|
Security 2001 Paper   
[Security '01 Tech Program Index]
Transient Addressing for Related Processes: Improved Firewalling by Using IPV6 and Multiple Addresses per Host*Work done while interning at AT&T Labs Research 1 IntroductionIn the simplest of inter-networked host models, a client or server host has a single network interface with a single network address identifying the host. Even under such an elementary set-up, firewalls have traditionally faced difficulty when confronted with application protocols needing to open secondary channels. Examples abound, most notably ftp, but also rsh, RealAudio, H.323, tftp and the X Window System. To operate with such popular applications, firewalls have been forced either to follow the application layer protocol and configure themselves appropriately or to keep open, sometimes unnecessarily, a range of ports. As an alternative to potentially complex, detailed, and often stateful firewall interaction, we propose a method using multiple network addresses per host to organize and simplify firewall decisions. Under our basic model, instead of trying to follow the unfolding application protocol details, the firewall makes an initial permissibility determination based on transport layer protocol and the endpoints’ ports and addresses. Assuming approval of the proposed transaction, the firewall subsequently permits all traffic between the approved address pairs, irrespective of port.
The security concerns arising from the firewall’s apparent loss of control over a session’s evolving ports will be alleviated by dynamic control of the protected host’s active addresses. Further, by segregating and controlling which addresses offer network services outside the firewall and which facilitate protected-host driven network requests, the architecture provides a natural address based division between potentially hostile requests from outside the bastion, and presumably benign outbound activities originating within the protected network. To help distinguish among a protected-host’s various client/server tasks, we will tie the address used by a client process to its process group identifier. This way, host addresses will come and go as part of the natural lifecycle of the processes that use them. For example, when a TARP client starts ftp from behind a firewall to a similar server, also protected by a firewall, it configures a new address. The client’s firewall recognizes the client’s new address and records it along with the destination address. The firewall presumably grants the FTP request based on port and address criteria and subsequently passes all outbound and inbound packets between the two addresses. Further port negotiations (using the same addresses) conducted by ftp will have no effect on what the client side firewall passes; it will simply pass along all the server packets addressed to the client’s ftp address. The server side behaves similarly (though there are some issues if different servers have different access policies); however, servers must use static addresses, largely so that clients can find them. The server’s firewall sees the client address opening a connection to the server’s fixed address and ftp port and decides whether to allow the connection. If the server’s firewall permits the connection, it passes all subsequent packets between the client and server addresses, independent of port, including incoming calls. Another way to view this is to adopt a virtual machine metaphor. That is, each client process group conceptually runs in a virtual machine, with an independent IP address space. This view can guide several design decisions, as shown in Figure 2.
Although IPv4 might conceivably support such a scheme, IPv6 and its large 128-bit addresses [HD98] provide a simpler opportunity to deploy TARP. Typically, the high order 64 bits of an IPv6 address provide the information needed to deliver a packet to the host’s LAN, and the low order 64 bits, assigned by the site administrator, are commonly the interface’s 48-bit Ethernet address padded with a constant to create a unique 64-bit interface identifier. TARP’s addressing scheme exploits these pad bits to create a family of 216 IPv6 addresses with the same basic uniqueness properties as Ethernet-based host identifiers. These addresses become associated with the sockets of a process group and form the foundation for TARP. A proof-of-concept implementation exists as a small set of kame/BSD [KAM] kernel enhancements, and allows socket programmers and applications transparent access to TARP’s advantages. The basic server/client programs supplied with the kame IPv6 package (tftp, ftp, rlogin, rsh, telnet, sendmail, ssh, mozilla, inetd) all work with the modified kernel, and none of them required modification to run. 2 Multiple Addresses per Host and a Large Family of IPv6 AddressesThe notion of assigning multiple IP addresses to a (non-routing) host is not new. In 1994, RFC 1681 [Bel94] described a number of potential uses and benefits of assigning multiple IP addresses to a host, and Stevens’ standard socket programming text [Ste90] is peppered with some of the subtleties of socket programming on multi-homed servers. The IPv6 addressing model specifically supports assignment of “multiple IPv6 addresses of any type (unicast, anycast, and multicast) or scope” [HD98] to a single interface. Although primarily intended to facilitate renumbering, this IPv6 feature can also be used to simplify firewall interaction via TARP. To form addresses based on unique interface identifiers, RFC 2373 [HD98] recommends using, whenever possible,the 48-bit IEEE 802/Ethernet MAC addresses, and the RFC offers the IEEE EUI-64 [Ins97] as one way to create a globally unique 64-bit identifier from the Ethernet address. The suggested method consists of inserting the two-byte padding pattern FFFE16 after the third byte of the the Ethernet address. To create TARP addresses, we actively vary the pad bytes to select from a large family of IPv6 addresses unique to an interface. We also use a slightly different ordering than the examples in RFC 2373. The first six bytes of our addresses are simply the interface’s Ethernet address with the last two bytes varying as needed. Re-ordering the bits that vary to the least significant bits has the benefit of allowing the last hop router to use a single table entry with a 112-bit prefix to send all such packets to a single host. 3 Some NotationWe now introduce some notation. Suppose a host interface has an Ethernet address of 00:10:4b:63:80:4b, a 64-bit network prefix length and a subnet prefix of 1111:2:3:4. Then the TARP address family available for this interface ranges from 1111:2:3:4:10:4b63:804b:0 through 1111:2:3:4:10:4b63:804b:ffff. We describe the general concept of an address like 1111:2:3:4:10:4b63:804b:0, as Subnet Prefix:Ethernet Address:0 and when the least significant bits of an address correspond to the process group identifier using the address, we write Subnet Prefix:Ethernet Address:pgid. When the discussion requires some notion of a port, a dot after the final address bytes delimits the port, as in Subnet Prefix:Ethernet Address:0.21 for port 21, or Subnet Prefix:Ethernet Address:0.