本文分析基于linux2.4.19 source,pxa 270 cpu.
ARM linux内核启动时,通过start_kernel()->trap_init()的调用关系,初始化内核的中断异常向量表.
/* arch/arm/kernel/traps.c */
void __init trap_init(void)
{
extern void __trap_init(unsigned long);
unsigned long base = vectors_base();
__trap_init(base);
if (base != 0)
oopsprintk(KERN_DEBUG "Relocating machine vectors to 0x%08lx/n", base);
#ifdef CONFIG_CPU_32
modify_domain(DOMAIN_USER, DOMAIN_CLIENT);
#endif
}
vectors_base是一个宏,它的作用是获取ARM异常向量的地址,该宏在include/arch/asm-arm/proc-armv/system.h中定义:
extern unsigned long cr_no_alignment; /* defined in entry-armv.S */
extern unsigned long cr_alignment; /* defined in entry-armv.S */
#if __LINUX_ARM_ARCH__ >= 4
#define vectors_base() ((cr_alignment & CR_V) ? 0xffff0000 : 0)
#else
#define vectors_base() (0)
#endif
对于ARMv4以下的版本,这个地址固定为0;ARMv4及其以上的版本,ARM异常向量表的地址受协处理器CP15的c1寄存器 (control register)中V位(bit[13])的控制,如果V=1,则异常向量表的地址为0x00000000~0x0000001C;如果V=0,则 为:0xffff0000~0xffff001C。(详情请参考ARM Architecture Reference Manual)
下面分析一下cr_alginment的值是在哪确定的,我们在arch/arm/kernel/entry-armv.S找到cr_alignment的定义:
.globl SYMBOL_NAME(cr_alignment)
.globl SYMBOL_NAME(cr_no_alignment)
SYMBOL_NAME(cr_alignment):
.space 4
SYMBOL_NAME(cr_no_alignment):
.space 4
分析过head-armv.S文件的朋友都会知道,head-armv.S是非压缩内核的入口:
1 .section ".text.init",#alloc,#execinstr
2 .type stext, #function
3ENTRY(stext)
4 mov r12, r0
5
6 mov r0, #F_BIT | I_BIT | MODE_SVC @ make sure svc mode
7 msr cpsr_c, r0 @ and all irqs disabled
8 bl __lookup_processor_type
9 teq r10, #0 @ invalid processor?
10 moveq r0, #'p' @ yes, error 'p'
11 beq __error
12 bl __lookup_architecture_type
13 teq r7, #0 @ invalid architecture?
14 moveq r0, #'a' @ yes, error 'a'
15 beq __error
16 bl __create_page_tables
17 adr lr, __ret @ return address
18 add pc, r10, #12 @ initialise processor
19 @ (return control reg)
20
21 .type __switch_data, %object
22__switch_data: .long __mmap_switched
23 .long SYMBOL_NAME(__bss_start)
24 .long SYMBOL_NAME(_end)
25 .long SYMBOL_NAME(processor_id)
26 .long SYMBOL_NAME(__machine_arch_type)
27 .long SYMBOL_NAME(cr_alignment)
28 .long SYMBOL_NAME(init_task_union)+8192
29
30 .type __ret, %function
31__ret: ldr lr, __switch_data
32 mcr p15, 0, r0, c1, c0
33 mrc p15, 0, r0, c1, c0, 0 @ read it back.
34 mov r0, r0
35 mov r0, r0
36 mov pc, lr
这里我们关心的是从17行开始,17行code处将lr放置为__ret标号处的相对地址,以便将来某处返回时跳转到31行继续运行;
18行,对于我所分析的pxa270平台,它将是跳转到arch/arm/mm/proc-xscale.S中执行__xscale_setup函数, 在__xscale_setup中会读取CP15的control register(c1)的值到r1寄存器,并在r1寄存器中设置相应的标志位(其中包括设置V位=1),但在__xscale_setup中,r1寄存 器并不立即写回到Cp15的control register中,而是在返回后的某个地方,接下来会慢慢分析到。__xscale_setup调用move pc, lr指令返回跳转到31行。
31行,在lr寄存器中放置__switch_data中的数据__mmap_switched,在36行程序会跳转到__mmap_switched处。
32,33行,把r0寄存器中的值写回到cp15的control register(c1)中,再读出来放在r0中。
接下来再来看一下跳转到__mmap_switched处的代码:
40 _mmap_switched:
41 adr r3, __switch_data + 4
42 ldmia r3, {r4, r5, r6, r7, r8, sp}@ r2 = compat
43 @ sp = stack pointer
44
45 mov fp, #0 @ Clear BSS (and zero fp)
46 1: cmp r4, r5
47 strcc fp, [r4],#4
48 bcc 1b
49
50 str r9, [r6] @ Save processor ID
51 str r1, [r7] @ Save machine type
52 bic r2, r0, #2 @ Clear 'A' bit
53 stmia r8, {r0, r2} @ Save control register values
54 b SYMBOL_NAME(start_kernel)
41~42行的结果是:r4=__bss_start,r5=__end,...,r8=cr_alignment,..,这里r8保存的是cr_alignment变量的地址.
