sys_open分析，从文件名找到文件信息(namei)

从文件名找到文件信息(namei)
 
本文档的Copyleft归yfydz所有，使用GPL发布，可以自由拷贝，转载，转载时请保持文档的完整性，严禁用于任何商业用途。
msn: yfydz_no1@hotmail.com
来源：http://yfydz.cublog.cn

1. 前言

inode是类Unix系统的文件系统的基本索引方法，每个文件都对应一个inode，再通过inode找到文件中的实际数据，因此根据文件路径名找到具体的inode节点就是一个很重要的处理步骤。系统会缓存用过的每个文件或目录对应的dentry结构, 从该结构可以指向相应的inode, 每次打开文件, 都会最终对应到文件的inode，中间查找过程称为namei。

本文介绍Linux下的路径到文件指针的转换过程，内核版本为2.6.19.2。

虚拟文件系统的转换源代码在fs/namei.c中，具体和文件系统相关的部分在fs/*/namei.c文件中。

2. 引子

由于这种转换是一个中间过程，在具体分析namei处理前，先看看系统的调用顺序是如何进入转换的：
当用户空间程序用open系统调用打开一个文件时，内核对应的处理是sys_open:
/* fs/open.c */
asmlinkage long sys_open(const char __user *filename, int flags, int mode)
{
 long ret;
 if (force_o_largefile())
  flags |= O_LARGEFILE;
 ret = do_sys_open(AT_FDCWD, filename, flags, mode);
 /* avoid REGPARM breakage on x86: */
 prevent_tail_call(ret);
 return ret;
}
真正的打开函数是do_sys_open:
/* fs/open.c */
// dfd为AT_FDCWD
long do_sys_open(int dfd, const char __user *filename, int flags, int mode)
{
// 通过该函数将用户空间的文件名传递到内核
// tmp是一个cache类的动态内存空间,用于保存文件路径名
//
 char *tmp = getname(filename);
 int fd = PTR_ERR(tmp);
 if (!IS_ERR(tmp)) {
// 获取一个未使用的文件描述符, 和inode无关
  fd = get_unused_fd();
  if (fd >= 0) {
// 打开文件,将文件名转换为文件结构
   struct file *f = do_filp_open(dfd, tmp, flags, mode);
   if (IS_ERR(f)) {
    put_unused_fd(fd);
    fd = PTR_ERR(f);
   } else {
    fsnotify_open(f->f_dentry);
    fd_install(fd, f);
   }
  }
  putname(tmp);
 }
 return fd;
}

// 文件打开
static struct file *do_filp_open(int dfd, const char *filename, int flags,
     int mode)
{
 int namei_flags, error;
// 注意这是结构而不是指针
 struct nameidata nd;
 namei_flags = flags;
 if ((namei_flags+1) & O_ACCMODE)
  namei_flags++;
// 根据文件名得到nameidata, nd作为namei空间保存结果
 error = open_namei(dfd, filename, namei_flags, mode, &nd);
 if (!error)
// 成功, nameidata再转换为file指针
  return nameidata_to_filp(&nd, flags);
 return ERR_PTR(error);
}

因此重点函数是open_namei函数, 实现了从文件名到inode的转换, 也是namei的处理入口.

在分析open_namei前, 再分析一下getname, 这用到了kmem_cache来处理的:
// 文件名转换, 从用户空间拷贝到内核空间
/* fs/namei.c */
char * getname(const char __user * filename)
{
 char *tmp, *result;
 result = ERR_PTR(-ENOMEM);
/* include/linux/fs.h */
// __getname和__putname的定义，实际就是内核cache的分配和释放
// #define __getname() kmem_cache_alloc(names_cachep, SLAB_KERNEL)
// #define __putname(name) kmem_cache_free(names_cachep, (void *)(name))
// 这里实际是分配names的cache, 该cache定义为
//  names_cachep = kmem_cache_create("names_cache", PATH_MAX, 0,
//   SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL, NULL);
 tmp = __getname();
 if (tmp)  {
// cache分配成功
// 进入实际操作函数
  int retval = do_getname(filename, tmp);
// 要返回结果指向cache
  result = tmp;
  if (retval < 0) {
// 操作失败,释放cache,返回错误
   __putname(tmp);
   result = ERR_PTR(retval);
  }
 }
// 编译内核时如果没有设置CONFIG_AUDITSYSCALL, 则audit_getname为空
// 审计系统调用结果
 audit_getname(result);
 return result;
}