* meaning any port on the address. As a further notational imprecision, quantities like Subnet Prefix and Ethernet Address can be enhanced to refer to some particular network, interface, or host context, as in Subnet Prefix Network 1, or Ethernet Address Host 2, etc. 4 Using a Large Family of Host AddressesHost usage of this this large address family depends on whether the host process acts as client or server. In general, client applications select actively from the pool of available addresses and dynamically configure the appropriate interface. Server processes will use a designated static address. Our decision to connect the address used by a client process to its process group was driven only by a desire to have a group of related processes use the same address. Process groups have no particular special properties, other than being somehow “related” (on Unix systems, they represent the processes descended from a single command line), and the basic concept of transient addressing extends to other similar, natural process relationships. As the basis for the UNIX process group indexing scheme, processes could, in principle, have their own IP addresses, although the clutter and possible performance penalties of establishing a per-process address seem excessive. In a multi-user environment, UIDs, credentials, or session identifiers could sensibly anchor similar addressing schemes, as could a new (as yet unspecified and only dimly imagined) set of kernel concepts and data structure for organizing related processes. More fundmanentally, the requirement is that the addresses assigned to “related programs”--ones that somehow “expect” to be at the same place--be stable, but that “different” programs have different addresses. Thus, a Web browser structured as multiple processes should have a single IP address, while individual ftp commands should have different addresses. Address organization by process group has the advantage that the kernel already supports the notion that a collection of processes is somehow related. The data structures are stable in the UNIX kernel, and bolting the addressing scheme onto them posed little threat to the processing flow for establishment, maintenance, and appropriate removal of the fundamental data structures managing process groups. We are unaware of any basic conflicts created by attaching our concepts to process groups, and found them lively and flexible enough to stress the kernel enhancements dealing with address management. 4.1 How a Client Uses the Large Family of AddressesWhen a client implicitly binds a local address to a socket, the kernel decides which interface to use. Our modified kernel determines what process group the binding client process belongs to, and uses the corresponding Subnet Prefix:Ethernet address:pgid address for the socket. For a process group’s first such socket, this involves assigning a new IPv6 address to the appropriate interface and joining the corresponding Solicited-Node multicast address required by the IPv6 specification [HD98]. Note that by only assigning a TARP address in the case of an implicit bind, the kernel still allows explicit binding to any valid host address a socket program can identify. With an address established for the process group, any other implicitly bound sockets arising from the same process group will use the same address. The address persists on the interface until the process group leader terminates, when the address and its Solicited-Node multicast address are removed. 4.2 How a Server Uses the Large Family of AddressesServer interactions passing through the firewall must occur on the reserved Subnet Prefix:Ethernet Address:0 address. This requirement serves several useful purposes. First, connecting to a server with a dynamic address is like hitting a moving target, so fixing the address allows clients to find servers. Next, fixing the server address (both listening and responding address) further assists the expected server/firewall interaction by omitting any need for server applications to inform interested firewalls of their addressing dynamics. Finally, it provides control for firewalled hosts running servers using wildcard listening sockets. The control requirements are the most subtle, and an example of what could be dubbed “coat tail riding” will clarify the importance of restricting external server activity to a fixed address. Consider an installation with a policy freely allowing outgoing ftp, but not incoming telnet from outside its firewall. A host inside the installation opens an FTP connection from Protected Installation Prefix:Ethernet Address:pgid.5040 to a hacker-compromised site, Compromised FTP Server.21. This complies with the installation’s security policy, and the firewall enables the rule allowing all traffic between the protected host and the compromised server, {Protected Installation Prefix:Ethernet Address:pgid.* <-> Compromised FTP Server.*}. A telnet server listening to a wildcard socket on the ftp client machine could then receive and serve connection requests from the compromised FTP server to Protected Installation Prefix:Ethernet Address:pgid.23 in violation of the installation’s telnet policy. We prevent a server from riding back on the client’s coat tails by forcing server activities to occur on the :0 address and preventing a server’s wild card listening sockets from connecting to inappropriate TARP addresses. Section 5.4 describes a set of socket matching rules that prevent coat tail riding. Note we cannot simply forbid connections to Protected Installation Prefix:Ethernet Address:pgid as active mode FTP will require that the compromised server connect back to the client. The address based firewall rule and designated server address principles lack fine discrimination concerning what services will be served outside the protected bastion. Once a client opens a connection through the firewall to any service, the firewall rule subsequently allows the same client through to any port. In another example, suppose the protected host’s security policy allows ftp connections from the outside, but not telnet. Any authorized user could start an ftp from the outside to Protected Installation Prefix:Ethernet Address:0.21, causing the firewall to allow all traffic between the ftp client and the protected server. With all ports now open to the FTP client, it can now start a telnet session from the outside, unimpeded by the firewall. This problem of finer discrimination of service access has some reasonable remedies short of following application-level protocols, and we discuss them in the security section below. 