全部博文（68）
分类： LINUX
2005-05-18 12:40:40
Traditionally, systems that support the POSIX (Portable Operating System Interface) family of standards [,] share a simple yet powerful file system permission model: Every file system object is associated with three sets of permissions that define access for the owner, the owning group, and for others. Each set may contain Read (r), Write (w), and Execute (x) permissions. This scheme is implemented using only nine bits for each object. In addition to these nine bits, the Set User Id, Set Group Id, and Sticky bits are used for a number of special cases. Many introductory and advanced texts on the UNIX operating system describe this model [].
Although the traditional model is extremely simple, it is sufficient for implementing the permission scenarios that usually occur on UNIX systems. System administrators have also found several workarounds for the model's limitations. Some of these workarounds require nonobvious group setups that may not reflect organizational structures. Only the root user can create groups or change group membership. Set-user-ID root utilities may allow ordinary users to perform some administrative tasks, but bugs in such utilities can easily lead to compromised systems. Some applications like FTP daemons implement their own extensions to the file system permission model []. The price of playing tricks with permissions is an increase in complexity of the system configuration. Understanding and maintaining the integrity of systems becomes more difficult.
Engineers have long recognized the deficiencies of the traditional permission model and have started to think about alternatives. This has eventually resulted in a number of Access Control List (ACL) implementations on UNIX, which are only compatible among each other to a limited degree.
This paper gives an overview of the most successful ACL scheme for UNIX-like systems that has resulted from the POSIX 1003.1e/1003.2c working group.
After briefly describing the concepts, some examples of how these are used are given for better understanding. Following that, the paper discusses Extended Attributes, the abstraction layer upon which ACLs are based on Linux. The rest of the paper deals with implementation, performance, interoperability, application support, and system maintenance aspects of ACLs.
The author was involved in the design and implementation of extended attributes and ACLs on Linux, which covered the user space tools and the kernel implementation for Ext2 and Ext3, Linux's most prominent file systems. Parts of the design of the system call interface are attributed to Silicon Graphics's Linux XFS project, particularly to Nathan Scott.
A need for standardizing other security relevant areas in addition to just ACLs was also perceived, so eventually a working group was formed to define security extensions within the POSIX 1003.1 family of standards. The document numbers 1003.1e (System Application Programming Interface) and 1003.2c (Shell and Utilities) were assigned for the working group's specifications. These documents are referred to as POSIX.1e in the remainder of this paper. The working group was focusing on the following extensions to POSIX.1: Access Control Lists (ACL), Audit, Capability, Mandatory Access Control (MAC), and Information Labeling.
Unfortunately, it eventually turned out that standardizing all these diverse areas was too ambitious a goal. In January 1998, sponsorship for 1003.1e and 1003.2c was withdrawn. While some parts of the documents produced by the working group until then were already of high quality, the overall works were not ready for publication as standards. It was decided that draft 17, the last version of the documents the working group had produced, should be made available to the public. Today, these documents can be found at Winfried Trümper's Web site [].
Several UNIX system vendors have implemented various parts of the security extensions, augmented by vendor-specific extensions. The resulting versions of their operating systems have often been labeled ``trusted'' operating systems, e.g., Trusted Solaris, Trusted Irix, Trusted AIX. Some of these ``trusted'' features have later been incorporated into the vendors' main operating systems.
ACLs are supported on different file system types on almost all UNIX-like systems nowadays. Some of these implementations are compatible with draft 17 of the specification, while others are based on older drafts. Unfortunately, this has resulted in a number of subtle differences among the different implementations.
The TrustedBSD project () lead by Robert Watson has implemented versions of the ACL, Capabilities, MAC, and Audit parts of POSIX.1e for FreeBSD. The ACL and MAC implementations appear in FreeBSD-RELEASE as of January, 2003. The MAC implementation is still considered experimental.
Patches that implement POSIX 1003.1e draft 17 ACLs have been available for various versions of Linux for several years now. They were added to version 2.5.46 of the Linux kernel in November 2002. Current Linux distributions are still based on the 2.4.x stable kernels series. SuSE and the United Linux consortium have integrated the 2.4 kernel ACL patches earlier than others, so their current products offer the most complete ACL support available for Linux to date. Other vendors apparently are still reluctant to make that important change, but experimental versions are expected to be available later this year.
