Design and Performance of the
OpenBSD Stateful Packet Filter (pf)
Daniel Hartmeier
[e-mail]
Systor AG
Usenix 2002  p.1/22
Introduction
part of a firewall, working on IP packet level (vs.
application level proxies or ethernet level bridges)
packet filter intercepting each IP packet that passes
through the kernel (in and out on each interface),
passing or blocking it
stateless inspection based on the fields of each
packet
stateful filtering keeping track of connections,
additional information makes filtering more
powerful (sequence number checks) and easier
(replies, random client ports)
Usenix 2002  p.2/22
Motivation
OpenBSD included IPFilter in the default install
what appeared to be a BSD license turned out to be
non-free
unlike other license problems discovered by the
ongoing license audit, this case couldn t be resolved,
IPFilter removed from the tree
existing alternatives were considered (ipfw), larger
code base, kernel dependencies
rewrite offers additional options, integrates better
with existing kernel features
Usenix 2002  p.3/22
Overview
Introduction
Motivation
Filter rules, skip steps
State table, trees, lookups, translations (NAT,
redirections)
Benchmarks
Conclusions
Usenix 2002  p.4/22
Filter rules
linear linked list, evaluated top to bottom for each
packet (unlike netfilter s chains tree)
rules contain parameters that match/mismatch a
packet
rules pass or block a packet
last matching rule wins (except for  quick , which
aborts rule evaluation)
rules can create state, further state matching packets
are passed without rule set evaluation
Usenix 2002  p.5/22
Skip steps
transparent optimization of rule set evaluation,
improve performance without affecting semantics
example: ten consecutive rules apply only to packets
from source address X, packet has source address Y,
first rule evaluated, next nine skipped
skipping is done on most parameters, in pre-defined
order
parameters like direction (in, out), interface or
address family (IPv4/IPv6) partition the rule set a
lot, performance increase is significant
worst case: consecutive rules have no equal
parameters, every rule must be evaluated, no
additional cost (linked list traversal)
Usenix 2002  p.6/22
State table
TCP (sequence number checks on each packet),
ICMP error messages match referred to packet
(simplifies rules without breaking PMTU etc.)
UDP, ICMP queries/replies, other protocols,
pseudo-connections with timeouts
adjustable timeouts, pseudo-connections for
non-TCP protocols
binary search tree (AVL, now Red-Black), O(log n)
even in worst-case
key is two address/port pairs
Usenix 2002  p.7/22
Translations (NAT, redirections)
translating source addresses: NAT/PAT to one
address using proxy ports
translating destination: redirections (based on
addresses/ports)
mapping stored in state table
application level proxies (ftp) in userland
Usenix 2002  p.8/22
State table keys
one state entry per connection, stored in two trees
example: 10.1.1.1:20000 -> 62.65.145.30:50001 ->
129.128.5.191:80
outgoing packets: 10.1.1.1:20000 ->
129.128.5.191:80, replace source address/port with
gateway
incoming packets: 129.128.5.191:80 ->
62.65.145.30:50001, replace destination
address/port with local host
three address/port pairs of one connection: lan, gwy,
ext
without translation, two pairs are equal
Usenix 2002  p.9/22
State table keys
two trees: tree-lan-ext (outgoing) and tree-ext-gwy
(incoming), contain the same state pointers
no addition translation map (and lookup) needed
Usenix 2002  p.10/22
Normalization
IP normalization (scrubbing) to remove
interpretation ambiguities, like overlapping
fragments (confusing IDSs)
reassembly (caching) of fragments before filtering,
only complete packets are filtered
sequence number modulation
Usenix 2002  p.11/22
Logging
through bpf, virtual network interface pflog0
link layer header used for pf related information
(rule, action)
binary log files, readable with tcpdump and other
tools
Usenix 2002  p.12/22
Benchmarks: Setup
two (old) i386 machines with two network interface
cards each, connected with two crosswire Cat5
cables, 10 mbit/s unidirectional
tester: generate TCP packets on ethernet level
through first NIC, capture incoming ethernet frames
on second NIC
firewall: OpenBSD and GNU/Linux (equal
hardware), IP forwarding enabled, packet filter
enabled, no other services, no other network traffic
(static arp table)
Usenix 2002  p.13/22
Benchmarks: Packet generation
TCP packets of variable size, random
source/destination addresses and ports
embedded timestamp to calculate latency,
incremental serial number to detect packet loss
send packets of specified size at specified rate for
several seconds, print throughput, latency and loss
verify that setup can handle maximum link rate
correctly
Usenix 2002  p.14/22
Local, reaching link limit
900
1518 bytes/packet
800
700
600
500
400
300
200
100
0
0 100 200 300 400 500 600 700 800 900
sending rate (packets/s)
Usenix 2002  p.15/22
receiving rate (packets/s)
Local, reaching link limit
900
1518 bytes/packet
812
800
700
600
500
400
300
200
100
812
0
0 100 200 300 400 500 600 700 800 900
sending rate (packets/s)
Usenix 2002  p.15/22
receiving rate (packets/s)
Local, varying packet sizes
1.4e+06
1518 bytes
1.2e+06
1e+06
800000
600000
400000
200000
812
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
1280 bytes
1.2e+06
1e+06
800000
600000
400000
200000
961
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
1024 bytes
1.2e+06
1e+06
800000
600000
400000
200000
1197
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
768 bytes
1.2e+06
1e+06
800000
600000
400000
200000
1586
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
512 bytes
1.2e+06
1e+06
800000
600000
400000
200000
2349
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
256 bytes
1.2e+06
1e+06
800000
600000
400000
200000
4528
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
1.2e+06
128 bytes
1e+06
800000
600000
400000
200000
8445
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
1.2e+06
1e+06 64 bytes
800000
600000
400000
200000
14880
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Local, varying packet sizes
1.4e+06
Local
OpenBSD
GNU/Linux
1.2e+06
1e+06
800000
600000
400000
200000
0
0 2000 4000 6000 8000 10000 12000 14000 16000
sending rate (packets/s)
Usenix 2002  p.16/22
throughput (bytes/s)
Stateless, 100 rules, throughput
5000
iptables
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0 1000 2000 3000 4000 5000
sending rate (packets/s)
Usenix 2002  p.17/22
throughput (packets/s)
Stateless, 100 rules, throughput
5000
iptables
ipf
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0 1000 2000 3000 4000 5000
sending rate (packets/s)
Usenix 2002  p.17/22
throughput (packets/s)
Stateless, 100 rules, throughput
5000
iptables
ipf
4500
pf
4000
3500
3000
2500
2000
1500
1000
500
0
0 1000 2000 3000 4000 5000
sending rate (packets/s)
Usenix 2002  p.17/22
throughput (packets/s)
Maximum throughput vs. rules
5000
iptables
ipf
4500
pf
4000
3500
3000
2500
2000
1500
1000
500
0 200 400 600 800 1000
number of rules
Usenix 2002  p.18/22
maximum throughput (packets/s)
Maximum throughput vs. states
7500
ipf
pf
7000
6500
6000
5500
5000
4500
4000
3500
3000
0 5000 10000 15000 20000
number of states
Usenix 2002  p.19/22
maximum throughput (packets/s)
Conclusions
rule set evaluation is expensive. State lookups are
cheap
filtering statefully not only improves filter decision
quality, it actually increases performance
memory cost: 64000 states with 64MB RAM
(without tuning), increasing linearly
binary search tree for states scales with O(log n)
Usenix 2002  p.20/22
Production results
Duron 700MHz, 128MB RAM, 3x DEC 21143
NICs
25000-40000 concurrent states
average of 5000 packets/s
fully stateful filtering (no stateless passing)
CPU load doesn t exceed 10 percent
(same box and filter policy with IPFilter was 90
percent load average)
Usenix 2002  p.21/22
Questions?
The OpenBSD Project: http://www.openbsd.org/
Paper and slides: http://www.benzedrine.cx/pf.html
[e-mail]
Usenix 2002  p.22/22