mod_firewall

Note: mod_firewall is in its very early stages. This documentation is liable to change, and some described functionality may be missing, incomplete or contain bugs.

Introduction

A firewall is an invaluable tool in the sysadmin’s toolbox. However while low-level firewalls such as iptables and pf are incredibly good at what they do, they are generally not able to handle application-layer rules.

The goal of mod_firewall is to provide similar services at the XMPP layer. Based on rule scripts it can efficiently block, bounce, drop, forward, copy, redirect stanzas and more! Furthermore all rules can be applied and updated dynamically at runtime without restarting the server.

Details

mod_firewall loads one or more scripts, and compiles these to Lua code that reacts to stanzas flowing through Prosody. The firewall script syntax is unusual, but straightforward.

A firewall script is dominated by rules. Each rule has two parts: conditions, and actions. When a stanza matches all of the conditions, all of the actions are executed in order.

Here is a simple example to block stanzas from spammer@example.com:

FROM: spammer@example.com
DROP.

FROM is a condition, and DROP is an action. This is about as simple as it gets. How about heading to the other extreme? Let’s demonstrate something more complex that mod_firewall can do for you:

%ZONE myorganisation: staff.myorg.example, support.myorg.example

ENTERING: myorganisation
KIND: message
TIME: 12am-9am, 5pm-12am, Saturday, Sunday
REPLY=Sorry, I am afraid our office is closed at the moment. If you need assistance, please call our 24-hour support line on 123-456-789.

This rule will reply with a short message whenever someone tries to send a message to someone at any of the hosts defined in the ‘myorganisation’ outside of office hours.

Specifying rule sets

Firewall rules should be written into text files, e.g. ruleset.pfw file. One or more rule files can be specified in the configuration using:

firewall_scripts = { "path/to/ruleset.pfw", "path/to/ruleset2.pfw" }

If multiple files are specified and they both add rules to the same chains, each file’s rules will be processed in order, but the order of files is undefined.

Reloading Prosody’s configuration also reloads firewall rules.

Make sure that firewall_scripts is in the global section of the configuration file and not below a virtual host or a component - unless you want per-vhost firewall rules.

Conditions

All conditions must come before any action in a rule block. The condition name is followed by a colon (‘:’), and the value to test for.

A condition can be preceded or followed by NOT to negate its match. For example:

NOT FROM: user@example.com
KIND NOT: message

Some conditions do not take parameters, and these should end with just a question mark, like:

IN ROSTER?

Zones

A ‘zone’ is one or more hosts or JIDs. It is possible to match when a stanza is entering or leaving a zone, while at the same time not matching traffic passing between JIDs in the same zone.

Zones are defined at the top of a script with the following syntax (they are not part of a rule block):

%ZONE myzone: host1, host2, user@host3, foo.bar.example

There is an automatic zone named $local, which automatically includes all of the current server’s active hosts (including components). It can be used to match stanzas entering or leaving the current server.

A host listed in a zone also matches all users on that host (but not subdomains).

The following zone-matching conditions are supported:

Lists

It is possible to create or load lists of strings for use in scripts. For example, you might load a list of blocked JIDs, malware URLs or simple words that you want to filter messages on.

List types

memory

%LIST name: memory (limit: number)

A memory-only list, with an optional limit. Supports addition and removal of items by scripts.

If a limit is provided, the oldest item will be discarded to make room for a new item if the list is full. The limit is useful to prevent infinite memory growth on busy servers.

file

%LIST name: file:/path/to/file (missing: string)

Reads a list from a file. The list can be added to and removed from by scripts, but these changes do not persist between restarts.

If the file is missing, an error will be raised. The optional ‘missing’ parameter can be set to ‘ignore’ (e.g. (missing: ignore)) to ignore a missing file.

http

%LIST name: http://example.com/ (ttl: number, pattern: pat, hash: sha1, checkcerts: when-sni)

Fetches a list from a HTTP or HTTPS URL. The following options are accepted:

The “when-sni” default disables certificate verification when Prosody’s HTTP client API doesn’t support SNI, as in Prosody 0.11.6 and earlier.

