What, then is PF? Let us start by looking briefly at the project's history to put things in their proper context.
OpenBSD's Packet Filter subsystem, which most people refer to simply by using the abbreviated form 'PF', was originally written in an effort of extremely rapid development during the northern hemisphere summer and autumn months of 2001 by Daniel Hartmeier and a number of OpenBSD developers, and was launched as a default part of the OpenBSD 3.0 base system in December of 2001.
The need for a new firewalling software subsystem for OpenBSD arose when Darren Reed announced to the world that IPFilter, which at that point had been rather intimately integrated in OpenBSD, was not after all BSD licensed. In fact quite to the contrary. The license itself was almost a word by word copy of the BSD license, omitting only the right to make changes to the code and distribute the result. The OpenBSD version of IPFilter contained quite a number of changes and customizations, which it turned out were not allowed according to the license. IPFilter was removed from the OpenBSD source tree on May 29th, 2001, and for a few weeks OpenBSD-current did not contain any firewalling software.
Fortunately, in Switzerland Daniel Hartmeier was already doing some limited experiments involving kernel hacking in the networking code.
His starting point was hooking a small function of his own into the networking stack, making packets pass through it, and after a while he had started thinking about filtering. Then the license crisis happened.
IPFilter was pruned from the source tree on May 29th. The first commit of the PF code happened Sunday, June 24 2001 at 19:48:58 UTC.
A few months of rather intense activity followed, and the version of PF that was released with OpenBSD 3.0 contained a rather complete implementation of packet filtering, including network address translation.
From the looks of it, Daniel Hartmeier and the other PF developers made good use of their experience with the IPFilter code. Under any circumstances Daniel presented a USENIX 2002 paper with performance tests which show that the OpenBSD 3.1 PF performed equally well as or better under stress than IPFilter on the same platform or iptables on Linux.
In addition, some tests were run on the original PF from OpenBSD 3.0. These tests showed mainly that the code had gained in efficiency from version 3.0 to version 3.1. The article which provides the details is available from Daniel Hartmeier's web, see http://www.benzedrine.ch/pf-paper.html.
I have not seen comparable tests performed recently, but in my own experience and that of others, the PF filtering overhead is pretty much negligible. As one data point, the machine which gateways between one of the networks where I've done a bit of work and the world is a Pentium III 450MHz with 384MB of RAM. When I've remembered to check, I've never seen the machine at less than 96 percent 'idle' according to top.
It is however worth noting that various optimisations have been introduced to OpenBSD's PF code during recent releases (mainly by the current main PF developers Henning Brauer and Ryan McBride with contributions from others), making each release from 4.4 through 5.6 perform better than its predecessors.
It is worth noting that the IPFilter copyright episode spurred the OpenBSD team to perform a license audit of the entire source tree and ports in order to avoid similar situations in the future. A number of potential problems were uncovered and resolved over the months that followed, removing a number of potential license pitfalls for everyone involved in free software development. Theo de Raadt summed up the effort in a message to the openbsd-misc mailing list on February 20th, 2003, available among others from the MARC mailing list archives.