static int do_getname(const char __user *filename, char *page)
{
 int retval;
 unsigned long len = PATH_MAX;
 if (!segment_eq(get_fs(), KERNEL_DS)) {
  if ((unsigned long) filename >= TASK_SIZE)
   return -EFAULT;
  if (TASK_SIZE - (unsigned long) filename < PATH_MAX)
   len = TASK_SIZE - (unsigned long) filename;
 }
// 将用户空间提供的文件名拷贝到cache中
 retval = strncpy_from_user(page, filename, len);
 if (retval > 0) {
  if (retval < len)
   return 0;
  return -ENAMETOOLONG;
 } else if (!retval)
  retval = -ENOENT;
 return retval;
}
 
3. namei相关数据结构
/* include/linux/namei.h */
struct nameidata {
// 路径点
 struct dentry *dentry;
// 虚拟系统挂接点
 struct vfsmount *mnt;
// 路径名中的最后的文件名或目录名
 struct qstr last;
 unsigned int flags;
 int  last_type;
// 目录深度
 unsigned depth;
 char *saved_names[MAX_NESTED_LINKS + 1]; // 9
 /* Intent data */
// 相关数据
 union {
// 包含打开的文件的指针
  struct open_intent open;
 } intent;
};

struct open_intent {
// 标志
 int flags;
// 创建模式
 int create_mode;
// 文件指针
 struct file *file;
};

// 路径结构, 属于中间处理结构, 将文件系统挂接点和dentry捆绑在一起而已
struct path {
 struct vfsmount *mnt;
 struct dentry *dentry;
};

/* include/linux/dcache.h */
// 文件目录项, 在系统cache中
struct dentry {
 atomic_t d_count;
 unsigned int d_flags;  /* protected by d_lock */
 spinlock_t d_lock;  /* per dentry lock */
 struct inode *d_inode;  /* Where the name belongs to - NULL is
      * negative */
 /*
  * The next three fields are touched by __d_lookup.  Place them here
  * so they all fit in a cache line.
  */
 struct hlist_node d_hash; /* lookup hash list */
 struct dentry *d_parent; /* parent directory */
 struct qstr d_name;
 struct list_head d_lru;  /* LRU list */
 /*
  * d_child and d_rcu can share memory
  */
 union {
  struct list_head d_child; /* child of parent list */
   struct rcu_head d_rcu;
 } d_u;
 struct list_head d_subdirs; /* our children */
 struct list_head d_alias; /* inode alias list */
 unsigned long d_time;  /* used by d_revalidate */
 struct dentry_operations *d_op;
 struct super_block *d_sb; /* The root of the dentry tree */
 void *d_fsdata;   /* fs-specific data */
#ifdef CONFIG_PROFILING
 struct dcookie_struct *d_cookie; /* cookie, if any */
#endif
 int d_mounted;
 unsigned char d_iname[DNAME_INLINE_LEN_MIN]; /* small names */
};

/* include/linux/fs.h */
// 文件结构
struct file {
 /*
  * fu_list becomes invalid after file_free is called and queued via
  * fu_rcuhead for RCU freeing
  */
 union {
  struct list_head fu_list;
  struct rcu_head  fu_rcuhead;
 } f_u;
// 文件的dentry
 struct dentry  *f_dentry;
// 虚拟文件系统挂接点
 struct vfsmount         *f_vfsmnt;
// 文件操作
 const struct file_operations *f_op;
 atomic_t  f_count;
 unsigned int   f_flags;
 mode_t   f_mode;
 loff_t   f_pos;
 struct fown_struct f_owner;
 unsigned int  f_uid, f_gid;
 struct file_ra_state f_ra;
 unsigned long  f_version;
#ifdef CONFIG_SECURITY
 void   *f_security;
#endif
 /* needed for tty driver, and maybe others */
 void   *private_data;
#ifdef CONFIG_EPOLL
 /* Used by fs/eventpoll.c to link all the hooks to this file */
 struct list_head f_ep_links;
 spinlock_t  f_ep_lock;
#endif /* #ifdef CONFIG_EPOLL */
 struct address_space *f_mapping;
};
 
4. namei操作

4.1 open_namei

/* fs/namei.c */
/*
 * open_namei()
 *
 * namei for open - this is in fact almost the whole open-routine.
 *
 * Note that the low bits of "flag" aren&apos;t the same as in the open
 * system call - they are 00 - no permissions needed
 *     01 - read permission needed
 *     10 - write permission needed
 *     11 - read/write permissions needed
 * which is a lot more logical, and also allows the "no perm" needed
 * for symlinks (where the permissions are checked later).
 * SMP-safe
 */
int open_namei(int dfd, const char *pathname, int flag,
  int mode, struct nameidata *nd)
{
 int acc_mode, error;
 struct path path;
 struct dentry *dir;
 int count = 0;
// #define ACC_MODE(x) ("/000/004/002/006"[(x)&O_ACCMODE])
// 审计模式
 acc_mode = ACC_MODE(flag);
 /* O_TRUNC implies we need access checks for write permissions */
// 截断标志, 基本上需要写权限, 除非要截断的长度实际大于文件本身长度
 if (flag & O_TRUNC)
  acc_mode |= MAY_WRITE;
 /* Allow the LSM permission hook to distinguish append
    access from general write access. */
// 添加标志, 也是需要写权限
 if (flag & O_APPEND)
  acc_mode |= MAY_APPEND;
 /*
  * The simplest case - just a plain lookup.
  */
// 不需要创建文件
 if (!(flag & O_CREAT)) {
// 直接找pathname的dentry和挂接点, 结果填在nd中
  error = path_lookup_open(dfd, pathname, lookup_flags(flag),
      nd, flag);
  if (error)
   return error;
  goto ok;
 }
 /*
  * Create - we need to know the parent.
  */
// 创建文件的dentry和挂接点, 数据填到nd中
 error = path_lookup_create(dfd,pathname,LOOKUP_PARENT,nd,flag,mode);
 if (error)
  return error;
 /*
  * We have the parent and last component. First of all, check
  * that we are not asked to creat(2) an obvious directory - that
  * will not do.
  */
 error = -EISDIR;
// 检查nameidata结构中的last参数是否合法
 if (nd->last_type != LAST_NORM || nd->last.name[nd->last.len])
  goto exit;
// 文件项dentry
 dir = nd->dentry;
// 去掉查询父目录标志
 nd->flags &= ~LOOKUP_PARENT;
 mutex_lock(&dir->d_inode->i_mutex);
// 填充path参数, 又根据nd的信息搜索一次当前的缓存的dentry
// 不过dir与path.dentry难道不相同么?
 path.dentry = lookup_hash(nd);
 path.mnt = nd->mnt;
do_last:
// 检查path.entry是否合法
 error = PTR_ERR(path.dentry);
 if (IS_ERR(path.dentry)) {
  mutex_unlock(&dir->d_inode->i_mutex);
  goto exit;
 }
// 检查nd->intent.open.file是否合法, 这是最终要返回的文件指针
 if (IS_ERR(nd->intent.open.file)) {
  mutex_unlock(&dir->d_inode->i_mutex);
  error = PTR_ERR(nd->intent.open.file);
  goto exit_dput;
 }
 /* Negative dentry, just create the file */
 if (!path.dentry->d_inode) {
// 创建新文件的inode, 然后返回
  error = open_namei_create(nd, &path, flag, mode);
  if (error)
   goto exit;
  return 0;
 }
// 现在是打开已经存在的文件
 /*
  * It already exists.
  */
 mutex_unlock(&dir->d_inode->i_mutex);
 audit_inode_update(path.dentry->d_inode);
 error = -EEXIST;
// O_EXCL标志是只必须打开的是不存在的文件, 文件已存在时错误
 if (flag & O_EXCL)
  goto exit_dput;
 if (__follow_mount(&path)) {
  error = -ELOOP;
  if (flag & O_NOFOLLOW)
   goto exit_dput;
 }
 error = -ENOENT;
 if (!path.dentry->d_inode)
  goto exit_dput;
// 如果dentry的具体FS的实现中定义了follow_link操作, 转
// 不过大多数FS的实现中都没有定义该函数
 if (path.dentry->d_inode->i_op && path.dentry->d_inode->i_op->follow_link)
  goto do_link;
// 从路径中的dentry和mnt信息赋值到nameidata
 path_to_nameidata(&path, nd);
 error = -EISDIR;
// 如果是一个目录, 返回错误
 if (path.dentry->d_inode && S_ISDIR(path.dentry->d_inode->i_mode))
  goto exit;
ok:
// 对nd中的dentry及其inode进行打开前的错误检查
 error = may_open(nd, acc_mode, flag);
 if (error)
  goto exit;
 return 0;
// 下面是错误处理, 释放掉已分配的资源, 返回错误
exit_dput:
 dput_path(&path, nd);
exit:
 if (!IS_ERR(nd->intent.open.file))
  release_open_intent(nd);
 path_release(nd);
 return error;
// 处理符号连接, 找到实际文件的inode,然后重新循环, 要注意回环情况的错误处理
do_link:
 error = -ELOOP;
 if (flag & O_NOFOLLOW)
  goto exit_dput;
 /*
  * This is subtle. Instead of calling do_follow_link() we do the
  * thing by hands. The reason is that this way we have zero link_count
  * and path_walk() (called from ->follow_link) honoring LOOKUP_PARENT.
  * After that we have the parent and last component, i.e.
  * we are in the same situation as after the first path_walk().
  * Well, almost - if the last component is normal we get its copy
  * stored in nd->last.name and we will have to putname() it when we
  * are done. Procfs-like symlinks just set LAST_BIND.
  */
// 设置查找LOOKUP_PARENT标志
 nd->flags |= LOOKUP_PARENT;
 error = security_inode_follow_link(path.dentry, nd);
 if (error)
  goto exit_dput;
// 处理符号链接
 error = __do_follow_link(&path, nd);
 if (error) {
  /* Does someone understand code flow here? Or it is only
   * me so stupid? Anathema to whoever designed this non-sense
   * with "intent.open".
   */
  release_open_intent(nd);
  return error;
 }
 nd->flags &= ~LOOKUP_PARENT;
// 检查最后一段文件或目录名的属性情况
 if (nd->last_type == LAST_BIND)
  goto ok;
 error = -EISDIR;
 if (nd->last_type != LAST_NORM)
  goto exit;
 if (nd->last.name[nd->last.len]) {
  __putname(nd->last.name);
  goto exit;
 }
 error = -ELOOP;
// 出现回环标志: 循环超过32次
 if (count++==32) {
  __putname(nd->last.name);
  goto exit;
 }
 dir = nd->dentry;
 mutex_lock(&dir->d_inode->i_mutex);
// 更新路径的挂接点和dentry
 path.dentry = lookup_hash(nd);
 path.mnt = nd->mnt;
 __putname(nd->last.name);
 goto do_last;
}

4.2  path_lookup_open和path_lookup_create

这两个函数找到路径名对应的挂接点和dentry结构, 赋值到nameidata结构中, create时如果文件不存在, 建立新文件:
/**
 * path_lookup_open - lookup a file path with open intent
 * @dfd: the directory to use as base, or AT_FDCWD
 * @name: pointer to file name
 * @lookup_flags: lookup intent flags
 * @nd: pointer to nameidata
 * @open_flags: open intent flags
 */
int path_lookup_open(int dfd, const char *name, unsigned int lookup_flags,
  struct nameidata *nd, int open_flags)
{
 return __path_lookup_intent_open(dfd, name, lookup_flags, nd,
   open_flags, 0);
}

/**
 * path_lookup_create - lookup a file path with open + create intent
 * @dfd: the directory to use as base, or AT_FDCWD
 * @name: pointer to file name
 * @lookup_flags: lookup intent flags
 * @nd: pointer to nameidata
 * @open_flags: open intent flags
 * @create_mode: create intent flags
 */
static int path_lookup_create(int dfd, const char *name,
         unsigned int lookup_flags, struct nameidata *nd,
         int open_flags, int create_mode)
{
 return __path_lookup_intent_open(dfd, name, lookup_flags|LOOKUP_CREATE,
   nd, open_flags, create_mode);
}

这两个函数都是调用__path_lookup_intent_open, 只是参数不同,create中加入了LOOKUP_CREATE标志和create_mode:

static int __path_lookup_intent_open(int dfd, const char *name,
  unsigned int lookup_flags, struct nameidata *nd,
  int open_flags, int create_mode)
{
// 找一个空闲的文件指针
 struct file *filp = get_empty_filp();
 int err;
// 找不到返回错误, 文件表溢出了
 if (filp == NULL)
  return -ENFILE;
// 在nameidate中填充打开的文件参数, 这是最终会返回的文件指针
 nd->intent.open.file = filp;
 nd->intent.open.flags = open_flags;
 nd->intent.open.create_mode = create_mode;
// 进行具体的路径查找, name是路径名
 err = do_path_lookup(dfd, name, lookup_flags|LOOKUP_OPEN, nd);
// 先检查nd->intent.open.file而不是err
 if (IS_ERR(nd->intent.open.file)) {
// 打开的文件指针错误
  if (err == 0) {
// do_path_lookup已经成功了, 释放path, err重新设置为错误值
   err = PTR_ERR(nd->intent.open.file);
   path_release(nd);
  }
 } else if (err != 0)
  release_open_intent(nd);
 return err;
}

