Pp. 247260 of the Proceedings |
Serge E. Hallyn
College of William and Mary
hallyn@cs.wm.edu, https://www.cs.wm.edu/~hallyn
Phil Kearns
College of William and Mary
kearns@cs.wm.edu, https://www.cs.wm.edu/~kearns
Access control in Linux currently consists of traditional Unix permissions and
POSIX capabilities[Caps-faq]. Domain and Type Enforcement (DTE) has been
presented [DTE95,DTE96] as a useful method for enhancing access
control. DTE groups processes into domains, and files into types, and
restricts access from domains to types as well as from domains to other
domains. Type access can be any of read, write, execute, create, and directory
descend. Domain access refers the right to send signals as well as that to
transition to a new domain. A process belongs to exactly one domain at any
particular time. A process transitions to a new domain by executing a file
which has been defined as an entry point to that domain. The three types
of domain transitions are auto, exec, or none. If Domain A has auto
access to domain B, and a process in domain A executes an entry point for domain
B, then the process will be automatically switched to domain B. If domain A has
exec access to domain B, then a process running under domain A can choose
whether to switch to domain B on execution of one of B's entry points.
DTE can be considered an abbreviated form of classical
capabilities[Dennis66]. In a system based upon classical
capabilities, a process carries with itself a set of access rights to
particular objects. At any point, a process can give up, or reclaim (if
permitted) some of its capabilities. POSIX capabilities work similarly, but
these capabilities are limited to a predefined subset of superuser access
rights such as the ability to nice a process, boot the system or open a
privileged () port. In DTE, a process carries with itself only an
indicator of the domain in which it runs, and this determines the process'
access rights. A process can enter a new domain (and hence change its access
rights) only upon file execution.
Trusted Information Systems has used DTE in its proprietary firewalls, but
details of its implementation were not publicly available, and TIS appears to
have stopped using DTE altogether. A group at SAIC has recently begun a DTE
for Linux implementation[SAIC-DTE]. Jonathon Tidswell and John
Potter[Tidswell97] submitted theoretical work on extending DTE to allow safe
dynamic policy changes, but have attempted no implementation.
Presented here is our prototype implementation of DTE for Linux version 2.3.
We have implemented a DTE prototype in the 2.3.28 Linux kernel. Our
implementation of DTE attaches type information to VFS inodes and domain
information to process descriptors (task structs). A DTE policy is read at
boot time from the text file /etc/dte.conf.
Traditional UNIX permissions are still enforced. There are several reason for
this, such as user and system administrator familiarity with traditional UNIX
protection. Most importantly, however, DTE is designed to provide mandatory access
control to protect a system from subverted superuser processes.
A DTE policy to replace traditional UNIX access control would be very large and
complex.
However, one could completely void traditional access control by simply
giving all users full access to all files. Similarly, one can bypass DTE by
creating a DTE policy with only one type and one domain, and full access from
the sole domain to the sole type.
At boot time, we build a structure for each domain as specified in the DTE
policy file. This structure contains information regarding permitted access to
types, permitted transitions and signal access to other domains, and entry
points. Every process' task structure will contain a pointer to the structure
for the domain to which it currently belongs.
At this time we also create an array containing the names of all types. Types
are then compared by the offset of the type name in this array. Every inode
contains three pointers which either are NULL or point into this array. The
three pointers represent the etype, rtype and utype values.
The etype value is the type of this particular file or directory. The
rtype represents the value of this directory and its children, whereas the
utype represents only the type of its children.
The type of a file is determined in one of three ways. First, if we have
previously determined the type, then the inode's etype will be set and we
simply use it. If this is the first time we are looking up this file, and a
rule exists assigning it a type, then that rule is used. Finally, if a rule
does not exist assigning a type to the file, then the values are inherited from
the parent's utype1. The etype of an inode is always set. If
there is no rule specifying the etype for the file, then the etype
is set to the parent's utype. The utype of an inode must therefore
also exist. It is set, in order of preference, to the assigned utype, the
assigned rtype, or the parent's utype. The rtype of an inode
is set only if a rule assigns an rtype to the inode's path.
