You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 12 Next »

Introduction

If files are accidentally or maliciously deleted from a file system, the user data may be permanently lost. The Trash Can is a useful feature in file systems that acts as a temporary holding area, allowing users to store deleted files for a short time before they are permanently deleted. It provides a mechanism to restore or retrieve deleted files if needed, and automatically deletes the files once they become too old or the filesystem is too full.

When the Trash Can feature is enabled, when a user deletes a file from a file system, it is not immediately deleted but moved to the Trash Can.  Deleted files and directories are temporarily stored in the Trash Can. Files and directories in the Trash Can may be restored or retrieved individually or in bulk if they are still available. The Trash Can may be manually emptied, or once the filesystem is nearly full the system will automatically empty files from the Trash, taking into account which users and projects are consuming the most space.

The Trash Can should including the following functionalities:

  • Files should be added to the Trash can during normal usage (e.g. rm or rmdir utility or unlink() and rmdir() syscalls, or if a file is renamed onto another one).
  • There should be a per-UID space for files in the Trash Can, so that administrative tools can easily find files for each user

  • Restore a file in the Trash Can. This will restore a file or directory to its original path. The corresponding User, Group, and Project quota account should also be increased again.

  • List "undeleted" files in the Trash Can.

  • After a file is deleted and moved into Trash Can, the User, Group, and Project quota for this file should be accounted and reduced accordingly.

  • A file in the Trash Can is not visible in the namespace of the file system.

  • A file or directory in the Trash Can is marked with a TRASH flag, so that tools like lipe_find3 can optionally skip/find deleted files.
  • A file in the Trash Can marked with the TRASH flag cannot be read by a user or application, to prevent applications continuing to read these files.

  • Permanently remove a file in the Trash Can. This will remove the file from the Trash can and destroy the OST objects to free the used space. The file is now unrecoverable.

  • Empty the Trash Can. This will remove all files in the Trash Can.

  • Files should be deleted from Trash Can after a specified retention period, such as 7 days. 

  • Files should be deleted from Trash Can when an OST approaches a capacity threshold (over 80% for HDD, or 90% for SSD) to avoid performance impact or the risk of running out of space if a large number of files are written at once.

  • Have a tunable parameter to enable/disable Trash Can feature on a entire file system.

  • A administrator can enable/disable Trash Can feature on a specified file or directory by setting the NOTRASH flag on the file.

Deleted files can no longer be restored from the Trash Can when:

  • A file (or directory) is deleted from the Trash Can. In other words it have been deleted twice. The first deletion only moves the file to the Trash Can. The second deletion actually removes the file from the file system.

  • The Trash Can is emptied of all of its contents.

Design and Implementation

The design for the Trash Can feature in Lustre is relatively straight forward.

On the server side, the MDS implements the basic functionalities such as moving the "deleted" files into the Trash Can, and the interface how to traverse them. On the client side, it implements the basic utility tools to interact with the Trash Can ("lfs trash set|clear|list|delete|restore FILE|DIR"), including:

  • Set or clear the TRASH flag on a given file or directory
  • list files in the Trash Can on a given MDT
  • Permanently delete a file or directory within the Trash Can on a given MDT
  • Empty the Trash Can on a given MDT
  • Restore a file within the Trash Can on a give MDT

Our mechanism only moves the regular files into the Trash Can upon its last unlink.

It borrows lots of ideas from orphan and volatile files in Lustre (which stores in "ROOT/PENDING" directory on each MDT). During the format and setup, each MDT creates a "ROOT/TRASH" directory as a Trash Can to store "undeleted" files.

The POSIX API is used to traverse the files under the Trash Can on a given MDT. First, a client can get the FID of Trash Can directory ROOT/TRASH on the MDT. Then the client can get the file handle via FID open: dir_fd=llapi_open_by_fid(). After that, the "undeleted" files within the Trash Can can be traversed via readdir(dir_fd); it can open by openat(dir_fd, ent->d_name) and obtain the "trusted.unrm" XATTR, which contains the necessary information to resotre, via fgetxattr(fd, "trusted.recyclebin"); The client can even read the data or swap layouts of the "undeleted" file on the Trash Can for restore: opendir()/readddir()/openat()/fgetxattr("trusted.recyclebin")/close()/closedir().

