Writing Rspamd rules

In this tutorial, we describe how to create new rules for Rspamd - both using Lua and regular expressions.

Introduction

Rules are the essential part of a spam filtering system and Rspamd ships with some prepared rules by default. However, if you run your own system you might want to have your own rules for better spam filtering or a better false positives rate. Rules are usually written in Lua, where you can specify both custom logic and generic regular expressions.

Configuration files

Since Rspamd ships with its own rules it is a good idea to store your custom rules and configuration in separate files to avoid clashing with the default rules which might change from version to version. There are some possibilities to achieve this:

• Local rules, both Lua and regular expressions, should be stored in the file named ${CONFDIR}/rspamd.local.lua where ${CONFDIR} is the directory where your configuration files are placed (e.g. /etc/rspamd, or /usr/local/etc/rspamd for some systems)

Lua local configuration can be used to both override and extend, for example if the main lua file has the following line:

rspamd.lua:

-- Regular expression rule defined in the main Rspamd configuration
config['regexp']['symbol'] = '/some_re/'

then you can define additional rules in rspamd.local.lua:

-- Regular expression rules defined in a local configuration
config['regexp']['symbol1'] = '/other_re/' -- add 'symbol1' key to the table
config['regexp']['symbol'] = '/override_re/' -- replace regexp for 'symbol'

Please bear in mind that this method is different comparing to the ordinary configuration files that use a different (UCL based) syntax and usually have two special includes:

.include(try=true,priority=1) "$CONFDIR/local.d/config.conf"
.include(try=true,priority=1) "$CONFDIR/override.d/config.conf"

In this case you can either enrich/rewrite (using local.d) or ultimately override (using override.d) settings in the Rspamd configuration.

For example, we can override some default symbols shipped with Rspamd. To do that we can create and edit etc/rspamd/local.d/metrics.conf:

Writing rules

There are two types of rules that are normally defined by Rspamd:

• Lua rules: code in written in Lua
• Regexp rules: regular expressions and combinations of regular expressions to match specific patterns

Lua rules are useful for some complex tasks: check DNS, query Redis or HTTP, examine some task-specific details. Regexp rules are useful since they are heavily optimized by Rspamd (especially when Hyperscan is enabled) and allow matching custom patterns in headers, URLs, text parts and even the entire message body.

There is another option called selectors that allows to combine data extraction and data transformation routines to skip custom Lua code writing. Selectors framework is also useful to reuse custom extraction/transformation routines in different plugins and even in the regular expressions constructions.

Rule weights

Rule weights are usually defined in the metrics section and contain the following data:

• score triggers for different actions
• symbol scores
• symbol descriptions
• symbol group definitions:
  • symbols in group
  • description of group
  • joint group score limit

For built-in rules scores are placed in the file called ${CONFDIR}/metrics.conf, however, you have two possibilities to define scores for your rules:

1. Define scores in local.d/groups.conf as following:

symbol "MY_SYMBOL" {
  description = "my cool rule";
  score = 1.5;
}
# Or, if you want to include it into a group:
group "mygroup" {
	symbol "MY_SYMBOL" {
	  description = "my cool rule";
	  score = 1.5;
	}
}

1. Define scores directly in Lua when describing symbol:

-- regexp rule
config['regexp']['MY_SYMBOL'] = {
	re = '/a/M & From=/blah/',
	score = 1.5,
	description = 'my cool rule',
	group = 'my symbols'
}

-- lua rule
rspamd_config.MY_LUA_SYMBOL = {
	callback = function(task)
		-- Do something
		return true
	end,
	score = -1.5,
	description = 'another cool rule',
	group = 'my symbols'
}

Please bear in mind that the scores you define directly from Lua have lower priority and are overriden by scores defined in the groups.conf file. WebUI defined scores have even higher priority.

