RFC 2525
Known TCP Implementation Problems
Pages: 61
Informational
Informational
Part 2 of 3 – Pages 17 to 40
2.6.
Name of Problem Extra additive constant in congestion avoidance Classification Congestion control / performance Description RFC 1122 section 4.2.2.15 states that TCP MUST implement Jacobson's "congestion avoidance" algorithm [Jacobson88], which calls for increasing the congestion window, cwnd, by: MSS * MSS / cwnd for each ACK received for new data [RFC2001]. This has the effect of increasing cwnd by approximately one segment in each round trip time. Some TCP implementations add an additional fraction of a segment (typically MSS/8) to cwnd for each ACK received for new data [Stevens94, Wright95]: (MSS * MSS / cwnd) + MSS/8 These implementations exhibit "Extra additive constant in congestion avoidance". Significance May be detrimental to performance even in completely uncongested environments (see Implications). In congested environments, may also be detrimental to the performance of other connections.
Implications
The extra additive term allows a TCP to more aggressively open its
congestion window (quadratic rather than linear increase). For
congested networks, this can increase the loss rate experienced by
all connections sharing a bottleneck with the aggressive TCP.
However, even for completely uncongested networks, the extra
additive term can lead to diminished performance, as follows. In
congestion avoidance, a TCP sender probes the network path to
determine its available capacity, which often equates to the
number of buffers available at a bottleneck link. With linear
congestion avoidance, the TCP only probes for sufficient capacity
(buffer) to hold one extra packet per RTT.
Thus, when it exceeds the available capacity, generally only one
packet will be lost (since on the previous RTT it already found
that the path could sustain a window with one less packet in
flight). If the congestion window is sufficiently large, then the
TCP will recover from this single loss using fast retransmission
and avoid an expensive (in terms of performance) retransmission
timeout.
However, when the additional additive term is used, then cwnd can
increase by more than one packet per RTT, in which case the TCP
probes more aggressively. If in the previous RTT it had reached
the available capacity of the path, then the excess due to the
extra increase will again be lost, but now this will result in
multiple losses from the flight instead of a single loss. TCPs
that do not utilize SACK [RFC2018] generally will not recover from
multiple losses without incurring a retransmission timeout
[Fall96,Hoe96], significantly diminishing performance.
Relevant RFCs
RFC 1122 requires use of the "congestion avoidance" algorithm.
RFC 2001 outlines the fast retransmit/fast recovery algorithms.
RFC 2018 discusses the SACK option.
Trace file demonstrating it
Recorded using tcpdump running on the same FDDI LAN as host A.
Host A is the sender and host B is the receiver. The connection
establishment specified an MSS of 4,312 bytes and a window scale
factor of 4. We omit the establishment and the first 2.5 MB of
data transfer, as the problem is best demonstrated when the window
has grown to a large value. At the beginning of the trace
excerpt, the congestion window is 31 packets. The connection is
never receiver-window limited, so we omit window advertisements
from the trace for clarity.
11:42:07.697951 B > A: . ack 2383006 11:42:07.699388 A > B: . 2508054:2512366(4312) 11:42:07.699962 A > B: . 2512366:2516678(4312) 11:42:07.700012 B > A: . ack 2391630 11:42:07.701081 A > B: . 2516678:2520990(4312) 11:42:07.701656 A > B: . 2520990:2525302(4312) 11:42:07.701739 B > A: . ack 2400254 11:42:07.702685 A > B: . 2525302:2529614(4312) 11:42:07.703257 A > B: . 2529614:2533926(4312) 11:42:07.703295 B > A: . ack 2408878 11:42:07.704414 A > B: . 2533926:2538238(4312) 11:42:07.704989 A > B: . 2538238:2542550(4312) 11:42:07.705040 B > A: . ack 2417502 11:42:07.705935 A > B: . 2542550:2546862(4312) 11:42:07.706506 A > B: . 2546862:2551174(4312) 11:42:07.706544 B > A: . ack 2426126 11:42:07.707480 A > B: . 2551174:2555486(4312) 11:42:07.708051 A > B: . 2555486:2559798(4312) 11:42:07.708088 B > A: . ack 2434750 11:42:07.709030 A > B: . 2559798:2564110(4312) 11:42:07.709604 A > B: . 2564110:2568422(4312) 11:42:07.710175 A > B: . 2568422:2572734(4312) * 11:42:07.710215 B > A: . ack 2443374 11:42:07.710799 A > B: . 2572734:2577046(4312) 11:42:07.711368 A > B: . 2577046:2581358(4312) 11:42:07.711405 B > A: . ack 2451998 11:42:07.712323 A > B: . 2581358:2585670(4312) 11:42:07.712898 A > B: . 2585670:2589982(4312) 11:42:07.712938 B > A: . ack 2460622 11:42:07.713926 A > B: . 2589982:2594294(4312) 11:42:07.714501 A > B: . 2594294:2598606(4312) 11:42:07.714547 B > A: . ack 2469246 11:42:07.715747 A > B: . 2598606:2602918(4312) 11:42:07.716287 A > B: . 2602918:2607230(4312) 11:42:07.716328 B > A: . ack 2477870 11:42:07.717146 A > B: . 2607230:2611542(4312) 11:42:07.717717 A > B: . 2611542:2615854(4312) 11:42:07.717762 B > A: . ack 2486494 11:42:07.718754 A > B: . 2615854:2620166(4312) 11:42:07.719331 A > B: . 2620166:2624478(4312) 11:42:07.719906 A > B: . 2624478:2628790(4312) ** 11:42:07.719958 B > A: . ack 2495118 11:42:07.720500 A > B: . 2628790:2633102(4312) 11:42:07.721080 A > B: . 2633102:2637414(4312) 11:42:07.721739 B > A: . ack 2503742 11:42:07.722348 A > B: . 2637414:2641726(4312)
11:42:07.722918 A > B: . 2641726:2646038(4312)
11:42:07.769248 B > A: . ack 2512366
The receiver's acknowledgment policy is one ACK per two packets
received. Thus, for each ACK arriving at host A, two new packets
are sent, except when cwnd increases due to congestion avoidance,
in which case three new packets are sent.
With an ack-every-two-packets policy, cwnd should only increase
one MSS per 2 RTT. However, at the point marked "*" the window
increases after 7 ACKs have arrived, and then again at "**" after
6 more ACKs.
