文件系统抽象层VFS

文件系统抽象层是把不同文件系统的对外共性接口提取出来,形成一个函数指针数组,这样,通用文件系统访问接口层只需访问文件系统抽象层,而不需关心具体文件系统的实现细节和接口。

file & dir 接口

file&dir 接口层定义了进程在内核中直接访问的文件相关信息,这定义在 file 数据结构中,具体描述如下:

// kern/fs/file.h
struct file {
    enum {
        FD_NONE, FD_INIT, FD_OPENED, FD_CLOSED,
    } status;                         //访问文件的执行状态
    bool readable;                    //文件是否可读
    bool writable;                    //文件是否可写
    int fd;                           //文件在filemap中的索引值
    off_t pos;                        //访问文件的当前位置
    struct inode *node;               //该文件对应的内存inode指针
    int open_count;                   //打开此文件的次数
};

而在 kern/process/proc.h中的 proc_struct 结构中加入了描述了进程访问文件的数据接口 files_struct,其数据结构定义如下:

// kern/fs/fs.h
struct files_struct {
    struct inode *pwd;                //进程当前执行目录的内存inode指针
    struct file *fd_array;            //进程打开文件的数组
    atomic_t files_count;             //访问此文件的线程个数
    semaphore_t files_sem;            //确保对进程控制块中fs_struct的互斥访问
};

当创建一个进程后,该进程的 files_struct 将会被初始化或复制父进程的 files_struct。当用户进程打开一个文件时,将从 fd_array 数组中取得一个空闲 file 项,然后会把此 file 的成员变量 node 指针指向一个代表此文件的 inode 的起始地址。

inode 接口

index node 是位于内存的索引节点,它是 VFS 结构中的重要数据结构,因为它实际负责把不同文件系统的特定索引节点信息(甚至不能算是一个索引节点)统一封装起来,避免了进程直接访问具体文件系统。其定义如下:

// kern/vfs/inode.h
struct inode {
    union {                                 //包含不同文件系统特定inode信息的union成员变量
        struct device __device_info;          //设备文件系统内存inode信息
        struct sfs_inode __sfs_inode_info;    //SFS文件系统内存inode信息
    } in_info;
    enum {
        inode_type_device_info = 0x1234,
        inode_type_sfs_inode_info,
    } in_type;                          //此inode所属文件系统类型
    atomic_t ref_count;                 //此inode的引用计数
    atomic_t open_count;                //打开此inode对应文件的个数
    struct fs *in_fs;                   //抽象的文件系统,包含访问文件系统的函数指针
    const struct inode_ops *in_ops;     //抽象的inode操作,包含访问inode的函数指针
};

在 inode 中,有一成员变量为 in_ops,这是对此 inode 的操作函数指针列表,其数据结构定义如下:

struct inode_ops {
    unsigned long vop_magic;
    int (*vop_open)(struct inode *node, uint32_t open_flags);
    int (*vop_close)(struct inode *node);
    int (*vop_read)(struct inode *node, struct iobuf *iob);
    int (*vop_write)(struct inode *node, struct iobuf *iob);
    int (*vop_getdirentry)(struct inode *node, struct iobuf *iob);
    int (*vop_create)(struct inode *node, const char *name, bool excl, struct inode **node_store);
int (*vop_lookup)(struct inode *node, char *path, struct inode **node_store);
……
 };

参照上面对 SFS 中的索引节点操作函数的说明,可以看出 inode_ops 是对常规文件、目录、设备文件所有操作的一个抽象函数表示。对于某一具体的文件系统中的文件或目录,只需实现相关的函数,就可以被用户进程访问具体的文件了,且用户进程无需了解具体文件系统的实现细节。

open系统调用的执行过程

下面我们通过打开文件的系统调用open()的执行过程, 看看文件系统的不同层次是如何交互的。

首先,经过syscall.c的处理之后,进入内核态,执行sysfile_open()函数

// kern/fs/sysfile.c
/* sysfile_open - open file */
int sysfile_open(const char *__path, uint32_t open_flags) {
    int ret;
    char *path;
    if ((ret = copy_path(&path, __path)) != 0) {
        return ret;
    }
    ret = file_open(path, open_flags);
    kfree(path);
    return ret;
}

