基于mykernel 2.0编写一个操作系统内核
本文为课程实验，based on mengning mykernel 2.0。
1. 准备工作
使用cat /proc/version使用查看系统版本
Develop your own OS kernel by reusing Linux infrastructure, based on x86-64/Linux Kernel 4.15.0. mykernel 1.0 based on IA32/Linux Kernel 3.9.4.
2. 配置并编译mykernel 2.0
wget https://raw.github.com/mengning/mykernel/master/mykernel-2.0_for_linux-5.4.34.patch
sudo apt install axel
axel -n 20 https://mirrors.edge.kernel.org/pub/linux/kernel/v5.x/linux-5.4.34.tar.xz
xz -d linux-5.4.34.tar.xz
tar -xvf linux-5.4.34.tar # or tar -xvf linux-5.4.34.tar.gz
cd linux-5.4.34
patch -p1 < ../mykernel-2.0_for_linux-5.4.34.patch
sudo apt install build-essential libncurses-dev bison flex libssl-dev libelf-dev
make defconfig # Default configuration is based on 'x86_64_defconfig'
make -j$(nproc) # 编译的时间比较久
sudo apt install qemu # install QEMU
qemu-system-x86_64 -kernel arch/x86/boot/bzImage
从qemu窗口中您可以看到my_start_kernel在执行，同时my_timer_handler时钟中断处理程序周期性执行。
可能报出的错误及解决方法：
执行wget https://raw.github.com/mengning/mykernel/master/mykernel-2.0_for_linux-5.4.34.patch时可能会报以下错误：
Connecting to raw.githubusercontent.com(raw.githubusercontent.com)|151.101.228.133|:443... failed: Connection refused.
原因是GitHub的raw.githubusercontent.com域名解析被污染了，可以在https://www.ipaddress.com/查询raw.githubusercontent.com的真实IP，然后修改hosts，在/etc/hosts/中绑定查到的host，例如
sudo vim /etc/hosts
#绑定host
199.232.28.133 raw.githubusercontent.com
修改完hosts如果还不能成功运行，执行以下命令，取消证书的检查即可
wget --no-check-certificate https://raw.github.com/mengning/mykernel/master/mykernel-2.0_for_linux-5.4.34.patch
3. 添加简单时间片轮转调度模块
在mymain.c基础上继续写进程描述PCB和进程链表管理等代码，在myinterrupt.c的基础上完成进程切换代码。首先在mykernel目录下增加一个mypcb.h 头文件，用来定义进程控制块（Process Control Block），也就是进程结构体的定义，在Linux内核中是struct tast_struct结构体。
/*
* linux/mykernel/mypcb.h
* [https://github.com/mengning/mykernel/blob/master/mypcb.h](https://github.com/mengning/mykernel/blob/master/mypcb.h)
*/
#define MAX_TASK_NUM 4
#define KERNEL_STACK_SIZE 1024*8
/* CPU-specific state of this task */
struct Thread {
unsigned long ip;
unsigned long sp;
};
typedef struct PCB{
int pid;
volatile long state; /* -1 unrunnable, 0 runnable, >0 stopped */
char stack[KERNEL_STACK_SIZE];
/* CPU-specific state of this task */
struct Thread thread;
unsigned long task_entry;
struct PCB *next;
}tPCB;
void my_schedule(void);
对mymain.c进行修改，这里是mykernel内核代码的入口，负责初始化内核的各个组成部分。在Linux内核源代码中，实际的内核入口是init/main.c中的start_kernel(void)函数。
/*
* linux/mykernel/mymain.c
*/
#include "mypcb.h"
tPCB task[MAX_TASK_NUM];
tPCB * my_current_task = NULL;
volatile int my_need_sched = 0;
void my_process(void);
void __init my_start_kernel(void)
{
int pid = 0;
int i;
/* Initialize process 0*/
task[pid].pid = pid;
task[pid].state = 0;/* -1 unrunnable, 0 runnable, >0 stopped */
task[pid].task_entry = task[pid].thread.ip = (unsigned long)my_process;
task[pid].thread.sp = (unsigned long)&task[pid].stack[KERNEL_STACK_SIZE-1];
task[pid].next = &task[pid];
/*fork more process */
for(i=1;i<MAX_TASK_NUM;i++)
{
memcpy(&task[i],&task[0],sizeof(tPCB));