5 TARP and Sockets5.1 OverviewOur TARP implementation rests on a set of kernel enhancements at the socket level and below. It relies extensively on the socket data structure’s process group identifier, and every AF_INET6 socket receives the process group identifier of the process group creating it. The main intent has been for socket programs to receive, without further programmer intervention, the appropriate TARP address when simply using the vanilla Socket API.1As illustrated above, it matters greatly to the addressing scheme’s security concept whether a process acts as a server or client. As a result the kernel tries to classify a process group as either a client process group or a server process group, and typically assigns a classification to a process group according to the first SOCK_STREAM or SOCK_DGRAM socket usage of a process in a process group. A process group typically exists in an indeterminate state until one of its processes first reads or writes a socket. The exception to the usual indeterminacy occurs when a server spawns a new process group. In this case, the new process group inherits its server state from the parent process group. If a process group’s first datagram or stream socket usage is a read, the enhanced kernel classifies that process group as a server process group. Similarly, if the first such socket usage writes a socket, then the process group receives a client classification. The implementation offers system calls to manipulate a process group’s client server state for socket applications ill-served by the default classification rules, but experience has so far shown them needless: except for auth (see Section 6.1), all tested applications worked without change. 5.2 Client SocketsClient applications wishing to use a TARP address simply create a socket and first write to it (using write, writev, sendmsg, or sendto) or call connect so that the socket receives an implicit bind. During the process of implicitly binding the socket to a local address, the kernel assigns and configures the appropriate TARP address.5.3 Server SocketsLike client sockets, server sockets can receive the source address of their outgoing packets via an implicit bind, but sometimes, this produces undesirable results. Under the general model that servers will create sockets somehow listening on some limited set of addresses for contact from clients, one of our goals is to insure that servers will listen to and respond from the same reserved server address that was contacted by the client. Achieving this goal depends on the transport layer protocol, the Unix flavor, and the server’s socket programming style.TCP servers pose little problem as the connection framework provides adequate structure to determine the correct address for a listening socket to bind to (presumably a reserved server address contacted by a client). We do not force the accept to occur on a server address, but rather view that decision as a administrative choice for the firewall. By allowing connections to valid addresses outside the TARP address family but sharing the same interface, we can also support multi-homed TCP servers. We expect (but do not force) that installations using these extra addresses will have firewalls routinely blocking all traffic involving these non-process group based addresses so that they will be served only from within the protected enclave. UDP client/server applications are slightly more problematic, mostly because the BSD derived kernels do nothing to force a multihomed UDP server to respond from the same address as the destination address of the prompting packet[Ste90, p. 220]. Coupled with the reality that many common UDP servers let the kernel choose the server’s responding address, this effectively causes problems for clients attempting to use UDP sockets to communicate with multi-homed UDP servers running a kernel supporting TARP addressing. The BSD based kernels choose a reply address from the outgoing interface which does not necessarily mean replying with either the same address as the client’s destination address or the designated :0 server address. If a client contacts a UDP server on Subnet Prefix:Ethernet Address:0, the kernel must insure that the response returns from the same server address or the firewall rule set will likely block the response. For UDP servers, the classification of a process group as a server process forces response on the server address, achieving the desired result. The whole classification scheme is mostly forced by BSD’s carefree attitude about ensuring that the UDP server responds from the same address contacted by the client. 5.4 Matching Protocol Control Blocks for Inbound Data to SocketsTo read incoming data, a process typically creates a socket and places it in a state (often blocking, but perhaps polling) awaiting network data. The kernel monitors incoming IP data, classifies it according to transport layer protocol and ports, and finally matches it to a list of available sockets. To work well with TARP addressing, the standard BSD rules for matching network data to sockets need refinement, and the new assignment rules are described below. These assignment rules enforce several operating principles. First, they must prevent coat tail riding by providing the control necessary to prevent wildcard listening sockets from providing network services on a process group based address. Second, the rules must also permit connections to servers on the Ethernet:0 address. Third, they need to connect the appropriate incoming packets to sockets listening specifically to a TARP address. Finally (and optionally), because servers may wish to provide separate services inside the protected bastion, the rules need to allow a server to match sockets listening to a valid address outside the TARP address family. The rule set below achieves these goals.
An example shows how these rules effectively prohibit
coat tail riding. Reconsider the example of Section 4.2,
where a protected client runs both telnet and FTP
servers on a wildcard address, but allows only FTP to
outside the firewall. After the protected client starts an
ftp session to the compromised FTP server, the
compromised site attempts to open a connection to
Protected Installation Prefix:Ethernet Address:pgid.23,
and the firewall passes the malicious SYN packet
according to the rule legitimately established by the ftp allowing {Protected Installation Prefix:Ethernet
Address:pgid.* <->
Compromised FTP Server.*}. As
the telnet server is listening with a wildcard socket,
the the first two rules (each requiring a specified
destination address) will fail to match the SYN packet.
The only remaining candidate is Rule 3(b), but this fails
to match because the telnet server’s listening socket
will belong to a different process group than the ftp.
