nginx.conf.txt

# -*- awk -*-
# eotk (c) 2017 Alec Muffett

# EMACS awk mode works quite well for nginx configs

# ---- BEGIN HARD/CLASSIC SWITCH ----
%%IF %HARD_MODE%
# *HARD* configuration
# swap domain names for onions via brute-force, with whitelisted repairs...
%%ELSE
# *CLASSIC* configuration
# swap domain names for onions via targeted regular expressions...
%%ENDIF
# ---- END HARD/CLASSIC SWITCH ----

# logs and pids
pid %PROJECT_DIR%/nginx.pid;
error_log %LOG_DIR%/nginx-error.log %NGINX_SYSLOG%;

# TODO: notes for custom 403 error-handling pages:
# https://www.cyberciti.biz/faq/unix-linux-nginx-custom-error-403-page-configuration/
# https://nginx.org/en/docs/http/ngx_http_core_module.html#error_page

# performance
%%IF %IS_SOFTMAP%
worker_processes %SOFTMAP_NGINX_WORKERS%; # softmap
%%ELSE
worker_processes %NGINX_WORKERS%; # hardmap
%%ENDIF
worker_rlimit_nofile %NGINX_RLIM%;
events {
  worker_connections %NGINX_RLIM%;
}

http {
  # nginx fails without large enough buckets (sigh)
  map_hash_bucket_size %NGINX_HASH_BUCKET_SIZE%;

  # dns for proxy (sigh)
  resolver %NGINX_RESOLVER% valid=%NGINX_TIMEOUT%s;
  resolver_timeout %NGINX_TIMEOUT%s;

  # we walk a line between keeping it small and flooding resources...
  proxy_buffering on;

  # for initial; impacts SSL header
  proxy_buffer_size %NGINX_BLOCK_SIZE%;

  # for rest of response
  proxy_buffers %NGINX_BLOCK_COUNT% %NGINX_BLOCK_SIZE%;

  # how much can be busy sending to client?
  proxy_busy_buffers_size %NGINX_BLOCK_BUSY_SIZE%;

  # where to stash oversize requests?
  client_body_temp_path /tmp/nginx-body-%PROJECT%;
  client_max_body_size 4m;

  # in case we want to start spooling responses locally
  proxy_temp_path /tmp/nginx-proxy-%PROJECT%;
  proxy_max_temp_file_size %NGINX_TMPFILE_SIZE%;
  proxy_temp_file_write_size %NGINX_BLOCK_SIZE%;

  %%IF %NGINX_CACHE_SECONDS%
  # nginx caching static responses for %NGINX_CACHE_SECONDS% seconds
  # - this is a lightweight cache to reduce "storms", hence the global
  # approch of "cache everything for a small number of seconds"
  # https://nginx.org/en/docs/http/ngx_http_proxy_module.html
  proxy_cache_path /tmp/nginx-cache-%PROJECT% levels=1:2 keys_zone=%PROJECT%:%NGINX_CACHE_SIZE%;
  proxy_cache %PROJECT%;
  proxy_cache_min_uses %NGINX_CACHE_MIN_USES%;
  proxy_cache_revalidate on;
  proxy_cache_use_stale timeout updating;
  proxy_cache_valid any %NGINX_CACHE_SECONDS%s; # "any" includes 404s, etc

  # content-types to not cache
  map $http_content_type $no_cache_content_type {
    %%CSV %NO_CACHE_CONTENT_TYPE%
    %1% 1;
    %%ENDCSV
    default 0;
  }

  # hosts not to cache
  map $http_host $no_cache_host {
    hostnames;
    %%CSV %NO_CACHE_HOST%
    %1% 1;
    %%ENDCSV
    default 0;
  }

  # so, should we skip caching this stuff for some reason?
  proxy_no_cache $no_cache_content_type $no_cache_host;
  proxy_cache_bypass $no_cache_content_type $no_cache_host;
  %%ELSE
  # nginx caching disabled
  %%ENDIF

  # logs
  access_log %LOG_DIR%/nginx-access.log;

  # global settings
  server_tokens off;

  # allow/deny (first wins)
  allow "unix:";
  deny all;

  # rewrite these content types; text/html is implicit
  subs_filter_types
  application/javascript
  application/json
  application/x-javascript
  text/css
  text/javascript
  text/xml
  %%IF %EXTRA_SUBS_FILTER_TYPES%
  # extra_subs_filter_types
  %EXTRA_SUBS_FILTER_TYPES%
  %%ELSE
  # no extra_subs_filter_types
  %%ENDIF
  ;