The Linux getfacl and setfacl command line utilities do not strictly follow POSIX 1003.2c draft 17, which shows mostly in the way they handle default ACLs. See section .
At the time of this writing, ACL support on Linux is available for the Ext2, Ext3, IBM JFS, ReiserFS, and SGI XFS file systems. Solaris-compatible ACL support for NFS version 3 exists since March 3, 2003.
The traditional POSIX file system object permission model defines three classes of users called owner, group, and other. Each of these classes is associated with a set of permissions. The permissions defined are read (r), write (w), and execute (x). In this model, the owner class permissions define the access privileges of the file owner, the group class permissions define the access privileges of the owning group, and the other class permissions define the access privileges of all users that are not in one of these two classes. The ls -l command displays the owner, group, and other class permissions in the first column of its output (e.g., ``-rw-r- --'' for a regular file with read and write access for the owner class, read access for the group class, and no access for others).
An ACL consists of a set of entries. The permissions of each file system object have an ACL representation, even in the minimal, POSIX.1-only case. Each of the three classes of users is represented by an ACL entry. Permissions for additional users or groups occupy additional ACL entries.
Table shows the defined entry types and their text forms. Each of these entries consists of a type, a qualifier that specifies to which user or group the entry applies, and a set of permissions. The qualifier is undefined for entries that require no qualification.
ACLs equivalent with the file mode permission bits are called minimal ACLs. They have three ACL entries. ACLs with more than the three entries are called extended ACLs. Extended ACLs also contain a mask entry and may contain any number of named user and named group entries.
|
These named group and named user entries are assigned to the group class, which already contains the owning group entry. Different from the POSIX.1 permission model, the group class may now contain ACL entries with different permission sets, so the group class permissions alone are no longer sufficient to represent all the detailed permissions of all ACL entries it contains. Therefore, the meaning of the group class permissions is redefined: under their new semantics, they represent an upper bound of the permissions that any entry in the group class will grant.
This upper bound property ensures that POSIX.1 applications that are unaware of ACLs will not suddenly and unexpectedly start to grant additional permissions once ACLs are supported.
In minimal ACLs, the group class permissions are identical to the owning group permissions. In extended ACLs, the group class may contain entries for additional users or groups. This results in a problem: some of these additional entries may contain permissions that are not contained in the owning group entry, so the owning group entry permissions may differ from the group class permissions.
This problem is solved by the virtue of the mask entry. With minimal ACLs, the group class permissions map to the owning group entry permissions. With extended ACLs, the group class permissions map to the mask entry permissions, whereas the owning group entry still defines the owning group permissions. The mapping of the group class permissions is no longer constant. Figure shows these two cases.
 |
When an application changes any of the owner, group, or other class permissions (e.g., via the chmod command), the corresponding ACL entry changes as well. Likewise, when an application changes the permissions of an ACL entry that maps to one of the user classes, the permissions of the class change.
The group class permissions represent the upper bound of the permissions granted by any entry in the group class. With minimal ACLs this is trivially the case. With extended ACLs, this is implemented by masking permissions (hence the name of the mask entry): permissions in entries that are a member of the group class which are also present in the mask entry are effective. Permissions that are absent in the mask entry are masked and thus do not take effect. See Table .
| ||||||||||||
The owner and other entries are not in the group class. Their permissions are always effective and never masked.
So far we have only looked at ACLs that define the current access permissions of file system objects. This type is called access ACL. A second type called default ACL is also defined. They define the permissions a file system object inherits from its parent directory at the time of its creation. Only directories can be associated with default ACLs. Default ACLs for non-directories would be of no use, because no other file system objects can be created inside non-directories. Default ACLs play no direct role in access checks.
When a directory is created inside a directory that has a default ACL, the new directory inherits the parent directory's default ACL both as its access ACL and default ACL. Objects that are not directories inherit the default ACL of the parent directory as their access ACL only.