Thus the packet intending to ride the coat tails of
the open firewall filter is never delivered to the
telnet server socket and the connection is never
opened.
Rules2(c) and 3(c) are optional and permit a host
to offer services on an address outside the TARP
family, presumably to machines within the protected
enclave. They should be filtered outside the protected
enclave and also admit the possibility of firewall
misconfiguration. For installations not requiring their
functionality, a compile option can remove them and
protect against misconfiguration dangers.
As a convenience and measure of control, the modified
kernel provides a system call allowing a program to
request either server or client status. So far, the
client/server applications tested have required no such
manipulations to use the correct addresses, suggesting
that the kernel rules generally provide the desired
results. During development, the modified kernel also
contained a setsockopt call allowing programmer
control of whether a socket uses the server addresss,
but this control appeared superfluous in light of the
server inheritance rule and was removed pending
demonstration of its utility.
Host applications may also explicitly bind to any address
they could otherwise bind to with an unmodified
kernel, although doing so will likely produce useless
results. For example, an application could bind to
the address used by a separate process group, but
the resulting socket would (absent process group
manipulation) have an address with the process group
component not equal to the socket’s process group.
This would always fail the protocol control block
matching algorithm joining sockets and protocol
control blocks, and so no data could flow to the
socket. (A simple code change could prohibit the
ability to misbind, to avoid violating the rule of least
surprise.)
The problems encountered by the unmodified
applications depend variously upon the session layer
protocol, the application layer protocol (auth),
and whether the application uses address based
authentication (e.g. rsh and rlogin). Cognizant of
the terrible security properties of address based
authentication [Bel89], we question the wisdom of
bothering to fix rsh and rlogin, especially with ssh
working, but their functionality can be restored with
modest DNS adaptations discussed in Section 7.3. We
have not performed exhaustive testing. (We do note
that ssh [Ylo96] can use cryptogrpahically-protected
name based authentication; similarly, rlogin et al. are
secure when protected by IPsec [KA98]. If use of such
protocols is considered desirable, we may need to
reopen this issue.)
Of the TCP applications, ssh, ftp, and telnet,
sendmail, and ssh appear to function normally with the modified kernel, while rsh and rlogin both
suffer mildly from authentication problems already
described.
The auth server presents insurmountable problems,
mainly because its poor interaction with TARP
addressing. Consider a server installation wishing
to fire an Ident query to a TARP client’s auth
server about a connection coming from, say, Subnet
Prefix:Ethernet Address:pgid . Ident queries consist
only of the port numbers; the query addresses are
implied by the source and destination address of
the queries. The querying host has two choices
for an address to query, neither of which work.
Querying the Subnet Prefix:Ethernet Address:pgid
will fail, as there will be no server on the process
group address. Querying the server address, Subnet
Prefix:Ethernet Address:0 yields a null result, as the
connection does not originate from that address. This
failure seems a small loss, as the quality of the
information returned from the auth server has
always been greatly disparaged (c.f.[CB94, GS96], and
even its defining RFC [Joh93]). As a workaround,
client installations feeling an obligation to reply to
Ident queries should have little trouble hacking
a small Ident server into client applications, so
that the application itself replies to Ident queries
on its process group address. Of course, such an
implementation would require a setuid helper
program.
By restricting UDP clients and servers to asking only
for services and responding to requests (respectively)
from an interface’s designated server address, then all
the kame/BSD UDP client/server applications work
correctly, unmodified. We wrote simple socket clients to
communicate with inetd’s built-in servers, and
(except for auth) succeeded with both the TCP
and UDP versions of: time, echo, daytime, and
chargen.
We should distinguish that our UDP clients used
unconnected UDP sockets when communicating with
the built-in servers. When requesting service from a
valid address outside the TARP Family, the clients
executed a sendto followed by a read. Following
the rule that processes first reading then writing
sockets are servers, the server responded from its
Subnet Prefix:Ethernet Address:0 address, which
the client successfully read from its unconnected
socket.
Had we instead used a client socket connected to the
valid non-TARP address where the UDP service
request was sent, the UDP client server transaction
would have failed when the reply returned from the
server address. This problem with connected sockets is
not isolated to the modified kernel, but also exists
when interacting with multihomed Berkeley-derived
UDP servers [Ste90, pp. 219-220]. Under BSD, the
server is only guaranteed to reply from an address
chosen from the same interface as the contacted server
address. The server can reply with an address other
than the one contacted by the client (but on the same
interface) and a connected UDP socket will fail to read
the server’s reply.
Our primary guiding principle is that of the virtual
machine, though for implementation reasons we cannot
always follow this. Thus, a “ping” message sent to a
TARP address should be replied to, because a real
machine with that address would respond.
Some ICMP messages create state on the receiving
host. For example, ICMP Redirect changes the local
routing table. Given the security risks posed by this
message [Bel89], it would be nice if these changes could
be restricted to the particular process group to
which the messages were addressed, but that would
require substantial kernel changes. Furthermore, there
are distinct advantages to making certain state
information global, such as that distributed by Path
MTU [MDM96]. Further work is needed in this
area.
For outbound data, the firewall only needs to examine
the outbound destination port, determine whether the
originating address belongs to a family of TARP
addresses, and know whether the protocol involved uses
auxiliary data connections involving other ports.