Regexp rules

Regexp rules are executed by the regexp module of Rspamd. You can find a detailed description of the syntax in the regexp module documentation

Here are some hints to maximise performance of your regexp rules:

• Prefer lightweight regexps, such as header or URL, to heavy ones, such as mime or body regexps (unless you are using Hyperscan, which is default for all Intel based platforms)
• Avoid complex regexps, avoid backtracing (e.g. /*+a?), avoid lookahead/lookbehind, avoid potentially empty patterns, avoid large boundary constraints a{1000,100000}, especially when using Hyperscan - all these constructions have to fallback to PCRE increasing scan complexity, but even PCRE can have exponential grow for some of the listed cases

Following these rules allows you to create fast and efficient rules. To add regexp rules you should use the config global table that is defined in any Lua file used by Rspamd:

local reconf = config['regexp'] -- Create alias for regexp configs

local re1 = 'From=/foo@/H' -- Mind local here
local re2 = '/blah/P'

reconf['SYMBOL'] = {
	re = string.format('(%s) && !(%s)', re1, re2), -- use string.format to create expression
	score = 1.2,
	description = 'some description',

	condition = function(task) -- run this rule only if some condition is satisfied
		return true
	end,
}

Lua rules

Lua rules are more powerful than regexp ones but they are not as heavily optimized and can cause performance issues if written incorrectly. All Lua rules accept a special parameter called task which represents a scanned message.

Return values

Each Lua rule can return 0, or false, meaning that the rule has not matched, or true if the symbol should be inserted. In fact, you can return any positive or negative number which would be multiplied by the rule’s static score, e.g. if the rule score is 1.2, then when your function returns 1 the symbol will have a score of 1.2, and when your function returns 0.5 then the symbol will have a score of 0.6. The common convention of the return values is to return confidence factor varying from 0 to 1.0. The remaining return values are treated as symbol’s options. They can be either in a single table:

return true,1.0,{'option1', 'option2'}

or as a list of return values:

return true,1.0,'option1','option2'

There is no difference in these notations. Tables are usually more convenient if you form list of options during the rule progressing.

Rule conditions

Like regexp rules, conditions are allowed for Lua regexps, for example:

rspamd_config.SYMBOL = {
	callback = function(task)
		return 1
	end,
	score = 1.2,
	description = 'some description',

	condition = function(task) -- run this rule only if some condition is satisfied
		return true
	end,
}

Useful task manipulations

There are a number of methods in task objects. For example, you can get any part of a message:

rspamd_config.HTML_MESSAGE = {
  callback = function(task)
    local parts = task:get_text_parts()

    if parts then
      for i,p in ipairs(parts) do
        if p:is_html() then
          return 1
        end
      end
    end

    return 0
  end,
  score = -0.1,
  description = 'HTML included in message',
}

You can get HTML information:

local function check_html_image(task, min, max)
  local tp = task:get_text_parts()

  for _,p in ipairs(tp) do
    if p:is_html() then
      local hc = p:get_html()
      local len = p:get_length()


      if len >= min and len < max then
        local images = hc:get_images()
        if images then
          for _,i in ipairs(images) do
            if i['embedded'] then
              return true
            end
          end
        end
      end
    end
  end
end

rspamd_config.HTML_SHORT_LINK_IMG_1 = {
  callback = function(task)
    return check_html_image(task, 0, 1024)
  end,
  score = 3.0,
  group = 'html',
  description = 'Short html part (0..1K) with a link to an image'
}

You can get message headers with full information passed:

rspamd_config.SUBJ_ALL_CAPS = {
  callback = function(task)
    local util = require "rspamd_util"
    local sbj = task:get_header('Subject')

    if sbj then
      local stripped_subject = subject_re:search(sbj, false, true)
      if stripped_subject and stripped_subject[1] and stripped_subject[1][2] then
        sbj = stripped_subject[1][2]
      end

      if util.is_uppercase(sbj) then
        return true
      end
    end

    return false
  end,
  score = 3.0,
  group = 'headers',
  description = 'All capital letters in subject'
}

You can also access HTTP headers, URLs and other useful properties of Rspamd tasks. Moreover, you can use global convenience modules exported by Rspamd, such as rspamd_util or rspamd_logger by requiring them in your rules:

rspamd_config.SUBJ_ALL_CAPS = {
  callback = function(task)
    local util = require "rspamd_util"
    local logger = require "rspamd_logger"
    ...
  end,
}