The permissions of inherited access ACLs are further modified by the mode parameter that each system call creating file system objects has. The mode parameter contains nine permission bits that stand for the permissions of the owner, group, and other class permissions. The effective permissions of each class are set to the intersection of the permissions defined for this class in the ACL and specified in the mode parameter.
If the parent directory has no default ACL, the permissions of the new file are determined as defined in POSIX.1. The effective permissions are set to the permissions defined in the mode parameter, minus the permissions set in the current umask.
The umask has no effect if a default ACL exists.
A process requests access to a file system object. Two steps are performed. Step one selects the ACL entry that most closely matches the requesting process. The ACL entries are looked at in the following order: owner, named users, (owning or named) groups, others. Only a single entry determines access. Step two checks if the matching entry contains sufficient permissions.
A process can be a member in more than one group, so more than one group entry can match. If any of these matching group entries contain the requested permissions, one that contains the requested permissions is picked (the result is the same no matter which entry is picked). If none of the matching group entries contains the requested permissions, access will be denied no matter which entry is picked.
The access check algorithm can be described in pseudo-code as follows.
Let us start by creating a directory and checking its permissions. The umask determines which permissions will be masked off when the directory is created. A umask of 027 (octal) disables write access for the owning group and read, write, and execute access for others.
$ umask 027 $ mkdir dir $ ls -dl dir drwxr-x--- ... agruen suse ... dir
The first character ls prints represents the file type (d for directory). The string ``rwxr-x--'' represents the resulting permissions for the new directory: read, write, and execute access for the owner and read and execute access for the owning group. The dots in the output of ls stand for text that is not relevant here and has been removed.
These base permissions have an equivalent representation as an ACL. ACLs are displayed using the getfacl command.
$ getfacl dir # file: dir # owner: agruen # group: suse user::rwx group::r-x other::---
The first three lines of output contain the file name, owner, and owning group of the file as comments. Each of the following lines contains an ACL entry for one of the three classes of users: owner, group, and other.
The next example grants read, write, and execute access to user Joe in addition to the existing permissions. For that, the -m (modify) argument of setfacl is used. The resulting ACL is again shown using the getfacl command. The -omit-header option to getfacl suppresses the three-line comment header containing the file name, owner, and owning group to shorten the examples shown.
$ setfacl -m user:joe:rwx dir $ getfacl --omit-header dir user::rwx user:joe:rwx group::r-x mask::rwx other::---
Two additional entries have been added to the ACL: one is for user Joe and the other is the mask entry. The mask entry is automatically created when needed but not provided. Its permissions are set to the union of the permissions of all entries that are in the group class, so the mask entry does not mask any permissions.
The mask entry now maps to the group class permissions. The output of ls changes as shown next.
$ ls -dl dir drwxrwx---+ ... agruen suse ... dir
An additional ``+'' character is displayed after the permissions of all files that have extended ACLs. This seems like an odd change, but in fact POSIX.1 allocates this character position to the optional alternate access method flag, which happens to default to a space character if no alternate access methods are in use.
The permissions of the group class permissions include write access. Traditionally such file permission bits would indicate write access for the owning group. With ACLs, the effective permissions of the owning group are defined as the intersection of the permissions of the owning group and mask entries. The effective permissions of the owning group in the example are still r-x, the same permissions as before creating additional ACL entries with setfacl.
The group class permissions can be modified using the setfacl or chmod command. If no mask entry exists, chmod modifies the permissions of the owning group entry as it does traditionally. The next example removes write access from the group class and checks what happens.
$ chmod g-w dir $ ls -dl dir drwxr-x---+ ... agruen suse ... dir $ getfacl --omit-header dir user::rwx user:joe:rwx #effective:r-x group::r-x mask::r-x other::---
As shown, if an ACL entry contains permissions that are disabled by the mask entry, getfacl adds a comment that shows the effective set of permissions granted by that entry. Had the owning group entry had write access, there would have been a similar comment for that entry. Now see what happens if write access is given to the group class again.
$ chmod g+w dir $ ls -dl dir drwxrwx---+ ... agruen suse ... dir $ getfacl --omit-header dir user::rwx user:joe:rwx group::r-x mask::rwx other::---
After adding the write permission to the group class, the ACL defines the same permissions as before taking the permission away. The chmod command has a nondestructive effect on the access permissions. This preservation of permissions is an important feature of POSIX.1e ACLs.