Assuming it approves the protected client’s transaction
and the underlying protocol requires dynamic port
negotiation, the firewall simply permits any incoming
traffic to that address, regardless of port. The firewall
no longer needs to know any protocol details; it simply
needs to know that the protocol involves secondary
channels. For services using only fixed ports, the
firewall can filter traditionally.
For service requests originating from outside the
protected bastion, the firewall typically would reject all
requests for the services of an address other than a
designated server address. For authorized services, the
firewall permits packets to flow freely between a server
address and the client outside. For example, an FTP
server host administratively configured to provide no
other services outside the firewall would reject all
inbound connection attempts and UDP packets other
than those to the FTP server port. Following a
successful connection, outbound packets to the client
from Subnet Prefix:Ethernet FTP Server Host:0.* are
freely allowed between the approved server and
client.
The firewall can use several approaches to revoke
authorization. TCP connections are easiest: If the
packet triggering authorization comes from a TCP
connection, the firewall simply disallows (pending
possible future authorization) packets between the
two addresses when all related connections (as
determined by address) terminate. Additionally, and
more appropriate for UDP and ICMP, a simple timer
mechanism can revoke authorization some number of
minutes after the last use of an address. Finally, a
protected host can explicitly release an address, upon
process group termination.
Having a host communicate its release of an address
provides the firewall the best general understanding of
when to terminate authorization, but it requires the
protected host to know how to reach its firewall(s). The
best control exists in situations where all relevant
firewalls are either embedded in the host (like a
Distributed Firewall [Bel99]) or or share the same link
layer and are able to see the link local ICMPv6 broadcast
as a result of the host leaving the Solicited-Node
multicast address. Here, the firewall either knows or
has an inkling of the address being removed at process
group termination. The embedded firewall can simply
de-authorize the canceled address from within the
kernel, while the link local firewall will need to
match the Multicast Listener Done message to its
authorized addresses, and cancel upon determining
their unreachability.
For installations where routers lie between the protected
host and the relevant firewall, hosts wishing to
communicate address revocation will need more
complex interaction mechanism with their firewalls.
Using a traditional proxy firewall to protect hosts with
TARP addresses works as before, but will require some
flexibility in configuration to accommodate the large
number of addresses that will be coming and going. For
example, certain outbound filter rules will need to
apply for Subnet Prefix:Ethernet:* where before
they perhaps pertained only to a single address per
host.
Note well that clients are usually well protected,
even if the firewall is slow revoking authorization.
This is because when a process group terminates,
its corresponding address is scrubbed from the
client’s interface, so even if the firewall is tardy
blocking the inbound packets, they fall on deaf ears
when reaching the host. Any nefarious or spurious
inbound packets to unconfigured addresses should
ultimately die an unconsumed death. This is the
fundamental reason why TARP-based hosts can simplify
firewalls.
Because we use only the initial address and port of a
connection, we have prevented new application
protocols from obsoleting otherwise adequate firewall
software. Instead of requiring complex (and possibly
buggy and insecure) new application protocol-aware
software to be written and installed in the firewall, to
keep up with new application protocol developments,
the firewall administrator need only know what port the service uses to start its transactions and filter
accordingly. Thus, only minor firewall configuration
changes will accommodate a large class of new,
unforseeable applications. (This work does not discuss
application-specific audit trails or intrusion detection
system events that an application-aware firewall may
collect today.)
At one level, routers are not affected by TARP
addresses. They will see an IP address for which they
have no corresponding link-layer address cached;
accordingly, they will resort to the Neighbor Discovery
Protocol [NNS98], in normal fashion. But TARP-aware
routers and protocols can do better.
Widespread use of TARP would require considerably
more storage on routers for link-layer addresses. But
this set of IP addresses all share a common prefix and a
common link-layer address. Accordingly, routers could
employ a single mapping of prefix to link-layer
address. Unfortunately, that is not supported by NDP.
If necessary in a given environment, this could
be faked by having a host pretend to be a stub
router; however, this would require the host to
participate in routing protocols, which is generally
considered to be a bad idea. A better solution would
be to extend NDP to handle host address prefix
lengths.
As evident in the discussion concerning rlogin
and rsh, TARP Addressing encounters problems
interacting with DNS. Of particular concern is the
dubious practice of so-called “double-reverse lookups”
(cf. [ZCC00] and [GS96]) where a DNS client attempts
address based authentication by first looking up the
hostname for an IP address and subsequently checking
the IP address(es) for that hostname, requiring a
match to proceed. But for double-reverse lookups, most
dynamic address clients would probably work happily
by simply registering their server address with the
appropriate name server; a wild-card PTR record
would yield the same name for all possible TARP
addresses.
IPv6-to-hostname lookups pose no problem, as
the DNS specification for IPv6 [TH95] provides
the necessary wildcarding. For example a name
server could associate a wildcard to a PTR record
for, say, *.12-nibble Ethernet Address.16-nibble
Subnet Prefix.IP6.ARPA (ordered appropriately) and
correctly return the hostname for a host’s Process
Group Dynamic Addresses. Another possibility is to
query the host directly with an ICMPv6 FQDN
query, as described in an IETF draft “IPv6 Node
Information Queries” [Cra99] and already implemented
by kame.