  #==================================================================
  %%IF %HARD_MODE%
  # ---- BEGIN HARD MODE CODE ----
  %%IF %PRESERVE_CSV%
  # preserve subs (save-phase): 1=description,2=re,3=i_or_empty,4=replacement
  %%CSV %PRESERVE_CSV%
  # saving regexp '%2%' as '%1%' for replacement with '%4%' (%3%)
  subs_filter
  (%PRESERVE_PREAMBLE%)(%2%)\\b
  $1%PRESERVE_BEFORE%%1%%PRESERVE_AFTER%
  g%3%r
  ;
  %%ENDCSV
  %%ELSE
  # no preserve subs (save-phase)
  %%ENDIF

  %%BEGIN
  # HARD-MODE: %DNS_DOMAIN% -> %ONION_ADDRESS%
  subs_filter
  \\b%DNS_DOMAIN_RE2%\\b
  %ONION_ADDRESS%
  gir
  ;
  %%IF %HARD_MODE% > 1
  # HARD-MODE-EXTRA: %DNS_DOMAIN_RE% -> %ONION_ADDRESS_RE%
  subs_filter
  \\b%DNS_DOMAIN_RERE2%\\b
  %ONION_ADDRESS_RE2%
  gir
  ;
  %%ENDIF
  %%END

  %%IF %FOREIGNMAP_CSV%
  # foreignmap subs: 1=onion,2=re,3=re2,4=dns,5=re,6=re2
  %%CSV %FOREIGNMAP_CSV%
  # for %4% -> %1%
  subs_filter
  \\b%6%\\b
  %1%
  gir
  ;
  %%ENDCSV
  %%ELSE
  # no foreignmap subs
  %%ENDIF

  %%IF %PRESERVE_CSV%
  # preserve subs (restore-phase): 1=description,2=re,3=i_or_empty,4=replacement
  %%CSV %PRESERVE_CSV%
  # restoring '%1%' with '%4%'
  subs_filter
  %PRESERVE_BEFORE%%1%%PRESERVE_AFTER%
  %4%
  g
  ;
  %%ENDCSV
  %%ELSE
  # no preserve subs (restore-phase)
  %%ENDIF
  # ---- END HARD MODE CODE ----
  #------------------------------------------------------------------
  %%ELSE
  #------------------------------------------------------------------
  # ---- BEGIN CLASSIC MODE CODE ----

  # subs_filter: these patterns bear some explanation; the goal is to
  # work regular expressions really hard in order to minimise the
  # number of expressions which are used in the basic config, so the
  # basic pattern is to capture zero/more "sub." in "//sub.foo.com"
  # and interpolate that into "//sub.xxxxxxxx.onion"; so far?

  # but it turns out that some JSON libraries like to "escape" the
  # forward slashes in JSON content, leading to input like (literal)
  # "http:\/\/sub.foo.com\/foo.html" - so you need to add the
  # backslashes, but then you need to escape the backslashes, except
  # they need double-escaping in the regexp because of string
  # interpolation; hence 4x backslash -> 1x matched character

  # likewise we use the "_RE2" form of the re-escaped domain name in
  # order to coerce the regexp to match literal dots, not wildcards.

  # there seems to be some sort of shortcut at play here; the trailing
  # "\\b" also seems to work as "\b" however that would apparently
  # break the double-escaping that is necessary/works everywhere else
  # in subs_filter.

  # also, regrettably, named capture groups appear not to work, we're
  # fortunate that there appear not to be more than 9 capture groups
  # by default, lest "$1" bleed into the subsequent digits of an onion
  # address: $1234567abcdefghij.onion

  # finally: some sites encode // with %-encoded "2F" in URIs...

  %%BEGIN
  # for %DNS_DOMAIN% -> %ONION_ADDRESS% anchored by // or \/\/
  subs_filter
  (/|\\\\/\\\\)/(([-0-9a-z]+\\.)+)?%DNS_DOMAIN_RE2%\\b
  $1/$2%ONION_ADDRESS%
  gir
  ;
  # for %DNS_DOMAIN% -> %ONION_ADDRESS% anchored with hex-encoded slashes
  subs_filter
  %%2F%%2F(([-0-9a-z]+\\.)+)?%DNS_DOMAIN_RE2%\\b
  %%2F%%2F$1%ONION_ADDRESS%
  gir
  ;
  %%END

  %%IF %FOREIGNMAP_CSV%
  # foreignmap subs: 1=onion,2=re,3=re2,4=dns,5=re,6=re2
  %%CSV %FOREIGNMAP_CSV%
  # for %4% -> %1% anchored by // or \/\/
  subs_filter
  (/|\\\\/\\\\)/(([-0-9a-z]+\\.)+)?%6%\\b
  $1/$2%1%
  gir
  ;
  # for %4% -> %1% anchored with hex-encoded slashes
  subs_filter
  %%2F%%2F(([-0-9a-z]+\\.)+)?%6%\\b
  %%2F%%2F$1%1%
  gir
  ;
  %%ENDCSV
  %%ELSE
  # no foreignmap subs
  %%ENDIF
  # ---- END CLASSIC MODE CODE ----
  %%ENDIF
  #==================================================================

  # o_to_d_lookup -> if cannot remap, return input.  note: old versions
  # of lua-plugin cannot cope with code like o_to_d_mappings[o[1]]
  # because of `long bracket syntax`; the `[o[` freaks it out.
  # See: https://github.com/openresty/lua-nginx-module/issues/748
  init_by_lua_block {
    -- helper functions for elsewhere

    slog = function (s) -- in case of manual debugging
      ngx.log(ngx.ERR, s)
      return
    end

    has_suffix = function (s, x)
      return string.sub(s, -string.len(x)) == x
    end

    -- mapping onions to dns

    o_to_d_mappings = {}
    %%BEGIN
    o_to_d_mappings["%ONION_ADDRESS%"] = "%DNS_DOMAIN%"
    %%END

    o_to_d_lookup = function (m)
      local k = m[1] -- see note above re: array syntax
      return ( o_to_d_mappings[k] or k )
    end

    onion_to_dns = function (i)
      if i == nil or i == "" then
        return i
      end
      if (type(i) == "table") then
        local j, k, result
        result = {}
        for j, k in ipairs(i) do
          table.insert(result, onion_to_dns(k))
        end
        return result
      end
      local o, num, errs = ngx.re.gsub(i, "\\b([a-z2-7]{16}(?:[a-z2-7]{40})?\\.onion)\\b", o_to_d_lookup, "io")
      return o
    end

    -- mapping dns to onions, for experimentation

    d_to_o_mappings = {}
    %%BEGIN
    d_to_o_mappings["%DNS_DOMAIN%"] = "%ONION_ADDRESS%"
    %%END

    d_to_o_lookup = function (m)
      local k = m[1] -- see note above re: array syntax
      return ( d_to_o_mappings[k] or k )
    end

    dns_to_onion = function (i)
      if i == nil or i == "" or i == "*" then
        return i
      end
      if (type(i) == "table") then
        local j, k, result
        result = {}
        for j, k in ipairs(i) do
          table.insert(result, dns_to_onion(k))
        end
        return result
      end
      local num, errs
      %%BEGIN
      i, num, errs = ngx.re.gsub(i, "\\b%DNS_DOMAIN_RE2%\\b", "%ONION_ADDRESS%", "io")
      %%END
      return i
    end

    -- a note for future maintainers; if we were being strictly orthogonal then
    -- the replacement with ONION_ADDRESS in much of this Lua block would have to
    -- be double-escaped for potential backslashes, because double-quotes;
    -- however this is not needed because DNS forbids backslash; the only code
    -- where this becomes evident/necessary is here, with "_RE2":

    dnsre_to_onionre = function (i)
      local num, errs
      -- TODO: BRING THIS INTO LINE WITH dns_to_onion --
      %%BEGIN
      i, num, errs = ngx.re.gsub(i, "\\b%DNS_DOMAIN_RERE2%\\b", "%ONION_ADDRESS_RE2%", "io")
      %%END
      return i
    end
  }

  # filter the response headers en-route back to the user
  header_filter_by_lua_block {
    local k, v
    -- ==================================================================
    %%IF ! %HARD_MODE%
    -- ---- BEGIN CLASSIC MODE CODE ----
    -- is this javascript/json? if so, extra processing:
    -- 1) set a processing flag to pick up in body_filter_by_lua_block
    -- 2) invalidate content-length, because we will change it
    k = "Content-Type"
    v = ngx.header[k]
    if v == "application/javascript" or
      v == "application/json" or
      v == "application/x-javascript" or
      v == "text/css" or
      v == "text/javascript" then
      ngx.ctx.needs_extra_processing = 1
      ngx.header.content_length = nil
    end
    %%IF %EXTRA_PROCESSING_CSV%
    -- run on `v` for further extra_processing_csv checks
    %%CSV %EXTRA_PROCESSING_CSV%
    if v == "%1%" then
    local m, err = ngx.re.match(ngx.var.uri, "%2%", "io")
      if m then
        ngx.ctx.needs_extra_processing = 1
        ngx.header.content_length = nil
      end
    end
    %%ENDCSV
    %%ELSE
    -- no extra_processing_csv checks
    %%ENDIF
    -- ---- END CLASSIC MODE CODE ----
    %%ENDIF
    -- ==================================================================

    local origin_rewrites = {
      "Access-Control-Allow-Origin",
      %%IF %SUPPRESS_HEADER_CSP%
      -- CSP headers are suppressed via SUPPRESS_HEADER_CSP
      %%ELSE
      "Content-Security-Policy",
      "Content-Security-Policy-Report-Only",
      %%ENDIF
      "Link",
      "Location",
      "Set-Cookie"
    }

    local i, k
    for i, k in ipairs(origin_rewrites) do
      local v = ngx.header[k]
      if v then
        ngx.header[k] = dns_to_onion(v)
      end
    end
  }

  # filter the response body en-route back to the user
  body_filter_by_lua_block {
    -- ==================================================================
    %%IF ! %HARD_MODE%
    -- ---- BEGIN CLASSIC MODE CODE ----
    -- rather than blindly replacing "foo.com" with "foo.onion" everywhere,
    -- instead we restrict such brute-force replacement to content that was
    -- flagged in header_filter_by_lua_block
    if ngx.ctx.needs_extra_processing == 1 then
      -- the flag was set; this content deserves brute-force search & replace
      local chunk = ngx.arg[1]
      -- subs_filter picked up the "//"-anchored strings; now we sub the rest
      chunk = dns_to_onion(chunk)
      -- and we sub the basic "foo\.com" regular-expressions, too
      chunk = dnsre_to_onionre(chunk)
      -- more complex regular expressions are out of scope.
      ngx.arg[1] = chunk
    end
    -- ---- END CLASSIC MODE CODE ----
    %%ENDIF
    -- ==================================================================

    %%IF %DEBUG_TRAP%
    -- debug traps
    local i = ngx.arg[1]
    local ct = ngx.header["Content-Type"]
    local uri = ngx.var.uri
    local iterator, err
    %%CSV %DEBUG_TRAP%
    iterator, err = ngx.re.gmatch(i, ".{0,32}(%1%).{0,32}")
    if not iterator then
      ngx.log(ngx.ERR, "gmatch error: ", err)
    else
      while true do
        local m, err = iterator()
        if err then
          ngx.log(ngx.ERR, "iterator error: ", err)
          break
        end
        if not m then
          break
        end
        slog(string.format("TRAP <<%s>> CONTEXT <<%s>> TYPE <<%s>> URI %s", m[1], m[0], ct, uri))
      end -- while true
    end -- if iterator
    %%ENDCSV
    %%ELSE
    -- no debug traps
    %%ENDIF
  }

  %%IF %SUPPRESS_HEADER_CSP%
  # csp suppression
  proxy_hide_header "Content-Security-Policy";
  proxy_hide_header "Content-Security-Policy-Report-Only";
  %%ELSE
  # csp not suppressed, will be rewritten instead, see below
  %%ENDIF

  %%IF %SUPPRESS_HEADER_HSTS%
  # hsts suppression
  proxy_hide_header "Strict-Transport-Security";
  %%ELSE
  # hsts not suppressed
  %%ENDIF

  %%IF %SUPPRESS_HEADER_HPKP%
  # hpkp suppression
  proxy_hide_header "Public-Key-Pins";
  proxy_hide_header "Public-Key-Pins-Report-Only";
  %%ELSE
  # hpkp not suppressed
  %%ENDIF

  # global proxy settings
  proxy_read_timeout %NGINX_TIMEOUT%;
  proxy_connect_timeout %NGINX_TIMEOUT%;

  # SSL config
  ssl_certificate %SSL_DIR%/%CERT_PREFIX%.cert;
  ssl_certificate_key %SSL_DIR%/%CERT_PREFIX%.pem;
  ssl_buffer_size 4k;
  #ssl_ciphers 'EECDH+CHACHA20:EECDH+AESGCM:EECDH+AES256'; ## LibreSSL, OpenSSL 1.1.0+
  ssl_ciphers 'EECDH+AESGCM:EECDH+AES256'; ## OpenSSL 1.0.1% to 1.0.2%
  ssl_ecdh_curve prime256v1;
  #ssl_ecdh_curve secp384r1:prime256v1; ## NGINX nginx 1.11.0 and later
  ssl_prefer_server_ciphers on;
  ssl_protocols TLSv1 TLSv1.1 TLSv1.2;
  ssl_session_cache shared:SSL:10m;
  ssl_session_timeout 10m;

  # websockets: on the basis of http_upgrade, set connection_upgrade:
  # empty -> empty
  # default -> "upgrade"
  map $http_upgrade $connection_upgrade {
    default "upgrade";
    "" "";
  }

  %%BEGIN
  %%IF %FORCE_HTTPS%
  # FORCE_HTTPS is in use; set up separate server for port 80 & force redirects
  server {
    %%IF %IS_SOFTMAP%
    %%RANGE I 1 %SOFTMAP_TOR_WORKERS%
    listen unix:%PROJECT_DIR%/%TOR_WORKER_PREFIX%-%I%.d/port-80.sock;
    %%ENDRANGE
    %%ELSE
    listen unix:%PROJECT_DIR%/%ONION_ADDRESS%.d/port-80.sock;
    %%ENDIF

    # subdomain regexp captures trailing dot, use carefully; does not need "~*"
    # NB: this regexp should be kept in-sync with the other FORCE_HTTPS copy
    server_name
    %ONION_ADDRESS%
    ~^(?<servernamesubdomain>([-0-9a-z]+\\.)+)%ONION_ADDRESS_RE2%$
    ;

    %%IF %SUPPRESS_TOR2WEB%
    # suppress tor2web traffic; "let them use clearnet"
    if ( $http_x_tor2web ) {
      return 403 "%BLOCK_ERR%";
    }
    %%ELSE
    # tor2web not suppressed
    %%ENDIF

    # tell the client to try again as HTTPS without ever leaving the onion
    # use 307 / temporary redirect because your URIs may change in future
    # use $host (not $server) to copy-over subdomains, etc, transparently
    # SEND BACK ORIGINAL PARAMS, FIX THEM ONLY UPON FORWARD TO THE PROXY.
    return 307 https://$host$request_uri;
  }
  %%ELSE
  # FORCE_HTTPS is not in use, cleartext data may traverse the internet
  %%ENDIF

  # for %ONION_ADDRESS% -> %DNS_DOMAIN%
  server {
    %%IF %IS_SOFTMAP%
    %%RANGE I 1 %SOFTMAP_TOR_WORKERS%
    %%IF not %FORCE_HTTPS%
    # FORCE_HTTPS is not in use, cleartext data may traverse the internet
    listen unix:%PROJECT_DIR%/%TOR_WORKER_PREFIX%-%I%.d/port-80.sock;
    %%ENDIF
    listen unix:%PROJECT_DIR%/%TOR_WORKER_PREFIX%-%I%.d/port-443.sock ssl;
    %%ENDRANGE
    %%ELSE
    # hardmap
    # unix sockets; use <ONION_ADDRESS>.d as a naming convention
    %%IF not %FORCE_HTTPS%
    # FORCE_HTTPS is not in use, cleartext data may traverse the internet
    listen unix:%PROJECT_DIR%/%ONION_ADDRESS%.d/port-80.sock;
    %%ENDIF
    listen unix:%PROJECT_DIR%/%ONION_ADDRESS%.d/port-443.sock ssl;
    %%ENDIF

    # subdomain regexp captures trailing dot, use carefully; does not need "~*"
    # NB: this regexp should be kept in-sync with the other FORCE_HTTPS copy
    server_name
    %ONION_ADDRESS%
    ~^(?<servernamesubdomain>([-0-9a-z]+\\.)+)%ONION_ADDRESS_RE2%$
    ;

    %%INCLUDE templates.d/nginx-generated-blocks.conf

    %%IF %COOKIE_LOCK%
    # if we are visiting the magic path, open the cookie-lock
    location "%COOKIE_LOCK%" {
      add_header Set-Cookie "eotk_lock=%COOKIE_LOCK%;Domain=.%ONION_ADDRESS%;Path=/;Max-Age=604800";
      return 200 "OK";
    }
    %%ELSE
    # no cookie_lock cookie setting
    %%ENDIF

    %%IF %NGINX_HELLO_ONION%
    # for test & to help SSL certificate acceptance
    location ~* ^/hello[-_]onion/?$ {
      return 200 "Hello, Onion User!";
    }
    %%ELSE
    # no "hello-onion" endpoint
    %%ENDIF

    %%IF %HARDCODED_ENDPOINT_CSV%
    # hardcoded_endpoints: 1=path_re,2=response
    %%CSV %HARDCODED_ENDPOINT_CSV%
    location "%1%" {
      return 200 %2%;
    }
    %%ENDCSV
    %%ELSE
    # no hardcoded_endpoints
    %%ENDIF

    # for traffic
    location / {
      %%INCLUDE templates.d/nginx-generated-checks.conf

      %%IF %COOKIE_LOCK%
      # check for cookie-lock
      if ( $cookie_eotk_lock != "%COOKIE_LOCK%" ) { %NGINX_ACTION_ABORT%; }
      %%ELSE
      # no cookie-lock checks
      %%ENDIF

      # deonionify the request_uri for forwarding (both path and args)
      set_by_lua_block $request_uri2 {
        local old = ngx.var.request_uri
        -- onion_to_dns is potentially expensive at scale, so do a cheap test
        local m, err = ngx.re.match(old, "\\b[a-z2-7]{16}(?:[a-z2-7]{40})?\\.onion\\b", "o")
        if not m then -- nothing to attempt to rewrite, quick return
          return old
        end
        return onion_to_dns(old)
      }

      # note use of both $scheme and the deonionified uri (both path and args)
      set $new_url "$scheme://${servernamesubdomain}%DNS_DOMAIN%$request_uri2";
      proxy_pass $new_url;
      proxy_http_version 1.1;

      # a note on proxy_set_header, add_header, similar methods, etc;
      # if you override *any* header then you will lose the other
      # headers inherited from the parent contexts:
      # https://blog.g3rt.nl/nginx-add_header-pitfall.html

      proxy_set_header X-From-Onion %X_FROM_ONION_VALUE%;
      proxy_set_header Host "${servernamesubdomain}%DNS_DOMAIN%";
      proxy_set_header Accept-Encoding "identity";
      proxy_set_header Connection $connection_upgrade; # SSL
      proxy_set_header Upgrade $http_upgrade; # SSL
      proxy_ssl_server_name on; # SSL

      # rewrite request referer
      set_by_lua_block $referer2 { return onion_to_dns(ngx.var.http_referer) }
      proxy_set_header Referer $referer2;

      # rewrite request origin
      set_by_lua_block $origin2 { return onion_to_dns(ngx.var.http_origin) }
      proxy_set_header Origin $origin2;

      # rewrite request cookies
      set_by_lua_block $cookie2 { return onion_to_dns(ngx.var.http_cookie) }
      proxy_set_header Cookie $cookie2;

      %%IF %SUPPRESS_METHODS_EXCEPT_GET%
      # suppress non-GET methods (e.g.: POST)
      limit_except GET {
        deny all;
      }
      %%ELSE
      # non-GET methods (e.g.: POST) are not suppressed
      %%ENDIF
    }
  }

  %%END

  # header purge
  more_clear_headers "Age";
  more_clear_headers "Server";
  more_clear_headers "Via";
  more_clear_headers "X-From-Nginx";
  more_clear_headers "X-NA";
  more_clear_headers "X-Powered-By";
  more_clear_headers "X-Request-Id";
  more_clear_headers "X-Runtime";
  more_clear_headers "X-Varnish";
}