Part I: Introduction

Why Rspamd? SpamAssassin → rspamd in “Rambler” Mail

Rspamd in actionReal-time stats

Rspamd in actionWeb UI

Rspamd in a nutshell• Written in plain C

• Uses non-blocking (event-driven) data processing model

• Oriented on performance

• Plugins and rules are written in Lua

• Supports all major filtering methods

• Packaged for the most of *nix systems

Spam filtering

Major spam classes• Fraud and scam:

• “Nigerian” mail

• Phishing (banks, social networks)

• Advertisement:

• Plain text

• Pictures, obfuscated HTML etc

• Spam from the social networks notifications

Scam:a typical example

Advertisement: images

Advertisement: images

Filtering methods Basic principles

• Policies (particularly efficient against fraud)

• DNS lists (for IP and URLs)

• SPF, DKIM, DMARC

• Sender’s reputation

• Static content filtering (patterns match)

• Statistics and ML (e.g. Bayesian)

Filtering methods Difficulties

• Policies:

• Require network requests with unpredictable delay

• False positive (e.g. when forwarding)

• Static Content filtering:

• CPU intensive

• Unlikely to be useful long-term

• Statistics:

• Not very precise

• Hard to adopt for multi-user systems

Rspamd special features • Text lemmatisation and normalisation (snowball stemmer)

• Rspamd converts all texts to utf-8

• Support for users settings in JSON

• Possibility to have multiple metrics for a message

• Web interface

• Rspamd can load SpamAssassin rules directly

• OpenSource (BSD 2-clause)

Rspamd integration Using Rmilter

MTA Rmilter Rspamd

Storage

Message and SMTP DATA HTTP request

Redis cache

JSON: action, symbolsDelay, modify or reject

Greylisting, limits Statistics, shared cache

✉✉✉✉

✉ ✉

🚮

Inbox Spam

✉Reject

✉

Rspamd integration milter

• rmilter + postfix/sendmail:

• supports all rspamd features (e.g. greylisting action)

• can split inbound and outbound messages

• can whitelist reply to own messages

• can add DKIM signatures to outgoing email

Rspamd integration Other MTA

• exim - supports rspamd internally (from 4.86)

• haraka - supports rspamd internally

• any other MTA - LDA mode when rspamc checks message on rspamd and calls delivery agent with altered message using Unix pipes

Part II: Architecture

Development principles

• Performance and scalability

• Filtering quality and extensions

• Security

• Simplicity and documentation coverage

Rules in Rspamd

Pre-filters Filters Post-filters

Rule Rule Rule Rule Rule Rule Rule Rule Rule Rule

Policy rules Content rules Statistics rules

Settings, ratelimit Main scan Postprocessing

Message lifecycle

RulesRulesFilters

RulesRulesPre-filters

RulesRulesPost-filters

📩 Message

📬 Result

Wait

Wait

Wait for dependencies

Event driven model Never blocks*

• Advantages:

✅ Can process network replies asynchronously

✅ Sends all network request simultaneously

✅ Can process many messages in parallel, scales well

• Disadvantages:

📛 Callbacks hell

⛔ Memory consumption can go wild

*almost never

Sequential processing Traditional approach

Rule 1 Rule 2 Rule 3

DNS Hashes

Wait Wait

Processing

Event-driven processing

Rule 1 Rule 2 Rule 3

DNS Hashes

Wait

Processing

Rule 1Rule 1Rules

DNS Hashes

Wait

Event-driven processing What really happens?

Event-driven processing• Rspamd can send thousands of DNS requests:

time: 5540.8ms real, 2427.4ms virtual, dns req: 120543

Event-driven processing

• Rspamd can send thousands of DNS requests: time: 5540.8ms real, 2427.4ms virtual, dns req: 120543

• For short messages waiting prevails on the processing time: time: 996.140ms real, 22.000ms virtual,

Statistics in rspamd• Uses Markov chains for statistic tokens (using

sparse bi-gramms model)

• Considers metadata in statistics: mime structure, headers and so on

• Statistics could be stored in redis or sqlite3

• Rspamd supports per-user, per-domain statistics, multiple classifiers and autolearning

Tokens generation Quick brown fox jumps over lazy dog

1234

1

2

Statistics example Hard case: images spam

0%

25%

50%

75%

100%

Spam Ham

8%5%

92%95%

Correctly recognized Unrecognized

Fuzzy hashes Introduction

• Can find similar or close to similar messages

• Rspamd uses fuzzy logic for text and exact hashes for attachments

• Storage engine - sqlite3

Shingles AlgorithmQuick brown fox jumps over lazy dog

w1 w2 w3

w2 w3 w4

w3 w4 w5

w4 w5 w6

w1 w2 w3

w2 w3 w4

w3 w4 w5

w4 w5 w6N hashes

h1

h2

h3

h4

h1’

h2’

h3’

h4’

Shingles Algorithmh1 h2 h3

h1’ h2’ h3’

h1’’ h2’’ h3’’

h1’’’’’ h2’’’’ h3’’’’

…

…

…

…

…

min

min

min

min

…

N shinglesN hash pipes

Shingles Algorithm

• Probabilistic in general

• Uses siphash with different keys

• Window size - 3 words, number of hashes - 32

Part III: Performance

Main Principles• Always consider performance.

• If something could be skipped, skip it

• Use more specific solution if it is faster than generic one (special FSM, asm fragments)

• If it is possible to do something fast but not 100% accurate, try faster method

Rules Optimisation Global optimisation

• Stop checks if a message has reached threshold

• Prioritise rules with negative weight

• Prioritise faster and more frequent rules:

• Check weight, hit probability and average time

• Use greedy sorting algorithm

Rules Optimisation Local optimisations

• Build Abstract Syntax Tree for each rule

• Tree is periodically resorted in respect of hit probability and execution time

• Regular expressions are optimised using PCRE+JIT and Hyperscan (from 1.1)

• Lua code is optimised by means of LuaJIT

AST Optimisations Branches cut

&

|C

! B

A

A = 0, B = 1, C = 0

0

110

1

0

Order of processing

4/6 branches skipped

(!A or B) and C

AST optimisations N-ary operations

>

+ 2

! B

A

Check order

C D E

What we compare?

Here is our limit

Stop processing

(!A+B+C+D+E) > 2

C Library optimisations• Accelerated base32/base64

• Optimised string functions

• Fixed size strings

• Own printf implementation

• Own logger layer

• Assembly versions for crypto ciphers and hashes

Hyperscan in Rspamd Multi-regexp match engine

Hyperscan in Rspamd Problem statement

• Need to match many regular expressions for the same data:

len: 610591, time: 2492.457ms real, 882.251ms virtual regexp statistics: 4095 pcre regexps scanned, 18 regexps matched, 694M bytes scanned using pcre

• No need to capture patterns

• Need to match both UTF8 and raw data

• Need to support Perl compatible syntax

Naive solution

• Join all expressions with ‘|’:

/abc/ + /cde/ => /(?:abc)|(?:cde)/

• Won’t work:

/abc/ + /a/ => /(?:abc)|(?:a)/ -> the second rule won’t match for ‘abc’

Hyperscan

Text

/re1/

/re2/

/re3/

/re4/

/re5/

/re6/

/re7/

Scan

Match

Match

Match

Rules and Hyperscan

From

Subject

Text Part

HTML Part

Attachment

Header RE classHeader RE class

Text Parts class

Raw Body class

Whole message

m1

m2

m3

m4

m5

m6

MIME message Hyperscan regexps

Problems with Hyperscan

• Slow compile time (like 100 times slower than PCRE+JIT)

• Many PCRE features are unsupported (lookahead, look-behind)

• Is not packaged for the major OSes (requires relatively fresh boost)

Lazy hyperscan compile

HS compiler

Main

Scanners

Compiled class

Compiled class

Compiled class

Use PCRE only

Compile to disk

Lazy hyperscan compile

HS compiler

Main

Scanners

Compiled class

Compiled class

Compiled class

Use PCRE only

All done Broadcast message

Lazy hyperscan compile

HS compiler

Main

Scanners

Compiled class

Compiled class

Compiled class

Switch to hyperscan

Loaded from cache

Wait for ‘recompile’ command

Lazy hyperscan compile

• Does not affect the starting time

• If nothing changes then scanners load pre-compiled expressions on startup (using crypto hashes)

• Recompile command is used to force recompilation process:rspamadm recompile

Unsupported features Pre-filter mode

Text

/re1/pre

/re2/

/re3/pre

/re4/

/re5/

/re6/

/re7/

Scan

Maybe match?

Match

Match

Check PCRE

Match!

Replaces unsupported constructs with "weaker" supported ones, guaranteeing a match superset:

Hyperscan resultsUse Hyperscan for the vast majority of RE rules:

len: 610591, time: 2492.457ms real, 882.251ms virtual regexp statistics: 4095 pcre regexps scanned, 18 regexps matched, 694M bytes scanned using pcre

len: 610591, time: 654.596ms real, 309.785ms virtual regexp statistics: 34 pcre regexps scanned, 41 regexps matched, 8.41M bytes scanned using pcre, 9.56M bytes scanned total

Part IV: Security

Main Principles • Use secure C programming guidelines

• Perform static analysis for the code including special clang plugin for own printf format safety

• Deal with malicious messages

• Limit and protect DNS subsystem

• Encrypt and protect all data passed over the network

Examples

• An email can contain thousands of objects to check (e.g. URLs)

• SPF records could have infinite recursion

• Need limits for network checks and memoization of the results

DNS protection

• Cryptographic DNS id generation

• Socket pool with constant rotation

• Source port randomisation

• Input filtering (+ IDN encoding)

Traffic Encryption

• Security -> performance -> simplicity

• TLS is hard in non-blocking IO

• TLS does a lot of unnecessary (in case of rspamd) operations increasing thus CPU load and latency

HTTPCrypt Protocol

Connection EstablishmentClientHelo

ServerHelo Certificates⚠

ClientChangeCipher

ServerChangeCipher

Data

TLS handshake

ClientPK Encrypted request

Encrypted reply Validation (optional)

HTTPCrypt handshake

Throughput Cached connections

Throughput New connections

Throughput New connections

Connection Latency Cached connections

Connection Latency New connections

Why HTTPCrypt is fast• Faster key exchange (ECDH)

• Server does not need to sign anything (needed for replay protection in TLS)

• Encryption is done in-place

• No replay protection for the first message - still suitable for idempotent scanning

Conclusions

Conclusions

• Rspamd is designed to work as fast as possible

• Rspamd has a rich and actively maintained collection of rules and plugins and can re-use existing SA rules

• It is possible to encrypt all data transferred over the network in rspamd ecosystem

Open issues

• No Debian maintainer (any volunteers?)

• No dynamic rules updates (planned for 1.2) yet

• Rapid development (hard to keep compatibility)

Questions?Vsevolod Stakhov

https://rspamd.com

Appendix

Rspamd processes

✉

Scanning processesScanning processesScanner processes Controller Service processes

Main process

✉✉

HTTP📬 ✉✉✉

Learn

Messages Results

Configuration evolution1. Grammar parser (lex + yacc)

⛔ Hard to manage

⛔ Hard to extend

2. XML

⛔ Unreadable

⛔ Problems with expressions (A > B)

3. UCL - universal configuration language

✅ Easy to manage (looks like nginx.conf)

✅ Macro support

✅ JSON data model (can be used as JSON parser)

UCL building blocks• Sections

• Arrays

• Variables

• Macros

• Comments

section { key = “value”; number = 10K; }

upstreams = [ “localhost:80”, “example.com:8080”,

]

static_dir = “${WWWDIR}/“; filepath = “${CURDIR}/data”;

.include “${CONFDIR}/workers.conf”

.include (glob=true,priority=2) “${CONFDIR}/conf.d/*.conf”

.lua { print(“hey!”); }

key = value; // Single line comment /* Multiline comment /* can also be nested */ */

Configuration components

Global options

Workers configuration

Scores (metrics)

Modules configuration

Statistics

• Each component is normally included to the main configuration

• rspamd.local.conf is used to extend configuration

• rspamd.override.conf is used to override values in the configuration

• It is possible to use numeric multipliers: “k/m/g” or “ms/s/m/h/d” for time values

Lua rules

• Most of rules are defined in Lua configuration

• Two types of Lua rules:

• Regexp rules (look like strings)

• Lua functions (pure Lua code)

Lua rules Some examples

-- Outlook versions that should be excluded from summary rulelocal fmo_excl_o3416 = 'X-Mailer=/^Microsoft Outlook, Build 10.0.3416$/H'local fmo_excl_oe3790 = 'X-Mailer=/^Microsoft Outlook Express 6.00.3790.3959$/H'-- Summary rule for forged outlookreconf['FORGED_MUA_OUTLOOK'] = string.format('(%s | %s) & !%s & !%s & !%s',

forged_oe, forged_outlook_dollars, fmo_excl_o3416, fmo_excl_oe3790, vista_msgid)

• Regexp rule

• Lua rulerspamd_config.R_EMPTY_IMAGE = function(task) local tp = task:get_text_parts() -- get text parts in a message for _,p in ipairs(tp) do -- iterate over text parts array using `ipairs` if p:is_html() then -- if the current part is html part local hc = p:get_html() -- we get HTML context local len = p:get_length() -- and part's length if len < 50 then -- if we have a part that has less than 50 bytes of text local images = hc:get_images() -- then we check for HTML images if images then -- if there are images for _,i in ipairs(images) do -- then iterate over images in the part if i['height'] + i['width'] >= 400 then -- if we have a large image return true -- add symbol end end end end end endend

Pure LUA functions Review

• Are very powerful

• Have access to all information from rspamd via lua API: https://rspamd.com/doc/lua/

• Are very fast since C <-> LUA interaction is cheap

• Can use zero-copy objects called rspamd{text} to avoid copying when moving data between C and LUA

Pure LUA functions• Variables:

• Conditionals:

• Loops:

• Tables:

• Functions:

• Closures:

local ret = false -- Generic variablelocal rules = {} -- Empty tablelocal rspamd_logger = require “rspamd_logger" -- Load rspamd module

if not ret then -- can use ‘not’, ‘and’, ‘or’ here…elseif ret ~= 10 then -- note ~= for ‘not equal’ operatorend

for k,m in pairs(opts) do … end -- Iterate over keyed table a[‘key’] = valuefor _,i in ipairs(images) do … end -- Iterate over array table a[1] = valuefor i=1,10 do … end -- Count from 1 to 10

local options = { [1] = ‘value’, [‘key’] = 1, -- Numbers starts from 1 another_key = function(task) … end, -- Functions can be values [2] = {} -- Other tables can be values} -- Can have both numbers and strings as key and anything as values

local function something(task) -- Normal definition local cb = function(data) -- Functions can be nested … endend

local function gen_closure(option) local ret = false -- Local variable return function(task) task:do_something(ret, option) -- Both ‘ret’ and ‘option’ are accessible here endendrspamd_config.SYMBOL = gen_closure(‘some_option’)

Pure LUA functions Generic recommendations

• Use local whenever possible (otherwise, global variables are expensive)

• Callbacks, closures and recursion are generally cheap (when using LuaJIT)

• Do not mix string and number keys in tables, that makes them hard to iterate

• ipairs and pairs are not equal

• Strings are constant in LUA

Regexp rules Types

• Can work with the following elements:

• Headers: Message-Id=/^something$/H

• Mime parts: /some word/P

• Raw messages: /some pattern/M

• URLs: /example.com/U

• Some new flags are added:

• UTF8 flag: /u

Regexp rules Generic information

• Can be combined using the following operators:

• AND: /something/P && Subject=/some/H

• OR: /something/P || Subject=/some/H

• NOT: !/something/P

• PLUS: /A/P + /B/P + /C/P >= 2

• Priority goes as following: NOT ➡ AND ➡ OR ➡ PLUS

• Braces can change priority: !A AND (B OR C)

Regexp rules Performance considerations

• Avoid message regexps at any cost (use trie instead)

• Regexp expressions are highly optimised in rspamd and unnecessary evaluations are not performed

• UTF regexps are more expensive than default ones (but could be useful sometimes)

• Always use the appropriate type of expression (e.g. url for links and part for textual content)

Trie matching• Perfect for fast raw message and text pattern

matching

• Scales almost linearly with input size (aho-corasic algorithm)

• Can handle thousands and hundreds of thousands of patterns (is a base for all antivirus scanners)

• Highly optimised for 64 bit systems