linux启动流程(从start_kernel中的rest_init函数到init进程(1))

时间:2021-03-11 15:47:25

linux启动流程(从start_kernel中的rest_init函数到init进程(1))

    在init/main.c文件中有个函数叫start_kernel,它是用来启动内核的主函数,我想大家都知道这个函数啦,而在该函数的最后将调用一个函数叫rest_init(),它执行完,内核就起来了,

    asmlinkage void __init start_kernel(void)

    {

    ......

            /* Do the rest non-__init'ed, we're now alive */

            rest_init();

    }

    现在我们来看一下rest_init()函数,它也在文件init/main.c中,它的前面几行是:

    static void noinline __init_refok rest_init(void)  __releases(kernel_lock)

    {

            int pid;

            kernel_thread(kernel_init, NULL, CLONE_FS | CLONE_SIGHAND);

    其

    中函数kernel_thread定义在文件arch/ia64/kernel/process.c中,用来启动一个内核线程,这里的

    kernel_init是要执行的函数的指针,NULL表示传递给该函数的参数为空,CLONE_FS |

    CLONE_SIGHAND为do_fork产生线程时的标志,表示进程间的fs信息共享,信号处理和块信号共享,然后我就屁颠屁颠地追随到

    kernel_init函数了,现在来瞧瞧它都做了什么好事,它的完整代码如下:

    static int __init kernel_init(void * unused)

    {

            lock_kernel();

            /*

             * init can run on any cpu.

             */

            set_cpus_allowed_ptr(current, CPU_MASK_ALL_PTR);

            /*

             * Tell the world that we're going to be the grim

             * reaper of innocent orphaned children.

             * We don't want people to have to make incorrect

             * assumptions about where in the task array this

             * can be found.

             */

            init_pid_ns.child_reaper = current;

            cad_pid = task_pid(current);

            smp_prepare_cpus(setup_max_cpus);

            do_pre_smp_initcalls();

            smp_init();

            sched_init_smp();

            cpuset_init_smp();

            do_basic_setup();

            /*

             * check if there is an early userspace init.  If yes, let it do all

             * the work

             */

            if (!ramdisk_execute_command)

                    ramdisk_execute_command = "/init";

            if (sys_access((const char __user *) ramdisk_execute_command, 0) != 0) {

                    ramdisk_execute_command = NULL;

                    prepare_namespace();

            }

            /*

             * Ok, we have completed the initial bootup, and

             * we're essentially up and running. Get rid of the

             * initmem segments and start the user-mode stuff..

             */

            init_post();

            return 0;

    }

    在

    kernel_init函数的一开始就调用了lock_kernel()函数,当编译时选上了CONFIG_LOCK_KERNEL,就加上大内核锁,否

    则啥也不做,紧接着就调用了函数set_cpus_allowed_ptr,由于这些函数对init进程的调起还是有影响的,我们还是一个一个来瞧瞧吧,

    不要忘了啥东东最好,

    static inline int set_cpus_allowed_ptr(struct task_struct *p,

                                           const cpumask_t *new_mask)

    {       

            if (!cpu_isset(0, *new_mask))

                    return -EINVAL;

            return 0;

    }  

    这函数其实就调用了cpu_isset宏,定义在文件"include/linux/cpumask.h中,如下:

    #define cpu_isset(cpu, cpumask) test_bit((cpu), (cpumask).bits)

    再来看看set_cpus_allowed_ptr的第二个参数类型吧,也定义在文件include/linux/cpumask.h中,具体如下:

    typedef struct { DECLARE_BITMAP(bits, NR_CPUS); } cpumask_t;

    接着尾随着DECLAR_BITMAP宏到文件include/linux/types.h中,定义如下:

    #define DECLARE_BITMAP(name,bits) \

            unsigned long name[BITS_TO_LONGS(bits)]

    而宏BITS_TO_LONGS定义在文件include/linux/bitops.h中,实现如下:

    #define BITS_TO_LONGS(nr)       DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long))

    DIV_ROUND_UP宏定义在文件include/linux/kernel.h中,BITS_PER_BYTE 宏定义在文件include/linux/bitops.h中,实现如下:

    #define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))

    #define BITS_PER_BYTE           8

    即当NR_CPUS为1~32时,cpumask_t类型为

    struct {

    unsigned long bits[1];

            }

    然后来看看在set_cpus_allowed_ptr(current, CPU_MASK_ALL_PTR);中的 CPU_MASK_ALL_PTR宏,定义在include/linux/cpumask.h中:

