Redwood is an internet content-filtering program. It adds flexibility and granularity to the filtering by classifying sites into multiple categories instead of just “Allow” and “Block.”

Basic Architecture

Redwood runs as an HTTP proxy server. It examines each HTTP message to determine if it should be allowed to proceed. If so, it passes the message on to its final destination. If not, it replaces the message with a customizable page stating that the request is not allowed (optionally giving the reason and providing a link for filing an overblock request).

Redwood’s filtering is based on URLs and also, where applicable, on page content.

Configuration File

By default, the main configuration file is located at /etc/redwood/redwood.conf. This path can be changed by using the -c command line switch. Configuration options may be specified either in the configuration file or as command-line switches. In the configuration file, they may be specified either as key = value or as key value. Comments are delimited with #. Values may be enclosed in double quotes, with the usual backslash escapes. Additional configuration files may be included by using the include directive.

An example configuration file:

# Listen for connections on port 8000.
http-proxy :8000

# the template for the block page
blockpage "/etc/redwood/block.html"

# directory of static files to be served by the internal web server
static-files-dir /etc/redwood/static

# directory of CGI scripts to run by the internal web server
cgi-bin /etc/redwood/cgi

# the directory containing the category information
categories /etc/redwood/categories

# the file containing the Access Control List configuration
acls /etc/redwood/acls.conf

# the minimum total score from a blocked category needed
# to block a page
threshold 275

# file configuring the content pruning
content-pruning /etc/redwood/pruning.conf

# file configuring URL query modification
query-changes /etc/redwood/safesearch.conf

# path to the access log
access-log /var/log/redwood/access.log

Categories

The configuration files for Redwood allow the user to establish any number of categories corresponding to the types of content that he wishes to block or to allow. As each HTTP message is processed, it is assigned a score in each category, based on the filter lists that are set up for that category. These scores are then used to determine whether the page should be blocked.

Each category is a assigned an action: allow, block, or ignore. A page will be blocked if the score for any category listed as block is higher than the highest score for any category listed as allow. If a category is listed as ignore, its score does not affect whether a page is blocked or not. However, a page is not blocked unless the score for the highest block category is greater than a certain configurable threshold. This prevents overblocks of pages with almost no textual content.

A category's action may also be set to acl. Then the category is ignored in the process of finding the top-scoring category for the page, but it is available for ACLs to act on it, whenever the page's score in that category is greater than zero.

The categories are stored in a directory whose location is specified in the configuration file. Each subdirectory of that directory defines a category (with the same name as the directory).

Each category’s directory contains a file named category.conf and any number of rule-list files. A category named “mechanical” might have a category.conf file like the following:

description: Auto Repair
action: allow

This configuration would mean that the category’s user-visible description would be “Auto Repair” rather than “mechanical,” and that pages that fall into the category would be allowed. The description defaults to the category name, and the action defaults to ignore. Actions can be overriden for specific users by the use of ACLs. A category.conf file may also have the entry invisible: true; this indicates that when a page is blocked because it belongs to that category, the response will be an invisible image instead of the usual block page.

The rule-list files define the rules used to calculate the category’s score. Each rule-list file must have an extension of .list. (This rule ensures that files ending in .bak, .orig, etc. are ignored.) It is a plain-text file encoded in UTF-8. Comments are delimited with #. Here is an example of a rule-list file that might be in the directory for the “mechanical” category mentioned earlier:

napaonline.com 200 # Give napaonline.com 200 points for this category.
www.napaonline.com/catalog/ 50 # bonus points for NAPA's catalog

default 150 # The following domains will each get 150 points.
carquest.com
autozone.com

/t[iy]re/ 75 # Any page with tire or tyre in the URL will get 75 points.
/parts/h 50 # A page with parts in the hostname will get 50 points

<grease gun> 25 # 25 points for each occurrence of "grease gun" in the content
<oil filter> 25 100 # 25 points for each occurrence, but no more than 100 total

%909841dcf4d4c000ff7f00fe30820000 100 # A hash of an image from napaonline.com

There are four kinds of filter rules:

• URL matching

  A URL matching rule consists of a domain name, optionally followed by a path. After it, separated by a space, is the weight—the number of points that get added to this category’s score for sites that match the rule.

  A rule for a domain will also match subdomains: napaonline.com also matches www.napaonline.com. A rule with a path will also match longer paths: www.napaonline.com/catalog also matches www.napaonline.com/catalog/result.aspx.

  If a domain and a subdomain (or a path and a subdirectory) are both listed, the subdomain will effectively get the sum of the two weights. For example, if xerox.com were listed with 100 points, and support.xerox.com were listed with 50 points, support.xerox.com would actually get a score of 150 points.

• URL regular expressions

  A regular expression to match the URL is listed between slashes. The points are added to the category score for each page whose URL matches the regular expression. The URL is converted to lower case before comparing it to the regular expressions. The regular expression syntax is that supported by the RE2 library.

  A regular expression can be restricted to matching a certain part of the URL by adding a one-character suffix immediately after the final slash. A suffix of h matches the hostname (e.g. www.google.com), d matches the base domain name (e.g. google), p matches the path, and q matches the query.

• Content phrases

  Unlike the other two kinds of rules, these apply to the content of the page, not the URL. Phrases are enclosed between angle brackets. Before testing to see if a phrase matches, both the phrase and the page are simplified: capital letters are converted to lowercase, all characters that are not letters or digits are replaced by spaces, and multiple spaces are replaced by single spaces. Then the phrase weight is added to the page’s score for the category for each time the phrase is found on the page. But if the phrase has a second weight listed, no more than that amount will be added no matter how many times the phrase occurs. (In the example, if “oil filter” occurred more than four times, the additional occurrences wouldn’t count.)

  The content of the page is scanned for phrases only if phrase scanning is selected with the phrase-scan ACL action.

• Image Hashes

  Redwood can hash images using the library at https://github.com/andybalholm/dhash. The rule consists of a percent sign (%) followed by the 32-character hash calculated by the dhash program. The hash my optionally be followed by a hyphen and a threshold, which is an integer specifying the number of bits that may be different for another hash to be considered to match this hash (this overrides the global dhash-threshold setting).

  Images are hashed only if hashing is selected with the hash-image ACL action.

There is also a default rule. It specifies what weight will be assigned to rules that don’t specify a weight. It applies to all rules without a specified weight between it and the next default rule or the end of the file. If there is no default rule, the default weight is zero.

Weights must be integers, but they may be negative. Negative weights can be used to offset short, general matches with long, more-specific ones, e.g.:

<grease> 10
<grease paint> -10

If a page is blocked based on its URL (i.e. by URL matching and/or URL regular expressions), its content will not be evaluated because the page will not be downloaded.

Access Control Lists (ACLs)

Much of Redwood’s functionality is configured with Access Control Lists (ACLs). Each request is assigned a number of ACL tags, and then an action is chosen based on those tags. For example:

acl no-web user-ip 192.168.1.25
block no-web

The first line creates an ACL tag no-web, and assigns it to all requests coming from IP address 192.168.1.25. The second line causes all requests with that tag to be blocked.

ACLs are checked at several points during the processing of a request: before sending the request to the origin server, after receiving a response, and after scanning the content for phrases. Each time, the request may have different ACL tags, since more information is available. Each stage also has a different set of possible actions, although there is some overlap. (The allow, block, and block-invisible actions are always available.)

Any number of ACL files can be loaded with the acl directive in the configuration file. An ACL file can load other ACL files with a line that contains include and the filename.

Assigning ACL Tags

ACL tags are assigned by lines starting with acl. These lines have the format:

acl tag-name attribute values

The tag-name can be any name that does not include spaces. The attribute refers to some property of the request or response (listed below). The values are a space-separated list; if any of them matches the attribute’s value, the tag will be assigned. If there is more than one acl line with the same tag name, the tag will be applied if any of them matches (logical OR). An ACL may have a description associated with it with a line like describe tag-name A long description.

In addition to the tags assigned by acl lines, a request is assigned a tag for its highest-scoring category (if the score is above the threshold). There is also an ACL invalid-ssl, which is automatically assigned to CONNECT requests when the data being sent over the connection is not valid SSL or TLS.

The following attributes are available:

• connect-port

  (request only) The destination port of a CONNECT request. This attribute never matches if the request method is not CONNECT.

• content-type

  (response only) The response’s media type, usually taken from the Content-Type header. This can also be a generic type, with an asterisk after the slash:

    acl images content-type image/*
  

• http-status

  (response only) The response's HTTP status code. If this is a multiple of 100, all status codes in that block of 100 will match.

• method

  The HTTP request method, such as GET or POST.

• referer

  The request’s Referer header. (This matches the same way as regular URL matching rules.)

• server-ip

  The server’s IP address, or a range of addresses (in CIDR format, or with a dash). This attribute only matches if the request URL contains a literal IP address; it does not do a DNS lookup.

    acl google server-ip 172.217.0.0/16
  

• time

  The current time.

    acl work-hours time MTWHF 9:00-17:00
  

  This attribute lets you select certain days of the week and/or ranges of times of the day. If the days of the week are specified, they must come first; they are abbreviated SMTWHFA. Any number of time ranges may be specified; the rule will match if the current time falls within any of them. Times must be in 24-hour format.

• url

  The URL requested. (This matches the same way as regular URL matching rules.)

• user-agent

  The User-Agent header. Instead of interpreting the remainder of the line as a list of values, this attribute interprets it as a single regular expression to be matched against the User-Agent string. The matching is case-insensitive.

• user-ip

  The user’s IP address, or a range of addresses (in CIDR format, or with a dash).

    acl managers 10.0.2.5 10.0.1.0/24 10.0.2.18-25
  

• user-name

  The username from HTTP proxy authentication.

ACL Actions

After the ACL tags are assigned, Redwood goes through the ACL files looking for an action to perform. An action will be selected only if it has all the tags specified in the action line. (And none of the negated tags; if a tag in an action line is preceded by an exclamation point, the request must not have that tag.) Since it goes through the files in order, earlier action lines take precedence over later ones. If it gets to the end of the file without finding a matching rule, it will use the default action of the highest-scoring category. If there is no category that scores over the threshold, the default action is allow.

An ACL action line may optionally have a description string at the end. This is a double-quoted string whose value will be available to the block page template as {{.RuleDescription}}.

• allow

  Allow the request to proceed.

• block

  Respond with an HTTP status code of 403, and send the standard block page.

• block-invisible

  Respond with HTTP 403, and send an invisible 1-pixel image instead of a block page.

• disable-proxy-headers

  Don't add headers that indicate that the request has passed through a proxy (X-Forwarded-For and Via).

• hash-image

  (response only) Calculate a hash of the image, and compare it to the hash rules. If the difference between this hash and the one in the rule is less than the number of bits specified with --dhash-threshold (or the hash's individual threshold), it matches.

  This action should only be applied when the content is an image:

    acl image content-type image/jpeg image/gif image/png
    hash-image image
  

• ignore-category

  Drop the highest-scoring category off the list of categories, and go through the ACL files again.

• log-content

  Log the page's content. The content-log-dir configuration directive must be set. The page's content will be saved in that directory, with its MD5 hash as the filename. A line will be added to index.csv in that directory, linking the page's URL to its MD5 hash.

• phrase-scan

  (response only) Run a phrase scan on the page content. Normally this will be configured to depend on the content type:

    acl text content-type text/* application/xhtml+xml
    acl css content-type text/css
    phrase-scan text !css
  

• require-auth

  (request only) Send an HTTP 407 response if the request doesn’t have a Proxy-Authorization header.

• ssl-bump

  (CONNECT requests only) Activate the SSLBump feature, to filter HTTPS connections. (Transparently intercepted HTTPS connections produce a virtual CONNECT request inside Redwood, so they can be filtered too.)

URL Query Modification

When processing an HTTP request, Redwood can modify the query parameters in the URL. The configuration file for these changes is specified with the query-changes keyword. Each line contains a URL-matching or URL-regular-expression rule, followed by a query expression. If the query in the URL already contains parameters with the same names as those specified in the file, they will be replaced with the new values. Otherwise the new values will be added.

# Force safe search on several search engines.
/www\.google\.[^/]+/search/ safe=vss
search.lycos.com adv=1&adf=on
search.yahoo.com vm=r
/hotbot/h adf=on
www.metacrawler.com familyfilter=1

Content Pruning

Between downloading a page and scanning its content for phrases, Redwood can perform “content pruning.” This is scanning the parsed HTML tree for elements matching certain criteria, and deleting those elements and their children.

Content pruning is controlled by a configuration file. Each line of the file contains a URL-matching or URL-regular-expression rule to specify what site or page the pruning applies to, and a CSS selector to specify what elements to delete. Between the two, there may be a threshold value. If a threshold is specified, the element and its children are deleted after the page is phrase-scanned if the score from the phrases found in a blocked category is at least the threshold.

# Craigslist personals and discussion forums
craigslist.org div#ppp, div#forums, option[value=ppp]

# Bing ad sidebar
bing.com div.sb_adsNv2

# Delete questionable forum topics.
talk.newagtalk.com/forums 50 td.messagecellbody > ul

Block Pages

When Redwood blocks access to a web page, it returns an HTTP response with a status of 404 Forbidden. Unless the category that caused the page to be blocked is configured as invisible, the body of the 404 response will be HTML rendered from a template file. The template file is specified with the blockpage configuration directive. The following placeholders may be used in the template file, to be replaced by the appropriate information when the block page is sent:

• {{.URL}}

  the URL of the page that was blocked

• {{.Categories}}

  the names of the categories that caused the page to be blocked

• {{.Conditions}}

  the conditions of the ACL rule that caused the page to be blocked

• {{.User}}

  the user’s IP address or username

• {{.Tally}}

  a list of the rules that matched, and how many times each one matched

• {{.Scores}}

  a list of categories, and how many points the page scored in each category

The block page is generated using the Go template package; see http://golang.org/pkg/text/template and http://golang.org/pkg/html/template for documentation.

There is one custom function defined for the templates to use, eq, which tests its parameters for equality.

Virtual Web Servers

Since the block page may need to refer to external resources (such as images, stylesheets, and scripts), Redwood includes an internal web server. This web server does not accept connections directly, but whenever Redwood processes a request with a server address of 203.0.113.1, it directs the request to the internal server instead of processing it normally. The content of the internal web server is configured with the static-files-dir, and cgi-bin directives.

If a more advanced virtual server is needed, you can use the virtual-host directive to transparently redirect requests for a given hostname to a different address, such as an Apache web server running on your gateway. If the server is running on your gateway, listening on port 8888, and you want it to be available as myserver.local, use virtual-host myserver.local localhost:8888. (Note: proxy settings are not set on the client, and Redwood is intercepting requests transparently, this will work only if the DNS server resolves the name to an IP address outside your local network. Any IP address will do, though. OpenDNS’s website-unavailable address works fine. Also note that virtual-host only works with HTTP, not with HTTPS.)

Test Mode

If Redwood is run with the -test switch, it does not run as a proxy server. Instead, it evaluates the URL given as an argument after the switch. It prints detailed debugging information about how the URL and its content would be rated if that page were requested in normal operation: how many times each rule matches, what the score is in each category, which categories would block the page, etc.

Log Files

Redwood has several categories of messages that can be logged:

General diagnostic messages are sent to standard error by default, and may be redirected to a file using normal shell redirection.

The access log has a line for each request processed. It is in CSV format and goes to standard output by default. It can be sent to a file by including the access-log directive in Redwood’s configuration file. The access log has the following fields: time, username or IP address, action (allow or block), URL, HTTP method (GET, PUT, etc.), HTTP response status (if an HTTP response was being processed), content type, content-length, whether the content was modified by Redwood, which rules matched (and how many times), the score for each category, the list of categories that caused the page to be blocked (if it was), the page title (if log-title is enabled), a list of the categories that were ignored even though they had higher scores than the one that determined the action, the User-Agent header (if log-user-agent is enabled), the HTTP version, the Referer header, and the client platform (such as Windows or iPad, found in the User-Agent header). The content length is meaningful only if a phrase scan was performed. The page title is available only if a phrase scan was performed and log-title was enabled in the configuration (logging the page title requires parsing the HTML, so it is disabled by default).

The TLS log has a line for each HTTPS connection that was intercepted. Like the access log, it goes to standard output by default, and it can be sent to a file with the tls-log directive. The TLS log has the following fields: time, username or client IP address, server name, server address, any error that was encountered, and whether the certificate used came from the certificate cache.

Authentication

Redwood can be configured (using the require-auth ACL action) to require HTTP basic proxy authentication, with a username and password. The usernames and passwords can come from a file that is specified by the --password-file configuration directive. Each line in the file consists of a username, a password, and some optional items, separated by spaces or tabs. Alternatively, a program can be specified to perform authentication with --authenticator. The program will be invoked with the username and password as command-line arguments. Its exit status determines whether the authentication is successful; if the exit status is zero, the user will be accepted.

The optional items in the password file are for setting up a custom proxy port for that individual user, to make authentication easier. The first optional item is a port number; if it is present, Redwood will listen for HTTP requests on that port. Only the specified user may use that port, but once a client has authenticated as that user, all further requests from that IP address to this port will be considered authenticated, whether they have the Proxy-Authorization header or not.

In addition, you can set up automatic authentication based on the device platform and the network the device is on. For example, to automatically authenticate an iPad on the Verizon network, you could have the following line in the password file:

ipad-user mySecurePassword 7500 iPad myvzw.com

Redwood uses the HTTP User-Agent string to determine a device's platform. The currently recognized platforms are Windows, Linux, Android, Macintosh, iPhone, iPad, and iPod.)

The network can be specified as an IP address range in CIDR notation (70.192.0.0/11) or as a domain name to be compared to an IP address's reverse DNS entry. Multiple networks can be specified, separated by commas. If a device successfully authenticates (using the username and password) from a network that is not on the list, that network will be added to the list of expected networks.

SSLBump

Redwood can be configured (using the ssl-bump ACL action) to perform Man-in-the-Middle filtering of HTTPS traffic. This feature is called SSLBump after the corresponding feature in Squid.

For SSLBump to work, Redwood must be configured with a root certificate that is trusted by the users’ browsers. Paths to the certificate and its private key are specified with the tls-cert and tls-key options. The certificate and key should be in PEM format.

Redwood uses the system root certificates to verify the identity of the sites it bumps. Other trusted root certificates can be specified with the trusted-root option.

The SSLBump feature only works with SSL version 3 and newer (including all TLS versions). By default, earlier versions are passed through unfiltered. It can be configured to block them instead with the block-obsolete-ssl option.

Transparent Proxy

With the proper firewall setup, Redwood can transparently intercept connections to web servers and filter them without needing to configure proxy settings on the client computers. Intercepted HTTP connections can use the same proxy port as is used for manually-configured proxy connections. For HTTPS connections, Redwood must be configured to listen for intercepted connections on a separate port, with the transparent-https directive.

The following configuration lines will set Redwood to listen on ports 6502 and 6510:

http-proxy :6502
transparent-https :6510

If Redwood is running on a Linux gateway/router system, the following iptables rules will enable transparent filtering for the computers on the LAN (assuming that the LAN interface is eth1):

iptables -A PREROUTING -i eth1 -p tcp --dport 80 -j REDIRECT --to-ports 6502
iptables -A PREROUTING -i eth1 -p tcp --dport 443 -j REDIRECT --to-ports 6510

To do the same thing with pf on a FreeBSD gateway (assuming that the LAN interface is re0):

table <filtered> { re0:network }
rdr pass inet proto tcp from <filtered> to any port 80 -> re0 port 6502
rdr pass inet proto tcp from <filtered> to any port 443 -> re0 port 6510

Classification Service

In addition to running as a proxy, Redwood can also be used as a URL classification service. It receives an HTTP request specifying a URL, and returns a JSON object that tells what categories it was classifed in, and the score for each category.

Categories can be excluded from the classification reports:

classifier-ignore sslbump
classifier-ignore masterwhitelist

The classification request must have the URL as an HTTP form parameter named "url." The JSON object in the response has the following keys:

• url: the URL being classified
• categories: an object with category names for keys, and their scores for values
• error: any error that was encountered fetching or processing the page

For example, if Redwood is running on port 6502 on 10.1.10.1, http://10.1.10.1:6502/classify?url=https%3A%2F%2Fgolang.org might return {"url":"https://golang.org","categories":{"computer":266}}.

It can also classify text directly. http://10.1.10.1:6502/classify-text?text=programming+language might return {"text":"programming language","categories":{"computer":27}}.

PAC Files

Redwood can provide PAC (Proxy Auto-Configuration) files to automatically configure client computers to use it as their proxy. Then, whenever Redwood receives a request for /proxy.pac, it sends a PAC file directing the client to proxy its requests.

PAC files also make another feature possible: listening on separate, pre-authenticated ports for individual users. This helps to address various authentication problems resulting from software that doesn't support proxy authentication properly. To enable this, specify a custom port number on the user's line in the password file. Then, put a base64-encoded username/password pair (just like in an HTTP basic authentication header) in the PAC request URL (e.g. /proxy.pac?a=dXNlcm5hbWU6cGFzc3dvcmQ=). (Generate it by typing a command like echo -n username:password | base64 at a UNIX command prompt.) All requests received on that port from the same IP address as the PAC file request will be automatically authenticated as that user.

Scripting

For more specialized rules than the ones that are supported by normal ACLs, you can write scripts (in JavaScript) that assign ACLs to requests. For example, suppose you want to block the sites that OpenDNS classifies as adult or phishing sites, but you want to have your own block page instead of the one OpenDNS provides. You could put a script like this in /etc/redwood/opendns.js:

var openDNSResult = lookupHost(request.URL.Host, "208.67.222.123");

if (openDNSResult == "146.112.61.106") {
	addACL("opendns-adult");
} else if (openDNSResult == "146.112.61.108") {
	addACL("opendns-phishing");
}

Put the following line in redwood.conf:

request-acl-script /etc/redwood/opendns.js

And in your ACL configuration file:

block opendns-adult
block opendns-phishing