码（PIC），所以我们需要加上一个偏移量。下面有每一个列表项的具体意义。
GOT表的初值是连接器指定的，当时程序并不知道代码在哪个地址执行。如果当前运行的地址已经和表上的地址不一样，还要修正GOT表。
.text
adr r0, LC0 //此时，r0 = 你下载的zImage的地址 + LC0 的偏移
ldmia r0, {r1, r2, r3, r4, r5, r6, ip, sp}
subs r0, r0, r1 @ calculate the delta offset //此时r0 = 你下载的zImage的地址，r1就是LC0的偏移
@ if delta is zero, we are
beq not_relocated @ running at the address we
@ were linked at.
/*
* We're running at a different address. We need to fix
* up various pointers:
* r5 - zImage base address
* r6 - GOT start
* ip - GOT end
*/
add r5, r5, r0
add r6, r6, r0
add ip, ip, r0
/*
* If we're running fully PIC === CONFIG_ZBOOT_ROM = n,
* we need to fix up pointers into the BSS region.
* r2 - BSS start
* r3 - BSS end
* sp - stack pointer
*/
add r2, r2, r0
add r3, r3, r0
add sp, sp, r0
修改GOT（全局偏移表）表。根据当前的运行地址，修正该表。
/*
* Relocate all entries in the GOT table.
*/
1: ldr r1, [r6, #0] @ relocate entries in the GOT
add r1, r1, r0 @ table. This fixes up the
str r1, [r6], #4 @ C references.
cmp r6, ip
blo 1b
清BSS段，所有的arm程序都需要做这些的。
not_relocated: mov r0, #0
1: str r0, [r2], #4 @ clear bss
str r0, [r2], #4
str r0, [r2], #4
str r0, [r2], #4
cmp r2, r3
blo 1b
正如下面的注释所说，C环境我们已经设置好了。下面我们要打开cache和mmu。为什么要这样做呢？这只是一个解压程序呀？为了速度。那为什么要开mmu呢，而且只是做一个平板式的映射？还是为了速度。如果不开mmu的话，就只能打开icache。因为不开mmu的话就无法实现内存管理，而io区是决不能开dcache的。
/*
* The C runtime environment should now be setup
* sufficiently. Turn the cache on, set up some
* pointers, and start decompressing.
*/
bl cache_on
是不是要跟读进去呢？对于只是对流程感兴趣的人只是知道打开cache就行了。不过跟进去是很有乐趣的，这就是为什么虽然Linux如此庞大，但仍有人会孜孜不倦的研究它的每一行代码的原因吧。反过来说，对于Linux内核的整体把握更加重要，要不然就成盲人摸象了。还有，想做ARM高手的人可以读Linux下的每一个汇编文件，因为Linux内核用ARM的东西还是比较全的。
mov r1, sp @ malloc space above stack
add r2, sp, #0x10000 @ 64k max
对下面这些地址的理解其实还是很麻烦，但有篇文档写得很清楚《About TEXTADDR, ZTEXTADDR, PAGE_OFFSET etc...》。下面程序的意义就是保证解压地址和当前程序的地址不重叠。上面分配了64KB的空间来做解压时的数据缓存。
/*
* Check to see if we will overwrite ourselves.
* r4 = final kernel address//内核执行的最终实地址
* r5 = start of this image//该程序的首地址
* r2 = end of malloc space (and therefore this image)
* We basically want:
* r4 >= r2 -> OK
* r4 + image length <= r5 -> OK
*/
cmp r4, r2
bhs wont_overwrite
add r0, r4, #4096*1024 @ 4MB largest kernel size
cmp r0, r5
bls wont_overwrite
如果空间不够了，只好解压到缓冲区地址后面。调用decompress_kernel进行解压缩，这段代码是用c实现的，和架构无关。
mov r5, r2 @ decompress after malloc space
mov r0, r5
mov r3, r7
bl decompress_kernel
完成了解压缩之后，由于空间不够，内核也没有解压到正确的地址，必须通过代码搬移来搬到指定的地址。搬运过程中有可能会覆盖掉现在运行的这段代码，所以必须将有可能会执行到的代码搬运到安全的地方，这里用的是解压缩了的代码的后面。
add r0, r0, #127
bic r0, r0, #127 @ align the kernel length
/*
* r0 = decompressed kernel length
* r1-r3 = unused
* r4 = kernel execution address
* r5 = decompressed kernel start
* r6 = processor ID
* r7 = architecture ID
* r8 = atags pointer
* r9-r14 = corrupted
*/
add r1, r5, r0 @ end of decompressed kernel
adr r2, reloc_start
ldr r3, LC1
add r3, r2, r3
1: ldmia r2!, {r9 - r14} @ copy relocation code
stmia r1!, {r9 - r14}
ldmia r2!, {r9 - r14}
stmia r1!, {r9 - r14}
cmp r2, r3
blo 1b
bl cache_clean_flush//因为有代码搬移，所以必须先清理（clean）清除（flush）cache。
add pc, r5, r0 @ call relocation code
decompress_kernel共有4个参数，解压的内核地址、缓存区首地址、缓存区尾地址、和芯片ID，返回解压缩代码的长度。
/*
* We're not in danger of overwriting ourselves. Do this the simple way.