While we do not have space to show the effect, this trace suffered
from repeated timeout retransmissions due to multiple packet
losses during a single RTT.
Trace file demonstrating correct behavior
Made using the same host and tracing setup as above, except now
A's TCP has been modified to remove the MSS/8 additive constant.
Tcpdump reported 77 packet drops; the excerpt below is fully
self-consistent so it is unlikely that any of these occurred
during the excerpt.
We again begin when cwnd is 31 packets (this occurs significantly
later in the trace, because the congestion avoidance is now less
aggressive with opening the window).
14:22:21.236757 B > A: . ack 5194679
14:22:21.238192 A > B: . 5319727:5324039(4312)
14:22:21.238770 A > B: . 5324039:5328351(4312)
14:22:21.238821 B > A: . ack 5203303
14:22:21.240158 A > B: . 5328351:5332663(4312)
14:22:21.240738 A > B: . 5332663:5336975(4312)
14:22:21.270422 B > A: . ack 5211927
14:22:21.271883 A > B: . 5336975:5341287(4312)
14:22:21.272458 A > B: . 5341287:5345599(4312)
14:22:21.279099 B > A: . ack 5220551
14:22:21.280539 A > B: . 5345599:5349911(4312)
14:22:21.281118 A > B: . 5349911:5354223(4312)
14:22:21.281183 B > A: . ack 5229175
14:22:21.282348 A > B: . 5354223:5358535(4312)
14:22:21.283029 A > B: . 5358535:5362847(4312)
14:22:21.283089 B > A: . ack 5237799
14:22:21.284213 A > B: . 5362847:5367159(4312)
14:22:21.284779 A > B: . 5367159:5371471(4312)
14:22:21.285976 B > A: . ack 5246423
14:22:21.287465 A > B: . 5371471:5375783(4312)
14:22:21.288036 A > B: . 5375783:5380095(4312) 14:22:21.288073 B > A: . ack 5255047 14:22:21.289155 A > B: . 5380095:5384407(4312) 14:22:21.289725 A > B: . 5384407:5388719(4312) 14:22:21.289762 B > A: . ack 5263671 14:22:21.291090 A > B: . 5388719:5393031(4312) 14:22:21.291662 A > B: . 5393031:5397343(4312) 14:22:21.291701 B > A: . ack 5272295 14:22:21.292870 A > B: . 5397343:5401655(4312) 14:22:21.293441 A > B: . 5401655:5405967(4312) 14:22:21.293481 B > A: . ack 5280919 14:22:21.294476 A > B: . 5405967:5410279(4312) 14:22:21.295053 A > B: . 5410279:5414591(4312) 14:22:21.295106 B > A: . ack 5289543 14:22:21.296306 A > B: . 5414591:5418903(4312) 14:22:21.296878 A > B: . 5418903:5423215(4312) 14:22:21.296917 B > A: . ack 5298167 14:22:21.297716 A > B: . 5423215:5427527(4312) 14:22:21.298285 A > B: . 5427527:5431839(4312) 14:22:21.298324 B > A: . ack 5306791 14:22:21.299413 A > B: . 5431839:5436151(4312) 14:22:21.299986 A > B: . 5436151:5440463(4312) 14:22:21.303696 B > A: . ack 5315415 14:22:21.305177 A > B: . 5440463:5444775(4312) 14:22:21.305755 A > B: . 5444775:5449087(4312) 14:22:21.308032 B > A: . ack 5324039 14:22:21.309525 A > B: . 5449087:5453399(4312) 14:22:21.310101 A > B: . 5453399:5457711(4312) 14:22:21.310144 B > A: . ack 5332663 *** 14:22:21.311615 A > B: . 5457711:5462023(4312) 14:22:21.312198 A > B: . 5462023:5466335(4312) 14:22:21.341876 B > A: . ack 5341287 14:22:21.343451 A > B: . 5466335:5470647(4312) 14:22:21.343985 A > B: . 5470647:5474959(4312) 14:22:21.350304 B > A: . ack 5349911 14:22:21.351852 A > B: . 5474959:5479271(4312) 14:22:21.352430 A > B: . 5479271:5483583(4312) 14:22:21.352484 B > A: . ack 5358535 14:22:21.353574 A > B: . 5483583:5487895(4312) 14:22:21.354149 A > B: . 5487895:5492207(4312) 14:22:21.354205 B > A: . ack 5367159 14:22:21.355467 A > B: . 5492207:5496519(4312) 14:22:21.356039 A > B: . 5496519:5500831(4312) 14:22:21.357361 B > A: . ack 5375783 14:22:21.358855 A > B: . 5500831:5505143(4312) 14:22:21.359424 A > B: . 5505143:5509455(4312) 14:22:21.359465 B > A: . ack 5384407
14:22:21.360605 A > B: . 5509455:5513767(4312)
14:22:21.361181 A > B: . 5513767:5518079(4312)
14:22:21.361225 B > A: . ack 5393031
14:22:21.362485 A > B: . 5518079:5522391(4312)
14:22:21.363057 A > B: . 5522391:5526703(4312)
14:22:21.363096 B > A: . ack 5401655
14:22:21.364236 A > B: . 5526703:5531015(4312)
14:22:21.364810 A > B: . 5531015:5535327(4312)
14:22:21.364867 B > A: . ack 5410279
14:22:21.365819 A > B: . 5535327:5539639(4312)
14:22:21.366386 A > B: . 5539639:5543951(4312)
14:22:21.366427 B > A: . ack 5418903
14:22:21.367586 A > B: . 5543951:5548263(4312)
14:22:21.368158 A > B: . 5548263:5552575(4312)
14:22:21.368199 B > A: . ack 5427527
14:22:21.369189 A > B: . 5552575:5556887(4312)
14:22:21.369758 A > B: . 5556887:5561199(4312)
14:22:21.369803 B > A: . ack 5436151
14:22:21.370814 A > B: . 5561199:5565511(4312)
14:22:21.371398 A > B: . 5565511:5569823(4312)
14:22:21.375159 B > A: . ack 5444775
14:22:21.376658 A > B: . 5569823:5574135(4312)
14:22:21.377235 A > B: . 5574135:5578447(4312)
14:22:21.379303 B > A: . ack 5453399
14:22:21.380802 A > B: . 5578447:5582759(4312)
14:22:21.381377 A > B: . 5582759:5587071(4312)
14:22:21.381947 A > B: . 5587071:5591383(4312) ****
"***" marks the end of the first round trip. Note that cwnd did
not increase (as evidenced by each ACK eliciting two new data
packets). Only at "****", which comes near the end of the second
round trip, does cwnd increase by one packet.
