本文為課程實驗,based on mengning mykernel 2.0,后續(xù)進階見調試跟蹤Linux內核的啟動過程
1. 準備工作
使用cat /proc/version使用查看系統(tǒng)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. 配置并編譯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) # 編譯的時間比較久
sudo apt install qemu # install QEMU
qemu-system-x86_64 -kernel arch/x86/boot/bzImage

從qemu窗口中您可以看到my_start_kernel在執(zhí)行,同時my_timer_handler時鐘中斷處理程序周期性執(zhí)行。
可能報出的錯誤及解決方法:
執(zhí)行wget https://raw.github.com/mengning/mykernel/master/mykernel-2.0_for_linux-5.4.34.patch時可能會報以下錯誤:
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/查詢raw.githubusercontent.com的真實IP,然后修改hosts,在/etc/hosts/中綁定查到的host,例如
sudo vim /etc/hosts
#綁定host
199.232.28.133 raw.githubusercontent.com
修改完hosts如果還不能成功運行,執(zhí)行以下命令,取消證書的檢查即可
wget --no-check-certificate https://raw.github.com/mengning/mykernel/master/mykernel-2.0_for_linux-5.4.34.patch
3. 添加簡單時間片輪轉調度模塊
在mymain.c基礎上繼續(xù)寫進程描述PCB和進程鏈表管理等代碼,在myinterrupt.c的基礎上完成進程切換代碼。首先在mykernel目錄下增加一個mypcb.h 頭文件,用來定義進程控制塊(Process Control Block),也就是進程結構體的定義,在Linux內核中是struct tast_struct結構體。
/*
* 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);
對mymain.c進行修改,這里是mykernel內核代碼的入口,負責初始化內核的各個組成部分。在Linux內核源代碼中,實際的內核入口是init/main.c中的start_kernel(void)函數(shù)。
/*
* 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函數(shù),用來作為進程的代碼模擬一個個進程,只是我們這里采用的是進程運行完一個時間片后主動讓出CPU的方式,沒有采用中斷的時機完成進程切換,因為中斷機制實現(xiàn)起來較為復雜,等后續(xù)部分再逐漸深入。
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);
}
}
}
進程運行過程中是怎么知道時間片消耗完了呢?這就需要時鐘中斷處理過程中記錄時間片。對myinterrupt.c中修改my_timer_handler用來記錄時間片。
/*
* 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;
}
對myinterrupt.c進行修改,主要是增加了進程切換的代碼my_schedule(void)函數(shù),在Linux內核源代碼中對應的是schedule(void)函數(shù)。
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;
}
修改完成之后運行以下命令重新編譯運行:
make clean
make
qemu-system-x86_64 -kernel arch/x86/boot/bzImage

可能報出的錯誤及解決方法:
在修改完代碼后執(zhí)行make會出現(xiàn)以下錯誤:
mykernel/mymain.c: Assembler messages:
mykernel/mymain.c:48: 錯誤: bad register name `%rsp'
mykernel/mymain.c:49: 錯誤: unsupported instruction `push'
mykernel/mymain.c:50: 錯誤: 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
看到錯誤的第一反應該是寄存器的位數(shù)不支持,所以我就把上述代碼中的所有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再重新編譯。結果倒是不報錯了,qemu正常啟動,但是卻一直卡在了Booting from ROM....
最后的做法是改回了原來的代碼,然后使用make clean和兩次make然后又運行成功了?!如果出現(xiàn)這種錯誤你也可以試一試......
4. 代碼分析
4.1 時間片輪轉調度模塊代碼分析
最終修改完成的mypcb.h、mymain.c和myinterrupt.c文件內容如下(使用注釋解釋):
/*
* 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]; //聲明tPCB類型的數(shù)組
tPCB * my_current_task = NULL; //聲明當前task的指針
volatile int my_need_sched = 0; //判斷是否需要調度
void my_process(void);
/** 從my_start_kernel開始執(zhí)行,實際的內核入口是init/main.c中的start_kernel(void)函數(shù) **/
void __init my_start_kernel(void)
{
int pid = 0;
int i;
/* Initialize process 0*/
task[pid].pid = pid; //初始化0號進程
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創(chuàng)建子進程時,也是使用fork復制進程,然后改變一些關鍵信息,比如進程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;
/** 進程運行體:用來作為進程的代碼模擬一個個進程 **/
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) //判斷是否需要調度
{
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
*
*/
/** 定義進程的最大數(shù)量和內核棧的大小 **/
#define MAX_TASK_NUM 4
#define KERNEL_STACK_SIZE 1024*2
/* CPU-specific state of this task */
//存儲ip,sp
struct Thread {
unsigned long ip;
unsigned long sp;
};
typedef struct PCB{
int pid; //進程的id
volatile long state; /* 定義進程狀態(tài): -1 unrunnable, 0 runnable, >0 stopped */
unsigned long stack[KERNEL_STACK_SIZE]; //進程堆棧
/* CPU-specific state of this task */
struct Thread thread;
unsigned long task_entry; //進程入口,
struct PCB *next; //指向下一個進程PCB
}tPCB;
//調度函數(shù)
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:任務隊列
* my_current_task:當前運行的進程
* my_need_sched:調度標示
* time_count:計數(shù)器
*/
extern tPCB task[MAX_TASK_NUM];
extern tPCB * my_current_task;
extern volatile int my_need_sched;
volatile int time_count = 0;
/* 模擬時鐘中斷 */
void my_timer_handler(void)
{
if(time_count%1000 == 0 && my_need_sched != 1) //控制時間片的大小,設置調度的標志
{
printk(KERN_NOTICE ">>>my_timer_handler here<<<\n");
my_need_sched = 1;
}
time_count ++ ;
return;
}
/* 進程切換 */
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 根據(jù)下一個進程的狀態(tài)來判斷是否切換*/
{
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指接下來的標號為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 操作系統(tǒng)內核核心功能及運行工作機制分析
4.2.0 準備工作:內嵌匯編語法