Since type information always comes from either the DTE policy or from an
inode's ancestors in the filesystem tree, no information needs to be added to
the filesystem on disk. The type assignment rules are represented in memory by
a tree of map nodes which is constructed at boot time from the type
assignment rules in the policy. A sample tree for a particular set of rules is
shown in Figure 1, along with the type information in
corresponding inodes. The map nodes are only used to determine whether a rule
exists binding a path to a type. Once an inode's type is set, subsequent
lookups will not cause us to check the map nodes again.
When a process performs an open system call, the modified kernel
checks for DTE permission before checking the standard UNIX permissions.
We use the domain structure pointed to by the current task structure to
check whether the current domain has the requested access to the type to
which the file being opened belongs. If this access is granted, then we proceed
to perform the normal UNIX checks. A check for DTE execute permission is
delayed until the actual call to execve. If the execution causes an
allowed domain transition, then this transition should occur before the check
for execute access, since the new domain may be the only one allowed to execute
the entry point.
Domain to type access information is kept in a set of hash tables. Each
domain structure has a hash table keyed by the type name, and each entry lists
the domain's access rights to the particular type. A type access check,
therefore, consists of simply calculating the hash value of the type name to
find the appropriate domain to type access entry for the current domain, and
comparing the requested access to the permitted access. This is done
regardless of username, so that the superuser is not exempt from the DTE
policy.
Our DTE implementation enforces restrictions on signals between processes in
different domains. Each domain structure contains a linked list of dte_signal_access structures, which contain the signal number and a pointer
to the domain to which it may be sent. One of these structures exists for
every signal which may be sent to another domain. However, for the sake of
abbreviation, setting the domain to null allows the specified signal to be
sent to all domains, and setting the signal to 0 allows all signals to be sent
to the specified domain.
Three types of domain transition are possible each time a file is executed. The
first is an auto, or mandatory, domain transition. Since this must be
automatic, it means that each time a process calls execve, we must check
whether the file being executed is an entry point into a domain to which the
current domain has auto access. The second type of transition is an exec, or user-requested, transition. This is facilitated by a new system call,
sys_dte_execve, which takes an additional argument over execve
containing the name of the requested domain. The third and default type of
transition is the NULL transition, wherein the domain is not changed.
Domain transition information is kept in two types of structures, both linked
from the domain structure. Since auto transitions must be checked for on
every execve system call, the search for a particular pathname must be
very quick. Therefore, each domain structure contains a hash table of the pathnames
whose execution lead to auto domain transitions, along with the domain to
be switched to.
The domain structure also has a linked list of structures representing allowed
exec transitions. Since a dte_exec is a relatively rare, and
user-requested, event, efficiency is not so critical, and we can elect for a more
memory-efficient representation. Therefore we do not keep a hash table of
every file which may cause an exec transition, but simply point to the
domains to which a voluntary (exec) transition is allowed. To check for
exec access to a domain, we must first check for an exec
entry for the desired domain, then check whether the file being executed is an
entry point for that domain.
Administering DTE consists of editing the policy, which is defined in the
file /etc/dte.conf. The system must be rebooted to effect the
changes2.
The DTE policy file consists of several sections. We first enumerate the
types and domains. Next we specify the default type for the filesystem
root ("/") and its children, and the domain in which to run the first
process (init). Following is the detailed definition of all domains. For each
domain we specify the entry points, permitted type access, permitted domain
transitions and permitted signals to processes in other domains. Finally we
list the type assignment rules.
