|Title||IPv4 Reassembly Errors at High Data Rates
|Author||J. Heffner, M. Mathis, B.
Network Working Group J. Heffner
Request for Comments: 4963 M. Mathis
Category: Informational B. Chandler
IPv4 Reassembly Errors at High Data Rates
Status of This Memo
This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.
Copyright (C) The IETF Trust (2007).
IPv4 fragmentation is not sufficiently robust for use under some
conditions in today's Internet. At high data rates, the 16-bit IP
identification field is not large enough to prevent frequent
incorrectly assembled IP fragments, and the TCP and UDP checksums are
insufficient to prevent the resulting corrupted datagrams from being
delivered to higher protocol layers. This note describes some easily
reproduced experiments demonstrating the problem, and discusses some
of the operational implications of these observations.
The IPv4 header was designed at a time when data rates were several
orders of magnitude lower than those achievable today. This document
describes a consequent scale-related failure in the IP identification
(ID) field, where fragments may be incorrectly assembled at a rate
high enough that it is likely to invalidate assumptions about data
integrity failure rates.
That IP fragmentation results in inefficient use of the network has
been well documented [Kent87]. This note presents a different kind
of problem, which can result not only in significant performance
degradation, but also frequent data corruption. This is especially
pertinent due to the recent proliferation of UDP bulk transport tools
that sometimes fragment every datagram.
Additionally, there is some network equipment that ignores the Don't
Fragment (DF) bit in the IP header to work around MTU discovery
problems [RFC2923]. This equipment indirectly exposes properly
implemented protocols and applications to corrupt data.
2. Wrapping the IP ID Field
The Internet Protocol standard [RFC0791] specifies:
"The choice of the Identifier for a datagram is based on the need
to provide a way to uniquely identify the fragments of a
particular datagram. The protocol module assembling fragments
judges fragments to belong to the same datagram if they have the
same source, destination, protocol, and Identifier. Thus, the
sender must choose the Identifier to be unique for this source,
destination pair and protocol for the time the datagram (or any
fragment of it) could be alive in the Internet."
Strict conformance to this standard limits transmissions in one
direction between any address pair to no more than 65536 packets per
protocol (e.g., TCP, UDP, or ICMP) per maximum packet lifetime.
Clearly, not all hosts follow this standard because it implies an
unreasonably low maximum data rate. For example, a host sending
1500-byte packets with a 30-second maximum packet lifetime could send
at only about 26 Mbps before exceeding 65535 packets per packet
lifetime. Or, filling a 1 Gbps interface with 1500-byte packets
requires sending 65536 packets in less than 1 second, an unreasonably
short maximum packet lifetime, being less than the round-trip time on
some paths. This requirement is widely ignored.
Additionally, it is worth noting that reusing values in the IP ID
field once per 65536 datagrams is the best case. Some
implementations randomize the IP ID to prevent leaking information
out of the kernel [Bellovin02], which causes reuse of the IP ID field
to occur probabilistically at all sending rates.
IP receivers store fragments in a reassembly buffer until all
fragments in a datagram arrive, or until the reassembly timeout
expires (15 seconds is suggested in [RFC0791]). Fragments in a
datagram are associated with each other by their protocol number, the
value in their ID field, and by the source/destination address pair.
If a sender wraps the ID field in less than the reassembly timeout,
it becomes possible for fragments from different datagrams to be
incorrectly spliced together ("mis-associated"), and delivered to the
upper layer protocol.
A case of particular concern is when mis-association is self-
propagating. This occurs, for example, when there is reliable
ordering of packets and the first fragment of a datagram is lost in
the network. The rest of the fragments are stored in the fragment
reassembly buffer, and when the sender wraps the ID field, the first
fragment of the new datagram will be mis-associated with the rest of
the old datagram. The new datagram will be now be incomplete (since
it is missing its first fragment), so the rest of it will be saved in
the fragment reassembly buffer, forming a cycle that repeats every
65536 datagrams. It is possible to have a number of simultaneous
cycles, bounded by the size of the fragment reassembly buffer.
IPv6 is considerably less vulnerable to this type of problem, since
its fragment header contains a 32-bit identification field [RFC2460].