This trace did not suffer any timeout retransmissions. It
transferred the same amount of data as the first trace in about
half as much time. This difference is repeatable between hosts A
and B.
References
[Stevens94] and [Wright95] discuss this problem. The problem of
Reno TCP failing to recover from multiple losses except via a
retransmission timeout is discussed in [Fall96,Hoe96].
How to detect
If source code is available, that is generally the easiest way to
detect this problem. Search for each modification to the cwnd
variable; (at least) one of these will be for congestion
avoidance, and inspection of the related code should immediately
identify the problem if present.
The problem can also be detected by closely examining packet
traces taken near the sender. During congestion avoidance, cwnd
will increase by an additional segment upon the receipt of
(typically) eight acknowledgements without a loss. This increase
is in addition to the one segment increase per round trip time (or
two round trip times if the receiver is using delayed ACKs).
Furthermore, graphs of the sequence number vs. time, taken from
packet traces, are normally linear during congestion avoidance.
When viewing packet traces of transfers from senders exhibiting
this problem, the graphs appear quadratic instead of linear.
Finally, the traces will show that, with sufficiently large
windows, nearly every loss event results in a timeout.
How to fix
This problem may be corrected by removing the "+ MSS/8" term from
the congestion avoidance code that increases cwnd each time an ACK
of new data is received.
2.7.
Name of Problem
Initial RTO too low
Classification
Performance
Description
When a TCP first begins transmitting data, it lacks the RTT
measurements necessary to have computed an adaptive retransmission
timeout (RTO). RFC 1122, 4.2.3.1, states that a TCP SHOULD
initialize RTO to 3 seconds. A TCP that uses a lower value
exhibits "Initial RTO too low".
Significance
In environments with large RTTs (where "large" means any value
larger than the initial RTO), TCPs will experience very poor
performance.
Implications
Whenever RTO < RTT, very poor performance can result as packets
are unnecessarily retransmitted (because RTO will expire before an
ACK for the packet can arrive) and the connection enters slow
start and congestion avoidance. Generally, the algorithms for
computing RTO avoid this problem by adding a positive term to the
estimated RTT. However, when a connection first begins it must
use some estimate for RTO, and if it picks a value less than RTT,
the above problems will arise.
Furthermore, when the initial RTO < RTT, it can take a long time
for the TCP to correct the problem by adapting the RTT estimate,
because the use of Karn's algorithm (mandated by RFC 1122,
4.2.3.1) will discard many of the candidate RTT measurements made
after the first timeout, since they will be measurements of
retransmitted segments.
Relevant RFCs
RFC 1122 states that TCPs SHOULD initialize RTO to 3 seconds and
MUST implement Karn's algorithm.
Trace file demonstrating it
The following trace file was taken using tcpdump at host A, the
data sender. The advertised window and SYN options have been
omitted for clarity.
07:52:39.870301 A > B: S 2786333696:2786333696(0)
07:52:40.548170 B > A: S 130240000:130240000(0) ack 2786333697
07:52:40.561287 A > B: P 1:513(512) ack 1
07:52:40.753466 A > B: . 1:513(512) ack 1
07:52:41.133687 A > B: . 1:513(512) ack 1
07:52:41.458529 B > A: . ack 513
07:52:41.458686 A > B: . 513:1025(512) ack 1
07:52:41.458797 A > B: P 1025:1537(512) ack 1
07:52:41.541633 B > A: . ack 513
07:52:41.703732 A > B: . 513:1025(512) ack 1
07:52:42.044875 B > A: . ack 513
07:52:42.173728 A > B: . 513:1025(512) ack 1
07:52:42.330861 B > A: . ack 1537
07:52:42.331129 A > B: . 1537:2049(512) ack 1
07:52:42.331262 A > B: P 2049:2561(512) ack 1
07:52:42.623673 A > B: . 1537:2049(512) ack 1
07:52:42.683203 B > A: . ack 1537
07:52:43.044029 B > A: . ack 1537
07:52:43.193812 A > B: . 1537:2049(512) ack 1
Note from the SYN/SYN-ACK exchange, the RTT is over 600 msec.
However, from the elapsed time between the third and fourth lines
(the first packet being sent and then retransmitted), it is
apparent the RTO was initialized to under 200 msec. The next line
shows that this value has doubled to 400 msec (correct exponential
backoff of RTO), but that still does not suffice to avoid an
unnecessary retransmission.
Finally, an ACK from B arrives for the first segment. Later two
more duplicate ACKs for 513 arrive, indicating that both the
original and the two retransmissions arrived at B. (Indeed, a
concurrent trace at B showed that no packets were lost during the
entire connection). This ACK opens the congestion window to two
packets, which are sent back-to-back, but at 07:52:41.703732 RTO
again expires after a little over 200 msec, leading to an
unnecessary retransmission, and the pattern repeats. By the end
of the trace excerpt above, 1536 bytes have been successfully
transmitted from A to B, over an interval of more than 2 seconds,
reflecting terrible performance.
Trace file demonstrating correct behavior
The following trace file was taken using tcpdump at host C, the
data sender. The advertised window and SYN options have been
omitted for clarity.
17:30:32.090299 C > D: S 2031744000:2031744000(0)
17:30:32.900325 D > C: S 262737964:262737964(0) ack 2031744001
17:30:32.900326 C > D: . ack 1
17:30:32.910326 C > D: . 1:513(512) ack 1
17:30:34.150355 D > C: . ack 513
17:30:34.150356 C > D: . 513:1025(512) ack 1
17:30:34.150357 C > D: . 1025:1537(512) ack 1
17:30:35.170384 D > C: . ack 1025
17:30:35.170385 C > D: . 1537:2049(512) ack 1
17:30:35.170386 C > D: . 2049:2561(512) ack 1
17:30:35.320385 D > C: . ack 1537
17:30:35.320386 C > D: . 2561:3073(512) ack 1
17:30:35.320387 C > D: . 3073:3585(512) ack 1
17:30:35.730384 D > C: . ack 2049
The initial SYN/SYN-ACK exchange shows that RTT is more than 800
msec, and for some subsequent packets it rises above 1 second, but
C's retransmit timer does not ever expire.