// 路径查找
/* Returns 0 and nd will be valid on success; Retuns error, otherwise. */
static int fastcall do_path_lookup(int dfd, const char *name,
    unsigned int flags, struct nameidata *nd)
{
 int retval = 0;
 int fput_needed;
 struct file *file;
// 文件系统指针从进程中获取
 struct fs_struct *fs = current->fs;
// 缺省情况last_type为绝对路径, 以"/"开头的格式
 nd->last_type = LAST_ROOT; /* if there are only slashes... */
 nd->flags = flags;
 nd->depth = 0;
// 下面只是用于增加某些变量的使用计数值, get是增加,put是减少
 if (*name==&apos;/&apos;) {
// 绝对路径格式
  read_lock(&fs->lock);
  if (fs->altroot && !(nd->flags & LOOKUP_NOALT)) {
// 检查是否更改了root, 即用chroot
// 增加altrootmnt的使用计数, 其为一vfsmount结构指针
   nd->mnt = mntget(fs->altrootmnt);
   nd->dentry = dget(fs->altroot);
   read_unlock(&fs->lock);
   if (__emul_lookup_dentry(name,nd))
    goto out; /* found in altroot */
   read_lock(&fs->lock);
  }
// 增加rootmnt的使用计数然后赋值到nd中
  nd->mnt = mntget(fs->rootmnt);
// 增加根的dentry的使用计数然后赋值到nd中
  nd->dentry = dget(fs->root);
  read_unlock(&fs->lock);
 } else if (dfd == AT_FDCWD) {
// 从sys_open调用来的话会到这里, 表示从当前工作目录的路径开始的相对路径
  read_lock(&fs->lock);
// 增加pwdmnt使用计数然后赋值到nd中
  nd->mnt = mntget(fs->pwdmnt);
// 增加pwd使用计数然后赋值到nd中
  nd->dentry = dget(fs->pwd);
  read_unlock(&fs->lock);
 } else {
  struct dentry *dentry;
// 轻量级的路径查找, fd不是共享的话不会增加引用计数
  file = fget_light(dfd, &fput_needed);
  retval = -EBADF;
  if (!file)
   goto out_fail;
  dentry = file->f_dentry;
  retval = -ENOTDIR;
  if (!S_ISDIR(dentry->d_inode->i_mode))
   goto fput_fail;
// 检查文件的执行权限
  retval = file_permission(file, MAY_EXEC);
  if (retval)
   goto fput_fail;
// 增加f_vfsmnt的使用计数
  nd->mnt = mntget(file->f_vfsmnt);
  nd->dentry = dget(dentry);
// 轻量级释放
  fput_light(file, fput_needed);
 }
// 清空总链接数
 current->total_link_count = 0;
// 变量路径表查询, 核心函数
 retval = link_path_walk(name, nd);
out:
 if (likely(retval == 0)) {
// 在大部分情况下都会执行到这,能正确打开路径
  if (unlikely(!audit_dummy_context() && nd && nd->dentry &&
    nd->dentry->d_inode))
  audit_inode(name, nd->dentry->d_inode);
 }
out_fail:
 return retval;
fput_fail:
 fput_light(file, fput_needed);
 goto out_fail;
}

do_path_lookup调用的核心函数是link_path_walk:

/*
 * Wrapper to retry pathname resolution whenever the underlying
 * file system returns an ESTALE.
 *
 * Retry the whole path once, forcing real lookup requests
 * instead of relying on the dcache.
 */
int fastcall link_path_walk(const char *name, struct nameidata *nd)
{
// 先备份一下
 struct nameidata save = *nd;
 int result;
 /* make sure the stuff we saved doesn&apos;t go away */
 dget(save.dentry);
 mntget(save.mnt);
 result = __link_path_walk(name, nd);
 if (result == -ESTALE) {
// ESTALE是失效的文件句柄错误
// 用备份的nameidate重新恢复, 设置LOOKUP_REVAL标志后重新查询
  *nd = save;
  dget(nd->dentry);
  mntget(nd->mnt);
  nd->flags |= LOOKUP_REVAL;
  result = __link_path_walk(name, nd);
 }
 dput(save.dentry);
 mntput(save.mnt);
 return result;
}