Mis-association will only be a problem at packet rates 65536 times
higher than for IPv4.
3. Effects of Mis-Associated Fragments
When the mis-associated fragments are delivered, transport-layer
checksumming should detect these datagrams as incorrect and discard
them. When the datagrams are discarded, it could create a
performance problem for loss-feedback congestion control algorithms,
particularly when a large congestion window is required, since it
will introduce a certain amount of non-congestive loss.
Transport checksums, however, may not be designed to handle such high
error rates. The TCP/UDP checksum is only 16 bits in length. If
these checksums follow a uniform random distribution, we expect mis-
associated datagrams to be accepted by the checksum at a rate of one
per 65536. With only one mis-association cycle, we expect corrupt
data delivered to the application layer once per 2^32 datagrams.
This number can be significantly higher with multiple concurrent
With non-random data, the TCP/UDP checksum may be even weaker still.
It is possible to construct datasets where mis-associated fragments
will always have the same checksum. Such a case may be considered
unlikely, but is worth considering. "Real" data may be more likely
than random data to cause checksum hot spots and increase the
probability of false checksum match [Stone98]. Also, some
applications or higher-level protocols may turn off checksumming to
increase speed, though this practice has been found to be dangerous
for other reasons when data reliability is important [Stone00].
4. Experimental Observations
To test the practical impact of fragmentation on UDP, we ran a series
of experiments using a UDP bulk data transport protocol that was
designed to be used as an alternative to TCP for transporting large
data sets over specialized networks. The tool, Reliable Blast UDP
(RBUDP), part of the QUANTA networking toolkit [QUANTA], was selected
because it has a clean interface which facilitated automated
experiments. The decision to use RBUDP had little to do with the
details of the transport protocol itself. Any UDP transport protocol
that does not have additional means to detect corruption, and that
could be configured to use IP fragmentation, would have the same
In order to diagnose corruption on files transferred with the UDP
bulk transfer tool, we used a file format that included embedded
sequence numbers and MD5 checksums in each fragment of each datagram.
Thus, it was possible to distinguish random corruption from that
caused by mis-associated fragments. We used two different types of
files. One was constructed so that all the UDP checksums were
constant -- we will call this the "constant" dataset. The other was
constructed so that UDP checksums were uniformly random -- the
"random" dataset. All tests were done using 400 MB files, sent in
1524-byte datagrams so that they were fragmented on standard Fast
Ethernet with a 1500-byte MTU.
The UDP bulk file transport tool was used to send the datasets
between a pair of hosts at slightly less than the available data rate
(100 Mbps). Near the beginning of each flow, a brief secondary flow
was started to induce packet loss in the primary flow. Throughout
the life of the primary flow, we typically observed mis-association
rates on the order of a few hundredths of a percent.
Tests run with the "constant" dataset resulted in corruption on all
mis-associated fragments, that is, corruption on the order of a few
hundredths of a percent. In sending approximately 10 TB of "random"
datasets, we observed 8847668 UDP checksum errors and 121 corruptions
of the data due to mis-associated fragments.
5. Preventing Mis-Association
The most straightforward way to avoid mis-association is to avoid
fragmentation altogether by implementing Path MTU Discovery [RFC1191]
[RFC4821]. However, this is not always feasible for all
applications. Further, as a work-around for MTU discovery problems
[RFC2923], some TCP implementations and communications gear provide
mechanisms to disable path MTU discovery by clearing or ignoring the
DF bit. Doing so will expose all protocols using IPv4, even those
that participate in MTU discovery, to mis-association errors.
If IP fragmentation is in use, it may be possible to reduce the
timeout sufficiently so that mis-association will not occur.
However, there are a number of difficulties with such an approach.