    #define CPU_MASK_ALL_PTR        (&CPU_MASK_ALL)

    而CPU_MASK_ALL宏也定义在文件include/linux/cpumask.h中:

    #define CPU_MASK_ALL                                                    \

    (cpumask_t) { {                                                         \

            [BITS_TO_LONGS(NR_CPUS)-1] = CPU_MASK_LAST_WORD                 \

    } }

    NR_CPUS宏定义在文件include/linux/threads.h中,实现如下:

    #ifdef CONFIG_SMP

    #define NR_CPUS         CONFIG_NR_CPUS                   

    #else

    #define NR_CPUS         1

    #endif

    CPU_MASK_LAST_WORD宏定义在文件include/linux/cpumask.h中,实现如下:

    #define CPU_MASK_LAST_WORD BITMAP_LAST_WORD_MASK(NR_CPUS)

    BITMAP_LAST_WORD_MASK(NR_CPUS)宏定义在文件include/linux/bitmap.h中,实现如下:

    #define BITMAP_LAST_WORD_MASK(nbits)                                    \

    (                                                                       \

            ((nbits) % BITS_PER_LONG) ?                                     \

                    (1ULcpu_isset(0,CPU_MASK_ALL_PTR)-->test_bit(0,CPU_MASK_ALL_PTR.bits)

    即当NR_CPUS为n时,就把usigned long bits[0]的第n位置1,应该就如注释所说的,init能运行在任何CPU上吧。

    现

    在kernel_init中的set_cpus_allowed_ptr(current, CPU_MASK_ALL_PTR);

    分析完了,我们接着往下看,首先 init_pid_ns.child_reaper = current;    

       init_pid_ns定义在kernel/pid.c文件中

    struct pid_namespace init_pid_ns = {

            .kref = {

                    .refcount       = ATOMIC_INIT(2),

            },

            .pidmap = {

                    [ 0 ... PIDMAP_ENTRIES-1] = { ATOMIC_INIT(BITS_PER_PAGE), NULL }

            },

            .last_pid = 0,

            .level = 0,

            .child_reaper = &init_task,

    };

    它是一个pid_namespace结构的变量,先来看看pid_namespace的结构,它定义在文件

    include/linux/pid_namespace.h中,具体定义如下:

    struct pid_namespace {

            struct kref kref;

            struct pidmap pidmap[PIDMAP_ENTRIES];

            int last_pid;

            struct task_struct *child_reaper;

            struct kmem_cache *pid_cachep;

            unsigned int level;

            struct pid_namespace *parent;

    #ifdef CONFIG_PROC_FS

            struct vfsmount *proc_mnt;

    #endif

    };

    即把当前进程设为接受其它孤儿进程的进程,然后取得该进程的进程ID,如:

    cad_pid = task_pid(current);

    然后调用 smp_prepare_cpus(setup_max_cpus);如果编译时没有指定CONFIG_SMP,它什么也不做,接着往下看,调用do_pre_smp_initcalls()函数,它定义在init/main.c文件中,实现如下:

    static void __init do_pre_smp_initcalls(void)

    {

            extern int spawn_ksoftirqd(void);

            migration_init();

            spawn_ksoftirqd();

            if (!nosoftlockup)

                    spawn_softlockup_task();

    }

    其中migration_init()定义在文件include/linux/sched.h中,具体实现如下:

    #ifdef CONFIG_SMP

    void migration_init(void);

    #else

    static inline void migration_init(void)

    {

    }

    #endif

    好像什么也没有做,然后是调用spawn_ksoftirqd()函数,定义在文件kernel/softirq.c中,代码如下:

    __init int spawn_ksoftirqd(void)

    {

            void *cpu = (void *)(long)smp_processor_id();

            int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);

            BUG_ON(err == NOTIFY_BAD);

            cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);

            register_cpu_notifier(&cpu_nfb);

            return 0;

    }      

    在该函数中,首先调用smp_processor_id函数获得当前CPU的ID并把它赋值给变量cpu,然后把cpu连同&cpu_nfb,CPU_UP_PREPARE传递给函数cpu_callback,我们先看cpu_callback的前几行:

    static int __cpuinit cpu_callback(struct notifier_block *nfb,

                                      unsigned long action,

                                      void *hcpu)

    {

            int hotcpu = (unsigned long)hcpu;

            struct task_struct *p;

            switch (action) {

            case CPU_UP_PREPARE:

            case CPU_UP_PREPARE_FROZEN:

                    p = kthread_create(ksoftirqd, hcpu, "ksoftirqd/%d", hotcpu);

                    if (IS_ERR(p)) {

                            printk("ksoftirqd for %i failed\n", hotcpu);