References
This problem is documented in [Paxson97].
How to detect
This problem is readily detected by inspecting a packet trace of
the startup of a TCP connection made over a long-delay path. It
can be diagnosed from either a sender-side or receiver-side trace.
Long-delay paths can often be found by locating remote sites on
other continents.
How to fix
As this problem arises from a faulty initialization, one hopes
fixing it requires a one-line change to the TCP source code.
2.8.
Name of Problem
Failure of window deflation after loss recovery
Classification
Congestion control / performance
Description
The fast recovery algorithm allows TCP senders to continue to
transmit new segments during loss recovery. First, fast
retransmission is initiated after a TCP sender receives three
duplicate ACKs. At this point, a retransmission is sent and cwnd
is halved. The fast recovery algorithm then allows additional
segments to be sent when sufficient additional duplicate ACKs
arrive. Some implementations of fast recovery compute when to
send additional segments by artificially incrementing cwnd, first
by three segments to account for the three duplicate ACKs that
triggered fast retransmission, and subsequently by 1 MSS for each
new duplicate ACK that arrives. When cwnd allows, the sender
transmits new data segments.
When an ACK arrives that covers new data, cwnd is to be reduced by
the amount by which it was artificially increased. However, some
TCP implementations fail to "deflate" the window, causing an
inappropriate amount of data to be sent into the network after
recovery. One cause of this problem is the "header prediction"
code, which is used to handle incoming segments that require
little work. In some implementations of TCP, the header
prediction code does not check to make sure cwnd has not been
artificially inflated, and therefore does not reduce the
artificially increased cwnd when appropriate.
Significance
TCP senders that exhibit this problem will transmit a burst of
data immediately after recovery, which can degrade performance, as
well as network stability. Effectively, the sender does not
reduce the size of cwnd as much as it should (to half its value
when loss was detected), if at all. This can harm the performance
of the TCP connection itself, as well as competing TCP flows.
Implications
A TCP sender exhibiting this problem does not reduce cwnd
appropriately in times of congestion, and therefore may contribute
to congestive collapse.
Relevant RFCs
RFC 2001 outlines the fast retransmit/fast recovery algorithms.
[Brakmo95] outlines this implementation problem and offers a fix.
Trace file demonstrating it
The following trace file was taken using tcpdump at host A, the
data sender. The advertised window (which never changed) has been
omitted for clarity, except for the first packet sent by each
host.
08:22:56.825635 A.7505 > B.7505: . 29697:30209(512) ack 1 win 4608
08:22:57.038794 B.7505 > A.7505: . ack 27649 win 4096
08:22:57.039279 A.7505 > B.7505: . 30209:30721(512) ack 1
08:22:57.321876 B.7505 > A.7505: . ack 28161
08:22:57.322356 A.7505 > B.7505: . 30721:31233(512) ack 1
08:22:57.347128 B.7505 > A.7505: . ack 28673
08:22:57.347572 A.7505 > B.7505: . 31233:31745(512) ack 1
08:22:57.347782 A.7505 > B.7505: . 31745:32257(512) ack 1
08:22:57.936393 B.7505 > A.7505: . ack 29185
08:22:57.936864 A.7505 > B.7505: . 32257:32769(512) ack 1
08:22:57.950802 B.7505 > A.7505: . ack 29697 win 4096
08:22:57.951246 A.7505 > B.7505: . 32769:33281(512) ack 1
08:22:58.169422 B.7505 > A.7505: . ack 29697
08:22:58.638222 B.7505 > A.7505: . ack 29697
08:22:58.643312 B.7505 > A.7505: . ack 29697
08:22:58.643669 A.7505 > B.7505: . 29697:30209(512) ack 1
08:22:58.936436 B.7505 > A.7505: . ack 29697
08:22:59.002614 B.7505 > A.7505: . ack 29697
08:22:59.003026 A.7505 > B.7505: . 33281:33793(512) ack 1
08:22:59.682902 B.7505 > A.7505: . ack 33281
08:22:59.683391 A.7505 > B.7505: P 33793:34305(512) ack 1
08:22:59.683748 A.7505 > B.7505: P 34305:34817(512) ack 1 ***
08:22:59.684043 A.7505 > B.7505: P 34817:35329(512) ack 1
08:22:59.684266 A.7505 > B.7505: P 35329:35841(512) ack 1
08:22:59.684567 A.7505 > B.7505: P 35841:36353(512) ack 1
08:22:59.684810 A.7505 > B.7505: P 36353:36865(512) ack 1
08:22:59.685094 A.7505 > B.7505: P 36865:37377(512) ack 1
The first 12 lines of the trace show incoming ACKs clocking out a
window of data segments. At this point in the transfer, cwnd is 7
segments. The next 4 lines of the trace show 3 duplicate ACKs
arriving from the receiver, followed by a retransmission from the
sender. At this point, cwnd is halved (to 3 segments) and
artificially incremented by the three duplicate ACKs that have
arrived, making cwnd 6 segments. The next two lines show 2 more
duplicate ACKs arriving, each of which increases cwnd by 1
segment. So, after these two duplicate ACKs arrive the cwnd is 8
segments and the sender has permission to send 1 new segment
(since there are 7 segments outstanding). The next line in the
trace shows this new segment being transmitted. The next packet
shown in the trace is an ACK from host B that covers the first 7
outstanding segments (all but the new segment sent during
recovery). This should cause cwnd to be reduced to 3 segments and
2 segments to be transmitted (since there is already 1 outstanding
segment in the network). However, as shown by the last 7 lines of
the trace, cwnd is not reduced, causing a line-rate burst of 7 new
segments.
Trace file demonstrating correct behavior
The trace would appear identical to the one above, only it would
stop after the line marked "***", because at this point host A
would correctly reduce cwnd after recovery, allowing only 2
segments to be transmitted, rather than producing a burst of 7
segments.