真正的名称解析函数__link_path_walk:
/*
 * Name resolution.
 * This is the basic name resolution function, turning a pathname into
 * the final dentry. We expect &apos;base&apos; to be positive and a directory.
 *
 * Returns 0 and nd will have valid dentry and mnt on success.
 * Returns error and drops reference to input namei data on failure.
 */
static fastcall int __link_path_walk(const char * name, struct nameidata *nd)
{
 struct path next;
 struct inode *inode;
 int err;
 unsigned int lookup_flags = nd->flags;
// 去掉起始多余的"/", 同时也说明系统可以允许你输入多个"/"而不报错
 while (*name==&apos;/&apos;)
  name++;
// 空路径
 if (!*name)
  goto return_reval;
// 路径对应的inode
 inode = nd->dentry->d_inode;
 if (nd->depth)
  lookup_flags = LOOKUP_FOLLOW | (nd->flags & LOOKUP_CONTINUE);
 /* At this point we know we have a real path component. */
 for(;;) {
// 循环处理,每个循环提取文件路径的一个目录名, &apos;/&apos;分隔
  unsigned long hash;
  struct qstr this;
  unsigned int c;
  nd->flags |= LOOKUP_CONTINUE;
// 检查文件权限, 包括读写执行权限, 用户/组/其他权限, 返回0为合法
  err = exec_permission_lite(inode, nd);
  if (err == -EAGAIN)
// EAGAIN表示该inode正在被操作, 检查其执行权限
// 而对于普通文件检查结果将是错误
   err = vfs_permission(nd, MAY_EXEC);
// 出错中断循环
   if (err)
   break;
// 填充quickstring结构
  this.name = name;
// name的第一个字符的数值
  c = *(const unsigned char *)name;
// 计算文件名的hash, 不包括&apos;/&apos;
  hash = init_name_hash();
  do {
   name++;
   hash = partial_name_hash(c, hash);
   c = *(const unsigned char *)name;
  } while (c && (c != &apos;/&apos;));
// 目录(如果有的话)的名称长度
  this.len = name - (const char *) this.name;
// hash
  this.hash = end_name_hash(hash);
  /* remove trailing slashes? */
// c为0表示是最后的具体文件名了
  if (!c)
   goto last_component;
// 跳过中间的&apos;/&apos;
  while (*++name == &apos;/&apos;);
// 到名称尾, 说明文件名最后一个字符是&apos;/&apos;
  if (!*name)
   goto last_with_slashes;
  /*
   * "." and ".." are special - ".." especially so because it has
   * to be able to know about the current root directory and
   * parent relationships.
   */
// 如果第一个字符是&apos;.&apos;
  if (this.name[0] == &apos;.&apos;) switch (this.len) {
   default:
// 是一个一&apos;.&apos;开头的文件或目录名称
    break;
   case 2:
// 第2 个字符不是".", 是普通文件或路径名
    if (this.name[1] != &apos;.&apos;)
     break;
// 以".."开头, 是父目录, 更新nd为父目录nameidata数据, inode相应更新重新循环
    follow_dotdot(nd);
    inode = nd->dentry->d_inode;
    /* fallthrough */
   case 1:
// 以&apos;.&apos;开头的当前目录, 忽略, 重新循环
    continue;
  }
  /*
   * See if the low-level filesystem might want
   * to use its own hash..
   */
// 底层FS实现中有自己的HASH算法
  if (nd->dentry->d_op && nd->dentry->d_op->d_hash) {
   err = nd->dentry->d_op->d_hash(nd->dentry, &this);
   if (err < 0)
    break;
  }
  /* This does the actual lookups.. */
// 根据文件/目录名进行具体的查找
  err = do_lookup(nd, &this, &next);
  if (err)
   break;
  err = -ENOENT;
// inode更新为本级文件目录的inode
  inode = next.dentry->d_inode;
// 找不到inode, 转错误处理
  if (!inode)
   goto out_dput;
  err = -ENOTDIR;
  if (!inode->i_op)
   goto out_dput;
  if (inode->i_op->follow_link) {
// 处理符号链接, 在其中考虑了递归互相链接的异常处理
   err = do_follow_link(&next, nd);
   if (err)
    goto return_err;
   err = -ENOENT;
// 更新inode为实际的inode
   inode = nd->dentry->d_inode;
   if (!inode)
    break;
   err = -ENOTDIR;
   if (!inode->i_op)
    break;
  } else
// nd中得到下一级路径信息
   path_to_nameidata(&next, nd);
  err = -ENOTDIR;
  if (!inode->i_op->lookup)
   break;
// 继续循环找下一目录文件名称
  continue;
  /* here ends the main loop */
// 最后的文件名了, 处理和前面类似
last_with_slashes:
// 最后一个字符是&apos;/&apos;, 是一个目录
  lookup_flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
last_component:
  /* Clear LOOKUP_CONTINUE iff it was previously unset */
  nd->flags &= lookup_flags | ~LOOKUP_CONTINUE;
  if (lookup_flags & LOOKUP_PARENT)
   goto lookup_parent;
  if (this.name[0] == &apos;.&apos;) switch (this.len) {
   default:
    break;
   case 2:
// 文件名不是"..", 继续
    if (this.name[1] != &apos;.&apos;)
     break;
// 文件名是"..", 到父目录
    follow_dotdot(nd);
    inode = nd->dentry->d_inode;
    /* fallthrough */
   case 1:
// 文件名就是".", 跳到返回处理
    goto return_reval;
  }
// 一般文件处理
// 底层FS实现中有自己的HASH算法
  if (nd->dentry->d_op && nd->dentry->d_op->d_hash) {
   err = nd->dentry->d_op->d_hash(nd->dentry, &this);
   if (err < 0)
    break;
  }
// 查找最后的文件名
  err = do_lookup(nd, &this, &next);
  if (err)
   break;
  inode = next.dentry->d_inode;
  if ((lookup_flags & LOOKUP_FOLLOW)
      && inode && inode->i_op && inode->i_op->follow_link) {
   err = do_follow_link(&next, nd);
   if (err)
    goto return_err;
   inode = nd->dentry->d_inode;
  } else
// 更新nameidata中的mnt, dentry值
   path_to_nameidata(&next, nd);
  err = -ENOENT;
  if (!inode)
   break;
  if (lookup_flags & LOOKUP_DIRECTORY) {
   err = -ENOTDIR;
   if (!inode->i_op || !inode->i_op->lookup)
    break;
  }
  goto return_base;
lookup_parent:
// 复制当前quickstring结构this信息到nd的last中
// 类型为LAST_NORM
  nd->last = this;
  nd->last_type = LAST_NORM;
  if (this.name[0] != &apos;.&apos;)
   goto return_base;
  if (this.len == 1)
   nd->last_type = LAST_DOT;
  else if (this.len == 2 && this.name[1] == &apos;.&apos;)
   nd->last_type = LAST_DOTDOT;
  else
   goto return_base;
return_reval:
// 返回
  /*
   * We bypassed the ordinary revalidation routines.
   * We may need to check the cached dentry for staleness.
   */
  if (nd->dentry && nd->dentry->d_sb &&
      (nd->dentry->d_sb->s_type->fs_flags & FS_REVAL_DOT)) {
   err = -ESTALE;
   /* Note: we do not d_invalidate() */
   if (!nd->dentry->d_op->d_revalidate(nd->dentry, nd))
    break;
  }
return_base:
  return 0;
out_dput:
  dput_path(&next, nd);
  break;
 }
// 到这里属于出错了
 path_release(nd);
return_err:
 return err;
}
 