Hostname-to-address queries are more problematic. A
standalone name server presumably has no information
about a host’s active addresses, and absent refinements
in the name server/resolver software and protocols, a
complete reply to a hostname-to-address query would
consist of all 216 Process Group Dynamic Addresses.
Unfortunately, simply listing a host’s server address is
insufficient, as a double-reverse lookup will find
inconsistency when comparing a client’s dynamic
address to the its server address.
Three straightforward solutions to the basic problem
come to mind. The simplest treats a Process Group
Dynamic Address client as a domain. Its parent DNS
server delegates authority to the client and simply
refers queries to the clients themselves for final
resolution. The clients can then run a pared down
name server which responds with the appropriate set of
addresses for a name-to-address query.
In another method, the TARP clients use the DNS
UPDATE message [VET+97] to inform the relevant
name servers of their evolving address state. It is not
clear how well this would work in practice, as problems
could arise in situations where the querying agent has a
faster path to the name server than the TARP
client. A clever TARP client could conceivably delay
outgoing datagrams from a new TARP address until a
DNS update could be verified, but this might prove
unnecessarily cumbersome.
A third method embeds the process group portion of
the address into a synthetic hostname and places the
corresponding 216 possible hostnames as entries in the
nameserver database. For example the fictitious TARP
client host foo.bar.com would have name server entries
for 0%foo.bar.com - 65535%foo.bar.com, where we use
the % to mean a meta-character (determined locally by
bar.com) denoting that the hostname uses Process
Group Dynamic Addressing. The name server would
conceptually have 216 PTR records mapping each Dynamic Address family member to its corresponding
expanded hostname (e.g. pgid%foo.bar.com) The
cross-check query invokes an address-to-hostname
query for foo.bar.com Subnet Prefix:foo.bar.com
Ethernet Address:pgid, which eventually returns
pgid%foo.bar.com; this is followed by a name-to-address
query about pgid%foo.bar.com, which will provide a
consistent check if the name servers are configured
correctly.
This third method scales poorly, taxing the DNS
servers serving TARP client domains by forcing them
to store 217 records per client. While some of this can
be dealt with by “syntactic sugar”--the servers need
not store all 217 records, but can simply generate them
on the fly--we are left without a canonical name
for the host. This affects host configuration files
(/etc/hosts, .rhosts, etc.), email, and other services
that require one true name for each machine. (In a
sense, this is carrying our virtual machine metaphor
too far.)
Finally, we note that the problem only arises because
hosts use name-based authentication, and hence need
the extra protection of the double-reverse lookup.
If this practice were to be abandoned--and we
strongly suggest that that be done in any event--the
name-to-address lookup could be omitted, thereby
eliminating the problem.
Our testbed implementation consists of a handful of
subroutines and about 400 lines of C code inside the
kame/FreeBSD 3.4 kernel. The kernel enhancements
appear to port easily to the other BSD families kame
supports (NetBSD, BSD/OS and OpenBSD). The
essential code flow remains unchanged, and the bulk
of the software changes sit inside a single kame
subroutine, in6_selectsrc (basically an IPv6 re-work
of parts of the function of in_pcbconnect [WS95]),
called when assigning a local IPv6 address to a
protocol control block.
The kame/FreeBSD code has a parameter for the
maximum process identifier, PID_MAX, which comes
configured at 99999. Our implementation reduces this
maximum to 65535 to insure the resulting process
group identifiers fit entirely into the corresponding two
bytes of the TARP address.
Given the uniqueness properties implicit the Ethernet
address based addresses, the implementation does not
perform Duplicate Address Detection [TN98] when
adding a TARP address to an interface. This enhances
performance by preventing unnecessary delays when
configuring an interface with a new address, and yet
still complies with the RFC’s notion that addresses
based on unique interface identifiers need only check
for duplicate addresses at initialization ([TN98],
Section 5.4). When bringing up an interface, the
kame IPv6 implementation configures it with a
proper Ethernet-derived, EUI-64 compliant link local
address. This address receives proper duplicate
address detection, and so any (presumably relatively
rare) problems caused by locally duplicate Ethernet
addresses should be caught when the interface goes
up.
Regrettably, the IPv6 Address Autoconfiguration
Standard [TN98] lacks a sub-net mask capability that
would allow hosts to reserve a range of addresses under
a mask. Were such a facility available, a TARP host
could make a single duplicate address detection query
to reserve a whole range of an address family with a
single query. In this case, hosts would be freed of
needing to use the Ethernet address-based connection
for its implied uniqueness properties, and could still
safely reserve a range of without fear of address
collisions. Local administrators would then have the
freedom to assign the host address portion according to
their own rules, and could conceivably use even larger
address families.
We could, in principle, do duplicate address detection
for each TARP address. However, the overhead, and
particularly the 1-second timeout, are prohibitive.
There is some conflict between this scheme and the
IPv6 privacy extension standard [ND01]. The easiest
solution is to use the suggested algorithms to generate
just a replacement for the MAC address portion
of the address, rather than the full low-order 64
bits.
An alternative would be to compare the prefix of the
address in incoming packets to the base address
of the interface, and use existing speedy look-up
mechanisms to ascertain if the associated process group
exists. That is, assume that a packet is destined
for a machine if the base address is valid and the
process group exists, rather than checking if such
a process group has actually performed network
operations.