In the following example, we add a default ACL to the directory. Then we check what getfacl shows.
$ setfacl -d -m group:toolies:r-x dir $ getfacl --omit-header dir user::rwx user:joe:rwx group::r-x mask::rwx other::--- default:user::rwx default:group::r-x default:group:toolies:r-x default:mask::r-x default:other::---
Following the access ACL, the default ACL is printed with each entry prefixed with ``default:''. This output format is an extension to POSIX.1e that is found on Solaris and Linux. A strict implementation of POSIX 1003.2c would only show the access ACL. The default ACL would be shown with the -d option to getfacl.
We have only specified an ACL entry for the toolies group in the setfacl command. The other entries required for a complete ACL have automatically been copied from the access ACL to the default ACL. This is a Linux-specific extension; on other systems all entries may need to be specified explicitly.
The default ACL contains no entry for Joe, so Joe will not have access (except possibly through group membership or the other class permissions).
A subdirectory inherits ACLs as shown next. Unless otherwise specified, the mkdir command uses a value of 0777 as the mode parameter to the mkdir system call, which it uses for creating the new directory. Observe that both the access and the default ACL contain the same entries.
$ mkdir dir/subdir $ getfacl --omit-header dir/subdir user::rwx group::r-x group:toolies:r-x mask::r-x other::--- default:user::rwx default:group::r-x default:group:toolies:r-x default:mask::r-x default:other::---
Files created inside dir inherit their permissions as shown next. The touch command passes a mode value of 0666 to the kernel for creating the file.
All permissions not included in the mode parameter are removed from the corresponding ACL entries. The same has happened in the previous example, but there was no noticeable effect because the value 0777 used for the mode parameter represents a full set of permissions.
$ touch dir/file $ ls -l dir/file -rw-r-----+ ... agruen suse ... dir/file $ getfacl --omit-header dir/file user::rw- group::r-x #effective:r-- group:toolies:r-x #effective:r-- mask::r-- other::---
No permissions have been removed from ACL entries in the group class; instead they are merely masked and thus made ineffective. This ensures that traditional tools like compilers will interact well with ACLs. They can create files with restricted permissions and mark the files executable later. The mask mechanism will cause the right users and groups to end up with the expected permissions.
In this section we begin detailing the implementation of ACLs in Linux.
ACLs are pieces of information of variable length that are associated with file system objects. Dedicated strategies for storing ACLs on file systems might be devised, as Solaris does on the UFS file system []. Each inode on a UFS file system has a field called i_shadow. If an inode has an ACL, this field points to a shadow inode. On the file system, shadow inodes are used like regular files. Each shadow inode stores an ACL in its data blocks. Multiple files with the same ACL may point to the same shadow inode.
Because other kernel and user space extensions in addition to ACLs benefit from being able to associate pieces of information with files, Linux and most other UNIX-like operating systems implement a more general mechanism called Extended Attributes (EAs). On these systems, ACLs are implemented as EAs.
Extended attributes are name and value pairs associated permanently with file system objects, similar to the environment variables of a process. The EA system calls used as the interface between user space and the kernel copy the attribute names and values between the user and kernel address spaces. The Linux attr(5) manual page contains a more complete description of EAs as found on Linux. A paper by Robert Watson discussing supporting infrastructure for security extensions in FreeBSD contains a comparison of different EA implementations on different systems [].
Other operating systems, such as Sun Solaris, Apple MacOS, and Microsoft Windows, allow multiple streams (or forks) of information to be associated with a single file. These streams support the usual file semantics. After obtaining a handle on the stream, it is possible to access the streams' contents using ordinary file operations like read and write. Confusingly, on Solaris these streams are called extended attributes as well. The EAs on Linux and several other UNIX-like operating systems have nothing to do with these streams. The more limited EA interface offers several advantages. They are easier to implement, EA operations are inherently atomic, and the stateless interface does not suffer from overheads caused by obtaining and releasing file handles. Efficiency is important for frequently accessed objects like ACLs.