References
This problem is documented and the performance implications
analyzed in [Brakmo95].
How to detect
Failure of window deflation after loss recovery can be found by
examining sender-side packet traces recorded during periods of
moderate loss (so cwnd can grow large enough to allow for fast
recovery when loss occurs).
How to fix
When this bug is caused by incorrect header prediction, the fix is
to add a predicate to the header prediction test that checks to
see whether cwnd is inflated; if so, the header prediction test
fails and the usual ACK processing occurs, which (in this case)
takes care to deflate the window. See [Brakmo95] for details.
2.9.
Name of Problem
Excessively short keepalive connection timeout
Classification
Reliability
Description
Keep-alive is a mechanism for checking whether an idle connection
is still alive. According to RFC 1122, keepalive should only be
invoked in server applications that might otherwise hang
indefinitely and consume resources unnecessarily if a client
crashes or aborts a connection during a network failure.
RFC 1122 also specifies that if a keep-alive mechanism is
implemented it MUST NOT interpret failure to respond to any
specific probe as a dead connection. The RFC does not specify a
particular mechanism for timing out a connection when no response
is received for keepalive probes. However, if the mechanism does
not allow ample time for recovery from network congestion or
delay, connections may be timed out unnecessarily.
Significance
In congested networks, can lead to unwarranted termination of
connections.
Implications
It is possible for the network connection between two peer
machines to become congested or to exhibit packet loss at the time
that a keep-alive probe is sent on a connection. If the keep-
alive mechanism does not allow sufficient time before dropping
connections in the face of unacknowledged probes, connections may
be dropped even when both peers of a connection are still alive.
Relevant RFCs
RFC 1122 specifies that the keep-alive mechanism may be provided.
It does not specify a mechanism for determining dead connections
when keepalive probes are not acknowledged.
Trace file demonstrating it
Made using the Orchestra tool at the peer of the machine using
keep-alive. After connection establishment, incoming keep-alives
were dropped by Orchestra to simulate a dead connection.
22:11:12.040000 A > B: 22666019:0 win 8192 datasz 4 SYN
22:11:12.060000 B > A: 2496001:22666020 win 4096 datasz 4 SYN ACK
22:11:12.130000 A > B: 22666020:2496002 win 8760 datasz 0 ACK
(more than two hours elapse)
00:23:00.680000 A > B: 22666019:2496002 win 8760 datasz 1 ACK
00:23:01.770000 A > B: 22666019:2496002 win 8760 datasz 1 ACK
00:23:02.870000 A > B: 22666019:2496002 win 8760 datasz 1 ACK
00:23.03.970000 A > B: 22666019:2496002 win 8760 datasz 1 ACK
00:23.05.070000 A > B: 22666019:2496002 win 8760 datasz 1 ACK
The initial three packets are the SYN exchange for connection
setup. About two hours later, the keepalive timer fires because
the connection has been idle. Keepalive probes are transmitted a
total of 5 times, with a 1 second spacing between probes, after
which the connection is dropped. This is problematic because a 5
second network outage at the time of the first probe results in
the connection being killed.
Trace file demonstrating correct behavior
Made using the Orchestra tool at the peer of the machine using
keep-alive. After connection establishment, incoming keep-alives
were dropped by Orchestra to simulate a dead connection.
16:01:52.130000 A > B: 1804412929:0 win 4096 datasz 4 SYN
16:01:52.360000 B > A: 16512001:1804412930 win 4096 datasz 4 SYN ACK
16:01:52.410000 A > B: 1804412930:16512002 win 4096 datasz 0 ACK
(two hours elapse)
18:01:57.170000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:03:12.220000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:04:27.270000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:05:42.320000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:06:57.370000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:08:12.420000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:09:27.480000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:10:43.290000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:11:57.580000 A > B: 1804412929:16512002 win 4096 datasz 0 ACK
18:13:12.630000 A > B: 1804412929:16512002 win 4096 datasz 0 RST ACK
In this trace, when the keep-alive timer expires, 9 keepalive
probes are sent at 75 second intervals. 75 seconds after the last
probe is sent, a final RST segment is sent indicating that the
connection has been closed. This implementation waits about 11
minutes before timing out the connection, while the first
implementation shown allows only 5 seconds.
References
This problem is documented in [Dawson97].
How to detect
For implementations manifesting this problem, it shows up on a
packet trace after the keepalive timer fires if the peer machine
receiving the keepalive does not respond. Usually the keepalive
timer will fire at least two hours after keepalive is turned on,
but it may be sooner if the timer value has been configured lower,
or if the keepalive mechanism violates the specification (see
Insufficient interval between keepalives problem). In this
example, suppressing the response of the peer to keepalive probes
was accomplished using the Orchestra toolkit, which can be
configured to drop packets. It could also have been done by
creating a connection, turning on keepalive, and disconnecting the
network connection at the receiver machine.
How to fix
This problem can be fixed by using a different method for timing
out keepalives that allows a longer period of time to elapse
before dropping the connection. For example, the algorithm for
timing out on dropped data could be used. Another possibility is
an algorithm such as the one shown in the trace above, which sends
9 probes at 75 second intervals and then waits an additional 75
seconds for a response before closing the connection.
2.10.
Name of Problem
Failure to back off retransmission timeout
Classification
Congestion control / reliability
Description
The retransmission timeout is used to determine when a packet has
been dropped in the network. When this timeout has expired
without the arrival of an ACK, the segment is retransmitted. Each
time a segment is retransmitted, the timeout is adjusted according
to an exponential backoff algorithm, doubling each time. If a TCP
fails to receive an ACK after numerous attempts at retransmitting
the same segment, it terminates the connection. A TCP that fails
to double its retransmission timeout upon repeated timeouts is
said to exhibit "Failure to back off retransmission timeout".
Significance
Backing off the retransmission timer is a cornerstone of network
stability in the presence of congestion. Consequently, this bug
can have severe adverse affects in congested networks. It also
affects TCP reliability in congested networks, as discussed in the
next section.
Implications
It is possible for the network connection between two TCP peers to
become congested or to exhibit packet loss at the time that a
retransmission is sent on a connection. If the retransmission
mechanism does not allow sufficient time before dropping
connections in the face of unacknowledged segments, connections
may be dropped even when, by waiting longer, the connection could
have continued.