Rspamd symbols

Rspamd rules fall under three categories:

1. Connection filters - are executed before a message has been processed (e.g. on a connection stage)
2. Pre-filters - run before other rules
3. Filters - run normally
4. Post-filters - run after all checks
5. Idempotent filters - performs statistical checks and are NOT allowed to change scan result in any way

The most common type of rules are generic filters. Each filter is basically a callback that is executed by Rspamd at some time, along with an optional symbol name associated with this callback. In general, there are three options to register symbols:

• register callback and associated symbol
• register just a plain callback (symbol is not expected to be inserted to result)
• register symbol with no callback (virtual symbol) and an associated callback rule

The last option is useful when you have a single callback but with different possible results; for example SYMBOL_ALLOW or SYMBOL_DENY. Filters are registered using the following method:

rspamd_config:register_symbol{
  type = 'normal', -- or virtual, callback, prefilter or postfilter
  name = 'MY_SYMBOL',
  callback = function(task) -- Main logic
  end,
  score = 1.0, -- Metric score
  group = 'some group', -- Metric group
  description = 'My super symbol',
  flags = 'fine', -- fine: symbol is always checked, skip: symbol is always skipped, empty: symbol allows to be executed with no message
  --priority = 2, -- useful for postfilters and prefilters to define order of execution
}

local id = rspamd_config:register_symbol({
  name = 'DMARC_CALLBACK',
  type = 'callback',
  callback = dmarc_callback
})
rspamd_config:register_symbol({
  name = dmarc_symbols['allow'],
  flags = 'nice',
  parent = id,
  type = 'virtual'
})
rspamd_config:register_symbol({
  name = dmarc_symbols['reject'],
  parent = id,
  type = 'virtual'
})
rspamd_config:register_symbol({
  name = dmarc_symbols['quarantine'],
  parent = id,
  type = 'virtual'
})
rspamd_config:register_symbol({
  name = dmarc_symbols['softfail'],
  parent = id,
  type = 'virtual'
})
rspamd_config:register_symbol({
  name = dmarc_symbols['dnsfail'],
  parent = id,
  type = 'virtual'
})
rspamd_config:register_symbol({
  name = dmarc_symbols['na'],
  parent = id,
  type = 'virtual'
})

rspamd_config:register_dependency(id, symbols['spf_allow_symbol'])
rspamd_config:register_dependency(id, symbols['dkim_allow_symbol'])

Numeric id is returned by a registration function with callback and can be used to link symbols:

• add virtual symbols associated with this callback
• correctly display average time for symbols without callbacks
• properly sort symbols
• register dependencies on virtual symbols (in fact, the true dependency is created based on the parent symbol but it is sometimes convenient to use virtual symbols for simplicity)

Asynchronous actions

For asynchronous actions, such as Redis access or DNS checks it is recommended to use dedicated callbacks, called symbol handlers. The difference to generic Lua rules is that dedicated callbacks are not obliged to return value but they use the method task:insert_result(symbol, weight) to indicate a match. All Lua plugins are implemented as symbol handlers. Here is a simple example of a symbol handler that checks DNS:

rspamd_config:register_symbol('SOME_SYMBOL', 1.0,
	function(task)
		local to_resolve = 'google.com'
		local logger = require "rspamd_logger"

		local dns_cb = function(resolver, to_resolve, results, err)
			if results then
				logger.infox(task, '<%1> host: [%2] resolved for symbol: %3',
					task:get_message_id(), to_resolve, 'RULE')
				task:insert_result(rule['symbol'], 1)
			end
		end
		task:get_resolver():resolve_a({
			task=task,
			name = to_resolve,
			callback = dns_cb})
	end)

You can also set the desired score and description:

rspamd_config:set_metric_symbol('SOME_SYMBOL', 1.2, 'some description')
-- Table version
if rule['score'] then
  if not rule['group'] then
    rule['group'] = 'whitelist'
  end
  rule['name'] = symbol
  rspamd_config:set_metric_symbol(rule)
end

You can also use coroutines to simplify your asynchronous code.