task[i].pid = i;
task[i].state = -1;
task[i].thread.sp = (unsigned long)&task[i].stack[KERNEL_STACK_SIZE-1];
task[i].next = task[i-1].next;
task[i-1].next = &task[i];
}
/* start process 0 by task[0] */
pid = 0;
my_current_task = &task[pid];
asm volatile(
"movq %1,%%rsp\n\t" /* set task[pid].thread.sp to rsp */
"pushq %1\n\t" /* push rbp */
"pushq %0\n\t" /* push task[pid].thread.ip */
"ret\n\t" /* pop task[pid].thread.ip to rip */
:
: "c" (task[pid].thread.ip),"d" (task[pid].thread.sp) /* input c or d mean %ecx/%edx*/
);
}
在mymain.c中添加了my_process函数，用来作为进程的代码模拟一个个进程，只是我们这里采用的是进程运行完一个时间片后主动让出CPU的方式，没有采用中断的时机完成进程切换，因为中断机制实现起来较为复杂，等后续部分再逐渐深入。
void my_process(void)
{
int i = 0;
while(1)
{
i++;
if(i%10000000 == 0)
{
printk(KERN_NOTICE "this is process %d -\n",my_current_task->pid);
if(my_need_sched == 1)
{
my_need_sched = 0;
my_schedule();
}
printk(KERN_NOTICE "this is process %d +\n",my_current_task->pid);
}
}
}
进程运行过程中是怎么知道时间片消耗完了呢？这就需要时钟中断处理过程中记录时间片。对myinterrupt.c中修改my_timer_handler用来记录时间片。
/*
* linux/mykernel/myinterrupt.c
*/
#include "mypcb.h"
extern tPCB task[MAX_TASK_NUM];
extern tPCB * my_current_task;
extern volatile int my_need_sched;
volatile int time_count = 0;
/*
* Called by timer interrupt.
*/
void my_timer_handler(void)
{
if(time_count%1000 == 0 && my_need_sched != 1)
{
printk(KERN_NOTICE ">>>my_timer_handler here<<<\n");
my_need_sched = 1;
}
time_count ++ ;
return;
}
对myinterrupt.c进行修改，主要是增加了进程切换的代码my_schedule(void)函数，在Linux内核源代码中对应的是schedule(void)函数。
void my_schedule(void)
{
tPCB * next;
tPCB * prev;
if(my_current_task == NULL
|| my_current_task->next == NULL)
{
return;
}
printk(KERN_NOTICE ">>>my_schedule<<<\n");
/* schedule */
next = my_current_task->next;
prev = my_current_task;
if(next->state == 0)/* -1 unrunnable, 0 runnable, >0 stopped */
{
my_current_task = next;
printk(KERN_NOTICE ">>>switch %d to %d<<<\n",prev->pid,next->pid);
/* switch to next process */
asm volatile(
"pushq %%rbp\n\t" /* save rbp of prev */
"movq %%rsp,%0\n\t" /* save rsp of prev */
"movq %2,%%rsp\n\t" /* restore rsp of next */
"movq $1f,%1\n\t" /* save rip of prev */
"pushq %3\n\t"
"ret\n\t" /* restore rip of next */
"1:\t" /* next process start here */
"popq %%rbp\n\t"
: "=m" (prev->thread.sp),"=m" (prev->thread.ip)
: "m" (next->thread.sp),"m" (next->thread.ip)
);
}
return;
}
修改完成之后运行以下命令重新编译运行：
make clean
make
qemu-system-x86_64 -kernel arch/x86/boot/bzImage
可能报出的错误及解决方法：
在修改完代码后执行make会出现以下错误：
mykernel/mymain.c: Assembler messages:
mykernel/mymain.c:48: 错误： bad register name `%rsp'
mykernel/mymain.c:49: 错误： unsupported instruction `push'
mykernel/mymain.c:50: 错误： unsupported instruction `push'
scripts/Makefile.build:265: recipe for target 'mykernel/mymain.o' failed
make[1]: *** [mykernel/mymain.o] Error 1
Makefile:1691: recipe for target 'mykernel' failed
make: *** [mykernel] Error 2
看到错误的第一反应该是寄存器的位数不支持，所以我就把上述代码中的所有rsp、rbp寄存器改成了esp、ebp寄存器。pushq和movq都改成了pushl和movl，然后使用make再重新编译。结果倒是不报错了，qemu正常启动，但是却一直卡在了Booting from ROM....
