path: root/lualib/lua_stat.lua

--[[
Copyright (c) 2022, Vsevolod Stakhov <vsevolod@rspamd.com>

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
]]--

--[[[
-- @module lua_stat
-- This module contains helper functions for supporting statistics
--]]

local logger = require "rspamd_logger"
local sqlite3 = require "rspamd_sqlite3"
local util = require "rspamd_util"
local lua_redis = require "lua_redis"
local lua_util = require "lua_util"
local exports = {}

local N = "stat_tools" -- luacheck: ignore (maybe unused)

-- Performs synchronous conversion of redis schema
local function convert_bayes_schema(redis_params, symbol_spam, symbol_ham, expire)

  -- Old schema is the following one:
  -- Keys are named <symbol>[<user>]
  -- Elements are placed within hash:
  -- BAYES_SPAM -> {<id1>: <num_hits>, <id2>: <num_hits> ...}
  -- In new schema it is changed to a more extensible schema:
  -- Keys are named RS[<user>]_<id> -> {'H': <ham_hits>, 'S': <spam_hits>}
  -- So we can expire individual records, measure most popular elements by zranges,
  -- add new fields, such as tokens etc

  local res, conn = lua_redis.redis_connect_sync(redis_params, true)

  if not res then
    logger.errx("cannot connect to redis server")
    return false
  end

  -- KEYS[1]: key to check (e.g. 'BAYES_SPAM')
  -- KEYS[2]: hash key ('S' or 'H')
  -- KEYS[3]: expire
  local lua_script = [[
redis.replicate_commands()
local keys = redis.call('SMEMBERS', KEYS[1]..'_keys')
local nconverted = 0
for _,k in ipairs(keys) do
  local cursor = redis.call('HSCAN', k, 0)
  local neutral_prefix = string.gsub(k, KEYS[1], 'RS')
  local elts
  while cursor[1] ~= "0" do
    elts = cursor[2]
    cursor = redis.call('HSCAN', k, cursor[1])
    local real_key
    for i,v in ipairs(elts) do
      if i % 2 ~= 0 then
        real_key = v
      else
        local nkey = string.format('%s_%s', neutral_prefix, real_key)
        redis.call('HSET', nkey, KEYS[2], v)
        if KEYS[3] and tonumber(KEYS[3]) > 0 then
          redis.call('EXPIRE', nkey, KEYS[3])
        end
        nconverted = nconverted + 1
      end
    end
  end
end
return nconverted
]]

  conn:add_cmd('EVAL', { lua_script, '3', symbol_spam, 'S', tostring(expire) })
  local ret
  ret, res = conn:exec()

  if not ret then
    logger.errx('error converting symbol %s: %s', symbol_spam, res)
    return false
  else
    logger.messagex('converted %s elements from symbol %s', res, symbol_spam)
  end

  conn:add_cmd('EVAL', { lua_script, '3', symbol_ham, 'H', tostring(expire) })
  ret, res = conn:exec()

  if not ret then
    logger.errx('error converting symbol %s: %s', symbol_ham, res)
    return false
  else
    logger.messagex('converted %s elements from symbol %s', res, symbol_ham)
  end

  -- We can now convert metadata: set + learned + version
  -- KEYS[1]: key to check (e.g. 'BAYES_SPAM')
  -- KEYS[2]: learn key (e.g. 'learns_spam' or 'learns_ham')
  lua_script = [[
local keys = redis.call('SMEMBERS', KEYS[1]..'_keys')

for _,k in ipairs(keys) do
  local learns = redis.call('HGET', k, 'learns') or 0
  local neutral_prefix = string.gsub(k, KEYS[1], 'RS')

  redis.call('HSET', neutral_prefix, KEYS[2], learns)
  redis.call('SADD', KEYS[1]..'_keys', neutral_prefix)
  redis.call('SREM', KEYS[1]..'_keys', k)
  redis.call('DEL', KEYS[1])
  redis.call('SET', k ..'_version', '2')
end
]]

