(Please note that most kernels provide the support necessary; it's probably safe to check back with this section only if you have problems.)
IP firewalling and firewall packet logging have to be compiled into the kernel. To see if IP firewalling is compiled into your kernel, type the command:
ls -al /proc/net/ip_fwchains /proc/net/ip_input
If ip_fwchains exists, you have ipchains compiled into your kernel. If ip_input exists, you have ipfwadmin firewalling compiled into your kernel. If neither file exists, one of the following is true:
When you recompile the kernel, I recommend you have all of the following enabled: network firewalls, ip firewalling, firewall packet logging, always defragment, proc filesystem, transparent proxy support, IP masquerading, and icmp masquerading.
To see if firewall packet logging is enabled in your kernel, type one of the following commands:
/sbin/ipfwadm -a deny -F -S 127.12.2.3/32 -o <Enter> /sbin/ipchains -A forward -s 127.12.2.3/32 -l <Enter> /sbin/iptables -A FORWARD -s 127.12.2.3/32 -j LOG<Enter>
The "-o" or "-l" at the end tells the kernel to log this particular packet type (one which should never show up). If your kernel does not support logging, I _think_ you would get an error. On the other hand, I've never had a kernel that has firewalling but does not have logging. The solution is the same - recompile your kernel to include both firewalling _and_ firewall packet logging.
(If recompiling a kernel is too daunting, try my automated kernel builder, "buildkernel", which can be found at http://www.stearns.org/buildkernel/).
Current versions of Mason handle ipfwadm, ipchains and iptables. It will accept log entries created under all three firewall types automatically. Mason automatically detects which kind of rule to create, although this can be overridden with environment variables set in /etc/masonrc. The masonrc file has comments describing these fields.
Make sure you have the ipfwadm, ipchains or iptables executable - one of these should be included with your distribution.
Mason has no support for ipfw firewalls (the firewalling used in kernels prior to 1.3.66). I don't intend to pursue this type of firewalling, but am not against integrating a patch if someone feels like adding the support. Does anyone still use this?
Mason does not try to look up the hostnames of any machines involved in DNS requests (unless they're in /etc/hosts). If it did, Mason could enter a situation where it issues a steady flow of DNS requests to resolve the machine names and each DNS request requires a new rule, which in turn requires more DNS requests... ugh.
The easy way to get machine names into your DNS rules is to make sure all your DNS servers are listed in /etc/hosts . If they're not listed there, Mason will just leave them as IP's.
When a packet needs to be processed (at entry, forwarding, or exit), the firewall scans the existing list of rules to decide whether to allow, deny or reject the packet. As this scans stops at the first rule that matches the packet, the order in which your final firewall rules are executed can make a difference. This document only provides basic coverage of how to order your rules - sorry. The best place to find out more about this is in the O'Reilly and associates books.
(If anyone would like to provide additional general guidelines as to how this is done, I would be glad to place them here with the appropriate disclaimers).
The packets Mason processes are data transfers between specific ports on specific machines. For example, here's a response packet from a specific FTP server (linux.kernel.org) to what is probably a machine on your LAN:
/sbin/ipfwadm -i accept -W ppp0 -I -P tcp -S linux.kernel.org/32 ftp -D \ devel1.goober.net/32 1024:65535 # ftp/tcp
The rule above (possibly along with others) would only allow devel1 to reach only linux.kernel.org, making for a ridiculously large ruleset if other machines wanted to ftp out to linux.kernel.org or wanted to reach other ftp servers.
By default, Mason _generalizes_ the source and destination IP addresses. For example, devel1.goober.net/32 is replaced with 18.104.22.168/24 (the fictitious network address block of which devel1 is a part). Since linux.kernel.org is not a part of any local network blocks, linux.kernel.org is replaced with 0/0 (which matches any machine anywhere).
This automatic generalization can be disabled by setting IPCONV="HOST" in /etc/masonrc.
Mason also does some generalization on the source and destination ports. Irc, X, realaudio, traceroute, and others use ranges of ports; Mason knows how to generalize many protocols to the appropriate range.
For the standard tcp and udp services, Mason generalizes the client port to 1024:65535. The connection that prompted this rule might have been, for example, port 1745 on devel1. As Mason didn't recognize 1745 as some special server, it assumed that the next connection might be from, say, port 1788. By using the entire range of high ports ("1024:65535" in the above rule), Mason uses a pretty standard approach to packet filtering to reduce the number of rules.