At the file system level, the obvious and straight-forward approach to implement EAs is to create an additional directory for each file system object that has EAs and to create one file for each extended attribute that has the attribute's name and contains the attribute's value. Because on most file systems allocating an additional directory plus one or more files requires several disk blocks, such a simple implementation would consume a lot of space, and it would not perform very well because of the time needed to access all these disk blocks. Therefore, most file systems use different mechanisms for storing EAs.
As described in the Linux kernel sources, each inode has a field that is called i_file_acl for historic reasons. If this field is not zero, it contains the number of the file system block on which the EAs associated with this inode are stored. This block contains both the names and values of all EAs associated with the inode. All EAs of an inode must fit on the same EA block.
For improved efficiency, multiple inodes with identical sets of EAs may point to the same EA block. The number of inodes referring to an EA block are tracked by a reference count in the EA block. EA block sharing is transparent for the user: Ext3 keeps an LRU list of recently accessed EA blocks and a table that has two indices (implemented as hash tables of double linked lists). One index is by block number. The other is by a checksum of the block's contents. Blocks that contain the same data with which a new inode shall be associated are reused until the block's reference count reaches an upper limit of 1024. This limits the damage a single disk block failure may cause. When an inode refers to a shared EA block and that inode's EAs are changed, a copy-on-write mechanism is used, unless another cached EA block already contains the same set of attributes, in which case that block is used.
The current implementation requires all EAs of an inode to fit on a single disk block, which is 1, 2, or 4 KiB. This also determines the maximum size of individual attributes.
If the sets of EAs tend to be unique among inodes, no sharing is possible and the time spent checking for potential sharing is wasted. If each inode has a unique set of EAs, each of these sets will be stored on a separate disk block, which can lead to a lot of slack space. The extreme case is applications that need to store unique EAs for each inode. Fortunately for many common workloads, the EA sharing mechanism is highly effective.
Alternative designs with fewer limitations have been proposed [], but it seems that they are not easy to actually implement. No alternatives to the existing scheme exist so far.
JFS stores all EAs of an inode in a consecutive range of blocks on the file system (i.e., in an extent) []. The extended attribute name and value pairs are stored consecutively in this extent. If the entire EAs are small enough, they are stored entirely within the inode to which they belong.
JFS does not implement an EA sharing mechanism. It does not have the one-disk-block limitation of Ext2 and Ext3. The size of individual attributes is limited to 64 KiB.
Of the file systems currently supported in Linux, XFS uses the most elaborate scheme for storing extended attributes []. Small sets of EAs are stored directly in inodes, medium-sized sets are stored on leaf blocks of B+ trees, and large sets of EAs are stored in full B+ trees. This results in performance characteristics similar to directories on XFS: although rarely needed, very large numbers of EAs can be stored efficiently.
XFS has a configurable inode size that is determined at file system create time. The minimum size is 256 bytes, which is also the default. The maximum size is one half of the file system block size. In the minimum case, the inodes are too small to hold ACLs, so they will be stored externally. If the inode size is increased, ACLs will fit directly in the inode. Since inodes and their ACLs are often accessed within a short period of time, this results in faster access checks, but also wastes more disk space.
XFS does not have an attribute sharing mechanism. The size of individual attributes is limited to 64 KiB.
ReiserFS supports tail merging of files, which means that several files can share the same disk block for storing their data. This makes the file system very efficient for many small files. One potential drawback is that tail merging can consume a noticeable amount of CPU time.
Since ReiserFS is so good at handling small files, EAs can directly use this mechanism. For each file that has EAs, a directory with a name derived from a unique inode identifier is created inside a special directory. The special directory is usually hidden from the file system namespace. Inside the inode specific directory, each EA is stored as a separate file. The file name equals the attribute name. The file's contents are the attribute value.
ReiserFS does not implement an attribute sharing mechanism, but such an extension will possibly be implemented in the future. Sharing could even be implemented on a per-attribute bases, so the result would be a highly efficient and flexible solution. The size of individual attributes is limited to 64 KiB.