                            return NOTIFY_BAD;

                    }

                    kthread_bind(p, hotcpu);

                    per_cpu(ksoftirqd, hotcpu) = p;

                    break;

    从

    上述代码可以看出当action为CPU_PREPARE时,将创建一个内核线程并把它赋值给p,该进程所要运行的函数为ksoftirqd,传递给该函

    数的参数为hcpu,而紧跟其后的”ksoftirqd/%d”,hotcpu为该进程的名字参数,这就是我们在终端用命令ps -ef | grep

    ksoftirqd所看到的线程;如果进程创建失败,打印出错信息,否则把创建的线程p绑定到当前CPU的ID上,这就是

    kthread_bind(p,hotcpu)所做的,接下来的几行为:

    case CPU_ONLINE:

            case CPU_ONLINE_FROZEN:

                    wake_up_process(per_cpu(ksoftirqd, hotcpu));

                    break;

    即

    在spawn_ksoftirqd函数中cpu_callback(&cpu_nfb, CPU_ONLINE,

    cpu);的action为CPU_ONLINE时,将调用wake_up_process函数来唤醒当前CPU上的ksoftirqd进程。最后调用

    register_cpu_notifier(&cpu_nfb);其实也没做什么,只是简单的返回0。返回到

    do_pre_smp_initcalls函数中,接着往下看:

    if (!nosoftlockup)

                    spawn_softlockup_task();

    spawn_softlockup_task()函数定义在文件include/linux/sched.h中,是个空函数。

    到

    现在为止,do_pre_smp_initcalls分析完了,它主要就是创建进程ksoftirqd,把它绑定到当前CPU上,然后再把该进程拷贝给每

    个CPU,并唤醒所有CPU上的进程ksoftirqd,就是当我们执行ps -ef | grep ksoftirqd的时候所看到的:

    root         4     2  0 08:30 ?        00:00:03 [ksoftirqd/0]

    root         7     2  0 08:30 ?        00:00:02 [ksoftirqd/1]

    革命尚未成功,同志仍需努力!接着享受吧,呵呵!

    现在到了kernel_init函数中的smp_init();了

    如果在编译时没有选择CONFIG_SMP,若定义CONFIG_X86_LOCAL_APIC则去调用APIC_init_uniprocessor()函数,否则什么也不做,具体代码定义在文件init/main.c中:

    #ifndef CONFIG_SMP

    #ifdef CONFIG_X86_LOCAL_APIC

    static void __init smp_init(void)

    {

            APIC_init_uniprocessor();

    }

    #else

    #define smp_init()      do { } while (0)

    #endif

    如果在编译时选择了CONFIG_SMP呢,那么它的实现就如下喽:

    /* Called by boot processor to activate the rest. */

    static void __init smp_init(void)

    {

            unsigned int cpu;

            /* FIXME: This should be done in userspace --RR */

            for_each_present_cpu(cpu) {

                    if (num_online_cpus() >= setup_max_cpus)

                            break;

                    if (!cpu_online(cpu))

                            cpu_up(cpu);

            }

            /* Any cleanup work */

            printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());

            smp_cpus_done(setup_max_cpus);

    }

    来看看这个函数的,for_each_present_cpu(cpu)宏在文件include/linux/cpumask.h中实现:

    #define for_each_present_cpu(cpu) for_each_cpu_mask((cpu), cpu_present_map)

    而for_each_cpu_mask(cpu,mask)宏也在文件include/linux/cpumask.h中实现:

    #if NR_CPUS > 1

    #define for_each_cpu_mask(cpu, mask)            \

            for ((cpu) = first_cpu(mask);           \

                    (cpu) < NR_CPUS;                \

                    (cpu) = next_cpu((cpu), (mask)))

    #else /* NR_CPUS == 1 */

    #define for_each_cpu_mask(cpu, mask)            \

            for ((cpu) = 0; (cpu) < 1; (cpu)++, (void)mask)

    #endif /* NR_CPUS */

    即对于每个cpu都要执行大括号里的语句,如果当前cpu没激活就把它激活的,该函数然后打印一些cpu信息,如当前激活的cpu数目。

    kernel_init中紧跟smp_init()函数后的是sched_init_smp()函数和do_basic_setup()函数,而其后便是最后一个函数init_post(),在该函数中将调起init进程。由于内容较多,下次分析......

    (如哪里有错误,请高手指出,不胜感激,刚接触内核不久)