*
* r4 = kernel execution address
* r7 = architecture ID
*/
wont_overwrite: mov r0, r4
mov r3, r7
bl decompress_kernel
b call_kernel
针对于不会出现代码覆盖的情况，就简单了。直接解压缩内核并且跳转到首地址运行。call_kernel这个函数我们会在下面分析它。
.type LC0, #object
LC0: .word LC0 @ r1
.word __bss_start @ r2
.word _end @ r3
.word zreladdr @ r4
.word _start @ r5
.word _got_start @ r6
.word _got_end @ ip
.word user_stack+4096 @ sp
LC1: .word reloc_end - reloc_start
.size LC0, . - LC0
上面这个就是刚才我们说过的地址表，里面有几个符号的地址定义。LC0是在这里定义的。Zreladdr是在当前目录下的Makfile里定义的。其他的符号是在lds里定义的。
下面我们来分析一下有关cache和mmu的代码。通过这些代码我们可以看到Linux的高手们是如何通过汇编来实现各个ARM处理器的识别，以达到通用的目的。
/*
* Turn on the cache. We need to setup some page tables so that we
* can have both the I and D caches on.
*
* We place the page tables 16k down from the kernel execution address,
* and we hope that nothing else is using it. If we're using it, we
* will go pop!
*
* On entry,
* r4 = kernel execution address
* r6 = processor ID
* r7 = architecture number
* r8 = atags pointer
* r9 = run-time address of "start" (???)
* On exit,
* r1, r2, r3, r9, r10, r12 corrupted
* This routine must preserve:
* r4, r5, r6, r7, r8
*/
.align 5
cache_on: mov r3, #8 @ cache_on function
b call_cache_fn
这里涉及到了很多MMU、cache、writebuffer、TLB的操作和协处理器的编程。具体编程的东西，我就不想多说了，可以对这ARM的手册逐行的理解。至于为什么要这样做，熟悉了他们的工作原理后也就不难理解了（《ARM嵌入式系统开发》这本书就有个比较好的说明）。因为这里包含了太多的代码搬运、解压等费时的操作，所以打开cache是有必要的。由于要用到数据cache所以需要对mmu进行配置。为了简单这里制作了一级映射，而且是物理地址和虚拟地址相同的1:1映射。
__setup_mmu: sub r3, r4, #16384 @ Page directory size
bic r3, r3, #0xff @ Align the pointer
bic r3, r3, #0x3f00
/*
* Initialise the page tables, turning on the cacheable and bufferable
* bits for the RAM area only.
*/
mov r0, r3
mov r9, r0, lsr #18
mov r9, r9, lsl #18 @ start of RAM
add r10, r9, #0x10000000 @ a reasonable RAM size
mov r1, #0x12
orr r1, r1, #3 << 10
add r2, r3, #16384
1: cmp r1, r9 @ if virt > start of RAM
orrhs r1, r1, #0x0c @ set cacheable, bufferable
cmp r1, r10 @ if virt > end of RAM
bichs r1, r1, #0x0c @ clear cacheable, bufferable
str r1, [r0], #4 @ 1:1 mapping
add r1, r1, #1048576
teq r0, r2
bne 1b
参考下面的注释，如果当前在flash中运行，我们再映射2MB。就算是当前在RAM中执行其实也没关系，只不过是做了重复工作。
/*
* If ever we are running from Flash, then we surely want the cache
* to be enabled also for our execution instance... We map 2MB of it
* so there is no map overlap problem for up to 1 MB compressed kernel.
* If the execution is in RAM then we would only be duplicating the above.
*/
mov r1, #0x1e
orr r1, r1, #3 << 10
mov r2, pc, lsr #20
orr r1, r1, r2, lsl #20
add r0, r3, r2, lsl #2
str r1, [r0], #4
add r1, r1, #1048576
str r1, [r0]
mov pc, lr
__armv4_cache_on:
mov r12, lr
bl __setup_mmu
mov r0, #0
mcr p15, 0, r0, c7, c10, 4 @ drain write buffer
mcr p15, 0, r0, c8, c7, 0 @ flush I,D TLBs
mrc p15, 0, r0, c1, c0, 0 @ read control reg
orr r0, r0, #0x5000 @ I-cache enable, RR cache replacement
orr r0, r0, #0x0030
bl __common_cache_on
mov r0, #0
mcr p15, 0, r0, c8, c7, 0 @ flush I,D TLBs
mov pc, r12
__common_cache_on:
#ifndef DEBUG
orr r0, r0, #0x000d @ Write buffer, mmu
#endif
mov r1, #-1
mcr p15, 0, r3, c2, c0, 0 @ load page table pointer
mcr p15, 0, r1, c3, c0, 0 @ load domain access control
mcr p15, 0, r0, c1, c0, 0 @ load control register
mov pc, lr
/*
* All code following this line is relocatable. It is relocated by
* the above code to the end of the decompressed kernel image and
* executed there. During this time, we have no stacks.