Since the sender controls the rate of packets sent and the selection
of IP ID, while the receiver controls the reassembly timeout, there
would need to be some mutual assurance between each party as to
participation in the scheme. Further, it is not generally possible
to set the timeout low enough so that a fast sender's fragments will
not be mis-associated, yet high enough so that a slow sender's
fragments will not be unconditionally discarded before it is possible
to reassemble them. Therefore, the timeout and IP ID selection would
need to be done on a per-peer basis. Also, it is likely NAT will
break any per-peer tables keyed by IP address. It is not within the
scope of this document to recommend solutions to these problems,
though we believe a per-peer adaptive timeout is likely to prevent
mis-association under circumstances where it would most commonly
A case particularly worth noting is that of tunnels encapsulating
payload in IPv4. To deal with difficulties in MTU Discovery
[RFC4459], tunnels may rely on fragmentation between the two
endpoints, even if the payload is marked with a DF bit [RFC4301]. In
such a mode, the two tunnel endpoints behave as IP end hosts, with
all tunneled traffic having the same protocol type. Thus, the
aggregate rate of tunneled packets may not exceed 65536 per maximum
packet lifetime, or tunneled data becomes exposed to possible mis-
association. Even protocols doing MTU discovery such as TCP will be
affected. Operators of tunnels should ensure that the receiving
end's reassembly timeout is short enough that mis-association cannot
occur given the tunnel's maximum rate.
6. Mitigating Mis-Association
It is difficult to concisely describe all possible situations under
which fragments might be mis-associated. Even if an end host
carefully follows the specification, ensuring unique IP IDs, the
presence of NATs or tunnels may expose applications to IP ID space
conflicts. Further, devices in the network that the end hosts cannot
see or control, such as tunnels, may cause mis-association. Even a
fragmenting application that sends at a low rate might possibly be
exposed when running simultaneously with a non-fragmenting
application that sends at a high rate. As described above, the
receiver might implement to reduce or eliminate the possibility of
conflict, but there is no mechanism in place for a sender to know
what the receiver is doing in this respect. As a consequence, there
is no general mechanism for an application that is using IPv4
fragmentation to know if it is deterministically or statistically
protected from mis-associated fragments.
Under circumstances when it is impossible or impractical to prevent
mis-association, its effects may be mitigated by use of stronger
integrity checking at any layer above IP. This is a natural side
effect of using cryptographic authentication. For example, IPsec AH
[RFC4302] will discard any corrupted datagrams, preventing their
deliver to upper layers. A stronger transport layer checksum such as
SCTP's, which is 32 bits in length [RFC2960], may help significantly.
At the application layer, SSH message authentication codes [RFC4251]
will prevent delivery of corrupted data, though since the TCP
connection underneath is not protected, it is considered invalid and
the session is immediately terminated. While stronger integrity
checking may prevent data corruption, it will not prevent the
potential performance impact described above of non-congestive loss
on congestion control at high congestion windows.
It should also be noted that mis-association is not the only possible
source of data corruption above the network layer [Stone00]. Most
applications for which data integrity is critically important should
implement strong integrity checking regardless of exposure to mis-
In general, applications that rely on IPv4 fragmentation should be
written with these issues in mind, as well as those issues documented
in [Kent87]. Applications that rely on IPv4 fragmentation while
sending at high speeds (the order of 100 Mbps or higher) and devices
that deliberately introduce fragmentation to otherwise unfragmented
traffic (e.g., tunnels) should be particularly cautious, and
introduce strong mechanisms to ensure data integrity.
7. Security Considerations
If a malicious entity knows that a pair of hosts are communicating
using a fragmented stream, it may be presented with an opportunity to
corrupt the flow. By sending "high" fragments (those with offset
greater than zero) with a forged source address, the attacker can
deliberately cause corruption as described above. Exploiting this
vulnerability requires only knowledge of the source and destination
addresses of the flow, its protocol number, and fragment boundaries.
It does not require knowledge of port or sequence numbers.
If the attacker has visibility of packets on the path, the attack
profile is similar to injecting full segments. Using this attack
makes blind disruptions easier and might possibly be used to cause
degradation of service. We believe only streams using IPv4
fragmentation are likely vulnerable. Because of the nature of the
problems outlined in this document, the use of IPv4 fragmentation for
critical applications may not be advisable, regardless of security
8. Informative References
[Kent87] Kent, C. and J. Mogul, "Fragmentation considered
harmful", Proc. SIGCOMM '87 vol. 17, No. 5, October
[RFC2923] Lahey, K., "TCP Problems with Path MTU Discovery", RFC
2923, September 2000.