/*
 *  It&apos;s more convoluted than I&apos;d like it to be, but... it&apos;s still fairly
 *  small and for now I&apos;d prefer to have fast path as straight as possible.
 *  It _is_ time-critical.
 */
static int do_lookup(struct nameidata *nd, struct qstr *name,
       struct path *path)
{
 struct vfsmount *mnt = nd->mnt;
// 从系统缓存的dentry的hash表中查找父dentry是nd->dentry, 名称为name的dentry
 struct dentry *dentry = __d_lookup(nd->dentry, name);
// 没找到dentry, 进行真正从存储硬盘中查找
 if (!dentry)
  goto need_lookup;
// 需要进行revalidate操作时先进行validate操作
 if (dentry->d_op && dentry->d_op->d_revalidate)
  goto need_revalidate;
done:
// 找到, 填充path参数: 挂接点mnt和目录项dentry
 path->mnt = mnt;
 path->dentry = dentry;
 __follow_mount(path);
 return 0;
need_lookup:
// 进行真正的查找, 不过read_lookup会重新调用__d_lookup, 找不到才调用底层的fs实现去查找
// 好象是重复操作了
// real_lookup中的操作才反映了各个fs底层和相关标志的区别处理
 dentry = real_lookup(nd->dentry, name, nd);
 if (IS_ERR(dentry))
  goto fail;
 goto done;
need_revalidate:
// 进行validate操作
 dentry = do_revalidate(dentry, nd);
 if (!dentry)
  goto need_lookup;
 if (IS_ERR(dentry))
  goto fail;
 goto done;
fail:
 return PTR_ERR(dentry);
}