Relevant RFCs
RFC 1122 specifies mandatory exponential backoff of the
retransmission timeout, and the termination of connections after
some period of time (at least 100 seconds).
Trace file demonstrating it
Made using tcpdump on an intermediate host:
16:51:12.671727 A > B: S 510878852:510878852(0) win 16384
16:51:12.672479 B > A: S 2392143687:2392143687(0)
ack 510878853 win 16384
16:51:12.672581 A > B: . ack 1 win 16384
16:51:15.244171 A > B: P 1:3(2) ack 1 win 16384
16:51:15.244933 B > A: . ack 3 win 17518 (DF)
<receiving host disconnected>
16:51:19.381176 A > B: P 3:5(2) ack 1 win 16384
16:51:20.162016 A > B: P 3:5(2) ack 1 win 16384
16:51:21.161936 A > B: P 3:5(2) ack 1 win 16384
16:51:22.161914 A > B: P 3:5(2) ack 1 win 16384
16:51:23.161914 A > B: P 3:5(2) ack 1 win 16384
16:51:24.161879 A > B: P 3:5(2) ack 1 win 16384
16:51:25.161857 A > B: P 3:5(2) ack 1 win 16384
16:51:26.161836 A > B: P 3:5(2) ack 1 win 16384
16:51:27.161814 A > B: P 3:5(2) ack 1 win 16384
16:51:28.161791 A > B: P 3:5(2) ack 1 win 16384
16:51:29.161769 A > B: P 3:5(2) ack 1 win 16384
16:51:30.161750 A > B: P 3:5(2) ack 1 win 16384
16:51:31.161727 A > B: P 3:5(2) ack 1 win 16384
16:51:32.161701 A > B: R 5:5(0) ack 1 win 16384
The initial three packets are the SYN exchange for connection
setup, then a single data packet, to verify that data can be
transferred. Then the connection to the destination host was
disconnected, and more data sent. Retransmissions occur every
second for 12 seconds, and then the connection is terminated with
a RST. This is problematic because a 12 second pause in
connectivity could result in the termination of a connection.
Trace file demonstrating correct behavior
Again, a tcpdump taken from a third host:
16:59:05.398301 A > B: S 2503324757:2503324757(0) win 16384
16:59:05.399673 B > A: S 2492674648:2492674648(0)
ack 2503324758 win 16384
16:59:05.399866 A > B: . ack 1 win 17520
16:59:06.538107 A > B: P 1:3(2) ack 1 win 17520
16:59:06.540977 B > A: . ack 3 win 17518 (DF)
<receiving host disconnected>
16:59:13.121542 A > B: P 3:5(2) ack 1 win 17520
16:59:14.010928 A > B: P 3:5(2) ack 1 win 17520
16:59:16.010979 A > B: P 3:5(2) ack 1 win 17520
16:59:20.011229 A > B: P 3:5(2) ack 1 win 17520
16:59:28.011896 A > B: P 3:5(2) ack 1 win 17520
16:59:44.013200 A > B: P 3:5(2) ack 1 win 17520
17:00:16.015766 A > B: P 3:5(2) ack 1 win 17520
17:01:20.021308 A > B: P 3:5(2) ack 1 win 17520
17:02:24.027752 A > B: P 3:5(2) ack 1 win 17520
17:03:28.034569 A > B: P 3:5(2) ack 1 win 17520
17:04:32.041567 A > B: P 3:5(2) ack 1 win 17520
17:05:36.048264 A > B: P 3:5(2) ack 1 win 17520
17:06:40.054900 A > B: P 3:5(2) ack 1 win 17520
17:07:44.061306 A > B: R 5:5(0) ack 1 win 17520
In this trace, when the retransmission timer expires, 12
retransmissions are sent at exponentially-increasing intervals,
until the interval value reaches 64 seconds, at which time the
interval stops growing. 64 seconds after the last retransmission,
a final RST segment is sent indicating that the connection has
been closed. This implementation waits about 9 minutes before
timing out the connection, while the first implementation shown
allows only 12 seconds.
References
None known.
How to detect
A simple transfer can be easily interrupted by disconnecting the
receiving host from the network. tcpdump or another appropriate
tool should show the retransmissions being sent. Several trials
in a low-rtt environment may be required to demonstrate the bug.
How to fix
For one of the implementations studied, this problem seemed to be
the result of an error introduced with the addition of the
Brakmo-Peterson RTO algorithm [Brakmo95], which can return a value
of zero where the older Jacobson algorithm always returns a
positive value. Brakmo and Peterson specified an additional step
of min(rtt + 2, RTO) to avoid problems with this. Unfortunately,
in the implementation this step was omitted when calculating the
exponential backoff for the RTO. This results in an RTO of 0
seconds being multiplied by the backoff, yielding again zero, and
then being subjected to a later MAX operation that increases it to
1 second, regardless of the backoff factor.
A similar TCP persist failure has the same cause.
2.11.
Name of Problem
Insufficient interval between keepalives
Classification
Reliability
Description
Keep-alive is a mechanism for checking whether an idle connection
is still alive. According to RFC 1122, keep-alive may be included
in an implementation. If it is included, the interval between
keep-alive packets MUST be configurable, and MUST default to no
less than two hours.
Significance
In congested networks, can lead to unwarranted termination of
connections.
Implications
According to RFC 1122, keep-alive is not required of
implementations because it could: (1) cause perfectly good
connections to break during transient Internet failures; (2)
consume unnecessary bandwidth ("if no one is using the connection,
who cares if it is still good?"); and (3) cost money for an
Internet path that charges for packets. Regarding this last
point, we note that in addition the presence of dial-on-demand
links in the route can greatly magnify the cost penalty of excess
keepalives, potentially forcing a full-time connection on a link
that would otherwise only be connected a few minutes a day.
If keepalive is provided the RFC states that the required inter-
keepalive distance MUST default to no less than two hours. If it
does not, the probability of connections breaking increases, the
bandwidth used due to keepalives increases, and cost increases
over paths which charge per packet.
Relevant RFCs
RFC 1122 specifies that the keep-alive mechanism may be provided.
It also specifies the two hour minimum for the default interval
between keepalive probes.
Trace file demonstrating it
Made using the Orchestra tool at the peer of the machine using
keep-alive. Machine A was configured to use default settings for
the keepalive timer.