到了53行,由于之前r0保存的是cp15的control register(c1)的值,这里把r0的值写入r8指向的地址,即cr_alignment=r0.到此为止,我们就看清楚了cr_alignment的赋值过程。
让我们回到trap_init()函数,经过上面的分析,我们知道vectors_base返回0xffff0000。函数__trap_init由汇编代码编写,在arch/arm/kernel/entry-arm.S:
.align 5
__stubs_start:
vector_IRQ:
...
vector_data:
....
vector_prefetch:
...
vector_undefinstr:
...
vector_FIQ: disable_fiq
subs pc, lr, #4
vector_addrexcptn:
b vector_addrexcptn
...
__stubs_end:
.equ __real_stubs_start, .LCvectors + 0x200
.LCvectors: swi SYS_ERROR0
b __real_stubs_start + (vector_undefinstr - __stubs_start)
ldr pc, __real_stubs_start + (.LCvswi - __stubs_start)
b __real_stubs_start + (vector_prefetch - __stubs_start)
b __real_stubs_start + (vector_data - __stubs_start)
b __real_stubs_start + (vector_addrexcptn - __stubs_start)
b __real_stubs_start + (vector_IRQ - __stubs_start)
b __real_stubs_start + (vector_FIQ - __stubs_start)
ENTRY(__trap_init)
stmfd sp!, {r4 - r6, lr} /* 压栈,保存数据*/
/* 复制异常向量表(.LCvectors起始的8个地址)到r0指向的地址(异常向量地址),r0就是__trap_init(base)函数调用时传递的参数,不明白的请参考ATPCS*/
adr r1, .LCvectors @ set up the vectors
ldmia r1, {r1, r2, r3, r4, r5, r6, ip, lr}
stmia r0, {r1, r2, r3, r4, r5, r6, ip, lr}
/* 在异常向量地址后的0x200偏移处,放置散转代码,即__stubs_start~__stubs_end之间的各个异常处理代码*/
add r2, r0, #0x200
adr r0, __stubs_start @ copy stubs to 0x200
adr r1, __stubs_end
1: ldr r3, [r0], #4
str r3, [r2], #4
cmp r0, r1
blt 1b
LOADREGS(fd, sp!, {r4 - r6, pc}) /*出栈,恢复数据,函数__trap_init返回*/
__trap_init函数填充后的向量表如下:
虚拟地址 异常 处理代码
0xffff0000 reset swi SYS_ERROR0
0xffff0004 undefined b __real_stubs_start + (vector_undefinstr - __stubs_start)
0xffff0008 软件中断 ldr pc, __real_stubs_start + (.LCvswi - __stubs_start)
0xffff000c 取指令异常 b __real_stubs_start + (vector_prefetch - __stubs_start)
0xffff0010 数据异常 b __real_stubs_start + (vector_data - __stubs_start)
0xffff0014 reserved b __real_stubs_start + (vector_addrexcptn - __stubs_start)
0xffff0018 irq b __real_stubs_start + (vector_IRQ - __stubs_start)
0xffff001c fiq b __real_stubs_start + (vector_FIQ - __stubs_start)
当有异常发生时,处理器会跳转到对应的0xffff0000起始的向量处取指令,然后,通过b指令散转到异常处理代码.因为ARM中b指令是相对跳转,而 且只有+/-32MB的寻址范围,所以把__stubs_start~__stubs_end之间的异常处理代码复制到了0xffff0200起始处.这 里可直接用b指令跳转过去,这样比使用绝对跳转(ldr)效率高。
-------------------------参考资料--------------------
1, 刘淼,嵌入式系统接口设计与Linux驱动程序开发,北京航天航空大学出版社,2006.
2, ARM Architecture Reference Manual, ARM limited,2000.
原文地址 http://hi.baidu.com/_%C5%CE%C8%FD%C4%EA_/blog/item/cceac934d9dd95b1d1a2d370.html