可以看到,sysfile_open 把路径复制了一份,然后调用了file_openfile_open调用了vfs_open, 使用了VFS的接口。

// kern/fs/file.c
// open file
int file_open(char *path, uint32_t open_flags) {
    bool readable = 0, writable = 0;
    switch (open_flags & O_ACCMODE) { //解析 open_flags
    case O_RDONLY: readable = 1; break;
    case O_WRONLY: writable = 1; break;
    case O_RDWR:
        readable = writable = 1;
        break;
    default:
        return -E_INVAL;
    }
    int ret;
    struct file *file;
    if ((ret = fd_array_alloc(NO_FD, &file)) != 0) { //在当前进程分配file descriptor
        return ret; 
    }
    struct inode *node;
    if ((ret = vfs_open(path, open_flags, &node)) != 0) { //打开文件的工作在vfs_open完成
        fd_array_free(file); //打开失败,释放file descriptor
        return ret;
    }
    file->pos = 0;
    if (open_flags & O_APPEND) {
        struct stat __stat, *stat = &__stat;
        if ((ret = vop_fstat(node, stat)) != 0) {
            vfs_close(node);
            fd_array_free(file);
            return ret;
        }
        file->pos = stat->st_size; //追加写模式,设置当前位置为文件尾
    }
    file->node = node;
    file->readable = readable;
    file->writable = writable;
    fd_array_open(file); //设置该文件的状态为“打开”
    return file->fd;
}
// fs_array_alloc - allocate a free file item (with FD_NONE status) in open files table
static int fd_array_alloc(int fd, struct file **file_store) {
    struct file *file = get_fd_array();
    if (fd == NO_FD) {
        for (fd = 0; fd < FILES_STRUCT_NENTRY; fd ++, file ++) {
            if (file->status == FD_NONE) {
                goto found;
            }
        }
        return -E_MAX_OPEN;
    }
    else {
        if (testfd(fd)) {
            file += fd;
            if (file->status == FD_NONE) {
                goto found;
            }
            return -E_BUSY;
        }
        return -E_INVAL;
    }
found:
    assert(fopen_count(file) == 0);
    file->status = FD_INIT, file->node = NULL;
    *file_store = file;
    return 0;
}

void fd_array_open(struct file *file) {
    assert(file->status == FD_INIT && file->node != NULL);
    file->status = FD_OPENED; //设置状态为“打开”
    fopen_count_inc(file); //增加文件的“打开计数”
}

vfs_open是一个比较复杂的函数,这里我们使用的打开文件的flags, 基本是参照linux,如果希望详细了解,可以阅读linux manual: open

// kern/fs/vfs/vfsfile.c

// open file in vfs, get/create inode for file with filename path.
int vfs_open(char *path, uint32_t open_flags, struct inode **node_store) {
    bool can_write = 0;
    // 解析open_flags并做合法性检查
    switch (open_flags & O_ACCMODE) {
    case O_RDONLY:
        break;
    case O_WRONLY:
    case O_RDWR:
        can_write = 1;
        break;
    default:
        return -E_INVAL;
    }

    if (open_flags & O_TRUNC) {
        if (!can_write) {
            return -E_INVAL;
        }
    }
/*
linux manual
       O_TRUNC
              If the file already exists and is a regular file and the
              access mode allows writing (i.e., is O_RDWR or O_WRONLY) it
              will be truncated to length 0.  If the file is a FIFO or ter‐
              minal device file, the O_TRUNC flag is ignored.  Otherwise,
              the effect of O_TRUNC is unspecified.
*/
    int ret; 
    struct inode *node;
    bool excl = (open_flags & O_EXCL) != 0;
    bool create = (open_flags & O_CREAT) != 0;
    ret = vfs_lookup(path, &node); // vfs_lookup根据路径构造inode

    if (ret != 0) {//要打开的文件还不存在,可能出错,也可能需要创建新文件
        if (ret == -16 && (create)) {
            char *name;
            struct inode *dir;
            if ((ret = vfs_lookup_parent(path, &dir, &name)) != 0) {
                return ret;//需要在已经存在的目录下创建文件,目录不存在,则出错
            }
            ret = vop_create(dir, name, excl, &node);//创建新文件
        } else return ret;
    } else if (excl && create) {
        return -E_EXISTS;
        /*
        linux manual
              O_EXCL Ensure that this call creates the file: if this flag is
              specified in conjunction with O_CREAT, and pathname already
              exists, then open() fails with the error EEXIST.
        */
    }
    assert(node != NULL);

    if ((ret = vop_open(node, open_flags)) != 0) { 
        vop_ref_dec(node);
        return ret;
    }

    vop_open_inc(node);
    if (open_flags & O_TRUNC || create) {
        if ((ret = vop_truncate(node, 0)) != 0) {
            vop_open_dec(node);
            vop_ref_dec(node);
            return ret;
        }
    }
    *node_store = node;
    return 0;
}