An interesting design decision is how ACLs should be passed between user space and the kernel, and inside the kernel, between the virtual file system (VFS) and the low-level file system layer. FreeBSD, Solaris, Irix, and HP-UX all have separate ACL system calls [,,,].
Linux does not have ACL system calls. Instead, ACLs are passed between the kernel and user space as EAs. This reduces the number of system interfaces, but with the same number of end operations. While the ACL system calls provide a more explicit system interface, the EA interface is easier to adapted to future requirements, such as non-numerical identifiers for users and groups in ACL entries.
The rationale for using separate ACL system calls in FreeBSD was that some file systems support EAs but not ACLs, and some file systems support ACLs but not EAs, so EAs are treated as pure binary data. EAs and ACLs only become related inside a file system. [,].
The rationale for the Linux design was to provide access to all meta data pertinent to a file system object through the same interface. Different classes of attributes that are recognized by name are reserved for system objects such as ACLs. The attribute names ``system.posix_acl_access'' and ``system.posix_acl_default'' are used for the access and default ACL of a file, respectively. The ACL attribute values are in a canonical, architecture-independent binary format. File systems that do not implement ACLs but do implement EAs, or ones that implement ACLs as something other than EAs, need to recognize the relevant attribute names.
While it is possible to manipulate ACLs directly as EAs, at the application level this is usually not done: since the EA system calls are Linux-specific, such applications would not be portable. Other systems support similar EA mechanisms, but with different system call interfaces. Applications that want to use POSIX.1 ACLs in a portable way are expected to use the libacl library, which implements the ACL-specific functions of POSIX.1e draft 17.
The access ACL of a file system object is accessed for every access decision that involves that object. Access checking is performed on the whole path from the namespace root to the file in question. It is important that ACL access checks are efficient. To avoid frequently looking up ACL attributes and converting them from the machine-independent attribute representation to a machine-specific representation, the Ext2, Ext3, JFS, and ReiserFS implementations cache the machine-specific ACL representations. This is done in addition to the normal file system caching mechanisms, which use either the page cache, the buffer cache, or both. XFS does not use this additional layer of caching.
Most UNIX-like systems that support ACLs limit the number of ACL entries allowed to some reasonable number. Table shows the limits on Linux.
|
ACLs with a high number of ACL entries tend to become more difficult to manage. More than a handful of ACL entries are usually an indication of bad application design. In most such cases, it makes more sense to make better use of groups instead of bloating ACLs.
The ReiserFS and JFS implementations define no limit on the number of ACL entries, so a limit is only imposed by the maximum size of EA values. The current EA size limit is 64 KiB, or 8191 ACL entries, which is too high for ACLs in practice: besides being impractical to work with, the time it would take to check access in such huge ACLs may be prohibitive.
An important aspect of introducing new file system features is how systems are upgraded, and how systems that do not support the new features are affected. File system formats evolve slowly. File systems are expected to continue to work with older versions of the kernel. In some situations, like in multiple boot environments or when booting from a rescue system, it may be necessary or preferable to use a kernel that does not have EA support.
All file systems that support ACLs on Linux are either inherently aware of EAs, or are upgraded to support EAs automatically, without user intervention. Depending on the file system, this is either done during mounting or when first using extended attributes.
On all file systems discussed in this paper, when using a kernel that does not support EAs on a file system with EAs, the EAs will be ignored. On ReiserFS, EA-aware kernels actively hide the system directory that contains the EAs, so in an EA-unaware kernel this directory becomes visible. It is still protected from ordinary users through file permissions.
Working with EA file systems with EA-unaware kernels will still lead to inconsistencies when files are deleted that have EAs. In that case, the EAs will not get removed and a disk space leak will result. At least on Ext2 and Ext3, such inconsistencies can be cleaned up later by running the file system checker.
ACL-unaware kernels will only see the traditional file permission bits and will not be able to check permissions defined in ACLs. The ACL inheritance algorithm will not work.
Since ACLs define a more sophisticated discretionary access control mechanism, they have an influence on all access decisions for file system objects. It is interesting to compare the time it takes to perform an access decision with and without ACLs.