The workflow for the Trash Can is as follows:

  • An administrator can enable/disable Trash Can feature on a specified MDT via: lctl set_param mdd.*.trash_can_enable

  • An adminstrator can enable/disable Trash Can feature on a specified directory or a file via the file flag: FS_UNRM_FL; All sub files under a directory flagged with FS_UNRM_FL can inherit this flag;    

     # lfs trash set $file|$dir
     # lfs trash clear $file|$dir

  • Move a deleting file into the Trash Can. When delete a regular file while Trash Can is enabled will mark it with FS_UNRM_FL upon its last unlink, and move the file into the Trash Can directory "ROOT/.lustre/Trash/MDTxxxx/UID/pFID". Then set a "trusted.unrm" XATTR on the deleted file on the Trash Can. The XATTR contains the following information:

     struct ll_trash_xattr {
        __u32 ltx_flags;      /* for future usage */
        __u32 ltx_uid;        /* original UID of the file, used to restore on unrm */
        __u32 ltx_gid;        /* original GID of the file, used to restore on unrm */
        __u32 ltx_projid;     /* original PROJID of the file, used to restore on unrm */
        __u64 ltx_timestamp;  /* time the file moved into the Trash Can, or we could use ctime here? */
     };

Where ltx_uid/ltx_gid/ltx_projid are the original UID/GID/PROJID of the deleted file, mainly used for the restore operation; @ltx_timestamp is the time that the file was moved into the Trash Can. It is used to determine whether the file is expired for the specified retention period and thus should be removed from the Trash Can finally (maybe we could also use the inode ctime for this purpose instead of storing a separate timestamp?). During deleting the file, we can get the full path information via the way similar to fid2path().

  • List "undeleted" files within a Trash Can.  By default it will list files/directories deleted relative to the current working directory.  If DIR is provided, then list deleted files/directories relative to that directory, in the same format as ls:

     # lfs trash {list|ls} [DIR|FILE]
     MDT index: 1
     uid gid size delete time  FID                    Fullpath
     0   0   4096 Nov 14 08:11 [0x200034021:0x1:0x0]->/mnt/lustre/f1
     0   0  32104 Nov 14 08:07 [0x200034021:0x2:0x0]->/mnt/lustre/dir/f2
     ...

Internally, the lfs trash list command is looking up the FID and MDT of the current directory, or the directory specified by DIR, and then listing the respective directory under $MOUNT/.lustre/trash/MDTxxxx/pFID/ or the directory file descriptor returned via llapi_recycle_fid_get(MNTPT, mdt) if the .lustre/trash directory is not available.  This is mainly for debugging, since users will generally use the virtual .Trash directory to interact with the Trash Can and restore files.



  • Deleting a file or directory in the Trash Can will remove the temporary file under "ROOT/.lustre/Trash" and free the data space on Lustre OSTs permanently:

     # lfs trash {delete|rm} [DIR/]FILE ...

  • Empty a Trash Can:

     # lfs trash clear DIR ...

  • Restore a file in the Trash Can on a given MDT. It will restore the file and its content according to the saved full path and then delete the stub on the Trash Can.

     # lfs trash {restore|unrm} [DIR/]FILE ...

  • A utility periodically scans the files under Trash Can directory "ROOT/TRASH" and delete the file with grace time expiration.

  • Provide the functionality to scan files in the trash on all MDTs that exceed the specified age manually:

     # lfs trash find -ctime +time [DIR]

  • Provide the functionality to restore/delete all files within a given directory. This can be achieved by using the command combination of "lfs trash list" and "lfs trash restore" or "lfs trash delete" to filter the files with the full path attribute under a given directory.

  • Provide .lustre/trash/MDTnnnn (where nnnn  is the MDT index) filesystem namespace. By this way, users can access the "undeleted" files with readonly mode under the Trash Can directory on a given MDTnnnn via POSIX file system API. However, we can not access these files from fileset sub directory mount. We can perform the following commands from a Lustre namespace (mount point of "/mnt/lustre") on a client:

     # ls /mnt/lustre/.lustre/trash/MDT0002
     0x200034021:0x1:0x0
     0x200034021:0x2:0x0
     ...
     # lfs trash ls /mnt/lustre/.lustre/recycle/MDT0002/0x200034021:0x1:0x0
     0 0 4096 Nov 14 08:11 [0x200034021:0x1:0x0]->/mnt/lustre/f1
     # lfs trash list /mnt/lustre/.lustre/recycle/MDT0002
     MDT index: 1
     uid gid   size delete time FID                    Fullpath
     0 0 4096  Nov 14 08:11     [0x200034021:0x1:0x0]->/mnt/lustre/f1
     0 0 32104 Nov 14 08:07     [0x200034021:0x2:0x0]->/mnt/lustre/dir/f2
     ...