A sample policy file is in Figure 2. First we specify that
there will be two types, root_t and log_t, and two domains,
common_d and log_d. We set the default root rtype, hence the
default type for the entire filesystem, to root_t. Next we set the type
of the first process to common_d. We specify that the log_d
domain will have one entry point, /sbin/syslogd, and should have read,
execute and directory descend access to files of type root_t and read,
write, execute, create and directory descend access to files of type log_t. For the domain common_d, we specify read, write, execute,
create and directory descend access to files of type root_t, but only
read access to files of type log_t. This domain also receives auto transition access to domain log_t, meaning that, on execution of
/sbin/syslogd, a process in domain common_d will automatically be
switched to domain log_d. Finally, the last statement assigns the type
log_t to the directory /var/adm/log and all files thereunder.
Three additional system calls are provided to allow software to interact with
DTE. The sys_dte_exec call was discussed earlier. A user may
invoke sys_dte_gettype to learn the type associated with a file.
Similarly, sys_dte_getdomain may be called to learn the domain
associated with a process.
We measured the performance of both a DTE-enabled and a DTE-free 2.3.28
kernel for the execve and lookup_dentry system calls, the overhead
imposed by the DTE-specific sys_dte_exec and dte_auto_switch
system calls, and a full kernel compile. The following tests were run on a
400Mhz Pentium II ( bogomips) with 512K L2 cache and 384M ram. Each test
was run on a kernel compiled without DTE and one with DTE using the simple
policy shown in Figure 3. We used the Pentium cycle clock
for timing. All confidence intervals are 95%.
The first time it is called on a particular file or directory, however, the
etype may not yet have been set. In this case, we must check for a type
assignment rule or, if such a rule does not exist, set the type from the parent
directory. Furthermore, if the parent directory does not exist then we must
first do the same for it, and so on until a directory is associated with
a type assignment rules or has its type set.3. The DTE type assignment
rules are kept in a tree format analogous to the filesystem tree, as shown in
Figure 1. The children are currently not sorted, so that
a large number of assignment rules for files under a single parent directory
could impact performance. However, for normal cases this lookup should be
reasonably quick.
We timed the upper part of the kernel function fs/namei.c:permission,
where the DTE code is located. Over the course of a boot sequence, several
repeats of the lookup test above, some general milling around, and a shutdown,
the DTE code added clock cycles to each permission call.
The time required to look up a given pathname greatly affects the subjective
performance of the system. The function fs/namei.c:lookup_dentry,
which performs this task in the Linux kernel, is affected by DTE in two
places. First, for each subdirectory in a pathname, lookup_dentry
calls permission to check for execute access. Second, if the types
for the deepest path element being looked up have not been set, then we must
set them, using the same function we use above in permission.
We timed lookup_dentry on a set of pathnames ranging in depth from 1 to
9 components, both for fully existing and fully nonexistent pathnames. For the
first execution, each component of each pathname was uncached. On subsequent
executions, all path components and their corresponding DTE type information
were (naturally) cached.
The results can be seen in figures 4, 5,
6 and 7. For the case of a lookup
for uncached filenames, results appear to be rather unpredictable. If this
appears to be more true for existing file lookup than for nonexistent files,
this is a result we should have predicted by our method of testing. We tested
the lookup for each set of pathnames, then rebooted, and repeated the test, eleven
times in all. However, for the nonexistent filename lookup, a part of the
pathname was legitimate. This piece was looked up uncached only for our first
test after reboot, which was for the first table entry in
figure 6. Since the DTE kernel was faster than the
plain kernel more often than it was slower, it appears safe to say that disk i/o
completely overshadows any time spent setting DTE types from map rules and
parents.
For cached lookups, the DTE kernel appears to do slightly better than twice as
long as the plain kernel.
Upon file execution, we must check whether the requested execution should cause
a mandatory domain switch. This is done using kernel/dte.c:dte_auto_switch. As previously mentioned, this function must be
fast as it is called with every file execution. Therefore, it simply hashes the
name of the executable to check for an entry in a table of gateways, or
executables which cause an automatic domain switch.