Measurements were performed on a PC running SuSE Linux 8.2, with the SuSE 2.4.20 kernel. The machine has an AMD Athlon processor clocked at 1.1 GHz and 512 MiB of RAM. The disk used was a 30 GB IBM Ultra ATA 100 hard drive with 7200 RPM, an average seek time of 9.8 ms, and 2 MiB of on-disk cache. The Ext2, Ext3, Reiserfs, and JFS file systems were created with default options on an 8 GiB partition. On XFS, to compare EAs that are stored in inodes and EAs that are stored externally, file systems with inode sizes of 256 bytes and 512 bytes were used. These file systems are labeled XFS-256 and XFS-512, respectively.
Table compares the times required for the initial access check to a file with and without an ACL after system restart. To exclude the time for loading the file's inode into the cache, a stat system call was performed before checking access. The time taken for the stat system call is not shown. The first access to the access ACL of a file may require one or more disk accesses, which are several orders of magnitude slower than accessing the cache. The actual times these disk accesses take vary widely depend on the disk speed and on the relative locations of the disk blocks that are accessed. The function used for measuring time has a resolution of 
microsecond. In the ACL case, the file that is checked has a five-entry access ACL.
|
XFS with 512-byte (or larger) inodes and JFS store the ACLs directly in the inodes. Therefore, no additional disk accesses are needed for retrieving the ACLs.
After the first repetition, all information is fully cached. Figures 
show micro-benchmarks of basic operations in this state. Each test repeats the same operation many times and averages the total time spent over the number of repetitions. In all configurations except XFS-256 with ACLs, the time per access check drops to around 
microseconds, as the leftmost measurements in Figures and show.
Figure compares the speed of various system calls. The getpid system call is included to show the overhead of switching between the user and kernel address spaces. The ls -l command indicates in its output if a file has an extended ACL. Internally, it uses the acl_extended_file function from the libacl library. This function is almost as fast as the stat system call, which ls -l also calls for each file, so the additional overhead is small. For comparison, the acl_get_file measurement shows the time it takes to retrieve a five-entry ACL.
 |
Figure shows the time taken for one access system call, depending on the number of directory levels in the pathname argument. Figure shows the performance of the same operation with a five-entry access ACL on each directory. XFS's overhead for converting the ACL from its EA representation to its in-memory representation results in a noticeable difference, particularly with 256 byte inodes.
 |
 |
Tables and show the overhead involved when copying files from one file system to another. All the files that are copied have ACLs, but no additional EAs. Figure shows the distribution of file sizes in the sample data sets. The tests show the time taken from starting the cp command to the return from the sync command, which we start immediately after the cp command. The sync command ensures that all files are immediately written. The first series uses a version of cp that does not support EAs or ACLs. The second series uses a version of cp that supports both EAs and ACLs. In both of the benchmarks, the source file system type is held constant, while the destination file system type is changed.
The sample size for the benchmark in Table is small enough for all files to fit in main memory. The time for first reading the data into memory is not included in the results. These tests were repeated five times. The results show the median and the standard deviation of the results. The sample size for the benchmark shown in Table is larger than either main memory or the file system journals. These tests were only executed once.
| ||||||||||||||||||||||||||||||||||||||||||||||||
|
 |
Note that both benchmarks use ACLs excessively, which is a worst-case scenario. The overheads for real workloads should be much smaller.
It can be observed that the overhead of ACLs varies widely among the supported file systems. The differences show more when the I/O load is low, and get smaller as the I/O load rises. For ACLs, the Ext2 and Ext3 implementations have little overhead. ReiserFS EAs have a relatively high overhead. This may improve if attribute sharing is implemented. For XFS, increasing the inode size so that ACLs can be stored directly in inodes makes a big difference.
It is unclear why JFS appears to be faster when copying ACLs than when not copying ACLs.
Full ACL support over NFS requires two things: First, a mechanism so that all access decisions are performed in a way that honors the ACLs. Second, extensions to the NFS protocol for manipulating ACLs on remote mounted file systems.
The NFS protocol performs client-side caching to improve efficiency. In version 2 of the protocol, decisions as to who gets read access to locally cached data are performed on the client. These decisions are made under the assumption that the file mode permissions bits and the IDs of the owner and owning group are sufficient to do that. This assumption is obviously wrong if an extended permission scheme like POSIX ACLs is used on the server.