Redis requests

Rspamd uses Redis heavily for different purposes. There are couple of useful functions that are defined in the file lua_redis.lua. These functions should be available globally in all Lua modules. Here is an example of parsing Redis config for a module and making requests subsequently:

local redis_params
local lua_redis = require "lua_redis"

local function symbol_cb(task)
  local function redis_set_cb(err)
    if err ~=nil then
      rspamd_logger.errx(task, 'redis_set_cb received error: %1', err)
    end
  end
  -- Create hash of message-id and store to redis
  local key = make_key(task)
  local ret = lua_redis.redis_make_request(task,
    redis_params, -- connect params
    key, -- hash key
    true, -- is write
    redis_set_cb, --callback
    'SETEX', -- command
    {key, tostring(settings['expire']), "1"} -- arguments
  )
end

-- Load redis server for module named 'module'
redis_params = lua_redis.parse_redis_server('module')
if redis_params then
  -- Register symbol
end

Using maps from Lua plugin

Maps hold dynamically loaded data like lists or IP trees. It is possible to use 3 types of maps:

• radix stores IP addresses
• hash stores plain strings and values
• set stores plain strings with no values
• regexp stores regular expressions (powered by hyperscan if possible)
• regexp_multi stores regular expressions and returns *all matches using get_key method
• glob stores glob expressions (powered by hyperscan if possible)
• glob_multi stores glob expressions and returns *all matches using get_key method
• callback call for a specified Lua callback when a map is loaded or changed, map’s content is passed to that callback as a parameter

Here is a sample of using maps from Lua API:

local rspamd_logger = require "rspamd_logger"

-- Add two maps in configuration section
local hash_map = rspamd_config:add_map{
  type = "hash",
  urls = ['file:///path/to/file'],
  description = 'sample map'
}
local radix_tree = rspamd_config:add_map{
  type = 'radix', 
  urls = ['http://somehost.com/test.dat', 'fallback+file:///path/to/file'], 
  description = 'sample ip map'
}
local generic_map = rspamd_config:add_map{
  type = 'callback',
  urls = ['file:///path/to/file']
  description = 'sample generic map',
  callback = function(str)
    -- This callback is called when a map is loaded or changed
    -- Str contains map content
    rspamd_logger.info('Got generic map content: ' .. str)
  end
}

local function sample_symbol_cb(task)
    -- Check whether hash map contains from address of message
    if hash_map:get_key(task:get_from()) then
        -- Check whether radix map contains client's ip
        if radix_map:get_key(task:get_from_ip_num()) then
        ...
        end
    end
end

Difference between config and rspamd_config

It might be confusing that there are two variables with a common meaning. Unfortunately, this is a legacy of older versions of Rspamd. However, currently rspamd_config represents an object that can be used for almost all configuration tasks:

• Get configuration options:

rspamd_config:get_all_opts('section')

• Add maps:

rule['map'] = rspamd_config:add_kv_map(rule['domains'],
            "Whitelist map for " .. symbol)

• Register callbacks for symbols:

rspamd_config:register_symbol('SOME_SYMBOL', 1.0, some_functions)

• Register lua rules (note that __newindex metamethod is actually used here):

rspamd_config.SYMBOL = {...}

• Register composites, pre-filters, post-filters and so on

On the other hand, the config global is extremely simple: it’s just a plain table of configuration options that is exactly the same as defined in rspamd.conf (and rspamd.conf.local or rspamd.conf.override). However, you can also use Lua tables and even functions for some options. For example, the regexp module also can accept a callback argument:

config['regexp']['SYMBOL'] = {
  callback = function(task) ... end,
  ...
}

Such syntax is discouraged, however, and is preserved mostly for compatibility reasons. Furthermore, you cannot use neither async requests nor coroutines in such callbacks - it will cause Rspamd crash.

Configuration order

There is a strict order of configuration application:

1. Configuration files are loaded
2. Lua rules are loaded and they can override everything from the previous steps, with the important exception of rules scores, which are NOT overridden if the relevant symbol is also defined in a metric section
3. Dynamic configuration options defined in the WebUI (normally) are loaded and can override rule scores or action scores from the previous steps

Rules check order

Rules in Rspamd are checked in the following order: