BeeGFS源码分析2-客户端概要分析

BeeGFS 的客户端是由一个内核模块和两个系统服务组成的，这里我们主要分析内核模块。内核模块主要实现了一个 Linux 的文件系统，因此注册了一个文件系统类型。因为 BeeGFS 的目录树解析，是在父目录 DEntry 里找子目录或文件 DEntry ，逐级迭代完成的，因此在 Mount 文件系统时，需要从管理节点获取根元数据节点的 ID ，然后再向根元数据节点查询根目录的 DEntry 的信息，为后续的目录解析打下基础。

注册文件系统类型

init_fhgfs_client

• 内核模块初始化：

// fhgfs_client_module\source\program\Main.c

#define BEEGFS_LICENSE "GPL v2"

static int __init init_fhgfs_client(void)
{
#define fail_to(target, msg) \
   do { \
      printk_fhgfs(KERN_WARNING, msg "\n"); \
      goto target; \
   } while (0)

   if (!beegfs_fault_inject_init() )
      fail_to(fail_fault, "could not register fault-injection debugfs dentry");

   if (!beegfs_native_init() )
      fail_to(fail_native, "could not allocate emergency pools");

   if (!FhgfsOpsCommKit_initEmergencyPools() )
      fail_to(fail_commkitpools, "could not allocate emergency pools");

   if (!SocketTk_initOnce() )
      fail_to(fail_socket, "SocketTk initialization failed");

   if (!FhgfsOps_initInodeCache() )
      fail_to(fail_inode, "Inode cache initialization failed");

   if (!RWPagesWork_initworkQueue() )
      fail_to(fail_rwpages, "Page work queue registration failed");

   if (!FhgfsOpsRemoting_initMsgBufCache() )
      fail_to(fail_msgbuf, "Message cache initialization failed");

   if (!FhgfsOpsPages_initPageListVecCache() )
      fail_to(fail_pagelists, "PageVec cache initialization failed");

   if (FhgfsOps_registerFilesystem() )
      fail_to(fail_register, "File system registration failed");

   ProcFs_createGeneralDir();

   printk_fhgfs(KERN_INFO, "File system registered. Type: %s. Version: %s\n",
      BEEGFS_MODULE_NAME_STR, App_getVersionStr() );

   return 0;

fail_register:
   FhgfsOpsPages_destroyPageListVecCache();
fail_pagelists:
   FhgfsOpsRemoting_destroyMsgBufCache();
fail_msgbuf:
   RWPagesWork_destroyWorkQueue();
fail_rwpages:
   FhgfsOps_destroyInodeCache();
fail_inode:
   SocketTk_uninitOnce();
fail_socket:
   FhgfsOpsCommKit_releaseEmergencyPools();
fail_commkitpools:
   beegfs_native_release();
fail_native:
   beegfs_fault_inject_release();
fail_fault:
   return -EPERM;
}

static void __exit exit_fhgfs_client(void)
{
   ProcFs_removeGeneralDir();
   BUG_ON(FhgfsOps_unregisterFilesystem() );
   FhgfsOpsPages_destroyPageListVecCache();
   FhgfsOpsRemoting_destroyMsgBufCache();
   RWPagesWork_destroyWorkQueue();
   FhgfsOps_destroyInodeCache();
   SocketTk_uninitOnce();
   FhgfsOpsCommKit_releaseEmergencyPools();
   beegfs_native_release();
   beegfs_fault_inject_release();

   printk_fhgfs(KERN_INFO, "BeeGFS client unloaded.\n");
}

module_init(init_fhgfs_client)
module_exit(exit_fhgfs_client)

MODULE_LICENSE(BEEGFS_LICENSE);
MODULE_DESCRIPTION("BeeGFS parallel file system client (http://www.beegfs.com)");
MODULE_AUTHOR("Fraunhofer ITWM, CC-HPC");

FhgfsOps_registerFilesystem

• 初始化时，向内核注册 BeeGFS 文件系统类型：

// fhgfs_client_module\source\filesystem\FhgfsOpsSuper.c

static struct file_system_type fhgfs_fs_type =
{
   .name       = BEEGFS_MODULE_NAME_STR,
   .owner      = THIS_MODULE,
   .kill_sb    = FhgfsOps_killSB,
   //.fs_flags   = FS_BINARY_MOUNTDATA, // not required currently

#ifdef KERNEL_HAS_GET_SB_NODEV
   .get_sb     = FhgfsOps_getSB,
#else
   .mount      = FhgfsOps_mount, // basically the same thing as get_sb before
#endif
};

int FhgfsOps_registerFilesystem(void)
{
   return register_filesystem(&fhgfs_fs_type);
}

挂载文件系统

FhgfsOps_mount

• Mount 文件系统时，间接调用 FhgfsOps_fillSuper 来填充文件系统超级块。

// fhgfs_client_module\source\filesystem\FhgfsOps_versions.c

#ifdef KERNEL_HAS_GET_SB_NODEV

int FhgfsOps_getSB(struct file_system_type *fs_type,
   int flags, const char *dev_name, void *data, struct vfsmount *mnt)
{
   return get_sb_nodev(fs_type, flags, data, FhgfsOps_fillSuper, mnt);
}

#else

/* kernel 2.6.39 switched from get_sb() to mount(), which provides similar functionality from our point of view. */

struct dentry* FhgfsOps_mount(struct file_system_type *fs_type,
   int flags, const char *dev_name, void *data)
{
   return mount_nodev(fs_type, flags, data, FhgfsOps_fillSuper);
}

#endif // LINUX_VERSION_CODE

FhgfsOps_fillSuper

• 初始化文件系统实例的超级块，并初始化根目录的 inode ，此时 ID 只是简单的初始化为 0 ，后面会更新成真正的 ID ：

// fhgfs_client_module\source\filesystem\FhgfsOpsSuper.c

/**
 * Fill the file system superblock (vfs object)
 */
int FhgfsOps_fillSuper(struct super_block* sb, void* rawMountOptions, int silent)
{
   App* app = NULL;
   Config* cfg = NULL;

   struct inode* rootInode;
   struct dentry* rootDentry;
   struct kstat kstat;
   EntryInfo entryInfo;

   FhgfsIsizeHints iSizeHints;

   // init per-mount app object

   if(__FhgfsOps_constructFsInfo(sb, rawMountOptions) )
      return -ECANCELED;

   app = FhgfsOps_getApp(sb);
   cfg = App_getConfig(app);

   // set up super block data

   sb->s_maxbytes = MAX_LFS_FILESIZE;
   sb->s_blocksize = PAGE_SIZE;
   sb->s_blocksize_bits = PAGE_SHIFT;
   sb->s_magic = BEEGFS_MAGIC;
   sb->s_op = &fhgfs_super_ops;
   sb->s_time_gran = 1000000000; // granularity of c/m/atime in ns
   sb->s_flags |= MS_NODIRATIME;

   if (Config_getSysXAttrsEnabled(cfg ) )
      sb->s_xattr = fhgfs_xattr_handlers_noacl; // handle only user xattrs

#ifdef KERNEL_HAS_POSIX_GET_ACL
   if (Config_getSysACLsEnabled(cfg) )
   {
      sb->s_xattr = fhgfs_xattr_handlers; // replace with acl-capable xattr handlers
      sb->s_flags |= MS_POSIXACL;
   }
#endif // KERNEL_HAS_POSIX_GET_ACL

   /* MS_ACTIVE is rather important as it marks the super block being successfully initialized and
    * allows the vfs to keep important inodes in the cache. However, it seems it is already
    * initialized in vfs generic mount functions.
      sb->s_flags |= MS_ACTIVE; // used in iput_final()  */

   // NFS kernel export is probably not worth the backport efforts for kernels before 2.6.29
#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,29)
   sb->s_export_op = &fhgfs_export_ops;
#endif

#if defined(KERNEL_HAS_SB_BDI)
   sb->s_bdi = FhgfsOps_getBdi(sb);
#endif

   // init root inode

   memset(&kstat, 0, sizeof(struct kstat) );

   kstat.ino = BEEGFS_INODE_ROOT_INO;
   kstat.mode = S_IFDIR | 0777; // allow access for everyone
   kstat.atime = kstat.mtime = kstat.ctime = current_fs_time(sb);
   kstat.uid = FhgfsCommon_getCurrentKernelUserID();
   kstat.gid = FhgfsCommon_getCurrentKernelGroupID();
   kstat.blksize = Config_getTuneInodeBlockSize(cfg);
   kstat.nlink = 1;

   // root entryInfo is always updated when someone asks for it (so we just set dummy values here)
   EntryInfo_init(&entryInfo, NodeOrGroup_fromGroup(0), StringTk_strDup(""), StringTk_strDup(""),
      StringTk_strDup(""), DirEntryType_DIRECTORY, 0);

   rootInode = __FhgfsOps_newInode(sb, &kstat, 0, &entryInfo, &iSizeHints);
   if(!rootInode || IS_ERR(rootInode) )
   {
      __FhgfsOps_destructFsInfo(sb);
      return IS_ERR(rootInode) ? PTR_ERR(rootInode) : -ENOMEM;
   }

   rootDentry = d_make_root(rootInode);
   if(!rootDentry)
   {
      __FhgfsOps_destructFsInfo(sb);
      return -ENOMEM;
   }

#ifdef KERNEL_HAS_S_D_OP
   // linux 2.6.38 switched from individual per-dentry to defaul superblock d_ops.
   /* note: Only set default dentry operations here, as we don't want those OPs set for the root
    * dentry. In fact, setting as before would only slow down everything a bit, due to
    * useless revalidation of our root dentry. */
   sb->s_d_op = &fhgfs_dentry_ops;
#endif // KERNEL_HAS_S_D_OP

   rootDentry->d_time = jiffies;
   sb->s_root = rootDentry;

   return 0;
}

初始化文件系统

__FhgfsOps_constructFsInfo

• 申请内存，构造文件系统基本数据结构：

// fhgfs_client_module\source\filesystem\FhgfsOpsSuper.c

/**
 * Initialize sb->s_fs_info
 *
 * @return 0 on success, negative linux error code otherwise
 */
int __FhgfsOps_constructFsInfo(struct super_block* sb, void* rawMountOptions)
{
   int res;
   int appRes;
   App* app;
   Logger* log;


#if defined(KERNEL_HAS_SB_BDI) && !defined(KERNEL_HAS_SUPER_SETUP_BDI_NAME)
   struct backing_dev_info* bdi;
#endif

   // use kzalloc to also zero the bdi
   FhgfsSuperBlockInfo* sbInfo = kzalloc(sizeof(FhgfsSuperBlockInfo), GFP_KERNEL);
   if (!sbInfo)
   {
      printk_fhgfs_debug(KERN_INFO, "Failed to allocate memory for FhgfsSuperBlockInfo");
      sb->s_fs_info = NULL;
      return -ENOMEM;
   }

   sb->s_fs_info = sbInfo;

   appRes = __FhgfsOps_initApp(sb, rawMountOptions);
   if(appRes)
   {
      printk_fhgfs_debug(KERN_INFO, "Failed to initialize App object");
      res = -EINVAL;
      goto outFreeSB;
   }

   app = FhgfsOps_getApp(sb);
   log = App_getLogger(app);
   IGNORE_UNUSED_VARIABLE(log);

#if defined(KERNEL_HAS_SB_BDI)

   #if defined(KERNEL_HAS_SUPER_SETUP_BDI_NAME) && !defined(KERNEL_HAS_BDI_SETUP_AND_REGISTER)
   {
      static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);

      res = super_setup_bdi_name(sb, BEEGFS_MODULE_NAME_STR "-%ld",
            atomic_long_inc_return(&bdi_seq));
   }
   #else
      bdi = &sbInfo->bdi;

      /* NOTE: The kernel expects a fully initialized bdi structure, so at a minimum it has to be
       *       allocated by kzalloc() or memset(bdi, 0, sizeof(*bdi)).
       *       we don't set the congest_* callbacks (like every other filesystem) because those are
       *       intended for dm and md.
       */
      bdi->ra_pages = BEEGFS_DEFAULT_READAHEAD_PAGES;

      #if defined(KERNEL_HAS_BDI_CAP_MAP_COPY) 
         res = bdi_setup_and_register(bdi, BEEGFS_MODULE_NAME_STR, BDI_CAP_MAP_COPY);
      #else
         res = bdi_setup_and_register(bdi, BEEGFS_MODULE_NAME_STR);
      #endif
   #endif

   if (res)
   {
      Logger_logFormatted(log, 2, __func__, "Failed to init super-block (bdi) information: %d",
         res);
      __FhgfsOps_uninitApp(app);
      goto outFreeSB;
   }
#endif

   // set root inode attribs to uninit'ed

   FhgfsOps_setHasRootEntryInfo(sb, false);
   FhgfsOps_setIsRootInited(sb, false);


   printk_fhgfs(KERN_INFO, "BeeGFS mount ready.\n");

   return 0; // all ok, res should be 0 here

outFreeSB:
   kfree(sbInfo);
   sb->s_fs_info = NULL;

   return res;
}

__FhgfsOps_initApp

• 解析参数，继续初始化：

// fhgfs_client_module\source\filesystem\FhgfsOpsSuper.c

/**
 * Creates and initializes the per-mount application object.
 */
int __FhgfsOps_initApp(struct super_block* sb, char* rawMountOptions)
{
   MountConfig* mountConfig;
   bool parseRes;
   App* app;
   int appRes;

   // create mountConfig (parse from mount options)
   mountConfig = MountConfig_construct();

   parseRes = MountConfig_parseFromRawOptions(mountConfig, rawMountOptions);
   if(!parseRes)
   {
      MountConfig_destruct(mountConfig);
      return APPCODE_INVALID_CONFIG;
   }

   //printk_fhgfs(KERN_INFO, "Initializing App...\n"); // debug in

   app = FhgfsOps_getApp(sb);
   App_init(app, mountConfig);

   appRes = App_run(app);

   if(appRes != APPCODE_NO_ERROR)
   { // error occurred => clean up
      printk_fhgfs_debug(KERN_INFO, "Stopping App...\n");

      App_stop(app);

      printk_fhgfs_debug(KERN_INFO, "Cleaning up...\n");

      App_uninit(app);

      printk_fhgfs_debug(KERN_INFO, "App unitialized.\n");

      return appRes;
   }

   ProcFs_createEntries(app);

   return appRes;
}

App_run

• 初始化客户端基本组件：

// fhgfs_client_module\source\app\App.c
int App_run(App* this)
{
   // init data objects & storage

   if(!__App_initDataObjects(this, this->mountConfig) )
   {
      printk_fhgfs(KERN_WARNING,
         "Configuration error: Initialization of common objects failed. "
         "(Log file may provide additional information.)\n");
      this->appResult = APPCODE_INVALID_CONFIG;
      return this->appResult;
   }

   if(!__App_initInodeOperations(this) )
   {
      printk_fhgfs(KERN_WARNING, "Initialization of inode operations failed.");
      this->appResult = APPCODE_INITIALIZATION_ERROR;
      return this->appResult;
   }


   if(!__App_initStorage(this) )
   {
      printk_fhgfs(KERN_WARNING, "Configuration error: Initialization of storage failed\n");
      this->appResult = APPCODE_INVALID_CONFIG;
      return this->appResult;
   }

   // init components

   if(!__App_initComponents(this) )
   {
      printk_fhgfs(KERN_WARNING, "Component initialization error. "
         "(Log file may provide additional information.)\n");
      this->appResult = APPCODE_INITIALIZATION_ERROR;
      return this->appResult;
   }

   __App_logInfos(this);


   // start components

   __App_startComponents(this);

   // Note: We wait some ms for the node downloads here because the kernel would like to
   //    check the properties of the root directory directly after mount.

   InternodeSyncer_waitForMgmtInit(this->internodeSyncer, 1000);

   if(!__App_mountServerCheck(this) )
   { // mount check failed => cancel mount
      printk_fhgfs(KERN_WARNING, "Mount sanity check failed. Canceling mount. "
         "(Log file may provide additional information. Check can be disabled with "
         "sysMountSanityCheckMS=0 in the config file.)\n");
      this->appResult = APPCODE_INITIALIZATION_ERROR;

      return this->appResult;
   }

   // mark: mount succeeded if we got here!

   return this->appResult;
}

__App_initInodeOperations

• 初始化 inode 基本操作，以备后面新建 inode 时使用：

// fhgfs_client_module\source\app\App.c

/**
 * Initialized the inode_operations structs depending on what features have been enabled in
 * the config.
 */
bool __App_initInodeOperations(App* this)
{
   Config* cfg = App_getConfig(this);

   this->fileInodeOps = os_kzalloc(sizeof(struct inode_operations) );
   this->symlinkInodeOps = os_kzalloc(sizeof(struct inode_operations) );
   this->dirInodeOps = os_kzalloc(sizeof(struct inode_operations) );
   this->specialInodeOps = os_kzalloc(sizeof(struct inode_operations) );

   if (!this->fileInodeOps || !this->symlinkInodeOps ||
       !this->dirInodeOps || !this->specialInodeOps)
   {
      SAFE_KFREE(this->fileInodeOps);
      SAFE_KFREE(this->symlinkInodeOps);
      SAFE_KFREE(this->dirInodeOps);
      SAFE_KFREE(this->specialInodeOps);

      return false;
   }

   this->fileInodeOps->getattr     = FhgfsOps_getattr;
   this->fileInodeOps->permission  = FhgfsOps_permission;
   this->fileInodeOps->setattr     = FhgfsOps_setattr;

#ifdef KERNEL_HAS_GENERIC_READLINK
   this->symlinkInodeOps->readlink    = generic_readlink; // default is fine for us currently
#endif
#ifdef KERNEL_HAS_GET_LINK
   this->symlinkInodeOps->get_link    = FhgfsOps_get_link;
#else
   this->symlinkInodeOps->follow_link = FhgfsOps_follow_link;
   this->symlinkInodeOps->put_link    = FhgfsOps_put_link;
#endif
   this->symlinkInodeOps->getattr     = FhgfsOps_getattr;
   this->symlinkInodeOps->permission  = FhgfsOps_permission;
   this->symlinkInodeOps->setattr     = FhgfsOps_setattr;

#ifdef KERNEL_HAS_ATOMIC_OPEN
   #ifdef BEEGFS_ENABLE_ATOMIC_OPEN
   this->dirInodeOps->atomic_open = FhgfsOps_atomicOpen;
   #endif // BEEGFS_ENABLE_ATOMIC_OPEN
#endif
   this->dirInodeOps->lookup      = FhgfsOps_lookupIntent;
   this->dirInodeOps->create      = FhgfsOps_createIntent;
   this->dirInodeOps->link        = FhgfsOps_link;
   this->dirInodeOps->unlink      = FhgfsOps_unlink;
   this->dirInodeOps->mknod       = FhgfsOps_mknod;
   this->dirInodeOps->symlink     = FhgfsOps_symlink;
   this->dirInodeOps->mkdir       = FhgfsOps_mkdir;
   this->dirInodeOps->rmdir       = FhgfsOps_rmdir;
   this->dirInodeOps->rename      = FhgfsOps_rename;
   this->dirInodeOps->getattr     = FhgfsOps_getattr;
   this->dirInodeOps->permission  = FhgfsOps_permission;
   this->dirInodeOps->setattr     = FhgfsOps_setattr;

   this->specialInodeOps->setattr = FhgfsOps_setattr;

   if (Config_getSysXAttrsEnabled(cfg) )
   {
      this->fileInodeOps->listxattr   = FhgfsOps_listxattr;
      this->dirInodeOps->listxattr   = FhgfsOps_listxattr;

#ifdef KERNEL_HAS_GENERIC_GETXATTR
      this->fileInodeOps->getxattr    = generic_getxattr;
      this->fileInodeOps->removexattr = FhgfsOps_removexattr;
      this->fileInodeOps->setxattr    = generic_setxattr;

      this->dirInodeOps->getxattr    = generic_getxattr;
      this->dirInodeOps->removexattr = FhgfsOps_removexattr;
      this->dirInodeOps->setxattr    = generic_setxattr;
#endif

      if (Config_getSysACLsEnabled(cfg) )
      {
#ifdef KERNEL_HAS_POSIX_GET_ACL
         this->fileInodeOps->get_acl = FhgfsOps_get_acl;
         this->dirInodeOps->get_acl  = FhgfsOps_get_acl;
         // Note: symlinks don't have ACLs
#ifdef KERNEL_HAS_SET_ACL
         this->fileInodeOps->set_acl = FhgfsOps_set_acl;
         this->dirInodeOps->set_acl  = FhgfsOps_set_acl;
#endif // LINUX_VERSION_CODE
#else
         Logger_logErr(this->logger, "Init inode operations",
            "ACLs activated in config, but not supported on this kernel version.");
         return false;
#endif // KERNEL_HAS_POSIX_GET_ACL
      }
   }

   return true;
}

创建和初始化Inode

由 FhgfsOps_fillSuper 函数初始化调用。

__FhgfsOps_newInode

• 创建新的 Inode 时，会调用此函数，根据父目录的 DEntry 信息（其中保存有父目录所在的元数据节点 ID ，以及目录 ID ），访问相应的元数据节点进行子目录或者文件的操作：

// fhgfs_client_module\source\filesystem\FhgfsOpsInode.h

/**
 * See __FhgfsOps_newInodeWithParentID for details. This is just a wrapper function.
 */
struct inode* __FhgfsOps_newInode(struct super_block* sb, struct kstat* kstat, dev_t dev,
   EntryInfo* entryInfo,  FhgfsIsizeHints* iSizeHints)
{
      return __FhgfsOps_newInodeWithParentID(sb, kstat, dev, entryInfo, (NumNodeID){0}, iSizeHints);
}

/**
 * Creates a new inode, inits it from the kstat, inits the ops (depending on the mode)
 * and hashes it.
 *
 * Note: Make sure everything is set in the kstat _before_ you call this, because we hash
 * the inode in here (so it can be found and accessed by others when this method returns).
 * Note: Consider using the _instantiateInode()-wrapper instead of calling this directly for new
 * files/dirs.
 *
 * @param kstat must have a valid .ino (inode number)
 * @param dev set to 0 if not required (only used for special files)
 * @param entryInfoPtr contained strings will just be moved to the new inode or free'd in case of an
 * error (or cached inode), so don't access the given entryInfoPtr anymore after calling this.
 * @param parentNodeID: usually 0, except for NFS export callers, which needs it to connect dentries
 *    with their parents. By default dentries are connected to their parents, so usually this
 *    is not required (nfs is an exception).
 * @return NULL if not successful
 */
struct inode* __FhgfsOps_newInodeWithParentID(struct super_block* sb, struct kstat* kstat,
   dev_t dev, EntryInfo* entryInfo, NumNodeID parentNodeID, FhgfsIsizeHints* iSizeHints)
{
   App* app = FhgfsOps_getApp(sb);
   Config* cfg = App_getConfig(app);

   FhgfsInode* fhgfsInode;

   FhgfsInodeComparisonInfo comparisonInfo =
   {
      .inodeHash = kstat->ino, // pre-set by caller
      .entryID = entryInfo->entryID,
   };


   // check inode cache for an existing inode with this ID (and get it) or allocate a new one

   struct inode* inode = iget5_locked(sb, kstat->ino,
      __FhgfsOps_compareInodeID, __FhgfsOps_initNewInodeDummy, &comparisonInfo);

   if(unlikely(!inode || IS_ERR(inode) ) )
      goto cleanup_entryInfo; // allocation of new inode failed

   fhgfsInode = BEEGFS_INODE(inode);

   if( !(inode->i_state & I_NEW) )
   {  // Found an existing inode, which is possibly actively used. We still need to update it.
      FhgfsInode_entryInfoWriteLock(fhgfsInode); // LOCK EntryInfo
      FhgfsInode_updateEntryInfoUnlocked(fhgfsInode, entryInfo);
      FhgfsInode_entryInfoWriteUnlock(fhgfsInode); // UNLOCK EntryInfo

      spin_lock(&inode->i_lock);

      __FhgfsOps_applyStatDataToInodeUnlocked(kstat, iSizeHints, inode); // already locked
      Time_setToNow(&fhgfsInode->dataCacheTime);
      spin_unlock(&inode->i_lock);

      goto outNoCleanUp; // we found a matching existing inode => no init needed
   }

   fhgfsInode->parentNodeID = parentNodeID;

   /* note: new inodes are protected by the I_NEW flag from access by other threads until we
    *       call unlock_new_inode(). */

   // init this fresh new inode...

   // no one can access inode yet => unlocked
   __FhgfsOps_applyStatDataToInodeUnlocked(kstat, iSizeHints, inode);

   inode->i_ino = kstat->ino; // pre-set by caller

   inode->i_flags |= S_NOATIME | S_NOCMTIME; // timestamps updated by server

   mapping_set_gfp_mask(&inode->i_data, GFP_USER); // avoid highmem for page cache pages

   // move values (no actual string copy)
   fhgfsInode->entryInfo = *entryInfo;

   switch (kstat->mode & S_IFMT)
   {
      case S_IFREG: // regular file
      {
         if(Config_getTuneFileCacheTypeNum(cfg) == FILECACHETYPE_Native)
         {
            inode->i_fop = &fhgfs_file_native_ops;
            inode->i_data.a_ops = &fhgfs_addrspace_native_ops;
         }
         else
         if(Config_getTuneFileCacheTypeNum(cfg) == FILECACHETYPE_Paged)
         { // with pagecache
            inode->i_fop = &fhgfs_file_pagecache_ops;
            inode->i_data.a_ops = &fhgfs_address_pagecache_ops;
         }
         else
         { // no pagecache (=> either none or buffered cache)
            inode->i_fop = &fhgfs_file_buffered_ops;
            inode->i_data.a_ops = &fhgfs_address_ops;
         }

         #ifdef KERNEL_HAS_ADDRESS_SPACE_BDI
            inode->i_data.backing_dev_info = FhgfsOps_getBdi(sb);
         #endif

         inode->i_op = App_getFileInodeOps(app);
      } break;

      case S_IFDIR: // directory
      {
         inode->i_op = App_getDirInodeOps(app);
         inode->i_fop = &fhgfs_dir_ops;
      } break;

      case S_IFLNK: // symlink
      {
         inode->i_op = App_getSymlinkInodeOps(app);
      } break;

      default: // pipes and other special files
      {
         inode->i_op = App_getSpecialInodeOps(app);
         init_special_inode(inode, kstat->mode, dev);
      } break;
   }


   unlock_new_inode(inode); // remove I_NEW flag, so the inode can be accessed by others

   return inode;


   // error occured
cleanup_entryInfo:
   EntryInfo_uninit(entryInfo);

   // found an existing inode
outNoCleanUp:
   return inode;
}

App_getFileInodeOps

• 最后根据文件类型，赋予之前初始化好的 inode 操作指针：

// fhgfs_client_module\source\app\App.h

struct inode_operations* App_getFileInodeOps(App* this)
{
   return this->fileInodeOps;
}

struct inode_operations* App_getSymlinkInodeOps(App* this)
{
   return this->symlinkInodeOps;
}

struct inode_operations* App_getDirInodeOps(App* this)
{
   return this->dirInodeOps;
}

struct inode_operations* App_getSpecialInodeOps(App* this)
{
   return this->specialInodeOps;
}