CHECK LIST

Checks whether a simple expression is found in a given list.

Example:

%LIST blocked_jids: file:/etc/prosody/blocked_jids.txt

# Rule to block presence subscription requests from blocked JIDs
KIND: presence
TYPE: subscribe
CHECK LIST: blocked_jids contains $<@from>
BOUNCE=policy-violation (Your JID is blocked)

SCAN

SCAN allows you to search inside a stanza for a given pattern, and check each result against a list. For example, you could scan a message body for words and check if any of the words are found in a given list.

Before using SCAN, you need to define a search location and a pattern. The search location uses the same ‘path’ format as documented under the ‘INSPECT’ condition. Patterns can be any valid Lua pattern.

To use the above example:

# Define a search location called 'body' which fetches the text of the 'body' element
%SEARCH body: body#
# Define a pattern called 'word' which matches any sequence of letters
%PATTERN word: [A-Za-z]+
# Finally, we also need our list of "bad" words:
%LIST badwords: file:/etc/prosody/bad_words.txt

# Now we can use these to SCAN incoming stanzas
# If it finds a match, bounce the stanza
SCAN: body for word in badwords
BOUNCE=policy-violation (This word is not allowed!)

COUNT

COUNT is similar to SCAN, in that it uses a defined SEARCH and breaks it up according to a PATTERN. Then it counts the number of results.

For example, to block every message with more than one URL:

# Define a search location called 'body' which fetches the text of the 'body' element
%SEARCH body: body#
# Define a pattern called 'url' which matches HTTP links
%PATTERN url: https?://%S+

COUNT: url in body > 1
BOUNCE=policy-violation (Up to one HTTP URL is allowed in messages)

Stanza matching

Stanza types

Note: The type ‘available’ for presence does not actually appear in the protocol. Available presence is signalled by the omission of a type. Similarly, a message stanza with no type is equivalent to one of type ‘normal’. mod_firewall handles these cases for you automatically.

Sender/recipient matching

The TO and FROM conditions both accept wildcards in the JID when it is enclosed in angle brackets (‘<…>’). For example:

# All users at example.com
FROM: <*>@example.com

# The user 'admin' on any subdomain of example.com
FROM: admin@<*.example.com>

You can also use Lua’s pattern matching for more powerful matching abilities. Patterns are a lightweight regular-expression alternative. Simply contain the pattern in double angle brackets. The pattern is automatically anchored at the start and end (so it must match the entire portion of the JID).

# Match admin@example.com, and admin1@example.com, etc.
FROM: <<admin%d*>>@example.com

Note: It is important to know that ‘example.com’ is a valid JID on its own, and does not match ‘user@example.com’. To efficiently match domains we recommend defining them as Zones.

These additional conditions do not support pattern matching, but are useful to match the exact to/from address on a stanza. For example, if no resource is specified then only bare JIDs will be matched. TO and FROM match all resources if no resource is specified to match.

Note: Some chains execute before Prosody has performed any normalisation or validity checks on the to/from JIDs on an incoming stanza. It is not advisable to perform access control or similar rules on JIDs in these chains (see the chain documentation for more info).

GeoIP matching

This condition uses a GeoIP database to look up the origin country of the IP attached to the current session.

For example:

# 3 letter country code
FROM COUNTRY: SWE

# or 2 letter
FROM COUNTRY: SE

# Explicit
FROM COUNTRY: code=SE
FROM COUNTRY: code3=SWE

Note: This requires that the lua-geoip and geoip-database packages are installed (on Debian, package names may differ on other operating systems).

INSPECT

INSPECT takes a ‘path’ through the stanza to get a string (an attribute value or text content). An example is the best way to explain. Let’s check that a user is not trying to register an account with the username ‘admin’. This stanza comes from XEP-0077: In-band Registration:

<iq type='set' id='reg2'>
  <query xmlns='jabber:iq:register'>
    <username>bill</username>
    <password>Calliope</password>
    <email>bard@shakespeare.lit</email>
  </query>
</iq>

KIND: iq
TYPE: set
PAYLOAD: jabber:iq:register
INSPECT: {jabber:iq:register}query/username#=admin
BOUNCE=not-allowed (The username 'admin' is reserved.)

That weird string deserves some explanation. It is a path, divided into segments by ‘/’. Each segment describes an element by its name, optionally prefixed by its namespace in curly braces (‘{…}’). If the path ends with a ‘#’ then the text content of the last element will be returned. If the path ends with ‘@name’ then the value of the attribute ‘name’ will be returned.

You can use INSPECT to test for the existence of an element or attribute, or you can check if it matches a specific value, e.g. by appending =VALUE (like in the example above, that checks if the content of username is ‘admin’).

INSPECT comparison operators

As well as checking for an exact string match, there are some other modifiers you can apply to the comparison:

Finally, if the comparison operator is preceded by a $ character, expressions will be interpreted in the string following the comparison operator.

e.g. INSPECT: {jabber:iq:register}query/username}$/=$(session.host) would match if the username of an account registration contained the session’s current hostname somewhere in it.

INSPECT performance

INSPECT can be somewhat slower than the other stanza matching conditions. To minimise performance impact, always place it below other faster condition checks where possible (e.g. in the example above we first checked KIND, TYPE and PAYLOAD matched what we wanted before reaching the INSPECT rule).

Roster

These conditions access the roster of the recipient (only). Therefore they cannot (currently) be used in some chains, such as for outgoing messages (the recipient may be on another server).

Performance note: these checks can potentially cause storage access (especially if the recipient is currently offline), so you may want to limit their use in high-traffic situations, and place rules containing a roster check below other rules (such as a rate limiter). The storage access is performed unconditionally just before evaluation of the first rule that contains a roster-based condition, even if earlier conditions in that rule do not match.

IN ROSTER

Tests whether the sender is in the recipient’s roster.

IN ROSTER?

IN ROSTER GROUP

Tests whether the sender is in the recipient’s roster, and in the named group.

IN ROSTER GROUP: Friends

SUBSCRIBED

Tests whether the recipient is subscribed to the sender, ie will receive presence updates from them.

Note that this does work, regardless of direction and which chain is used, since both the sender and the recipient will have mirrored roster entries.

Groups

Using Prosody’s mod_groups it is possible to define groups of users on the server. You can match based on these groups in firewall rules.

CROSSING GROUPS

The CROSSING GROUPS condition takes a comma-separated list of groups to check. If the sender and recipient are not in the same group (only the listed groups are checked), then the this condition matches and the stanza is deemed to be crossing between groups.

For example, if you had three groups: Engineering, Marketing and Employees. All users are members of the ‘Employees’ group, and the others are for employees of the named department only.

To prevent employees in the marketing department from communicating with engineers, you could use the following rule:

CROSSING GROUPS: Marketing, Engineering
BOUNCE=policy-violation (no communication between these groups is allowed!)

This works, even though both the users are in the ‘Employees’ group, because that group is not listed in the condition.

In the above example, a user who is member of both groups is not restricted.

SENT DIRECTED PRESENCE TO SENDER

This condition matches if the recipient of a stanza has previously sent directed presence to the sender of the stanza. This is often done in XMPP to exchange presence information with JIDs that are not on your roster, such as MUC rooms.

This condition does not take a parameter - end the condition name with a question mark:

# Rule to bounce messages from senders not in the roster who haven't been sent directed presence
NOT IN ROSTER?
NOT SENT DIRECTED PRESENCE TO SENDER?
BOUNCE=service-unavailable

Permissions

Rules can consult Prosody’s internal role and permissions system to check whether a certain action may be performed. The acting entity, their role, and appropriate context is automatically inferred. All you need to do is provide the identifier of the permission that should be checked.

As with all other conditions, MAY can be combined with NOT to negate the result of the check.

Example, blocking outgoing stanzas from users with roles that do not allow the ‘xmpp:federate’ permission:

::deliver_remote
MAY NOT: xmpp:federate
BOUNCE=policy-violation (You are not allowed access to the federation)

Roles

Note: In most cases, you should avoid checking for specific roles, and instead check for permissions granted by those roles (using the ‘MAY’ condition).