*
* r0 = decompressed kernel length
* r1-r3 = unused
* r4 = kernel execution address
* r5 = decompressed kernel start
* r6 = processor ID
* r7 = architecture ID
* r8 = atags pointer
* r9-r14 = corrupted
*/
下面这段代码是在解压空间不够的情况下需要重新定位的，具体原因上面已经说明。
.align 5
reloc_start: add r9, r5, r0
debug_reloc_start
mov r1, r4
1:
.rept 4
ldmia r5!, {r0, r2, r3, r10 - r14} @ relocate kernel
stmia r1!, {r0, r2, r3, r10 - r14}
.endr
cmp r5, r9
blo 1b
debug_reloc_end
这是最后一个函数了，这个时候一切实质性的工作已经做完。关闭cache，并跳转到真正的内核入口。
call_kernel: bl cache_clean_flush
bl cache_off
mov r0, #0 @ must be zero
mov r1, r7 @ restore architecture number
mov r2, r8 @ restore atags pointer
mov pc, r4 @ call kernel
/*
* Here follow the relocatable cache support functions for the
* various processors. This is a generic hook for locating an
* entry and jumping to an instruction at the specified offset
* from the start of the block. Please note this is all position
* independent code.
*
* r1 = corrupted
* r2 = corrupted
* r3 = block offset
* r6 = corrupted
* r12 = corrupted
*/
通过下面函数我们可以通过proc_types结构体数组我们可以顺利的找到现在的处理器型号，并且会根据R3的偏移量跳转到相应的函数中。里面涉及到协处理器CP15中c0的操作，如果有疑问，可以参考ARM相关手册。
call_cache_fn: adr r12, proc_types
mrc p15, 0, r6, c0, c0 @ get processor ID
1: ldr r1, [r12, #0] @ get value
ldr r2, [r12, #4] @ get mask
eor r1, r1, r6 @ (real ^ match)
tst r1, r2 @ & mask
addeq pc, r12, r3 @ call cache function
add r12, r12, #4*5
b 1b
/*
* Table for cache operations. This is basically:
* - CPU ID match
* - CPU ID mask
* - 'cache on' method instruction
* - 'cache off' method instruction
* - 'cache flush' method instruction
*
* We match an entry using: ((real_id ^ match) & mask) == 0
*
* Writethrough caches generally only need 'on' and 'off'
* methods. Writeback caches _must_ have the flush method
* defined.
*/
.type proc_types,#object
proc_types:
.word 0x41560600 @ ARM6/610
.word 0xffffffe0
b __arm6_cache_off @ works, but slow
b __arm6_cache_off
mov pc, lr
@ b __arm6_cache_on @ untested
@ b __arm6_cache_off
@ b __armv3_cache_flush
.word 0x00000000 @ old ARM ID
.word 0x0000f000
mov pc, lr
mov pc, lr
mov pc, lr
.word 0x41007000 @ ARM7/710
.word 0xfff8fe00
b __arm7_cache_off
b __arm7_cache_off
mov pc, lr
.word 0x41807200 @ ARM720T (writethrough)
.word 0xffffff00
b __armv4_cache_on
b __armv4_cache_off
mov pc, lr
.word 0x00007000 @ ARM7 IDs
.word 0x0000f000
mov pc, lr
mov pc, lr
mov pc, lr
@ Everything from here on will be the new ID system.
.word 0x4401a100 @ sa110 / sa1100
.word 0xffffffe0
b __armv4_cache_on
b __armv4_cache_off
b __armv4_cache_flush
.word 0x6901b110 @ sa1110
.word 0xfffffff0
b __armv4_cache_on
b __armv4_cache_off
b __armv4_cache_flush
@ These match on the architecture ID
.word 0x00020000 @ ARMv4T
.word 0x000f0000
b __armv4_cache_on
b __armv4_cache_off
b __armv4_cache_flush
.word 0x00050000 @ ARMv5TE
.word 0x000f0000
b __armv4_cache_on
b __armv4_cache_off
b __armv4_cache_flush
.word 0x00060000 @ ARMv5TEJ
.word 0x000f0000
b __armv4_cache_on
b __armv4_cache_off
b __armv4_cache_flush
.word 0x00070000 @ ARMv6
.word 0x000f0000
b __armv4_cache_on
b __armv4_cache_off
b __armv6_cache_flush
.word 0 @ unrecognised type
.word 0
mov pc, lr
mov pc, lr
mov pc, lr
.size proc_types, . - proc_types
/*
* Turn off the Cache and MMU. ARMv3 does not support
* reading the control register, but ARMv4 does.
*
* On entry, r6 = processor ID
* On exit, r0, r1, r2, r3, r12 corrupted
* This routine must preserve: r4, r6, r7
*/
.align 5
cache_off: mov r3, #12 @ cache_off function
b call_cache_fn
//代码略
这里分配了4K的空间用来做堆栈。
reloc_end:
.align
.section ".stack", "w"
user_stack: .space 4096