There is one other area of some concern. As noted
earlier, for each new TARP address allocated it is
necessary to join the appropriate Solicited-Node
multicast address group. For some hardware designs, it
is necessary to load the group address onto the
controller chip. Depending on the chip and driver
design, this may be an expensive operation.
Successful IPSEC interaction essentially depends on the
ability of hosts running TARP addressing to conduct
the necessary key exchanges. Fortunately, ISAKMP
and IKE provide an adequate framework to support
TARP. A simple solution uses either of ISAKMP’s
[Pip98] address range or subnet identification payloads
to designate that the ISAKMP peer negotiating the
key exchange is the address range or subnet of the
TARP Family. Using keys negotiated for the address
family thus permits processes to use IPSEC in
conjunction with TARP addresses.
Depending on the threat model, installations may wish
to eschew IKE’s Base Quick Mode [HC98, p. 16].
Lacking perfect forward secrecy, Base Quick Mode
admits the possibility that an authorized host process
able to obtain its keying material can use that
knowledge to determine keying material for other
processes, including those belonging to other users. The
Quick Mode key exchange payload option [HC98]
prevents this problem by providing the necessary
perfect forward secrecy at the cost of an additional
exponentiation. Alternatively, use of cryptographically
strong random number generators, ciphers resistant to
chosen-plaintext attacks, and suitable crypto-APIs
(i.e., those that will not, under any circumstances,
disclose a session key to an application) can prevent
this attack.
At this point in the evolution of IPv6, it would not be
acceptable to introduce a new scheme that would break
compatibility with existing systems. For the most part,
we have not done so.
Servers do not notice anything different, with the
possible exception of contacts from many more clients.
That depends on whether the server is noting addresses
or names, and in the latter case, on what mechanism is
selected for address-to-name resolution.
Servers that do rely on host names for authentication
may have problems. As noted, we recommend that that
practice be abandoned in any event; we do not
consider its difficulties to be a disadvantage of our
scheme.
In some sense, the simplicity of the security enabled
depends on the assortment and type of network
services a protected host needs to provide. TARP
addressing offers simpler, more robust protection for
hosts acting primarily as clients rather than servers.
This does not mean that many server configurations
cannot receive adequate protection using the method
and a compatible firewall, rather that servers will need
to be much more careful about their configuration to
achieve results comparable to hosts acting entirely as
clients.
Again, consider a server that wishes to permit incoming
access to ftpd services from the network, but not
telnetd. A devious client could start an FTP session
to the server, which the server firewall permits, and
once the firewall enables the rule {Devious Host.* <->
Server:0.*} the client can now connect to the server’s
telnet port from the network in violation of the
intended security policy.
One obvious solution to this problem is to use an
embedded server access control mechanism (something
like TCP Wrappers [Ven92]) to regulate what will be
served to the outside on the Ethernet:0 address.
After breezing by the firewall with an attempted
telnet request, the previous paragraph’s devious
client could be refused a connection to telnetd
according to its source address. Similarly, the inetd
super-server has a address option that will bind
the server to a particular address, and this should
provide adequate control for services running under
inetd.
A simpler approach would be for the firewall to
continue to monitor connection attempts to well-known
services and filter accordingly. This deviates from the
concept of filtering solely based on address, but not
necessarily to the point of continuously following
application level protocols. In our example, connection
attempts to all privileged ports other than the ftp
ports would be blocked at the firewall, and the ftp
could proceed.
Clearly, relying solely on a packet filtering approach to
provide server security becomes more unworkable as
the variety of externally-accessible services grows. As
the number of services increases, sensible packet
filter rules become more difficult to specify, and
the likelihood of consistency problems caused by
interaction effects increases correspondingly. This is an
old and familiar firewalling problem, traditionally best
solved with a mixture of strategies combining packet
filters and proxies. We make no claims that TARP
Addressing solves any of these complexities, only that
there should be cases where firewalls can use the
addressing scheme to provide equivalent security
without following application level protocols.
Server facilities using a multi-homed server strategy
to serve a variety of domains from the same host
[Ste90, p. 93], may not be a good match for TARP’s
preference to serve from a single fixed address per
interface. At the least, they seem incompatible with
inetd’s wildcard listening sockets and would seem
better off if the relevant servers bound only to
the the address they are serving. This is more a
limitation of BSD, rather than TARP addressing, per
se.
The security concepts make no contributions to
solving problems of inside threats, but this is a
recognized limitation of firewalls in general ([Cha92],
[CB94]).
Our implementation cannot support more than a single
TARP address per interface, and doing so would
require extensive kernel modifications. This is for two
reasons. First, when faced with an outgoing address
decision, the kernel already knows which interface
to use, and the implementation determines the
Ethernet address of the outgoing interface for use in
address computation. Even if we had an extra, valid Ethernet address to use, it would be difficult for
the kernel to determine when to assign it. The
second set of problems is that all of the problems of
multi-homed BSD servers described above would
occur.
Although our implementation was built on the UNIX
notion of “process groups”, it is clearly not necessary
to do it that way. Two preconditions are necessary for
a substitute mechanism.
First, and most obvious, there has to be some way to
generate a 16-bit number not currently in use. Clearly,
a counter and an in-use list will suffice. The harder
problem is somehow assigning this number to a
program or group of related programs. All “related”
programs--we define “related” as meaning “they would
expect to have the same IP address”--must somehow
be linked to this number. The notion of a process
group in UNIX captures these semantics quite well; the
fact that process groups have the right sort of number
is simply a happy accident.