最后的做法是改回了原来的代码，然后使用make clean和两次make然后又运行成功了？！如果出现这种错误你也可以试一试......
4. 代码分析
4.1 时间片轮转调度模块代码分析
最终修改完成的mypcb.h、mymain.c和myinterrupt.c文件内容如下（使用注释解释）：
/*
* linux/mykernel/mymain.c
*
* Kernel internal my_start_kernel
* Change IA32 to x86-64 arch, 2020/4/26
*
* Copyright (C) 2013, 2020 Mengning
*
*/
#include <linux/types.h>
#include <linux/string.h>
#include <linux/ctype.h>
#include <linux/tty.h>
#include <linux/vmalloc.h>
#include "mypcb.h"
tPCB task[MAX_TASK_NUM]; //声明tPCB类型的数组
tPCB * my_current_task = NULL; //声明当前task的指针
volatile int my_need_sched = 0; //判断是否需要调度
void my_process(void);
/** 从my_start_kernel开始执行,实际的内核入口是init/main.c中的start_kernel(void)函数 **/
void __init my_start_kernel(void)
{
int pid = 0;
int i;
/* Initialize process 0*/
task[pid].pid = pid; //初始化0号进程
task[pid].state = 0;/* -1 unrunnable, 0 runnable, >0 stopped */
task[pid].task_entry = task[pid].thread.ip = (unsigned long)my_process; //入口
task[pid].thread.sp = (unsigned long)&task[pid].stack[KERNEL_STACK_SIZE-1];
task[pid].next = &task[pid];
/*
======= fork more process =======
Linux创建子进程时,也是使用fork复制进程,然后改变一些关键信息,比如进程PID等
*/
for(i=1;i<MAX_TASK_NUM;i++)
{
memcpy(&task[i],&task[0],sizeof(tPCB));
task[i].pid = i;
task[i].thread.sp = (unsigned long)(&task[i].stack[KERNEL_STACK_SIZE-1]);
task[i].next = task[i-1].next;
task[i-1].next = &task[i];
}
/* start process 0 by task[0] */
pid = 0;
my_current_task = &task[pid];
asm volatile(
"movq %1,%%rsp\n\t" /* set task[pid].thread.sp to rsp */
"pushq %1\n\t" /* push rbp */
"pushq %0\n\t" /* push task[pid].thread.ip */
"ret\n\t" /* pop task[pid].thread.ip to rip */
:
: "c" (task[pid].thread.ip),"d" (task[pid].thread.sp) /* input c or d mean %ecx/%edx*/
);
}
int i = 0;
/** 进程运行体:用来作为进程的代码模拟一个个进程 **/
void my_process(void)
{
while(1)
{
i++;
if(i%10000000 == 0)
{
printk(KERN_NOTICE "this is process %d -\n",my_current_task->pid);
if(my_need_sched == 1) //判断是否需要调度
{
my_need_sched = 0;
my_schedule();
}
printk(KERN_NOTICE "this is process %d +\n",my_current_task->pid);
}
}
}
/*
* linux/mykernel/mypcb.h
*
* Kernel internal PCB types
*
* Copyright (C) 2013 Mengning
*
*/
/** 定义进程的最大数量和内核栈的大小 **/
#define MAX_TASK_NUM 4
#define KERNEL_STACK_SIZE 1024*2
/* CPU-specific state of this task */
