Книга: Iptables Tutorial 1.2.2
What is an IP filter
What is an IP filter
It is important to fully understand what an IP filter is. Iptables is an IP filter, and if you don't fully understand this, you will get serious problems when designing your firewalls in the future.
An IP filter operates mainly in layer 2, of the TCP/IP reference stack. Iptables however has the ability to also work in layer 3, which actually most IP filters of today have. But per definition an IP filter works in the second layer.
If the IP filter implementation is strictly following the definition, it would in other words only be able to filter packets based on their IP headers (Source and Destionation address, TOS/DSCP/ECN, TTL, Protocol, etc. Things that are actually in the IP header.) However, since the Iptables implementation is not perfectly strict around this definition, it is also able to filter packets based on other headers that lie deeper into the packet (TCP, UDP, etc), and shallower (MAC source address).
There is one thing however, that iptables is rather strict about these days. It does not "follow" streams or puzzle data together. This would simply be too processor- and memoryconsuming . The implications of this will be discussed a little bit more further on. It does keep track of packets and see if they are of the same stream (via sequence numbers, port numbers, etc.) almost exactly the same way as the real TCP/IP stack. This is called connection tracking, and thanks to this we can do things such as Destination and Source Network Address Translation (generally called DNAT and SNAT), as well as state matching of packets.
As I implied above, iptables can not connect data from different packets to each other (per default), and hence you can never be fully certain that you will see the complete data at all times. I am specifically mentioning this since there are constantly at least a couple of questions about this on the different mailing lists pertaining to netfilter and iptables and how to do things that are generally considered a really bad idea. For example, every time there is a new windows based virus, there are a couple of different persons asking how to drop all streams containing a specific string. The bad idea about this is that it is so easily circumvented. For example if we match for something like this:
Now, what happens if the virus/exploit writer is smart enough to make the packet size so small that cmd winds up in one packet, and .exe winds up in the next packet? Or what if the packet has to travel through a network that has this small a packet size on its own? Yes, since these string matching functions is unable to work across packet boundaries, the packet will get through anyway.
Some of you may now be asking yourself, why don't we simply make it possible for the string matches, etcetera to read across packet boundaries? It is actually fairly simple. It would be too costly on processor time. Connection tracking is already taking way to much processor time to be totally comforting. To add another extra layer of complexity to connection tracking, such as this, would probably kill more firewalls than anyone of us could expect. Not to think of how much memory would be used for this simple task on each machine.
There is also a second reason for this functionality not being developed. There is a technology called proxies. Proxies were developed to handle traffic in the higher layers, and are hence much better at fullfilling these requirements. Proxies were originally developed to handle downloads and often used pages and to help you get the most out of slow Internet connections. For example, Squid is a webproxy. A person who wants to download a page sends the request, the proxy either grabs the request or receives the request and opens the connection to the web browser, and then connects to the webserver and downloads the file, and when it has downloaded the file or page, it sends it to the client. Now, if a second browser wants to read the same page again, the file or page is already downloaded to the proxy, and can be sent directly, and saves bandwidth for us.
As you may understand, proxies also have quite a lot of functionality to go in and look at the actual content of the files that it downloads. Because of this, they are much better at looking inside the whole streams, files, pages etc.
Now, after warning you about the inherent problems of doing level 7 filtering in iptables and netfilter, there is actually a set of patches that has attacked these problems. This is called http://l7-filter.sourceforge.net/. It can be used to match on a lot of layer 7 protocols but is mainly to be used together with QoS and traffic accounting, even though it can be used for pure filtering as well. The l7-filter is still experimental and developed outside the kernel and netfilter coreteam, and hence you will not hear more about it here.