Clean up files from the trash

It needs to automatically clean up files from the trash can when the filesystem becomes full. It cannot be that the user has to delete every file twice, and it cannot be that the filesystem is allowed to get 100% full (or even 90% full) due to files in the trash. There needs to be an automatic mechanism to clean up the trash to ensure that the filesystem performance does not degrade when users though they deleted files.

It can assign the UID/GID/PROJID to a trash user so that this quota is not accounted against the end user, and keep the original UID/GID/PROJID in a the XATTR "trusted.unrm".

In our design, it does not depend on a userspace utility for such a critical function to clean up files from the trash when FS is nearly full, since that utility may never be started, or the client is evicted, or similar. If that happens, the filesystem would become full and unusable, even though the user had already deleted files from the filesystem. This needs to be bulletproof and run automatically when the OSTs (or MDTs) are getting full.

The MDS is already monitoring the OST fullness every 5s to make object allocation decisions, so it can also make decisions about files to delete. Therefore, the MDT can periodically monitor the space usage of the trash user (quota) and space usage for the entire file system with the additional consideration of the retention period and deleted timestamp for the files, choose the candidates to be deleted permanently to free up the space.

Also, there needs to be some accounting of files in the Trash Can, so that "df" does not show the filesystem as 90% full all the time, but only show the non-trash space usage (= real usage - trash usage).

Per-User Trash Can

A per-user Trash/MDTxxxx/UID/ directory that is owned by that UID and mode 0600 should always be created in the top-level directory to avoid world readable access to deleted files, and to de-conflict files/directories of the same name created by users (e.g. tmp/ or data/ or Documents/ or similar. That avoids exposing files to other users that may be private, and also allows tracking space usage more clearly for each UID, so that a user's data can be found and purged more quickly if they are exceeding their quotas.

Per-Tenant Trash Can

Files and directories deleted from within a subdirectory mount of a Nodemap should be stored in a Trash/MDTxxxx/NODEMAP/UID/ directory to isolate the files/directories from different tenants.  The NODEMAP/ directoryname is the configured name of the nodemap for that tenant, and can be found from the client export used to perform the final unlink operation. The UID/ directory name should be the unmapped ID of the user, so that the visible directory name matches the user expectation.  The UID directory ownership should be the mapped ID of the user, so that proper file access controls can be maintained.  By having the multi-level NODEMAP/UID/ naming, it isolates the UID directory names from other tenants that may have the same mapped UID directory name.

The Nodemap for a tenant should allow configuring the UID/GID/PROJID to which files in the Trash are assigned.  These IDs should be within the ID offset range of the Tenant (e.g. 99999) so that they can be accessed and mapped correctly, but are unlikely to cause conflicts with other IDs used by the Tenant.  This will also allow the Tenant project quota group to account for all space used by the tenancy, while still separating Trash Can usage for the regular UID/GID/PROJID of the Tenant users.

In a multi-tenant environment, it would be desirable to have a more sophisticated policy engine to manage Garbage Collection of files within the trash, in order to provide maximum utility to each Tenant.  For example, if Tenant 1 has deleted files an hour ago, but Tenant 2 has written and deleted TB of data since that time, the Tenant 1 files may have expired out of the Trash Can.  Developing a complex policy engine to manage GC in an MTFL environment is out of scope for the initial TCU implementation.  We likely want to leverage and enhance the lpurge utility from Hot Pools to actively monitor the space usage of tenants on different OSTs to decide which objects (files) should be removed.

Repeated deletion of same filename

If the same filename is repeatedly created and deleted within the same parent directory, then the deleted files will have conflicts when moved into the pFID directory in the Trash Can.  To disambiguate the files in Trash, the conflicting filenames should be disambiguated by appending a timestamp to the filename, like filename.2025-04-03-00:11:24, possibly adding .microseconds if there is still a conflict.  It isn't totally clear whether it would be better to use the timestamp from when the file was deleted, or when the file was created.  Both have some value to help users distinguish between the different versions.

In order to avoid overwhelming the Trash Can with files that are rapidly created and deleted (e.g. short-lived temporary files), it would be desirable to impose an upper limit on the number of versions that will be saved in the trash can.