This program was originally intended for use on a traditional firewall - a packet filtering router (linux box that connects 2 or more networks through one or more interfaces). It works equally well on Linux boxes with only one interface. These could be workstations on a LAN, servers outside of your firewall, or even slip or ppp connected workstations. The number of interfaces and their type and speed are irrelevant to the firewall creation process.
This would be great for locking down a web or mail server outside your firewall, for example. Start up Mason and make sure you make one of every kind of connection you want to that machine. Mason will create the corresponding rules. Generalize these and add a default policy of "deny". _Only_ the connection types you specified will be allowed to that machine. The difficulty of setting up the rules has been the major impediment to this kind of hardened end node in the past. Now that Mason is here, there's no reason why every machine on your LAN can't have packet filtering enabled and active.
Note that on an end node (Linux box with a single NIC connected to a single IP network) you should never see forwarding rules created - this makes sense if you think about it.
You could technically create a firewall on a machine with only the loopback interface, but this would be more for instructional value about internal tcp connections than for any security goal. On the other hand, if you wanted to stop shell account users from getting to an internal Web server, you certainly could; just make sure you put in blocking rules for all interfaces, not just the loopback interface.
As a shell script, Mason is much less efficient at its work than a C app would be. On a slow machine, it can take a couple of seconds from the time the log entry is fed into it until the firewall rule is implemented. If the system is slow, if it has a lot of packets traveling through it, or if it simply has a great deal of log file traffic it can take Mason a long time to catch up. If this is the case, start slow. Try one connection type at a time and give the system a chance to settle before you move on.
If Mason _cannot_ catch up, choose the "EL" (End Learning) option in mason-gui-text. Wait until Mason stops, then restart learning.
If at all possible, try to set up these rules in a controlled environment. Hook up your firewall to machines that simulate the routers and networks that will be used in its final location. It is not a good idea to create a firewall in an environment not completely under your control.
If you must create the firewall rules in a live environment, be warned: Mason simply creates rules based on what traffic is passing through it. IT CANNOT DISTINGUISH BETWEEN THE TRAFFIC YOU'RE CREATING TO TEACH IT AND SOMEONE ACTIVELY TRYING TO HACK THROUGH YOUR FIREWALL. IF THIS HAPPENS, MASON WILL CREATE RULES THAT _SPECIFICALLY_ _ALLOW_ PEOPLE TO GET BACK IN LATER. _Please_ read and try to understand the rules before you put them to use in a production environment.
(I hate all caps too, but the "boldface" button on my keyboard is jammed :-).
The "hacker" mentioned above does not need to be a computer criminal in a far-off country looking to crash your machines. This individual could be someone in accounting that is (without malicious intent) connecting to an Internet IRC server, when this doesn't fit in the security policy you're trying to implement. If you don't read and understand the rules Mason spits out, you may very well leave an explicit opening for this user's future IRC use.
One more time: Mason _does_ _not_ understand the traffic flowing through your firewall; it just creates the rules that you can later use to specifically allow or disallow this traffic. This is why it is a good idea to delete any rules that look even vaguely suspicious. If it turns out these rules are needed for normal operation, they will be relearned when you restart Mason.
One of the common uses for Linux firewalling is to act not only as a packet filter but also as a masquerading host, allowing multiple machines to share a single IP address.
As of Mason 0.13.0, Mason will automatically masquerade traffic from RFC 1918 (also called "reserved") addresses. Since you probably don't want to masquerade between internal lans, you need to list all the interfaces leading _out_ to the real world (_not_ the interfaces that use these reserved addresses).
If you are especially cautious, you might not want Mason actively creating rules on your production server. Or maybe you think you've created a good firewall, but keep getting log messages and don't know how to keep your log files from filling your disk. Or perhaps your CPU can't keep up. Or maybe you just don't trust Mason's author - no offense taken :-). In all of the above circumstances, Mason can create the commands without actually being fed the log messages live. For example, if you have packet logging entries in /var/log/messages, try this:
cat /var/log/messages | grep ' I=' | DOCOMMAND="none" mason <Enter>
The output can, of course, be tee'd, redirected to a file, piped to less, etc. "... | sort | uniq" can be useful too when you're not converting it live.
Obviously, the source file can be one that has been transferred from another machine.
There is one caveat to the offline approach. The specific case is when one has a "deny" or "reject" policy in place for the input logging rule. Let's say I try to telnet through the firewall. My packet arrives at the firewall, is stopped and logged (so Mason can successfully create the correct input rule later). The firewall never has a rule implemented that allows me to get any further than that, however, so there is never a log entry created for any of the remaining 5 packet checks.
