tipidee
Software
skarnet.org

The tipideed program

tipideed is the binary that actually does what you want from a web server package: it serves files over HTTP.

     tipideed [ -f cdbfile ] [ -d basedir ] [ -R ] [ -U ]

• tipideed reads a stream of HTTP (1.0 or 1.1) requests on its stdin, and tries to fulfill them, sending answers to stdout and logs to stderr.
• tipideed only speaks plaintext HTTP. It supports HTTPS, but the TLS layer must be handled upstream by a program such as s6-tlsd.
• tipideed stays alive until the client closes the connection, or times out, or (in HTTP 1.1) sends a request with a Connection: close header; or an error occurs that makes it nonsensical to keep the connection open.
• The documents it serves must be in subdirectories of its current working directory, one subdirectory for every domain it hosts.

tipideed is intended to be run under a TCP super-server such as s6-tcpserver, for plain text HTTP, or s6-tlsserver, for HTTPS. It delegates to the super-server the job of binding and listening to the socket, accepting connections, spawning a separate process to handle a given connection, and potentially establishing a TLS tunnel with the client for secure communication.

As such, a command line for tipideed, running as user www, listening on address ${ip}, would typically look like this, for HTTP:

     s6-envuidgid www s6-tcpserver -U -- ${ip} 80 s6-tcpserver-access -- tipideed

or, for HTTPS:

     s6-envuidgid www env KEYFILE=/path/to/private/key CERTFILE=/path/to/certificate s6-tlsserver -U -- ${ip} 443 tipideed

Most users will want to run these command lines as services, i.e. daemons run in the background when the machine starts. The examples/ subdirectory of the tipidee package provides service templates to help you run tipideed under OpenRC, s6 and s6-rc.

0

Clean exit. There was a successful series of HTTP exchanges, that either tipideed or the client decided to end in a way that is permitted by HTTP.

1

Illicit client behaviour. tipideed exited because it could not serve the client in good faith.

2

Illicit CGI script behaviour. tipideed exited because the invoked CGI script made it impossible to continue. Before exiting, tipideed likely has sent a 502 (Bad Gateway) response to the client.

100

Bad usage. tipideed was run in an incorrect way: bad command line options, or missing environment variables, etc.

101

Cannot happen. This signals a bug in tipideed, and comes with an error message asking you to report the bug. Please do so, on the skaware mailing-list.

102

Misconfiguration. tipideed found something in its configuration data or in the document layout that it does not like. This can happen, for instance, when a document is a symbolic link pointing outside of the server's root.

111

System call failed. This usually signals an issue with the the underlying operating system. Before exiting, if in the middle of processing a request, tipideed likely has sent a 500 (Internal Server Error) response to the client.

Reading - mandatory

tipideed expects the following variables in its environment, and will exit with an error message if they are undefined. When tipideed is run under s6-tcpserver or s6-tlsserver, these variables are automatically set by the super-server. This is the way tipidee gets its network information without having to perform network operations itself.

PROTO

The network protocol, normally TCP.

TCPLOCALIP

The IP address the server is bound to. It will be passed as SERVER_ADDR to CGI scripts.

TCPLOCALPORT

The port the server is bound to. It will be passed as SERVER_PORT to CGI scripts unless the requested URI explicitly mentions a different port.

TCPREMOTEIP

The IP address of the client. It will be passed as REMOTE_ADDR to CGI scripts.

TCPREMOTEPORT

The port of the client socket. It will be passed as REMOTE_PORT to CGI scripts.

Reading - optional

tipideed can function without these variables, but if they're present, it uses them to get more information. They're typically obtained by calling s6-tcpserver-access before tipideed in the s6-tcpserver command line. (For HTTPS, s6-tlsserver calls it implicitly.)

TCPLOCALHOST

The default domain name associated to the local IP address. It will be passed as SERVER_NAME to CGI scripts when the requested URI does not mention a Host, i.e. in HTTP/1.0 requests without a full request URL. If this variable is absent, the default will be set to the local IP address itself (between square brackets if IPv6).

TCPREMOTEHOST

The domain name associated to the IP address of the client. It will be passed as REMOTE_HOST to CGI scripts; if absent, the value of TCPREMOTEIP will be used instead.

TCPREMOTEINFO

The name provided by an IDENT server running on the client, if any. This is obsolete and not expected to be present; but if present, it will be passed as REMOTE_IDENT to CGI scripts.

SSL_PROTOCOL

The version of the TLS protocol used to cipher communications between the client and the server. If present, tipideed will assume that the client connection is secure, and will pass HTTPS=on to CGI scripts; otherwise, it will assume it is running plaintext HTTP.

Writing

When spawning a CGI or NPH script, tipideed clears all the previous variables, so the passed environment is as close as possible to the environment of the super-server; and it adds all the variables that are required by the CGI 1.1 specification. As an exception, it does not add PATH_TRANSLATED, which cannot be used by CGI scripts in a portable way.

-f file

Use file as the compiled configuration database, typically obtained by running tipidee-config -o file. The default is /etc/tipidee.conf.cdb; /etc may be something else if the --sysconfdir option has been given to configure at build time.

-d docroot

Change the working directory to docroot before serving. Default is serving from the current working directory. Note that documents need to be located in subdirectories of docroot, one subdirectory per virtual domain tipideed is serving.

-R

chroot. If the underlying operating system has the chroot() system call, use it before serving. This always happens after opening the configuration database, after changing the working directory, and before dropping privileges. The idea is that chrooting helps with security, but the configuration database should be located outside of the document space.

-U

Drop root privileges. If this option is given, tipideed expects two additional environment variables, UID and GID, containing the uid and gid it should run as; it will drop its privileges to $UID:$GID before serving. This option is mainly useful when paired with -R, because chrooting can only be performed as root, so root privileges need to be kept all the way to tipideed then dropped after tipideed has chrooted. In a non-chrooted setup, it is simpler and more secure to run the super-server with the -U option instead: root privileges will be dropped as soon as the super-server has bound to its socket, and all the subsequent operations, including the spawning of tipideed processes, are performed as a normal user.

The way to organize your documents so they can be served by tipideed may look a little weird, but there's a logic to it.

tipideed serves documents from subdirectories of its working directory, and these subdirectories are named according to the host and the port of the request.

• A request for https://example.com:1234/doc/u/ment will result in a lookup in the filesystem for ./example.com:1234/doc/u/ment.
• A request for https://example.com/doc/u/ment will result in a lookup in the filesystem for ./example.com:443/doc/u/ment.

The fact that the port is always specified allows you to have different document sets for the same host on different ports: more flexibility.

However, most of the time, you don't want different document sets for different ports. You want the same document sets for ports 80 and 443, and that's it. And you don't want to have both a domain example.com:80 section and a domain example.com:443 section in your /etc/tipidee.conf, with duplicate information.

That is why you are allowed to make your document roots symbolic links, and resource attributes declared in the configuration file are always looked up with the canonical path. In other words, the common case would be:

• Have your document root in ./example.com, a real directory.
• Declare your resource attributes under a domain example.com section in your configuration file.
• Have a ./example.com:80 symlink pointing to example.com, if you want to serve example.com under plaintext HTTP.
• Have a ./example.com:443 symlink pointing to example.com, if you want to serve example.com under HTTPS.

This system allows you to share documents across virtual hosts without fear of misconfiguration. You can symlink any document under example.com to any name under example.org; if the path via example.com is the canonical path, then your resource will still get the correct attributes, defined in a domain example.com section, even if it is accessed via an example.org URL. You will not inadvertently expose source code for CGI scripts, for instance.

You can do wild things with symbolic links. However, anything that does not resolve to a file in a document root under tipideed's current working directory will be rejected. If an attacker symlinks your /etc/passwd file, tipideed will keep it safe.

HTTP/1.0 does not have the concepts of virtual hosts. For HTTP/1.0 requests that do not provide a full URL, tipideed will use the value it reads from the TCPLOCALHOST variable, which is normally the result of a reverse DNS lookup on the server's address. You can override the lookup and provide your own value by giving the -l option to s6-tcpserver-access or s6-tlsserver. If TCPLOCALHOST does not exist or is empty, a fallback value of @ (at), will be used. So if you aren't calling s6-tcpserver-access at all, your documents will most likely be accessible for HTTP/1.0 clients under @:80 or @:443.

• tipideed uses stderr for all its logging. All its log lines are prefixed with "tipideed: pid pid: ".
• The log lines continue with "fatal: " for fatal error messages (meaning that tipideed exits right after writing the message), or "warning: " for warnings (meaning that tipideed continues operating after writing the message). In normal operation, you should not see any fatal or warning line.
• In normal operation, tipideed can log informational lines, and the continuing prefix is "info: ". It can potentially log:
  • One line when it starts (i.e. a client has connected)
  • Up to three lines for every request:
    • One when the request is received
    • One when a suitable resource is found. In rare cases (namely: the resource is a CGI script answering with a local redirection), there may be more than one resource line.
    • One when an answer is sent
  • One line when it exits normally
• What to log is configured via the log directive in the configuration file. By default, only a minimal request line and an answer line are printed.
• The log format is designed to be readable by a human, but still easily processable by automation. For instance, the regular prefix structure makes it easy for s6-log to select different lines to send them to various backends for archiving or processing.

• tipideed reads its compiled configuration file. Then:
  • If the -d option has been given, it changes its working directory.
  • If the -R option has been given, it chroots to its current directory.
  • If the -U option has been given, it drops root privileges.
• It checks that its environment is valid, and that its configuration has some minimal defaults it can use.
• tipideed listens to a stream of HTTP requests on its standard input. For every HTTP request:
  • It parses the request line and checks it's HTTP/1.0 or 1.1
  • It parses the headers into a quick access structure
  • It checks header consistency with the request
  • If the method is OPTIONS * or TRACE, it answers here and continues the loop
  • It reads the request body, if any
  • It checks in its configuration if a redirection has been defined for the wanted resource or a prefix (by directory) of the wanted resource. If it's the case, it answers with that redirection and continues the loop.
  • It looks for a suitable resource in the filesystem, completing the request with index files if necessary, or extracting CGI INFO_PATHs if necessary
  • It uses the canonical path of the resource in the filesystem to look for resource attributes in its configuration. (Is this a CGI script? a NPH script? Does it have a customized Content-Type? etc.)
  • If the method is a targeted OPTIONS, it answers here and continues the loop
  • If the resource is a CGI script:
    • If it is an NPH script, tipideed execs into the script (possibly after spawning a helper child if there is a request body to feed to the script) with the appropriate environment; and the connection will close when the script exits.
    • Else, tipideed spawns the CGI script as a child with the appropriate environment, feeds it the request body if any, reads its output, and answers the client.
    • If a problem occurs server-side, the client will receive a 502 answer ("Bad Gateway"), and tipideed will write an error message to its stderr, so that administrators can see what went wrong with their setup. tipideed trusts its CGI scripts more than its clients, but it does not give them its full trust either — lots of sites are running third-party backends.
  • Else, the resource is a regular ("static") file, and tipideed serves it on its stdout, to the client.
• tipideed exits on EOF (when the client closes the connection), or when the client times out before sending a request, or after tipideed receives a single HTTP/1.0 request, or when it has executed into an NPH script, or when it has answered a request with a Connection: close header. It also exits when it encounters an error making it likely that the client will have no use for the connection anymore anyway and exiting is simpler and cheaper; in which case tipideed adds Connection: close to its last answer.

On systems that implement posix_spawn(), the s6-tcpserver super-server (and the s6-tlsserver one as well, since both use the same underlying program) uses it instead of fork(), and that partly alleviates the performance penalty usually associated with servers that spawn one process per connection.

One of tipidee's stated goals is to explore what kind of performance is achievable for a fully compliant Web server within the limits of that model. To that effect, tipideed is meant to be fast. It should serve static files as fast as any server out there, especially on Linux (or other systems supporting splice()) where it uses zero-copy transfer. CGI performance should be limited by the performance of the CGI script itself, never by tipideed.

tipideed itself does not use fork() if the system supports posix_spawn() — with one exception, that you will not hit, and if you do, fork() will not be the bottleneck. (Can you guess which case it is, without looking at the code?) tipideed does not parse its configuration file itself, delegating the task to the offline tipidee-config program and directly mapping a binary file instead. To parse a client request, it uses a deterministic finite automaton, only reading the request once, and only backtracking in pathological cases. This should streamline request processing as much as possible.

If you have benchmarks, results of comparative testing of tipideed against other Web servers, please share them on the skaware mailing-list.

• tipideed sometimes answers 400, or even does not answer at all (it just exits), when receiving some malformed or weirdly paced client requests, despite what the HTTP RFC says. This is on purpose. HTTP servers are very heavily solicited, they can run very hot, the Web is a cesspool of badly written bots and bad actors, and every legitimate browser knows how to speak HTTP properly and without abusing corner cases in the protocol. It makes no sense to try to follow the book to the letter, expending precious resources, when the client can't even be bothered to pretend it's legit. Knowing when to exit early is crucial for good resource management.
• tipideed does not have any configuration defaults baked in; it reads all its configuration values, including the defaults, from the cdb file created by tipidee-config. That is why having such a cdb file is mandatory for tipideed to run.
• tipideed is pronounced tipi-deed. You can say tipi-dee-dee, but only if you're the type of person who also says PC computer, NIC card or ATM machine.
• tipidee is the name of the package, the software suite implementing a Web server. tipideed is the name of the program doing the HTTP serving part.