Admins

Deprecated: These conditions should no longer be used. Prefer ‘MAY’, ‘TO ROLE’ or ‘FROM ROLE’.

Prosody allows certain JIDs to be declared as administrators of a host, component or the whole server.

Time and date

TIME

Matches stanzas sent during certain time periods.

TIME: 10pm-6am, 14:00-15:00
REPLY=Zzzz.

DAY

It is also possible to match only on certain days of the week.

Example:

DAY: Sat-Sun, Wednesday
REPLY=Sorry, I'm out enjoying life!

All times and dates are handled in the server’s local time.

Rate-limiting

It is possible to selectively rate-limit stanzas, and use rules to decide what to do with stanzas when over the limit.

First, you must define any rate limits that you are going to use in your script. Here we create a limiter called ‘normal’ that will allow 2 stanzas per second, and then we define a rule to bounce messages when over this limit. Note that the RATE definition is not part of a rule (multiple rules can share the same limiter).

%RATE normal: 2 (burst 3)

KIND: message
LIMIT: normal
BOUNCE=policy-violation (Sending too fast!)

The ‘burst’ parameter on the rate limit allows you to spread the limit check over a given time period. For example the definition shown above will allow the limit to be temporarily surpassed, as long as it is within the limit after 3 seconds. You will almost always want to specify a burst factor.

Both the rate and the burst can be fractional values. For example a rate of 0.1 means only one event is allowed every 10 seconds.

The LIMIT condition actually does two things; first it counts against the given limiter, and then it checks to see if the limiter over its limit yet. If it is, the condition matches, otherwise it will not.

Note: Reloading mod_firewall resets the current state of any limiters.

Dynamic limits

Sometimes you may want to have multiple throttles in a single condition, using some property of the session or stanza to determine which throttle to use. For example, you might have a limit for incoming stanzas, but you want to limit by sending JID, instead of all incoming stanzas sharing the same limit.

You can use the ‘on’ keyword for this, like so:

LIMIT: normal on EXPRESSION

For more information on expressions, see the section later in this document.

Each value of ‘EXPRESSION’ has to be tracked individually in a table, which uses a small amount of memory. To prevent memory exhaustion, the number of tracked values is limited to 1000 by default. You can override this by setting the maximum number of table entries when you define the rate:

%RATE normal: 2 (burst 3) (entries 4096)

Old values are automatically removed from the tracking table. However if the tracking table becomes full, new entries will be rejected - it will behave as if the rate limit was reached, even for values that have not been seen before. Since this opens up a potential denial of service (innocent users may be affected if malicious users can fill up the tracking table within the limit period). You can choose to instead “fail open”, and allow the rate limit to be temporarily bypassed when the table is full. To choose this behaviour, add (allow overflow) to the RATE definition.

Session marking

It is possible to ‘mark’ sessions (see the MARK_ORIGIN action below). To match stanzas from marked sessions, use the ORIGIN_MARKED condition.

Example usage:

# This rule drops messages from sessions that have been marked as spammers in the past hour
ORIGIN MARKED: spammer (3600s)
DROP.

# This rule marks the origin session as a spammer if they send a message to a honeypot JID
KIND: message
TO: honeypot@example.com
MARK ORIGIN=spammer

Actions

Actions come after all conditions in a rule block. There must be at least one action, though conditions are optional.

An action without parameters ends with a full-stop/period (‘.’), and one with parameters uses an equals sign (‘=’):

# An action with no parameters:
DROP.

# An action with a parameter:
REPLY=Hello, this is a reply.

Route modification

The following common actions modify the stanza’s route in some way. These rules will halt further processing of the stanza - no further actions will be executed, and no further rules will be checked.

Note: It is incorrect behaviour to reply to an ‘error’ stanza with another error, so BOUNCE will simply act the same as ‘DROP’ for stanzas that should not be bounced (error stanzas and iq results).