  conn:add_cmd('EVAL', { lua_script, '2', symbol_spam, 'learns_spam' })
  ret, res = conn:exec()

  if not ret then
    logger.errx('error converting metadata for symbol %s: %s', symbol_spam, res)
    return false
  end

  conn:add_cmd('EVAL', { lua_script, '2', symbol_ham, 'learns_ham' })
  ret, res = conn:exec()

  if not ret then
    logger.errx('error converting metadata for symbol %s', symbol_ham, res)
    return false
  end

  return true
end

exports.convert_bayes_schema = convert_bayes_schema

-- It now accepts both ham and spam databases
-- parameters:
-- redis_params - how do we connect to a redis server
-- sqlite_db_spam - name for sqlite database with spam tokens
-- sqlite_db_ham - name for sqlite database with ham tokens
-- symbol_ham - name for symbol representing spam, e.g. BAYES_SPAM
-- symbol_spam - name for symbol representing ham, e.g. BAYES_HAM
-- learn_cache_spam - name for sqlite database with spam learn cache
-- learn_cache_ham - name for sqlite database with ham learn cache
-- reset_previous - if true, then the old database is flushed (slow)
local function convert_sqlite_to_redis(redis_params,
                                       sqlite_db_spam, sqlite_db_ham, symbol_spam, symbol_ham,
                                       learn_cache_db, expire, reset_previous)
  local nusers = 0
  local lim = 1000 -- Update each 1000 tokens
  local users_map = {}
  local converted = 0

  local db_spam = sqlite3.open(sqlite_db_spam)
  if not db_spam then
    logger.errx('Cannot open source db: %s', sqlite_db_spam)
    return false
  end
  local db_ham = sqlite3.open(sqlite_db_ham)
  if not db_ham then
    logger.errx('Cannot open source db: %s', sqlite_db_ham)
    return false
  end

  local res, conn = lua_redis.redis_connect_sync(redis_params, true)

  if not res then
    logger.errx("cannot connect to redis server")
    return false
  end

  if reset_previous then
    -- Do a more complicated cleanup
    -- execute a lua script that cleans up data
    local script = [[
local members = redis.call('SMEMBERS', KEYS[1]..'_keys')

for _,prefix in ipairs(members) do
  local keys = redis.call('KEYS', prefix..'*')
  redis.call('DEL', keys)
end
]]
    -- Common keys
    for _, sym in ipairs({ symbol_spam, symbol_ham }) do
      logger.messagex('Cleaning up old data for %s', sym)
      conn:add_cmd('EVAL', { script, '1', sym })
      conn:exec()
      conn:add_cmd('DEL', { sym .. "_version" })
      conn:add_cmd('DEL', { sym .. "_keys" })
      conn:exec()
    end

    if learn_cache_db then
      -- Cleanup learned_cache
      logger.messagex('Cleaning up old data learned cache')
      conn:add_cmd('DEL', { "learned_ids" })
      conn:exec()
    end
  end

  local function convert_db(db, is_spam)
    -- Map users and languages
    local what = 'ham'
    if is_spam then
      what = 'spam'
    end

    local learns = {}
    db:sql('BEGIN;')
    -- Fill users mapping
    for row in db:rows('SELECT * FROM users;') do
      if row.id == '0' then
        users_map[row.id] = ''
      else
        users_map[row.id] = row.name
      end
      learns[row.id] = row.learns
      nusers = nusers + 1
    end

    -- Workaround for old databases
    for row in db:rows('SELECT * FROM languages') do
      if learns['0'] then
        learns['0'] = learns['0'] + row.learns
      else
        learns['0'] = row.learns
      end
    end

    local function send_batch(tokens, prefix)
      -- We use the new schema: RS[user]_token -> H=ham count
      --                                          S=spam count
      local hash_key = 'H'
      if is_spam then
        hash_key = 'S'
      end
      for _, tok in ipairs(tokens) do
        -- tok schema:
        -- tok[1] = token_id (uint64 represented as a string)
        -- tok[2] = token value (number)
        -- tok[3] = user_map[user_id] or ''
        local rkey = string.format('%s%s_%s', prefix, tok[3], tok[1])
        conn:add_cmd('HINCRBYFLOAT', { rkey, hash_key, tostring(tok[2]) })

        if expire and expire ~= 0 then
          conn:add_cmd('EXPIRE', { rkey, tostring(expire) })
        end
      end

      return conn:exec()
    end
    -- Fill tokens, sending data to redis each `lim` records

    local ntokens = db:query('SELECT count(*) as c FROM tokens')['c']
    local tokens = {}
    local num = 0
    local total = 0

    for row in db:rows('SELECT token,value,user FROM tokens;') do
      local user = ''
      if row.user ~= 0 and users_map[row.user] then
        user = users_map[row.user]
      end

      table.insert(tokens, { row.token, row.value, user })
      num = num + 1
      total = total + 1
      if num > lim then
        -- TODO: we use the default 'RS' prefix, it can be false in case of
        -- classifiers with labels
        local ret, err_str = send_batch(tokens, 'RS')
        if not ret then
          logger.errx('Cannot send tokens to the redis server: ' .. err_str)
          db:sql('COMMIT;')
          return false
        end

        num = 0
        tokens = {}
      end

      io.write(string.format('Processed batch %s: %s/%s\r', what, total, ntokens))
    end
    -- Last batch
    if #tokens > 0 then
      local ret, err_str = send_batch(tokens, 'RS')
      if not ret then
        logger.errx('Cannot send tokens to the redis server: ' .. err_str)
        db:sql('COMMIT;')
        return false
      end

      io.write(string.format('Processed batch %s: %s/%s\r', what, total, ntokens))
    end
    io.write('\n')

    converted = converted + total

    -- Close DB
    db:sql('COMMIT;')
    local symbol = symbol_ham
    local learns_elt = "learns_ham"

    if is_spam then
      symbol = symbol_spam
      learns_elt = "learns_spam"
    end

    for id, learned in pairs(learns) do
      local user = users_map[id]
      if not conn:add_cmd('HSET', { 'RS' .. user, learns_elt, learned }) then
        logger.errx('Cannot update learns for user: ' .. user)
        return false
      end
      if not conn:add_cmd('SADD', { symbol .. '_keys', 'RS' .. user }) then
        logger.errx('Cannot update learns for user: ' .. user)
        return false
      end
    end
    -- Set version
    conn:add_cmd('SET', { symbol .. '_version', '2' })
    return conn:exec()
  end

  logger.messagex('Convert spam tokens')
  if not convert_db(db_spam, true) then
    return false
  end

  logger.messagex('Convert ham tokens')
  if not convert_db(db_ham, false) then
    return false
  end

  if learn_cache_db then
    logger.messagex('Convert learned ids from %s', learn_cache_db)
    local db = sqlite3.open(learn_cache_db)
    local ret = true
    local total = 0

    if not db then
      logger.errx('Cannot open cache database: ' .. learn_cache_db)
      return false
    end

    db:sql('BEGIN;')

    for row in db:rows('SELECT * FROM learns;') do
      local is_spam
      local digest = tostring(util.encode_base32(row.digest))

      if row.flag == '0' then
        is_spam = '-1'
      else
        is_spam = '1'
      end

      if not conn:add_cmd('HSET', { 'learned_ids', digest, is_spam }) then
        logger.errx('Cannot add hash: ' .. digest)
        ret = false
      else
        total = total + 1
      end
    end
    db:sql('COMMIT;')

    if ret then
      conn:exec()
    end

    if ret then
      logger.messagex('Converted %s cached items from sqlite3 learned cache to redis',
          total)
    else
      logger.errx('Error occurred during sending data to redis')
    end
  end

  logger.messagex('Migrated %s tokens for %s users for symbols (%s, %s)',
      converted, nusers, symbol_spam, symbol_ham)
  return true
end

exports.convert_sqlite_to_redis = convert_sqlite_to_redis

-- Loads sqlite3 based classifiers and output data in form of array of objects:
-- [
--  {
--  symbol_spam = XXX
--  symbol_ham = YYY
--  db_spam = XXX.sqlite
--  db_ham = YYY.sqlite
--  learn_cache = ZZZ.sqlite
--  per_user = true/false
--  label = str
--  }
-- ]
local function load_sqlite_config(cfg)
  local result = {}

  local function parse_classifier(cls)
    local tbl = {}
    if cls.cache then
      local cache = cls.cache
      if cache.type == 'sqlite3' and (cache.file or cache.path) then
        tbl.learn_cache = (cache.file or cache.path)
      end
    end

    if cls.per_user then
      tbl.per_user = cls.per_user
    end

    if cls.label then
      tbl.label = cls.label
    end

    local statfiles = cls.statfile
    for _, stf in ipairs(statfiles) do
      local path = (stf.file or stf.path or stf.db or stf.dbname)
      local symbol = stf.symbol or 'undefined'

      if not path then
        logger.errx('no path defined for statfile %s', symbol)
      else

        local spam
        if stf.spam then
          spam = stf.spam
        else
          if string.match(symbol:upper(), 'SPAM') then
            spam = true
          else
            spam = false
          end
        end

        if spam then
          tbl.symbol_spam = symbol
          tbl.db_spam = path
        else
          tbl.symbol_ham = symbol
          tbl.db_ham = path
        end
      end
    end

    if tbl.symbol_spam and tbl.symbol_ham and tbl.db_ham and tbl.db_spam then
      table.insert(result, tbl)
    end
  end

  local classifier = cfg.classifier

  if classifier then
    if classifier[1] then
      for _, cls in ipairs(classifier) do
        if cls.bayes then
          cls = cls.bayes
        end
        if cls.backend and cls.backend == 'sqlite3' then
          parse_classifier(cls)
        end
      end
    else
      if classifier.bayes then
        classifier = classifier.bayes
        if classifier[1] then
          for _, cls in ipairs(classifier) do
            if cls.backend and cls.backend == 'sqlite3' then
              parse_classifier(cls)
            end
          end
        else
          if classifier.backend and classifier.backend == 'sqlite3' then
            parse_classifier(classifier)
          end
        end
      end
    end
  end

  return result
end

exports.load_sqlite_config = load_sqlite_config

-- A helper method that suggests a user how to configure Redis based
-- classifier based on the existing sqlite classifier
local function redis_classifier_from_sqlite(sqlite_classifier, expire)
  local result = {
    new_schema = true,
    backend = 'redis',
    cache = {
      backend = 'redis'
    },
    statfile = {
      [sqlite_classifier.symbol_spam] = {
        spam = true
      },
      [sqlite_classifier.symbol_ham] = {
        spam = false
      }
    }
  }

  if expire then
    result.expire = expire
  end

  return { classifier = { bayes = result } }
end

exports.redis_classifier_from_sqlite = redis_classifier_from_sqlite

-- Reads statistics config and return preprocessed table
local function process_stat_config(cfg)
  local opts_section = cfg:get_all_opt('options') or {}

  -- Check if we have a dedicated section for statistics
  if opts_section.statistics then
    opts_section = opts_section.statistics
  end

  -- Default
  local res_config = {
    classify_headers = {
      "User-Agent",
      "X-Mailer",
      "Content-Type",
      "X-MimeOLE",
      "Organization",
      "Organisation"
    },
    classify_images = true,
    classify_mime_info = true,
    classify_urls = true,
    classify_meta = true,
    classify_max_tlds = 10,
  }

  res_config = lua_util.override_defaults(res_config, opts_section)

  -- Postprocess classify_headers
  local classify_headers_parsed = {}

  for _, v in ipairs(res_config.classify_headers) do
    local s1, s2 = v:match("^([A-Z])[^%-]+%-([A-Z]).*$")

    local hname
    if s1 and s2 then
      hname = string.format('%s-%s', s1, s2)
    else
      s1 = v:match("^X%-([A-Z].*)$")

      if s1 then
        hname = string.format('x%s', s1:sub(1, 3):lower())
      else
        hname = string.format('%s', v:sub(1, 3):lower())
      end
    end

    if classify_headers_parsed[hname] then
      table.insert(classify_headers_parsed[hname], v)
    else
      classify_headers_parsed[hname] = { v }
    end
  end

  res_config.classify_headers_parsed = classify_headers_parsed

  return res_config
end

local function get_mime_stat_tokens(task, res, i)
  local parts = task:get_parts() or {}
  local seen_multipart = false
  local seen_plain = false
  local seen_html = false
  local empty_plain = false
  local empty_html = false
  local online_text = false

  for _, part in ipairs(parts) do
    local fname = part:get_filename()

    local sz = part:get_length()

    if sz > 0 then
      rawset(res, i, string.format("#ps:%d",
          math.floor(math.log(sz))))
      lua_util.debugm("bayes", task, "part size: %s",
          res[i])
      i = i + 1
    end

    if fname then
      rawset(res, i, "#f:" .. fname)
      i = i + 1

      lua_util.debugm("bayes", task, "added attachment: #f:%s",
          fname)
    end

    if part:is_text() then
      local tp = part:get_text()

      if tp:is_html() then
        seen_html = true

        if tp:get_length() == 0 then
          empty_html = true
        end
      else
        seen_plain = true

        if tp:get_length() == 0 then
          empty_plain = true
        end
      end

      if tp:get_lines_count() < 2 then
        online_text = true
      end

      rawset(res, i, "#lang:" .. (tp:get_language() or 'unk'))
      lua_util.debugm("bayes", task, "added language: %s",
          res[i])
      i = i + 1

      rawset(res, i, "#cs:" .. (tp:get_charset() or 'unk'))
      lua_util.debugm("bayes", task, "added charset: %s",
          res[i])
      i = i + 1

    elseif part:is_multipart() then
      seen_multipart = true;
    end
  end

  -- Create a special token depending on parts structure
  local st_tok = "#unk"
  if seen_multipart and seen_html and seen_plain then
    st_tok = '#mpth'
  end

  if seen_html and not seen_plain then
    st_tok = "#ho"
  end

  if seen_plain and not seen_html then
    st_tok = "#to"
  end

  local spec_tok = ""
  if online_text then
    spec_tok = "#ot"
  end

  if empty_plain then
    spec_tok = spec_tok .. "#ep"
  end

  if empty_html then
    spec_tok = spec_tok .. "#eh"
  end

  rawset(res, i, string.format("#m:%s%s", st_tok, spec_tok))
  lua_util.debugm("bayes", task, "added mime token: %s",
      res[i])
  i = i + 1

  return i
end

local function get_headers_stat_tokens(task, cf, res, i)
  --[[
  -- As discussed with Alexander Moisseev, this feature can skew statistics
  -- especially when learning is separated from scanning, so learning
  -- has a different set of tokens where this token can have too high weight
  local hdrs_cksum = task:get_mempool():get_variable("headers_hash")

  if hdrs_cksum then
    rawset(res, i, string.format("#hh:%s", hdrs_cksum:sub(1, 7)))
    lua_util.debugm("bayes", task, "added hdrs hash token: %s",
        res[i])
    i = i + 1
  end
  ]]--

  for k, hdrs in pairs(cf.classify_headers_parsed) do
    for _, hname in ipairs(hdrs) do
      local value = task:get_header(hname)

      if value then
        rawset(res, i, string.format("#h:%s:%s", k, value))
        lua_util.debugm("bayes", task, "added hdrs token: %s",
            res[i])
        i = i + 1
      end
    end
  end

  local from = (task:get_from('mime') or {})[1]

  if from and from.name then
    rawset(res, i, string.format("#F:%s", from.name))
    lua_util.debugm("bayes", task, "added from name token: %s",
        res[i])
    i = i + 1
  end