One way around this might be to use a policy of "accept" on your logging rules while you're creating /var/log/messages for later consumption by Mason. I'm not saying this is appropriate for you, but might be one way to handle this. Be warned; this can create very large log files as every packet passing through the system can create 6 log entries!
One final use for this technique is creating the rules when you're not root. Simply edit /etc/masonrc to set DOCOMMAND="NO" and the script will still output the appropriate ipfwadm/ipchains commands but won't try to execute them, allowing non-root users to create the firewall rules. Note that you still need to be root long enough to turn on some kind of logging, or /var/log/messages will never contain any entries to convert. Root privileges are also required to implement the rules once you've created them.
Mason converts the protocol number and type (i.e. 53, udp) into the more common name (domain, in this example). It uses the /etc/services file to do make this conversion. Before you start, make sure all the protocols you will work with are listed there. If a particular protocol is not in that file, Mason will have serious problems producing accurate rules.
Having this entry is especially important if you are working with services whose ports are >= 1024 (nfs, X, squid, irc, vdolive, etc.). If a service >= 1024 is not found in /etc/services, it will be automatically (and incorrectly) generalized to the port range of 1024-65535. If your favourite service isn't in there, simply edit the file and add it in the same format as the other entries.
These services whose ports are >=1024 can occasionally show up in your rules where Mason should have used 1024:65535 instead. Well, you know how to fix this, right? Just delete the rule, add the service to /etc/services, and relearn it.
The entries in /etc/services should only be for well-known server ports. Client ports (which are usually just random ports between 1024 and 65535 anyways) should not be listed in here. The specific example of something that should be missing is the ssh client port.
If you plan to do the conversion on one machine and actually run the firewall on another, make sure all of the protocols used are listed in the /etc/services on both machines.
The authoritative source for these ports is the Internet Assigned Numbers Authority (IANA). A list of these ports can be found at: ftp://ftp.isi.edu/in-notes/iana/assignments/port-numbers . Mason includes what seems to be an even more up-to-date reference; see /var/lib/mason/nmap-services. Many thanks to the authors of nmap.
Ipfwadm has two ways of adding rules: at the beginning of the rule list using insert ("-i"), or at the end of the list using append ("-a"). The usual way of creating the firewall is to flush the existing rules and then add each of the rules using append so they will be scanned in the same order in which they were implemented. For this reason, the rules that Mason spits out to stdout use "append" so they can easily be put in a shell script.
Mason needs some way to tell the kernel to not log already logged packets anymore. The way to do this is to put a matching rule before the logging rule. Unfortunately, that means one of two things: deleting the logging rule at the end, implementing the new rule at the end, and reinstating the logging rule, or simply inserting the new rule at the top of the list. The first option is tricky to do well. It's also a bad choice because the user using Mason may not be logging everything, so mason doesn't know what logging rule to reinstate. That leaves using "-i" to insert the rule at the very top of the list.
The end effect is that the rules that Mason displays use "-a" to match how that would be put into a rule file, but the rules that are actually implemented while Mason is running use "-i" to avoid relogging those packets again in this Mason run.
The major side effect of this approach is that the rule set in memory as Mason is running is almost certainly _not_ in the order you'd want. The final firewall rule set you put in place should flush whatever is in memory before starting so as to clean out these incorrectly ordered rules.
As ipchains and iptables support additional user defined chains, we can throw all the temporary rules in user defined chains (called inputN, outputN, and forwardN; the "N" stands for Nolog). These chains get called just before the logging rules.
During the course of a Mason run, it's quite reasonable that the firewall creator might want to spend some time working with traffic types that he/she wants to allow, and then switch over to other traffic types that he/she wants to reject or deny (see man ipfwadm for the subtle difference between deny and reject). If you change any settings by choosing "Change Settings" in mason-gui-text, it will automatically signal a running Mason to re-read its configuration file. You can do the same if running mason manually by typing "killall -USR1 mason".
Changing the target of a single rule to Accept, Deny, Reject, or Masquerade can be done right in the menu under that rule without having to go back to the main menu and changing the global settings.
To implement packet filtering, the Linux kernel needs to inspect each packet at at least one of the following three times: when the packet enters the system, as it passes through the system on the way to its exit interface, and as it leaves the system.