Replying and forwarding

These actions cause a new stanza to be generated and sent somewhere. Processing of the original stanza will continue beyond these actions.

Reporting

Only the jid is mandatory. The reason parameter should be either abuse, spam or a custom URI. If not specified, it defaults to abuse. After the reason, some human-readable text may be included to explain the report.

Example:

KIND: message
TO: honeypot@example.com
REPORT TO=antispam.example.com spam Caught by the honeypot!
DROP.

Stanza modification

These actions make it possible to modify the content and structure of a stanza.

Sessions

It is possible to mark sessions, and then use these marks to match rules later on.

Note: Marks apply to sessions, not JIDs. E.g. if marking in a rule that matches a stanza received over s2s, it is the s2s session that is marked.

It is possible to have multiple marks on an origin at any given time.

Informational

You can include expressions in log messages, using $(...) syntax. For example, to log the stanza that matched the rule, you can use $(stanza), or to log just the top tag of the stanza, use $(stanza:top_tag()). To fetch the sender JID, use $(stanza.attr.from).

Example:

# Log all stanzas to user@example.com:
TO: user@example.com
LOG=[debug] User received: $(stanza)

More info about expressions can be found below.

Chains

Rules are grouped into “chains”, which are injected at particular points in Prosody’s routing code.

Available built-in chains are:

A chain is begun by a line ::name where ‘name’ is the name of the chain you want the following rules to be inserted into. If no chain is specified, rules are put into the ‘deliver’ chain.

It is possible to create custom chains (useful with the JUMP CHAIN action described below). User-created chains must begin with “user/”, e.g. “user/spam_filtering”.

Example of chain use:

# example.com's firewall script

# This line is optional, because 'deliver' is the default chain anyway:
::deliver

# This rule matches any stanzas delivered to our local user bob:
TO: bob@example.com
DROP.

# Oops! This rule will never match, because alice is not a local user,
# and only stanzas to local users go through the 'deliver' chain:
TO: alice@remote.example.com
DROP.

# Create a 'preroute' chain of rules (matched for incoming stanzas from local clients):
::preroute
# These rules are matched for outgoing stanzas from local clients

# This will match any stanzas sent to alice from a local user:
TO: alice@remote.example.com
DROP.

It is possible to jump to chains defined by other scripts and modules.

Expressions

Some conditions and actions in rules support “expressions” in their parameters (their documentation will indicate if this is the case). Most parameters are static once the firewall script is loaded and compiled internally, however parameters that allow expressions can be dynamically calculated when a rule is being run.

There are two kinds of expression that you can use: stanza expressions, and code expressions.

Stanza expressions

Stanza expressions are of the form $<...>, where ... is a stanza path. For syntax of stanza paths, see the documentation for the ‘INSPECT’ condition above.

Example:

LOG=Matched a stanza from $<@from> to $<@to>

There are built in functions which can be applied to the output of a stanza expression, by appending the pipe (‘|’) operator, followed by the function name. These functions are:

For example, to apply a rate limit to stanzas per sender domain:

LIMIT normal on $<@from|host>

If the path does not match (e.g. the element isn’t found, or the attribute doesn’t exist) or any of the functions fail to produce an output (e.g. an invalid JID was passed to a function that only handles valid JIDs) the expression will return the text <undefined>. You can override this by ending the expression with a double pipe (‘||’) followed by a quoted string to use as a default instead. E.g. to default to the string “normal” when there is no ‘type’ attribute:

LOG=Stanza type is $<@type||"normal">

Code expressions

Code expressions use $(...) syntax. Code expressions are powerful, and allow unconstrained access to Prosody’s internal environment. Therefore code expressions are typically for advanced use-cases only. You may want to refer to Prosody’s developer documentation for more information. In particular, within code expressions you may access the ‘session’ object, which is the session object of the origin of the stanza, and the ‘stanza’ object, which is the stanza being considered within the current rule. Whatever value the expression returns will be converted to a string.

Example to limit stanzas per session type:

LIMIT: normal on $(session.type)

sudo prosodyctl install --server=https://modules.prosody.im/rocks/ mod_firewall