  return i
end

local function get_meta_stat_tokens(task, res, i)
  local day_and_hour = os.date('%u:%H',
      task:get_date { format = 'message', gmt = true })
  rawset(res, i, string.format("#dt:%s", day_and_hour))
  lua_util.debugm("bayes", task, "added day_of_week token: %s",
      res[i])
  i = i + 1

  local pol = {}

  -- Authentication results
  if task:has_symbol('DKIM_TRACE') then
    -- Autolearn or scan
    if task:has_symbol('R_SPF_ALLOW') then
      table.insert(pol, 's=pass')
    end

    local trace = task:get_symbol('DKIM_TRACE')
    local dkim_opts = trace[1]['options']
    if dkim_opts then
      for _, o in ipairs(dkim_opts) do
        local check_res = string.sub(o, -1)
        local domain = string.sub(o, 1, -3)

        if check_res == '+' then
          table.insert(pol, string.format('d=%s:%s', "pass", domain))
        end
      end
    end
  else
    -- Offline learn
    local aur = task:get_header('Authentication-Results')

    if aur then
      local spf = aur:match('spf=([a-z]+)')
      local dkim, dkim_domain = aur:match('dkim=([a-z]+) header.d=([a-z.%-]+)')

      if spf then
        table.insert(pol, 's=' .. spf)
      end
      if dkim and dkim_domain then
        table.insert(pol, string.format('d=%s:%s', dkim, dkim_domain))
      end
    end
  end

  if #pol > 0 then
    rawset(res, i, string.format("#aur:%s", table.concat(pol, ',')))
    lua_util.debugm("bayes", task, "added policies token: %s",
        res[i])
    i = i + 1
  end

  --[[
  -- Disabled.
  -- 1. Depending on the source the message has a different set of Received
  -- headers as the receiving MTA adds another Received header.
  -- 2. The usefulness of the Received tokens is questionable.
  local rh = task:get_received_headers()

  if rh and #rh > 0 then
    local lim = math.min(5, #rh)
    for j =1,lim do
      local rcvd = rh[j]
      local ip = rcvd.real_ip
      if ip and ip:is_valid() and ip:get_version() == 4 then
        local masked = ip:apply_mask(24)

        rawset(res, i, string.format("#rcv:%s:%s", tostring(masked),
            rcvd.proto))
        lua_util.debugm("bayes", task, "added received token: %s",
            res[i])
        i = i + 1
      end
    end
  end
  ]]--

  return i
end

local function get_stat_tokens(task, cf)
  local res = {}
  local E = {}
  local i = 1

  if cf.classify_images then
    local images = task:get_images() or E

    for _, img in ipairs(images) do
      rawset(res, i, "image")
      i = i + 1
      rawset(res, i, tostring(img:get_height()))
      i = i + 1
      rawset(res, i, tostring(img:get_width()))
      i = i + 1
      rawset(res, i, tostring(img:get_type()))
      i = i + 1

      local fname = img:get_filename()

      if fname then
        rawset(res, i, tostring(img:get_filename()))
        i = i + 1
      end

      lua_util.debugm("bayes", task, "added image: %s",
          fname)
    end
  end

  if cf.classify_mime_info then
    i = get_mime_stat_tokens(task, res, i)
  end

  if cf.classify_headers and #cf.classify_headers > 0 then
    i = get_headers_stat_tokens(task, cf, res, i)
  end

  if cf.classify_urls then
    local urls = lua_util.extract_specific_urls { task = task, limit = 5, esld_limit = 1 }

    if urls then
      for _, u in ipairs(urls) do
        rawset(res, i, string.format("#u:%s", u:get_tld()))
        lua_util.debugm("bayes", task, "added url token: %s",
            res[i])
        i = i + 1
      end
    end
  end

  if cf.classify_meta then
    i = get_meta_stat_tokens(task, res, i)
  end

  return res
end

exports.gen_stat_tokens = function(cfg)
  local stat_config = process_stat_config(cfg)

  return function(task)
    return get_stat_tokens(task, stat_config)
  end
end

return exports