We compiled a kernel which timed the execution of dte_auto_switch. If a
particular domain has no gateways, then dte_auto_switch does not bother
to hash the typename, so that the dte_auto_switch call during execve takes clock cycles. If there are gateways, then we must
search the hash tables. While we tested using domains with a variable number of
permitted auto switches,4 this number does not affect the running time of dte_auto_switch, which is clock cycles. Since this function
does not exist in the plain Linux kernel, its running time must be considered
pure overhead to file execution.
The least efficient of all the code added with DTE certainly sits in kernel/dte.c:sys_dte_exec. First, the user provides the name of a domain to switch to. Since domains are currently not kept hashed or in any order, the lookup for the corresponding domain structure is , where is the number of defined domains and is the maximum length of any domain name. Next, we search another unsorted list, containing the domains to which the current domain may voluntarily switch, to check whether the domain switch is legal. Then we search a third list, containing valid entry points for the destination domain.
In order to measure the amount of time required to check for an exec
domain switch, we set up 12 domains, with entry points numbering
and , where two domains had 30 entry points.
Then we performed an exec domain switch into each of these domains, and
measured the time between the start of sys_dte_exec and its call to
sys_execve. Since entry points are stored unsorted, the 12th domain's
list of entry points contained the entry point which we actually executed last,
whereas all others listed it first. The results for 10 trials (excluding the
first of eleven since the entry point needed to be read from disk) are shown in
Figure 8.
The first 11 domains each executed the first file in the respective domains'
linked list of entry points. The difference in performance is due to the
analogous problem with the list of allowed exec transitions.
For domains with what we believe to be a realistic number of entry points (1-8),
sys_dte_exec takes about 4 times as long as dte_auto_switch.
Clearly, performance will be greatly improved when we store entry points,
domains, and allowed exec transitions in a data structure which allows
quicker lookups. This will be a simple but low priority improvement, since a
policy must be quite large for the effects to become noticeable, and a sys_dte_exec call is a rare event.
To time the kernel function fs/exec.c:do_execve, we wrappered it and
took a timestamp before and after the real function call. In this way we
measure the full time for file execution including such details as the time to
load library files. For more fine-grained measurements of specific parts of
this process, we later measure the time to check for the auto domain
switch, a user-requested domain switch and filename lookup.
The command
/bin/echo -n .was executed 500 times. Execution time for the first run was an order of magnitude larger than for subsequent runs, both with and without DTE. This is to be expected since some library files as well as executable /bin/echo may not yet have been loaded from disk. The same thing occurs for later performance tests. Since this is independent of the DTE code and serves only to hide the performance impact of DTE, we will, in all subsequent tests, ignore the first execution after boot.
The DTE code introduces a overhead. The table in
Figure 9 shows the timing results.
Finally we turned off all micro-performance measurements and used /usr/bin/time to determine the performance on a kernel make on both DTE
and non-DTE enabled kernels. The plain 2.3.28 kernel took 5 minutes
and 55 seconds for the first compile, and 5:35 0.384387 for 14 subsequent
compiles, while the DTE-enabled kernel required 5 minutes and
56 seconds for the first compile and 5:36 0.205464 for subsequent compiles.
Clearly a new access control system cannot be added without affecting
performance. The above sections, in testing specifically the areas of the
kernel where code was added, might make the performance impact of DTE seem more
significant than it really is. This result shows that, when amortized over the
course of a realistic activity, which includes heavy file opening, creation and
execution as well as heavy computation and file i/o, the amount of overhead, one
second for every six minutes, is negligible.
To show the effectiveness of our DTE implementation, we picked a recent,
high-profile vulnerability, the buffer overflow in wu-ftpd[CERT-ftpd], and showed how our implementation of DTE can prevent
an attacker from obtaining a root shell. Our goal was to show that we could
protect the system from the wu-ftpd vulnerability (the posted exploits as
well as future or hand-crafted ones) without modifying the binary. In order for
ftp to retain its full functionality, it would need to be made DTE-aware so that
it could, like login, allow ftp to transition into the domain associated with a
user being authenticated5. We did not do this, but set protections such
that users can retrieve files from, if not deposit files onto, the server.