11:36:32.910000 A > B: 3288354305:0 win 28672 datasz 4 SYN
11:36:32.930000 B > A: 896001:3288354306 win 4096 datasz 4 SYN ACK
11:36:32.950000 A > B: 3288354306:896002 win 28672 datasz 0 ACK
11:50:01.190000 A > B: 3288354305:896002 win 28672 datasz 0 ACK
11:50:01.210000 B > A: 896002:3288354306 win 4096 datasz 0 ACK
12:03:29.410000 A > B: 3288354305:896002 win 28672 datasz 0 ACK
12:03:29.430000 B > A: 896002:3288354306 win 4096 datasz 0 ACK
12:16:57.630000 A > B: 3288354305:896002 win 28672 datasz 0 ACK
12:16:57.650000 B > A: 896002:3288354306 win 4096 datasz 0 ACK
12:30:25.850000 A > B: 3288354305:896002 win 28672 datasz 0 ACK
12:30:25.870000 B > A: 896002:3288354306 win 4096 datasz 0 ACK
12:43:54.070000 A > B: 3288354305:896002 win 28672 datasz 0 ACK
12:43:54.090000 B > A: 896002:3288354306 win 4096 datasz 0 ACK
The initial three packets are the SYN exchange for connection
setup. About 13 minutes later, the keepalive timer fires because
the connection is idle. The keepalive is acknowledged, and the
timer fires again in about 13 more minutes. This behavior
continues indefinitely until the connection is closed, and is a
violation of the specification.
Trace file demonstrating correct behavior
Made using the Orchestra tool at the peer of the machine using
keep-alive. Machine A was configured to use default settings for
the keepalive timer.
17:37:20.500000 A > B: 34155521:0 win 4096 datasz 4 SYN
17:37:20.520000 B > A: 6272001:34155522 win 4096 datasz 4 SYN ACK
17:37:20.540000 A > B: 34155522:6272002 win 4096 datasz 0 ACK
19:37:25.430000 A > B: 34155521:6272002 win 4096 datasz 0 ACK
19:37:25.450000 B > A: 6272002:34155522 win 4096 datasz 0 ACK
21:37:30.560000 A > B: 34155521:6272002 win 4096 datasz 0 ACK
21:37:30.570000 B > A: 6272002:34155522 win 4096 datasz 0 ACK
23:37:35.580000 A > B: 34155521:6272002 win 4096 datasz 0 ACK
23:37:35.600000 B > A: 6272002:34155522 win 4096 datasz 0 ACK
01:37:40.620000 A > B: 34155521:6272002 win 4096 datasz 0 ACK
01:37:40.640000 B > A: 6272002:34155522 win 4096 datasz 0 ACK
03:37:45.590000 A > B: 34155521:6272002 win 4096 datasz 0 ACK
03:37:45.610000 B > A: 6272002:34155522 win 4096 datasz 0 ACK
The initial three packets are the SYN exchange for connection
setup. Just over two hours later, the keepalive timer fires
because the connection is idle. The keepalive is acknowledged,
and the timer fires again just over two hours later. This
behavior continues indefinitely until the connection is closed.
References
This problem is documented in [Dawson97].
How to detect
For implementations manifesting this problem, it shows up on a
packet trace. If the connection is left idle, the keepalive
probes will arrive closer together than the two hour minimum.
2.12.
Name of Problem
Window probe deadlock
Classification
Reliability
Description
When an application reads a single byte from a full window, the
window should not be updated, in order to avoid Silly Window
Syndrome (SWS; see [RFC813]). If the remote peer uses a single
byte of data to probe the window, that byte can be accepted into
the buffer. In some implementations, at this point a negative
argument to a signed comparison causes all further new data to be
considered outside the window; consequently, it is discarded
(after sending an ACK to resynchronize). These discards include
the ACKs for the data packets sent by the local TCP, so the TCP
will consider the data unacknowledged.
Consequently, the application may be unable to complete sending
new data to the remote peer, because it has exhausted the transmit
buffer available to its local TCP, and buffer space is never being
freed because incoming ACKs that would do so are being discarded.
If the application does not read any more data, which may happen
due to its failure to complete such sends, then deadlock results.
Significance
It's relatively rare for applications to use TCP in a manner that
can exercise this problem. Most applications only transmit bulk
data if they know the other end is prepared to receive the data.
However, if a client fails to consume data, putting the server in
persist mode, and then consumes a small amount of data, it can
mistakenly compute a negative window. At this point the client
will discard all further packets from the server, including ACKs
of the client's own data, since they are not inside the
(impossibly-sized) window. If subsequently the client consumes
enough data to then send a window update to the server, the
situation will be rectified. That is, this situation can only
happen if the client consumes 1 < N < MSS bytes, so as not to
cause a window update, and then starts its own transmission
towards the server of more than a window's worth of data.
Implications
TCP connections will hang and eventually time out.
Relevant RFCs
RFC 793 describes zero window probing. RFC 813 describes Silly
Window Syndrome.
Trace file demonstrating it
Trace made from a version of tcpdump modified to print out the
sequence number attached to an ACK even if it's dataless. An
unmodified tcpdump would not print seq:seq(0); however, for this
bug, the sequence number in the ACK is important for unambiguously
determining how the TCP is behaving.
[ Normal connection startup and data transmission from B to A.