We have shown TARP to be a useful new possibility
made available by IPv6’s Addressing Architecture.
Using TARP addresses greatly simplifies firewalling
decisions for the machines protecting either clients or
servers. Furthermore, by partitioning the address space
into client and server addresses and only configuring
client network addresses as needed by client activity, it
looks to provide particular advantages for hosts acting
mostly as network service consumers.
A number of people involved in the IETF’s IPv6 effort
(Steve Deering, Thomas Narten, Erik Nordmark, Keith
Moore, and others) provided valuable advice. Jon
Crowcroft noted that similar facilities could be built on
top of Linux’s IP Address Masquerading facility. Also,
great thanks are due to John Ioannidis for helpful and
generous use of the computer laboratory facilties he
maintains. We gratefully acknowledge the code,
excellent assistance and cheerful bug fixes provided by
the kame project.
[Bel89] Steven M. Bellovin. Security problems in
the TCP/IP protocol suite. Computer
Communications Review, 19(2):32-48,
April 1989.
[Bel94] S. Bellovin. On many addresses per host.
Request for Comments 1681, Internet
Engineering Task Force, August 1994.
[Bel99] Steven M. Bellovin. Distributed firewalls.
;login:, pages 39-47, November 1999.
[CB94] William R. Cheswick
and Steven M. Bellovin. Firewalls and
Internet Security: Repelling the Wily
Hacker. Addison-Wesley, Reading, MA,
1994.
[Cha92] D. Brent
Chapman. Network (in)security through
IP packet filtering. In Proceedings of the
Third Usenix UNIX Security Symposium,
pages 63-76, Baltimore, MD, September
1992.
[Cra99] Matt Crawford. IPv6 node information
queries, 1999. Work in progress.
[GS96] Simson Garfinkel and Gene Spafford.
Practical Unix and Internet Security.
O’Reilly, Sebastopol, CA, 1996.
[HC98] D. Harkins and D. Carrel. The internet
key exchange (IKE). Request for
Comments 2409, Internet Engineering
Task Force, November 1998.
[HD98] R. Hinden and S. Deering. IP version 6 addressing architecture. Request for
Comments 2373, Internet Engineering
Task Force, July 1998.
[Ins97] Institute of Electrical and Electronics
Engineers. Guidelines for 64-bit global
identifier EUI-64 registration authority,
1997.
[Joh93] M. St. Johns. Identification protocol.
Request for Comments 1413, Internet
Engineering Task Force, January 1993.
[KA98] S. Kent and R. Atkinson. Security
architecture for the internet protocol.
Request for Comments 2401, Internet
Engineering Task Force, November 1998.
[MDM96] J. McCann, S. Deering, and J. Mogul.
Path MTU discovery for IP version 6.
Request for Comments 1981, Internet
Engineering Task Force, August 1996.
[ND01] T. Narten and R. Draves.
Privacy extensions for stateless address
autoconfiguration in IPv6. Request for
Comments 3041, Internet Engineering
Task Force, January 2001.
[NNS98] T. Narten, E. Nordmark, and
W. Simpson. Neighbor discovery for IP
version 6 (ipv6). Request for Comments
2461, Internet Engineering Task Force,
December 1998.
[Pip98] D. Piper. The internet IP security
domain of interpretation for ISAKMP.
Request for Comments 2407, Internet
Engineering Task Force, November 1998.
[Ste90] W. Richard Stevens. UNIX Network
Programming: Networking APIs: Sockets
and XTI, volume 1. Prentice-Hall,
Englewood Cliffs, NJ, second edition,
1990.
[TH95] S. Thomson and C. Huitema. DNS
extensions to support IP version 6.
Request for Comments 1886, Internet
Engineering Task Force, December 1995.
[TN98] S. Thomson and T. Narten. IPv6 stateless
address autoconfiguration. Request for
Comments 2462, Internet Engineering
Task Force, December 1998.
[Ven92] Wietse Venema. TCP WRAPPER:
Network monitoring, access control and
booby traps. In Proceedings of the Third
Usenix UNIX Security Symposium, pages
85-92, Baltimore, MD, September 1992.
[VET+97] P. Vixie, Ed., S. Thomson, Y. Rekhter,
and J. Bound. Dynamic updates in the
domain name system (DNS UPDATE).
Request for Comments 2136, Internet
Engineering Task Force, April 1997.
[WS95] Gary R. Wright and W. Richard Stevens.
TCP/IP Illustrated: The
Implementation, volume 2.
Addison-Wesley, Reading, MA, 1995.
[Ylo96] Tatu
Ylonen. SSH - secure login connections
over the internet. In Proceedings of the
Sixth Usenix UNIX Security Symposium,
pages 37-42, July 1996.
[ZCC00] Elizabeth D. Zwicky, Simon Cooper, and
D. Brent Chapman. Building Internet
Firewalls. O’Reilly, Sebastopol, CA,
second edition, 2000.
|
This paper was originally published in the
Proceedings of the 10th USENIX Security Symposium,
August 1317, 2001, Washington, D.C., USA
Last changed: 2 Jan. 2002 ml |
|