Anonymous ftp is fully functional.
The policy shown in Figure 10 prevents domain ftpd_d
from executing any system binaries other than /usr/sbin/in.ftpd and binaries
located under ~ftp/bin/ (lines 19-21). These files are defined to be
of the type ftpd_xt (lines 29 and 30), which the domain ftpd_d
may execute but not write (line 20). Only ftpd_d may execute this type
(lines 9-21), and root_d automatically switches to ftpd_d on
execution of /usr/sbin/in.ftpd (line 12), since that is an entry point to
ftpd_d (line 19). The exploits to be found on the internet
to take advantage of this vulnerability will therefore fail, as they expect to
be allowed to run /bin/sh. Nor can a script be written to upload and run
a Trojan horse, since the only types which ftpd_d is allowed to write
may not be executed by anyone.
The script which we tested was wuftpd2600, which can be found at https://www.securityfocus.com. It connected to our test machine, and exploited
the buffer overflow. However, the DTE-enabled kernel refused to allow the ftpd_d domain to execute /bin/sh. The script therefore hung, and the
system was not compromised. The error messages in Figure 11 were
sent to syslog. In contrast, the plain 2.3.28 kernel happily
provided a root shell.
Our implementation of DTE for Linux is functional. It reads a policy file
at boot time and enforces domain to type access as well as domain
transitions. We have not implemented DTE for networking.
First, we must extend our policy parser to allow easier and abbreviated entry
of more complicated policies. Next, we plan to create tools to help a system
administrator graphically create and view DTE policies and detect possible
security risks. Examples of such risks might include a domain which is
permitted to enter another domain as well as write one of the other domain's
entry points, or a domain which has auto access to two domains which
share an entry point.
Badger et al.[DTE95] suggest dynamically changing the DTE policy as
certain events occur. For example, a particular file (/var/adm/topsecretlog) might be tightly protected by a particular type. If
this file is then moved to /tmp, then Badger et al. suggest that a rule
should be added to keep the file under its original type. Alternatively, they
suggest that renames across type boundaries could be forbidden.
We currently go the lazy route. If a domain has permission to two types,
and a process running in that domain chooses to move a file from a
directory belonging to one type to that belonging to another, then the
file's type simply changes. Since the person (or process) moving the file
had the permission to do so, we trust it to understand the
implications.
The more dangerous problem lies with hard links. Since hard links provide no
notion of one name being superior to another, the type of an inode with multiple
corresponding filenames is currently determined based upon the name first looked
up.6 We can prevent creation of hard
links across type boundaries, however a change in policy can thwart this
defense quite easily. We will, in future versions, allow the system to add its
own type assignment rules (which will be necessary for several other desirable
features), and plan to use this capability to implement a better resolution of the
problem with hard links.
When a partition is mounted into the filesystem tree, it fits into the tree
defined by type rules depending on where it is mounted. For example,
mounting a partition under /tmp instead of /mnt might
completely change access permissions to the partition. It seems helpful to
allow the filesystem on a partition to specify certain type assignment
rules which apply only to the partition. Badger et al[DTE96] also
added a DTE configuration section allowing a DTE administrator to limit mount
points for partitions, which should be trivial for us to add as well.
DTE for Linux is freely available as a patch to 2.3.28 at
https://www.cs.wm.edu/~hallyn/dte.
This document was generated using the LaTeX2HTML translator Version 99.2beta6 (1.42)
Copyright © 1993, 1994, 1995, 1996,
Nikos Drakos,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999,
Ross Moore,
Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -split 0 -nonavigation paper.tex
The translation was initiated by Serge E. Hallyn on 2000-08-24
This paper was originally published in the
Proceedings of the 4th Annual Linux Showcase and Conference, Atlanta,
October 10-14, 2000, Atlanta, Georgia, USA
Last changed: 8 Sept. 2000 bleu |
|