//存储ip,sp
struct Thread {
unsigned long ip;
unsigned long sp;
};
typedef struct PCB{
int pid; //进程的id
volatile long state; /* 定义进程状态: -1 unrunnable, 0 runnable, >0 stopped */
unsigned long stack[KERNEL_STACK_SIZE]; //进程堆栈
/* CPU-specific state of this task */
struct Thread thread;
unsigned long task_entry; //进程入口，
struct PCB *next; //指向下一个进程PCB
}tPCB;
//调度函数
void my_schedule(void);
/*
* linux/mykernel/myinterrupt.c
*
* Kernel internal my_timer_handler
* Change IA32 to x86-64 arch, 2020/4/26
*
* Copyright (C) 2013, 2020 Mengning
*
*/
#include <linux/types.h>
#include <linux/string.h>
#include <linux/ctype.h>
#include <linux/tty.h>
#include <linux/vmalloc.h>
#include "mypcb.h"
/* task:任务队列
* my_current_task:当前运行的进程
* my_need_sched:调度标示
* time_count:计数器
*/
extern tPCB task[MAX_TASK_NUM];
extern tPCB * my_current_task;
extern volatile int my_need_sched;
volatile int time_count = 0;
/* 模拟时钟中断 */
void my_timer_handler(void)
{
if(time_count%1000 == 0 && my_need_sched != 1) //控制时间片的大小，设置调度的标志
{
printk(KERN_NOTICE ">>>my_timer_handler here<<<\n");
my_need_sched = 1;
}
time_count ++ ;
return;
}
/* 进程切换 */
void my_schedule(void)
{
tPCB * next;
tPCB * prev;
if(my_current_task == NULL
|| my_current_task->next == NULL){
return;
}
printk(KERN_NOTICE ">>>my_schedule<<<\n");
/* schedule */
next = my_current_task->next;
prev = my_current_task;
if(next->state == 0)/* -1 unrunnable, 0 runnable, >0 stopped 根据下一个进程的状态来判断是否切换*/
{
my_current_task = next;
printk(KERN_NOTICE ">>>switch %d to %d<<<\n",prev->pid,next->pid);
/* switch to next process */
asm volatile(
"pushq %%rbp\n\t" /* save rbp of prev */
"movq %%rsp,%0\n\t" /* save rsp of prev */
"movq %2,%%rsp\n\t" /* restore rsp of next */
"movq $1f,%1\n\t" /* save rip of prev ,%1f指接下来的标号为1的位置*/
"pushq %3\n\t"
"ret\n\t" /* restore rip of next */
"1:\t" /* next process start here */
"popq %%rbp\n\t"
: "=m" (prev->thread.sp),"=m" (prev->thread.ip)
: "m" (next->thread.sp),"m" (next->thread.ip)
);
}
return;
}
4.2 操作系统内核核心功能及运行工作机制分析
4.2.0 准备工作：内嵌汇编语法
============================ 示例 ==============================
/* 下面的%0和%1代表第一个参数和第二个参数，其index是从输出部分算起，到输入部分结束
"=m" 表示内存变量只写
"r"表示将输入变量放入通用寄存器
%%表示转移字符
$0表示立即数0
*/
int main(void){
int input, output, temp;
input = 1;