Avoid preserving temporary files

Files that only exist for a very short time (e.g. temporary files) should not necessarily be preserved in the Trash Can, or they can quickly overwhelm the available capacity of the filesystem, and result in important files being purged from trash and/or filling the trash faster than files can be cleaned up. Files marked with the I_LINKABLE flag on the MDS (from O_TMPFILE or Lustre Volatile files, see LU-18844) should not be preserved in the Trash Can.  It would be useful to have a tunable parameter that sets a minimum age for files to be preserved in the Trash Can (e.g. 65 minutes?) so that files that are frequently created and deleted are not preserved since they could consume a considerable amount of space.

JobID of process deleting a file

In LU-13031 the JobID of the process that first creates a file is stored in the user.job xattr on the MDT inode, for diagnostic purposes and to allow determining provenance of each file later on.  For the Trash Can, LU-17648 describes storing the JobID of the process that is deleting the file, for diagnostics such as determining rogue processes that are deleting files in the filesystem.  Something like user.del would be a reasonable default xattr name.  The actual xattr name can be configured with the mdt.*.job_xattr_del parameter.

.Trash Virtual Directory Support 

.Trash virtual directory

It would useful to implement a virtual ".Trash" subdirectory accessible in each directory in the filesystem that can be used to browse files/directories in the Trash Can and access them for recovery.

The FID of the ".Trash" directory is derived from the FID of the parent directory (pFID), by looking up the corresponding "stub" directory with the FID-named directory: ".lustre/trash/MDTXXXX/UID/pFID". Essentially ".Trash" under each normal directory is just a virtual shortcut to the stub directory (if the parent is not a striped dir) that is accessible in each directory if specified by name ".Trash". The files/directories under ".Trash" directory can be access via normal POSIX file system API such as via readdir()/stat()/getxattr() so that it can be used by normal tools such as "ls -l .Trash/" or "find .Trash" to locate files for restoration or permanent removal.  If there are no deleted files under a specific directory, then the virtual .Trash directory will not be accessible, and will return -ENOENT for any lookup

.Trash pFID name lookup

The FID-based names of stub directories stored as .lustre/trash/MDTxxxx/UID/pFID directory are needed for efficient lookup of the parent FID during unlink.  However, these directory names are not very user-friendly when browsing the virtual .Trash directory in the filesystem namespace.  Rather than showing the pFID name to users during readdir() calls (ls, find, etc.) it would be better to look up the actual parent directory name via the FID→trusted.link xattr on the parent and return this to clients.  The FID number of the directory entry would be the FID of the stub directory itself, not the pFID that is used internally for identification.  While copying the trusted.link xattr over to the stub directory at creation would simplify this lookup, there is some risk that the name in the trusted.link xattr would become stale if the parent directory is renamed.  On the other hand, this may also be useful to preserve the original name of the directory in case some automated tool is renaming the original directory to a temporary name before deletion?

.Trash striped directory

For a striped directory, its ".Trash" directory is also a vitual striped directory with each stripe on the same location (MDTs) where the shard FID is the FID of the corresponding stub directory on that MDT. If the stub directory on a certain MDT does not exist (or not create yet), the client lookup() or readdir() under ".Trash" directory should skip the stripe. The master FID of the virtual ".Trash" directory could be same with the FID of the parent directory but with f_ver setting with 1 (FID_VERSION_TRASH = 1) to distinguish them.

To avoid the inconsistent problem, each access on the virtual striped ".Trash" should check and revalidate the virtual stripe LMV EA. For example, It should add the new shards into the stripe EA after a new stub directory on a certain MDT was created.

.Trash directory migration

It should handle the case that a directory was restriped and the LMV layout was changed. In this case, the files under the directory will be migrated to another MDT. To simplify the implementation, we do not migrate the files according to the new LMV layout in the Trash Can. This may result in the lookup() operation will be issued to a wrong MDT and return -ENOENT  wrongly (after files in the trash can are restored). However, the readdir() operations will still return all the dir entries in the striped trash even if the parent LMV layout was re-striped and changed, since the parent directory FID (pFID) will remain the same as before restriping. Maybe it needs to migrate the files restored from the trash can to the appropriate shard according to their name hash once the LMV layout has been changed.

Orphans in Trash Can

For an orphan file, it means the file is still opened (but not closed) by a certain user. Upon its last unlink, it can directly move into the trash can and mark with LUSTRE_ORPHAN_FL | LUSTRE_TRASH_FL. And the orphans file can not be permanently deleted from the trash can until its last close(). 





  • No labels