Options, including MSS of 16344 in both directions, omitted
for clarity. ]
16:07:32.327616 A > B: S 65360807:65360807(0) win 8192
16:07:32.327304 B > A: S 65488807:65488807(0) ack 65360808 win 57344
16:07:32.327425 A > B: . 1:1(0) ack 1 win 57344
16:07:32.345732 B > A: P 1:2049(2048) ack 1 win 57344
16:07:32.347013 B > A: P 2049:16385(14336) ack 1 win 57344
16:07:32.347550 B > A: P 16385:30721(14336) ack 1 win 57344
16:07:32.348683 B > A: P 30721:45057(14336) ack 1 win 57344
16:07:32.467286 A > B: . 1:1(0) ack 45057 win 12288
16:07:32.467854 B > A: P 45057:57345(12288) ack 1 win 57344
[ B fills up A's offered window ]
16:07:32.667276 A > B: . 1:1(0) ack 57345 win 0
[ B probes A's window with a single byte ]
16:07:37.467438 B > A: . 57345:57346(1) ack 1 win 57344
[ A resynchronizes without accepting the byte ]
16:07:37.467678 A > B: . 1:1(0) ack 57345 win 0
[ B probes A's window again ]
16:07:45.467438 B > A: . 57345:57346(1) ack 1 win 57344
[ A resynchronizes and accepts the byte (per the ack field) ]
16:07:45.667250 A > B: . 1:1(0) ack 57346 win 0
[ The application on A has started generating data. The first
packet A sends is small due to a memory allocation bug. ]
16:07:51.358459 A > B: P 1:2049(2048) ack 57346 win 0
[ B acks A's first packet ]
16:07:51.467239 B > A: . 57346:57346(0) ack 2049 win 57344
[ This looks as though A accepted B's ACK and is sending
another packet in response to it. In fact, A is trying
to resynchronize with B, and happens to have data to send
and can send it because the first small packet didn't use
up cwnd. ]
16:07:51.467698 A > B: . 2049:14337(12288) ack 57346 win 0
[ B acks all of the data that A has sent ]
16:07:51.667283 B > A: . 57346:57346(0) ack 14337 win 57344
[ A tries to resynchronize. Notice that by the packets
seen on the network, A and B *are* in fact synchronized;
A only thinks that they aren't. ]
16:07:51.667477 A > B: . 14337:14337(0) ack 57346 win 0
[ A's retransmit timer fires, and B acks all of the data.
A once again tries to resynchronize. ]
16:07:52.467682 A > B: . 1:14337(14336) ack 57346 win 0
16:07:52.468166 B > A: . 57346:57346(0) ack 14337 win 57344
16:07:52.468248 A > B: . 14337:14337(0) ack 57346 win 0
[ A's retransmit timer fires again, and B acks all of the data.
A once again tries to resynchronize. ]
16:07:55.467684 A > B: . 1:14337(14336) ack 57346 win 0
16:07:55.468172 B > A: . 57346:57346(0) ack 14337 win 57344
16:07:55.468254 A > B: . 14337:14337(0) ack 57346 win 0
Trace file demonstrating correct behavior
Made between the same two hosts after applying the bug fix
mentioned below (and using the same modified tcpdump).
[ Connection starts up with data transmission from B to A.
Note that due to a separate bug (the fact that A and B
are communicating over a loopback driver), B erroneously
skips slow start. ]
17:38:09.510854 A > B: S 3110066585:3110066585(0) win 16384
17:38:09.510926 B > A: S 3110174850:3110174850(0)
ack 3110066586 win 57344
17:38:09.510953 A > B: . 1:1(0) ack 1 win 57344
17:38:09.512956 B > A: P 1:2049(2048) ack 1 win 57344
17:38:09.513222 B > A: P 2049:16385(14336) ack 1 win 57344
17:38:09.513428 B > A: P 16385:30721(14336) ack 1 win 57344
17:38:09.513638 B > A: P 30721:45057(14336) ack 1 win 57344
17:38:09.519531 A > B: . 1:1(0) ack 45057 win 12288
17:38:09.519638 B > A: P 45057:57345(12288) ack 1 win 57344
[ B fills up A's offered window ]
17:38:09.719526 A > B: . 1:1(0) ack 57345 win 0
[ B probes A's window with a single byte. A resynchronizes
without accepting the byte ]
17:38:14.499661 B > A: . 57345:57346(1) ack 1 win 57344
17:38:14.499724 A > B: . 1:1(0) ack 57345 win 0
[ B probes A's window again. A resynchronizes and accepts
the byte, as indicated by the ack field ]
17:38:19.499764 B > A: . 57345:57346(1) ack 1 win 57344
17:38:19.519731 A > B: . 1:1(0) ack 57346 win 0
[ B probes A's window with a single byte. A resynchronizes
without accepting the byte ]
17:38:24.499865 B > A: . 57346:57347(1) ack 1 win 57344
17:38:24.499934 A > B: . 1:1(0) ack 57346 win 0
[ The application on A has started generating data.
B acks A's data and A accepts the ACKs and the
data transfer continues ]
17:38:28.530265 A > B: P 1:2049(2048) ack 57346 win 0
17:38:28.719914 B > A: . 57346:57346(0) ack 2049 win 57344
17:38:28.720023 A > B: . 2049:16385(14336) ack 57346 win 0
17:38:28.720089 A > B: . 16385:30721(14336) ack 57346 win 0
17:38:28.720370 B > A: . 57346:57346(0) ack 30721 win 57344
17:38:28.720462 A > B: . 30721:45057(14336) ack 57346 win 0
17:38:28.720526 A > B: P 45057:59393(14336) ack 57346 win 0
17:38:28.720824 A > B: P 59393:73729(14336) ack 57346 win 0
17:38:28.721124 B > A: . 57346:57346(0) ack 73729 win 47104
17:38:28.721198 A > B: P 73729:88065(14336) ack 57346 win 0
17:38:28.721379 A > B: P 88065:102401(14336) ack 57346 win 0
17:38:28.721557 A > B: P 102401:116737(14336) ack 57346 win 0
17:38:28.721863 B > A: . 57346:57346(0) ack 116737 win 36864
References
None known.
How to detect
Initiate a connection from a client to a server. Have the server
continuously send data until its buffers have been full for long
enough to exhaust the window. Next, have the client read 1 byte
and then delay for long enough that the server TCP sends a window
probe. Now have the client start sending data. At this point, if
it ignores the server's ACKs, then the client's TCP suffers from
the problem.
How to fix
In one implementation known to exhibit the problem (derived from
4.3-Reno), the problem was introduced when the macro MAX() was
replaced by the function call max() for computing the amount of
space in the receive window:
tp->rcv_wnd = max(win, (int)(tp->rcv_adv - tp->rcv_nxt));
When data has been received into a window beyond what has been
advertised to the other side, rcv_nxt > rcv_adv, making this
negative. It's clear from the (int) cast that this is intended,
but the unsigned max() function sign-extends so the negative
number is "larger". The fix is to change max() to imax():
tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt));
4.3-Tahoe and before did not have this bug, since it used the
macro MAX() for this calculation.
(page 40 continued on part 3)