armv8(aarch64)linux内核中flush_dcache_all函数详细分析-LMLPHP

/*

*  __flush_dcache_all()

*  Flush the wholeD-cache.

* Corrupted registers: x0-x7, x9-x11

*/

ENTRY(__flush_dcache_all)

//保证之前的访存指令的顺序

dsb sy

//读cache level id register

mrs x0, clidr_el1           // read clidr

//取bits[26:24](Level of Coherency for the cache hierarchy.)

//需要遵循cache一致性的cache层级(例如有3级cache,但2级需要做一致性)

and x3, x0, #0x7000000      // extract loc from clidr

//逻辑右移23位,把bits[26:24]放到bits[2:0]

lsr x3, x3, #23         // left align loc bit field

//如果需要做cache一致性的层级为0,则不需要flush,跳转到finished标记处。

cbz x3, finished            // if loc is 0, then no need toclean

//x10存放cache级,从level0 cache开始做flush

//以下三个循环loop3是set/way(x9),

//loop2是index(x7),loop1是cache level(x10)

mov x10, #0             // start clean at cache level 0

loop1:

//x10+2后右移一位正好等于1,再加上x10本身正好等于3

//每执行一次loop1,x2+3*执行次数,目的在于把x0(clidr_el1)右移3位,

//取下一个cache的ctype type fields字段,clidr_el1的格式见《ARMv8 ARM》

add x2, x10, x10, lsr #1        /

//x0逻辑右移x2位,给x1,提取cache类型放到x1中,x0中存放:clidr_el1

lsr x1, x0, x2

//掩掉高位,只取当前cache类型

and x1, x1, #7

/* 判断当前cache是什么类型:

* 000  No cache.

* 001  Instruction cache only.

* 010  Data cache only.

* 011  Separate instruction and data caches.

* 100  Unified cache.

*/

//小于2说明data cache不存在或者只有icache,

//跳转skip执行,大于等于2继续执行

cmp x1, #2

b.lt   skip

/*

*  Save/disableand restore interrupts.

* .macro save_and_disable_irqs, olddaif

* mrs \olddaif,daif

* disable_irq

* .endm

*/

//保存daif到x9寄存器中,关闭中断

save_and_disable_irqs x9        // make CSSELR and CCSIDR access atomic

//选择当前cache级进行操作,csselr_el1寄存器bit[3:1]选择要操作的cache级

//第一次执行时x10=0,选择level 0级cache

msr csselr_el1,x10

//isb用于同步新的cssr和csidr寄存器

isb

//因为执行了“msr csselr_el1,x10”,所以要重新读取ccsidr_el1

mrs x1, ccsidr_el1          // read the new ccsidr

/*

* .macro  restore_irqs, olddaif

* msrdaif, \olddaif

. * endm

*/

restore_irqs x9

//x1存储ccsidr_el1内容,低三位是(Log2(Number of bytes in cache line)) – 4

后x2=(Log2(Numberof bytes in cache line))

and x2, x1, #7          // extract the length of the cachelines

add x2, x2, #4          // add 4 (line length offset)

mov x4, #0x3ff

//逻辑右移3位,提取bits[12:3](Associativityof cache) – 1,

//x4存储cache的way数

and x4, x4, x1, lsr #3     // find maximum number on the way size

//计算x4前面0的个数,存到x5

clz x5, x4              // find bit position of way sizeincrement

//提取bits[27:13]位:(Number of sets in cache) - 1

mov x7, #0x7fff

//x7中存储cache中的set数

and x7, x7, x1, lsr #13     // extract max number of the index size

loop2:

//把x4值备份

mov x9, x4              // create working copy of max waysize

loop3:

//把需要操作哪个way存储到x6

lsl x6, x9, x5

//确定操作哪一级的哪个way(x10指定操作哪一级cache)

orr x11, x10, x6            // factor way and cache number intox11

//确定操作哪个set

lsl x6, x7, x2

orr x11, x11, x6            // factor index number into x11

//x11中存储了哪一级cache(10),哪一路cache(x9),哪个set(x7)

dc  cisw, x11           // clean & invalidate by set/way

//way数-1

subs   x9, x9, #1          // decrementthe way

b.ge   loop3

subs   x7, x7, #1          // decrementthe index

b.ge   loop2

skip:

add x10, x10, #2            // increment cache number,

//为什么加2不是1?见loop1标号处解释

cmp x3, x10

b.gt   loop1

finished:

mov x10, #0             // swith back to cache level 0

msr csselr_el1, x10         // select current cache level incsselr

dsb sy

isb

ret

ENDPROC(__flush_dcache_all)

如果你对此有疑问,欢迎留言讨论。

05-02 23:31