__asm__ __volatile__(
"movl $0, %%eax; \n\t" // eax = 0
"movl %%eax, %1; \n\t" // temp = 0
"movl %2, %%eax; \n\t" // eax = input = 1
"movl %%eax, %0; \n\t" // output = eax = 1
:"=m"(output), "=m"(temp) // $0 = output $1 = temp
:"r"(input) // $2 = input
:"eax");
// 输出为 0， 1
printf("%d, %d \n",temp, output);
return 0;
}
4.2.1 启动执行第一个进程的关键汇编代码分析
asm volatile(
"movq %1,%%rsp\n\t" /* 将进程原堆栈栈顶的地址存入RSP寄存器 */
"pushq %1\n\t" /* 将当前RBP寄存器值压栈 */
"pushq %0\n\t" /* 将当前进程的RIP压栈 */
"ret\n\t" /* ret命令正好可以让压栈的进程RIP保存到RIP寄存器中 */
:
: "c" (task[pid].thread.ip),"d" (task[pid].thread.sp)
);
由于启动的第一个进程是我们自己初始化好的0号进程，所以上面代码的task[pid].thread.ip和 task[pid].thread.sp分别为0号进程ip和sp。
movq %1,%%rsp:将RSP寄存器指向进程0的堆栈栈底pushq %1:本来应该压栈当前进程的RBP，因为是空栈，所以RSP与RBP相同，这里简化起见，直接使用进程的堆栈栈顶的值task[pid].thread.sp，之后RSP = RSP - 8(堆栈地址空间从高到低，位数为64位)pushq %0:将当前进程的RIP（这里是初始化的值my_process(void)函数的位置）入栈，RSP = RSP - 8ret:将栈顶位置的task[0].thread.ip，也就是my_process(void)函数的地址放入RIP寄存器中，RSP = RSP + 8，修改IP的内容，从而实现近转移。
4.2.2 进程上下文切换的关键代码分析
printk(KERN_NOTICE ">>>switch %d to %d<<<\n",prev->pid,next->pid);
/* switch to next process */
asm volatile(
"pushq %%rbp\n\t" /* save rbp of prev */
"movq %%rsp,%0\n\t" /* save rsp of prev */
"movq %2,%%rsp\n\t" /* restore rsp of next */
"movq $1f,%1\n\t" /* save rip of prev */
"pushq %3\n\t"
"ret\n\t" /* restore rip of next */
"1:\t" /* next process start here */
"popq %%rbp\n\t"
: "=m" (prev->thread.sp),"=m" (prev->thread.ip)
: "m" (next->thread.sp),"m" (next->thread.ip)
);
为了简便，假设系统只有两个进程，分别是进程0和进程1。进程0由内核启动时初始化执行，然后需要进程调度和进程切换，然后开始执行进程1。进程切换过程中进程0和进程1的堆栈和相关寄存器的变化过程大致如下：
pushq %%rbp：保存0号进程的rbp，rsp = rsp - 8(x86向下增长)movq %%rsp,%0：把0号进程的rsp保存在prev->thread.sp变量中movq %2,%%rsp：rsp指向1号进程的栈顶movq $1f,%1：其中的$1f是magic number，指的是下面的1:的地址，这句的作用是prev->thread.ip = $1fpushq %3：把即将执行的next进程的指令地址next->thread.ip入栈，这时的next->thread.ip可能是进程1的起点my_process(void)函数，也可能是$1f（标号1）。第一次被执行从头开始为进程1的起点my_process(void)函数，其余的情况均为$1f（标号1），因为next进程如果之前运行过那么它就一定曾经也作为prev进程被进程切换过ret：就是将压入栈中的next->thread.ip放入RIP寄存器，为什么不直接放入RIP寄存器呢？因为程序不能直接使用RIP寄存器，只能通过call、ret等指令间接改变RIP寄存器。使用ret从而实现近转移。1:：标号1是一个特殊的地址位置，该位置的地址是$1f。popq %%rbp：像这段代码的第一句pushq %%rbp一样，进程切换时会把当前栈的基址RBP保存在栈顶。所以本条命令的作用就是将1号进程堆栈基地址从堆栈中恢复到RBP寄存器中，从而完成进程的切换，即RBP和RSP都指向了进程1的堆栈。
参考文章：