Because NFSv2 clients perform some access decisions locally, they will incorrectly grant read access to file and directory contents cached on the client to users who are a member in the owning group in two cases. First, if the group class permissions include read access, but the owning group does not have read access. Second, if the owning group does have read access, but a named user entry for that user exists that does not allow read access. Both situations are rare. Workarounds exist that reduce the permissions on the server side so that clients only see a safe subset of the real permissions [,]. No anomalies exist for users who are not a member in the owning group.
There are two ways to solve this problem. One is to extend the access check algorithm used on the client. The other is to delegate access decisions to the server and possibly cache those decisions for a defined period of time on the client. The first solution would probably scale better to a high number of readers on the client side, as long as the server and all clients can agree on the access check algorithms use. Unfortunately, this approach falls apart as soon as servers implement different permission schemes.
Version 3 of the NFS protocol therefore defines a new remote procedure call (RPC) called ACCESS for delegating access decisions to the server. This RPC is similar to the access system call. NFSv3 clients are expected to use this RPC for determining to whom to grant access to cached contents.
The NFSv3 protocol unfortunately does not define mechanisms for transferring ACLs. As a consequence, different vendors have implemented proprietary protocol extensions that are incompatible with each other. Solaris implements an NFSv3 protocol extension called NFS ACL that supports ACLs only. Irix implements a more general protocol that supports EAs and passes ACLs as special EAs.
NFSv4 defines the structure and semantics of its own kind of ACLs, along with RPCs for transferring them between clients and servers. NFSv4 ACLs are similar to Microsoft Windows ACLs []. Unfortunately, the designers of NFSv4 have mostly ignored the existence of POSIX ACLs, so NFSv4 ACLs are not compatible with POSIX ACLs. Marius Aamodt Eriksen describes a one-way mapping between POSIX ACLs and NFSv4 ACLs [], but this mapping is impractical. One of the central concepts in POSIX ACLs, which is needed to ensure compatibility with legacy POSIX.1 applications, is the mask entry. The NFSv4 ACL model could be extended by the mask concept. Although this would greatly improve interoperability with POSIX ACLs, proposals to extend the NFSv4 specification have so far been rejected.
Partial NFSv3 support has been added in the 2.2 Linux kernel series. The ACCESS RPC was added to the kernel NFS daemon in version 2.2.18, but the NFS client only correctly uses the ACCESS RPC in the 2.5 kernel series. A patch for older kernels exists since kernel version 2.4.19, which is included in the SuSE and UnitedLinux products.
Because the ACCESS RPC can lead to noticeable network overhead even on file systems that are known not to include any ACLs, the Linux NFSv3 client allows to mount file systems with the noacl mount option. Then the NFS client will use neither the ACCESS RPC nor the GETACL or SETACL RPCs. To ensure that no ACLs can be set on the server, the Ext2, Ext3, JFS, and ReiserFS file systems can be mounted on the server without ACL support by omitting the acl mount option.
Since March 3, 2003, an implementation of Sun's NFS ACL protocol for Linux (which is also included in SuSE Linux 8.2) is available at , with friendly permission from Sun to use it. The NFS ACL protocol was chosen because it is simple and supports POSIX 1003.1e draft 17 ACLs well enough. Solaris ACLs are based on an earlier draft of POSIX 1003.1e, so its handling of the mask ACL entry is slightly different than in draft 17 for ACLs with only four ACL entries. This is a corner case that occurs only rarely, so the semantic differences may not be noticeable.
Microsoft Windows supports ACLs on its NTFS file system, and in its Common Internet File System (CIFS) protocol [], which formerly has been known as the Server Message Block (SMB) protocol. CIFS is used to offer file and print services over a network. Samba is an Open Source implementation of CIFS. It is used to offer UNIX file and print services to Windows users. Samba allows POSIX ACLs to be manipulated from Windows. This feature adds a new quality of interoperability between UNIX and Windows.
The ACL model of Windows differs from the POSIX ACL model in a number of ways, so it is not possible to offer entirely seamless integration. The most significant differences between these two kinds of ACLs are:
3