/*
 * This is called when everything else fails, and we actually have
 * to go to the low-level filesystem to find out what we should do..
 *
 * We get the directory semaphore, and after getting that we also
 * make sure that nobody added the entry to the dcache in the meantime..
 * SMP-safe
 */
static struct dentry * real_lookup(struct dentry * parent, struct qstr * name, struct nameidata *nd)
{
 struct dentry * result;
 struct inode *dir = parent->d_inode;
 mutex_lock(&dir->i_mutex);
 /*
  * First re-do the cached lookup just in case it was created
  * while we waited for the directory semaphore..
  *
  * FIXME! This could use version numbering or similar to
  * avoid unnecessary cache lookups.
  *
  * The "dcache_lock" is purely to protect the RCU list walker
  * from concurrent renames at this point (we mustn&apos;t get false
  * negatives from the RCU list walk here, unlike the optimistic
  * fast walk).
  *
  * so doing d_lookup() (with seqlock), instead of lockfree __d_lookup
  */
// 查找缓存中的dentry项
 result = d_lookup(parent, name);
 if (!result) {
// 没找到, 新建dentry项
  struct dentry * dentry = d_alloc(parent, name);
  result = ERR_PTR(-ENOMEM);
  if (dentry) {
// 调用inode的查找操作, 这是和具体文件系统相关
   result = dir->i_op->lookup(dir, dentry, nd);
   if (result)
// 失败, 释放dentry
    dput(dentry);
   else
// 成功, 找到的dentry作为结果返回
    result = dentry;
  }
  mutex_unlock(&dir->i_mutex);
  return result;
 }
 /*
  * Uhhuh! Nasty case: the cache was re-populated while
  * we waited on the semaphore. Need to revalidate.
  */
// 在缓存中找到dentry项, 进行validate操作
 mutex_unlock(&dir->i_mutex);
 if (result->d_op && result->d_op->d_revalidate) {
  result = do_revalidate(result, nd);
  if (!result)
   result = ERR_PTR(-ENOENT);
 }
 return result;
}

小结一下函数调用顺序:
path_lookup_open    path_lookup_create
     |                     |
     V                     V
   __path_lookup_intent_open
               |
               V
        do_path_lookup
               |
               V
        link_path_walk
               |
               V
      __link_path_walk
               |
               V
           do_lookup
               |
               V
          real_lookup

这些函数操作都属于虚拟文件系统操作, 对所有类型的文件系统都适用, 而从各个FS的具体实现才能看出差异和相关标志的作用.

4.3 open_namei_create

static int open_namei_create(struct nameidata *nd, struct path *path,
    int flag, int mode)
{
 int error;
// nd当前的dentry
 struct dentry *dir = nd->dentry;
 if (!IS_POSIXACL(dir->d_inode))
  mode &= ~current->fs->umask;
 error = vfs_create(dir->d_inode, path->dentry, mode, nd);
 mutex_unlock(&dir->d_inode->i_mutex);
 dput(nd->dentry);
 nd->dentry = path->dentry;
 if (error)
  return error;
 /* Don&apos;t check for write permission, don&apos;t truncate */
 return may_open(nd, 0, flag & ~O_TRUNC);
}

4.4 path_to_nameidata

// 将路径参数赋值到nameidata结构中
static inline void path_to_nameidata(struct path *path, struct nameidata *nd)
{
// 释放原来的目录项
 dput(nd->dentry);
// 如果挂接点也不同,释放掉原来的
 if (nd->mnt != path->mnt)
  mntput(nd->mnt);
// 将新路径参数赋值到nameidata结构中
 nd->mnt = path->mnt;
 nd->dentry = path->dentry;
}

5. 结论

打开文件时, 目的是要生成一个struct file的结构的指针, 该结构中有相关文件名的名称, dentry指针, 挂接点文件系统等信息, 而struct nameidata作为一个中间过程结构保存相关的处理结果, 最终返回需要的文件信息。

本文来自CSDN博客，转载请标明出处：http://blog.csdn.net/air_snake/archive/2008/07/22/2690554.aspx