[RFC0791] Postel, J., "Internet Protocol", STD 5, RFC 791,
[RFC1191] Mogul, J. and S. Deering, "Path MTU discovery", RFC
1191, November 1990.
[Stone98] Stone, J., Greenwald, M., Partridge, C., and J. Hughes,
"Performance of Checksums and CRC's over Real Data",
IEEE/ ACM Transactions on Networking vol. 6, No. 5,
[Stone00] Stone, J. and C. Partridge, "When The CRC and TCP
Checksum Disagree", Proc. SIGCOMM 2000 vol. 30, No. 4,
[QUANTA] He, E., Alimohideen, J., Eliason, J., Krishnaprasad, N.,
Leigh, J., Yu, O., and T. DeFanti, "Quanta: a toolkit
for high performance data delivery over photonic
networks", Future Generation Computer Systems Vol. 19,
No. 6, August 2003.
[Bellovin02] Bellovin, S., "A Technique for Counting NATted Hosts",
Internet Measurement Conference, Proceedings of the 2nd
ACM SIGCOMM Workshop on Internet Measurement, November
[RFC2460] Deering, S. and R. Hinden, "Internet Protocol, Version 6
(IPv6) Specification", RFC 2460, December 1998.
[RFC2960] Stewart, R., Xie, Q., Morneault, K., Sharp, C.,
Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M.,
Zhang, L., and V. Paxson, "Stream Control Transmission
Protocol", RFC 2960, October 2000.
[RFC4251] Ylonen, T. and C. Lonvick, "The Secure Shell (SSH)
Protocol Architecture", RFC 4251, January 2006.
[RFC4301] Kent, S. and K. Seo, "Security Architecture for the
Internet Protocol", RFC 4301, December 2005.
[RFC4302] Kent, S., "IP Authentication Header", RFC 4302, December
[RFC4459] Savola, P., "MTU and Fragmentation Issues with In-the-
Network Tunneling", RFC 4459, April 2006.
[RFC4821] Mathis, M. and J. Heffner, "Packetization Layer Path MTU
Discovery", RFC 4821, March 2007.
Appendix A. Acknowledgements
This work was supported by the National Science Foundation under
Grant No. 0083285.
John W. Heffner
Pittsburgh Supercomputing Center
4400 Fifth Avenue
Pittsburgh, PA 15213
Pittsburgh Supercomputing Center
4400 Fifth Avenue
Pittsburgh, PA 15213
Pittsburgh Supercomputing Center
4400 Fifth Avenue
Pittsburgh, PA 15213
Full Copyright Statement
Copyright (C) The IETF Trust (2007).
This document is subject to the rights, licenses and restrictions
contained in BCP 78, and except as set forth therein, the authors
retain all their rights.
This document and the information contained herein are provided on an
"AS IS" basis and THE CONTRIBUTOR, THE ORGANIZATION HE/SHE REPRESENTS
OR IS SPONSORED BY (IF ANY), THE INTERNET SOCIETY, THE IETF TRUST AND
THE INTERNET ENGINEERING TASK FORCE DISCLAIM ALL WARRANTIES, EXPRESS
OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF
THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED
WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
The IETF takes no position regarding the validity or scope of any
Intellectual Property Rights or other rights that might be claimed to
pertain to the implementation or use of the technology described in
this document or the extent to which any license under such rights
might or might not be available; nor does it represent that it has
made any independent effort to identify any such rights. Information
on the procedures with respect to rights in RFC documents can be
found in BCP 78 and BCP 79.
Copies of IPR disclosures made to the IETF Secretariat and any
assurances of licenses to be made available, or the result of an
attempt made to obtain a general license or permission for the use of
such proprietary rights by implementers or users of this
specification can be obtained from the IETF on-line IPR repository at
The IETF invites any interested party to bring to its attention any
copyrights, patents or patent applications, or other proprietary
rights that may cover technology that may be required to implement
this standard. Please address the information to the IETF at
Funding for the RFC Editor function is currently provided by the