我们看看vfs_look_up的实现

/*
 * get_device- Common code to pull the device name, if any, off the front of a
 *             path and choose the inode to begin the name lookup relative to.
 */

static int get_device(char *path, char **subpath, struct inode **node_store) {
    int i, slash = -1, colon = -1;
    for (i = 0; path[i] != '\0'; i ++) {
        if (path[i] == ':') { colon = i; break; }
        if (path[i] == '/') { slash = i; break; }
    }
    if (colon < 0 && slash != 0) {
      /* *
      * No colon before a slash, so no device name specified, and the slash isn't leading
      * or is also absent, so this is a relative path or just a bare filename. Start from
      * the current directory, and use the whole thing as the subpath.
      * */
        *subpath = path;
        return vfs_get_curdir(node_store); //把当前目录的inode存到node_store
    }
    if (colon > 0) {
        /* device:path - get root of device's filesystem */
        path[colon] = '\0';

        /* device:/path - skip slash, treat as device:path */
        while (path[++ colon] == '/');
        *subpath = path + colon;
        return vfs_get_root(path, node_store);
    }

    /* *
     * we have either /path or :path
     * /path is a path relative to the root of the "boot filesystem"
     * :path is a path relative to the root of the current filesystem
     * */
    int ret;
    if (*path == '/') {
        if ((ret = vfs_get_bootfs(node_store)) != 0) {
            return ret;
        }
    }
    else {
        assert(*path == ':');
        struct inode *node;
        if ((ret = vfs_get_curdir(&node)) != 0) {
            return ret;
        }
        /* The current directory may not be a device, so it must have a fs. */
        assert(node->in_fs != NULL);
        *node_store = fsop_get_root(node->in_fs);
        vop_ref_dec(node);
    }

    /* ///... or :/... */
    while (*(++ path) == '/');
    *subpath = path;
    return 0;
}

/*
 * vfs_lookup - get the inode according to the path filename
 */
int vfs_lookup(char *path, struct inode **node_store) {
    int ret;
    struct inode *node;
    if ((ret = get_device(path, &path, &node)) != 0) {
        return ret;
    }
    if (*path != '\0') {
        ret = vop_lookup(node, path, node_store);
        vop_ref_dec(node);
        return ret;
    }
    *node_store = node;
    return 0;
}

/*
 * vfs_lookup_parent - Name-to-vnode translation.
 *  (In BSD, both of these are subsumed by namei().)
 */
int vfs_lookup_parent(char *path, struct inode **node_store, char **endp){
    int ret;
    struct inode *node;
    if ((ret = get_device(path, &path, &node)) != 0) {
        return ret;
    }
    *endp = path;
    *node_store = node;
    return 0;
}

我们注意到,这个流程中,有大量以vop开头的函数,它们都通过一些宏和函数的转发,最后变成对inode结构体里的inode_ops结构体的“成员函数”(实际上是函数指针)的调用。对于SFS文件系统的inode来说,会变成对sfs文件系统的具体操作。sfs的这些具体接口的实现较为繁琐,可以在kern/fs/sfs/sfs_inode.c具体查看。我们的练习要求在kern/fs/sfs/sfs_io.c填写一个函数。

// kern/fs/sfs/sfs_inode.c

// The sfs specific DIR operations correspond to the abstract operations on a inode.
static const struct inode_ops sfs_node_dirops = {
    .vop_magic                      = VOP_MAGIC,
    .vop_open                       = sfs_opendir,
    .vop_close                      = sfs_close,
    .vop_fstat                      = sfs_fstat,
    .vop_fsync                      = sfs_fsync,
    .vop_namefile                   = sfs_namefile,
    .vop_getdirentry                = sfs_getdirentry,
    .vop_reclaim                    = sfs_reclaim,
    .vop_gettype                    = sfs_gettype,
    .vop_lookup                     = sfs_lookup,
};
/// The sfs specific FILE operations correspond to the abstract operations on a inode.
static const struct inode_ops sfs_node_fileops = {
    .vop_magic                      = VOP_MAGIC,
    .vop_open                       = sfs_openfile,
    .vop_close                      = sfs_close,
    .vop_read                       = sfs_read,
    .vop_write                      = sfs_write,
    .vop_fstat                      = sfs_fstat,
    .vop_fsync                      = sfs_fsync,
    .vop_reclaim                    = sfs_reclaim,
    .vop_gettype                    = sfs_gettype,
    .vop_tryseek                    = sfs_tryseek,
    .vop_truncate                   = sfs_truncfile,
};

最后更新于