At each of those three times, the kernel can decide to allow or deny/reject the packet. The rules can be different at each stage - it's perfectly legal to, for example, allow it in, allow it to be forwarded, but then block it at the last second before it leaves the system.
A simple firewall could be implemented using just, say, input rules(*). It's when you get complex firewalls that having rules at all three stages is useful. You might want to allow hosts from eth0 to get to a pop-3 server on eth1, but not allow hosts from eth2 to get to the same server. This kind of restriction might be impossible to do without forwarding rules, especially if eth2 hosts _should_ be allowed to get to a pop-3 server on eth0.
For simpler firewalls, or if you want less than the imposing grandeur of a firewall ruleset that goes on for pages and pages, Mason can accomodate you. If you just want input rules, add the following to /var/lib/mason/baserules :
if [ -f /proc/net/ip_fwchains ]; then /sbin/ipchains -A forward -j ACCEPT /sbin/ipchains -A output -j ACCEPT elif [ -f /proc/net/ip_input ]; then /sbin/ipfwadm -F -a accept /sbin/ipfwadm -O -a accept fi
Place any general traffic types you don't care about in baserules.
Please note that I am _not_ advocating the above, but pointing out that the technique is available for those that feel the reduced security is appropriate for them.
(*) The exceptions to this are the special rules for redirecting packets (which must be done as an input rule), and masquerading packets, (which must be done as a forwarding rule). Even in the cases where you wish to use these facilities, it's still legal to implement packet filtering using another rule type.
Please note that the above does not apply to iptables. In iptables, packets are not inspected multiple times in multiple chains.
If you're creating this firewall rule set and you're telnetting, ssh'ing, or rsh'ing (collectively, "telnetting") in to the firewall, be careful. Some of the first rules to be created will be for the telnet packet flow you're using. If you are so unfortunate as to start this process with a policy of deny, guess what packet flow will be stopped almost immediately? That's right, your telnet session(s). Your machine will be completely locked down with no way to remotely reach it. (Now where were my car keys? <grrrr>)
If you want to put the rules allowing your remote access before starting Mason, great. If not, just make sure that your startup policy is allow or it's remote reboot time! Logging in on any of the console's virtual terminals does not require TCP/IP packets, so you can never lock yourself out completely.
You did read the section above on "simulating the working environment under controlled conditions", didn't you? Are you still sure you want to be creating a firewall not directly under your control? Just a thought...
Let's look at some standard rules that allows a telnet connection to a server somewhere (these are only two of the 6 possible rules).
allow LAN_IP's, ports 1024-65535 -> Outside_world_IP's, port 23 allow Outside_world_IP's, port 23 -> LAN_IP's, ports 1024-65535
It looks pretty safe, right? Hmmm....
Let's say that one of your LAN machines runs a squid server. This sits waiting for connections on port 3128. Additionally, consider the possibility that the root user on some Outside_world_IP machine writes some program that starts a connection _from_ port 23. This user starts this program and connects to your LANs squid server.
All with your firewalls full consent. Ugh.
The way to avoid this problem is to be able to identify the _direction_ in which the connection is created. We want to allow connections that start _from_ LAN:1024-65535 _to_ Outside:23, but block connections that start _from_ Outside:23 _to_ LAN:1024-65535.
The TCP ACK flag comes to the rescue. The first packet in a connection does _not_ have this flag set. Every packet after the first _does_ have this flag set. If we require all packets coming from the server port have their ACK flag set, we can stop the bogus connection from port 23 back to port 3128.
In short, by requiring all packets from a server port have their ACK flag set, we block connections that originate from those server ports.
Three notes. Only TCP uses ACK flags, so we can't use this to control the direction in which icmp or udp conversations are initiated. Secondly, DNS may be a problem. Tcp domain transfers and large dns requests can be from port 53 to port 53, depending on what dns software you're using. FTP-data connections do not have their ACK flag set because they can be created in either direction. Finally, there may be issues from ssh low ports if /etc/services has entries up near 1023.
Mason is able to automatically set the ack flag if your /etc/services lists all of the services you use.
I specifically avoided the "-b" (bidirectional) flag so that I could use "-k" to control the direction.
Iptables uses the state of the connection as a more dependable way of handling the above problem. I'd generally encourage you to use the "-m state --state ESTABLISHED,RELATED" lines in baserules. If you do, then Mason hands you a single rule for any given type of traffic; the opening packet. The ESTABLISHED,RELATED lines handle all the other packets.