============================ 示例 ==============================
/* 下面的%0和%1代表第一個參數(shù)和第二個參數(shù),其index是從輸出部分算起,到輸入部分結束
"=m" 表示內存變量只寫
"r"表示將輸入變量放入通用寄存器
%%表示轉移字符
$0表示立即數(shù)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");
// 輸出為 0, 1
printf("%d, %d \n",temp, output);
return 0;
}
4.2.1 啟動執(zhí)行第一個進程的關鍵匯編代碼分析
asm volatile(
"movq %1,%%rsp\n\t" /* 將進程原堆棧棧頂?shù)牡刂反嫒隦SP寄存器 */
"pushq %1\n\t" /* 將當前RBP寄存器值壓棧 */
"pushq %0\n\t" /* 將當前進程的RIP壓棧 */
"ret\n\t" /* ret命令正好可以讓壓棧的進程RIP保存到RIP寄存器中 */
:
: "c" (task[pid].thread.ip),"d" (task[pid].thread.sp)
);
由于啟動的第一個進程是我們自己初始化好的0號進程,所以上面代碼的task[pid].thread.ip和 task[pid].thread.sp分別為0號進程ip和sp。
-
movq %1,%%rsp:將RSP寄存器指向進程0的堆棧棧底 -
pushq %1:本來應該壓棧當前進程的RBP,因為是空棧,所以RSP與RBP相同,這里簡化起見,直接使用進程的堆棧棧頂?shù)闹祎ask[pid].thread.sp,之后RSP = RSP - 8(堆棧地址空間從高到低,位數(shù)為64位) -
pushq %0:將當前進程的RIP(這里是初始化的值my_process(void)函數(shù)的位置)入棧,RSP = RSP - 8 -
ret:將棧頂位置的task[0].thread.ip,也就是my_process(void)函數(shù)的地址放入RIP寄存器中,RSP = RSP + 8,修改IP的內容,從而實現(xiàn)近轉移。
4.2.2 進程上下文切換的關鍵代碼分析
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)
);
為了簡便,假設系統(tǒng)只有兩個進程,分別是進程0和進程1。進程0由內核啟動時初始化執(zhí)行,然后需要進程調度和進程切換,然后開始執(zhí)行進程1。進程切換過程中進程0和進程1的堆棧和相關寄存器的變化過程大致如下:
-
pushq %%rbp:保存0號進程的rbp,rsp = rsp - 8(x86向下增長) -
movq %%rsp,%0:把0號進程的rsp保存在prev->thread.sp變量中 -
movq %2,%%rsp:rsp指向1號進程的棧頂 -
movq $1f,%1:其中的$1f是magic number,指的是下面的1:的地址,這句的作用是prev->thread.ip = $1f -
pushq %3:把即將執(zhí)行的next進程的指令地址next->thread.ip入棧,這時的next->thread.ip可能是進程1的起點my_process(void)函數(shù),也可能是$1f(標號1)。第一次被執(zhí)行從頭開始為進程1的起點my_process(void)函數(shù),其余的情況均為$1f(標號1),因為next進程如果之前運行過那么它就一定曾經也作為prev進程被進程切換過 -
ret:就是將壓入棧中的next->thread.ip放入RIP寄存器,為什么不直接放入RIP寄存器呢?因為程序不能直接使用RIP寄存器,只能通過call、ret等指令間接改變RIP寄存器。使用ret從而實現(xiàn)近轉移。 -
1::標號1是一個特殊的地址位置,該位置的地址是$1f。 -
popq %%rbp:像這段代碼的第一句pushq %%rbp一樣,進程切換時會把當前棧的基址RBP保存在棧頂。所以本條命令的作用就是將1號進程堆?;刂窂亩褩V谢謴偷絉BP寄存器中,從而完成進程的切換,即RBP和RSP都指向了進程1的堆棧。
參考文章: