thttpd - tiny/turbo/throttling HTTP server

thttpd [-C configfile] [-p port] [-d dir] [-r|-nor] [-s|-nos] [-v|-nov] [-g|-nog] [-u user] [-c cgipat] [-t throttles] [-h host] [-l logfile] [-i pidfile] [-T charset] [-P P3P] [-M maxage] [-V] [-D]

thttpd is a simple, small, fast, and secure HTTP server. It doesn't have a lot of special features, but it suffices for most uses of the web, it's about as fast as the best full-featured servers (Apache, NCSA, Netscape), and it has one extremely useful feature (URL-traffic-based throttling) that no other server currently has.

-C

Specifies a config-file to read. All options can be set either by command-line flags or in the config file. See below for details.

-p

Specifies an alternate port number to listen on. The default is 80. The config-file option name for this flag is "port", and the config.h option is DEFAULT_PORT.

-d

Specifies a directory to chdir() to at startup. This is merely a convenience - you could just as easily do a cd in the shell script that invokes the program. The config-file option name for this flag is "dir", and the config.h options are WEBDIR, USE_USER_DIR.

-r

Do a chroot() at initialization time, restricting file access to the program's current directory. If -r is the compiled-in default, then -nor disables it. See below for details. The config-file option names for this flag are "chroot" and "nochroot", and the config.h option is ALWAYS_CHROOT.

-nos

Don't do explicit symbolic link checking. Normally, thttpd explicitly expands any symbolic links in filenames, to check that the resulting path stays within the original document tree. If you want to turn off this check and save some CPU time, you can use the -nos flag, however this is not recommended. Note, though, that if you are using the chroot option, the symlink checking is unnecessary and is turned off, so the safe way to save those CPU cycles is to use chroot. The config-file option names for this flag are "symlink" and "nosymlink".

-v

Do el-cheapo virtual hosting. If -v is the compiled-in default, then -nov disables it. See below for details. The config-file option names for this flag are "vhost" and "novhost", and the config.h option is ALWAYS_VHOST.

-g

Use a global passwd file. This means that every file in the entire document tree is protected by the single .htpasswd file at the top of the tree. Otherwise the semantics of the .htpasswd file are the same. If this option is set but there is no .htpasswd file in the top-level directory, then thttpd proceeds as if the option was not set - first looking for a local .htpasswd file, and if that doesn't exist either then serving the file without any password. If -g is the compiled-in default, then -nog disables it. The config-file option names for this flag are "globalpasswd" and "noglobalpasswd", and the config.h option is ALWAYS_GLOBAL_PASSWD.

-u

Specifies what user to switch to after initialization when started as root. The default is "nobody". The config-file option name for this flag is "user", and the config.h option is DEFAULT_USER.

-c

Specifies a wildcard pattern for CGI programs, for instance "**.cgi" or "/cgi-bin/*". See below for details. The config-file option name for this flag is "cgipat", and the config.h option is CGI_PATTERN.

-t

Specifies a file of throttle settings. See below for details. The config-file option name for this flag is "throttles".

-h

Specifies a hostname to bind to, for multihoming. The default is to bind to all hostnames supported on the local machine. See below for details. The config-file option name for this flag is "host", and the config.h option is SERVER_NAME.

-l

Specifies a file for logging. If no -l argument is specified, thttpd logs via syslog(). If "-l /dev/null" is specified, thttpd doesn't log at all. The config-file option name for this flag is "logfile".

-i

Specifies a file to write the process-id to. If no file is specified, no process-id is written. You can use this file to send signals to thttpd. See below for details. The config-file option name for this flag is "pidfile".

-T

Specifies the character set to use with text MIME types. The default is iso-8859-1. The config-file option name for this flag is "charset", and the config.h option is DEFAULT_CHARSET.

-P

Specifies a P3P server privacy header to be returned with all responses. See http://www.w3.org/P3P/ for details. Thttpd doesn't do anything at all with the string except put it in the P3P: response header. The config-file option name for this flag is "p3p".

-M

Specifies the number of seconds to be used in a "Cache-Control: max-age" header to be returned with all responses. An equivalent "Expires" header is also generated. The default is no Cache-Control or Expires headers, which is just fine for most sites. The config-file option name for this flag is "maxage".

-V

Shows the current version info.

-D

This was originally just a debugging flag, however it's worth mentioning because one of the things it does is prevent thttpd from making itself a background daemon. Instead it runs in the foreground like a regular program. This is necessary when you want to run thttpd wrapped in a little shell script that restarts it if it exits.

All the command-line options can also be set in a config file. One advantage of using a config file is that the file can be changed, and thttpd will pick up the changes with a restart.

The syntax of the config file is simple, a series of "option" or "option=value" separated by whitespace. The option names are listed above with their corresponding command-line flags.

chroot() is a system call that restricts the program's view of the filesystem to the current directory and directories below it. It becomes impossible for remote users to access any file outside of the initial directory. The restriction is inherited by child processes, so CGI programs get it too. This is a very strong security measure, and is recommended. The only downside is that only root can call chroot(), so this means the program must be started as root. However, the last thing it does during initialization is to give up root access by becoming another user, so this is safe.

The program can also be compile-time configured to always do a chroot(), without needing the -r flag.

Note that with some other web servers, such as NCSA httpd, setting up a directory tree for use with chroot() is complicated, involving creating a bunch of special directories and copying in various files. With thttpd it's a lot easier, all you have to do is make sure any shells, utilities, and config files used by your CGI programs and scripts are available. If you have CGI disabled, or if you make a policy that all CGI programs must be written in a compiled language such as C and statically linked, then you probably don't have to do any setup at all.

However, one thing you should do is tell syslogd about the chroot tree, so that thttpd can still generate syslog messages. Check your system's syslodg man page for how to do this. In FreeBSD you would put something like this in /etc/rc.conf:

    syslogd_flags="-l /usr/local/www/data/dev/log"

Relevant config.h option: ALWAYS_CHROOT.

thttpd supports the CGI 1.1 spec.

In order for a CGI program to be run, its name must match the pattern specified either at compile time or on the command line with the -c flag. This is a simple shell-style filename pattern. You can use * to match any string not including a slash, or ** to match any string including slashes, or ? to match any single character. You can also use multiple such patterns separated by |. The patterns get checked against the filename part of the incoming URL. Don't forget to quote any wildcard characters so that the shell doesn't mess with them.

Restricting CGI programs to a single directory lets the site administrator review them for security holes, and is strongly recommended. If there are individual users that you trust, you can enable their directories too.

If no CGI pattern is specified, neither here nor at compile time, then CGI programs cannot be run at all. If you want to disable CGI as a security measure, that's how you do it, just comment out the patterns in the config file and don't run with the -c flag.

Note: the current working directory when a CGI program gets run is the directory that the CGI program lives in. This isn't in the CGI 1.1 spec, but it's what most other HTTP servers do.

Relevant config.h options: CGI_PATTERN, CGI_TIMELIMIT, CGI_NICE, CGI_PATH, CGI_LD_LIBRARY_PATH, CGIBINDIR.

Basic Authentication is available as an option at compile time. If enabled, it uses a password file in the directory to be protected, called .htpasswd by default. This file is formatted as the familiar colon-separated username/encrypted-password pair, records delimited by newlines. The protection does not carry over to subdirectories. The utility program thtpasswd(1) is included to help create and modify .htpasswd files.

Relevant config.h option: AUTH_FILE

The throttle file lets you set maximum byte rates on URLs or URL groups. There is no provision for setting a maximum request rate throttle, because throttling a request uses as much cpu as handling it, so there would be no point.

The format of the throttle file is very simple. A # starts a comment, and the rest of the line is ignored. Blank lines are ignored. The rest of the lines should consist of a pattern, whitespace, and a number. The pattern is a simple shell-style filename pattern, using ?/**/*, or multiple such patterns separated by |.

The numbers in the file are byte rates, specified in units of bytes per second. For comparison, a v.32b/v.42b modem gives about 1500/2000 B/s depending on compression, a double-B-channel ISDN line about 12800 B/s, and a T1 line is about 150000 B/s.

Example:

  # throttle file for www.acme.com

** 100000 # limit total web usage to 2/3 of our T1 **.jpg|**.gif 50000 # limit images to 1/3 of our T1 **.mpg 20000 # and movies to even less jef/** 20000 # jef's pages are too popular

Throttling is implemented by checking each incoming URL filename against all of the patterns in the throttle file. The server accumulates statistics on how much bandwidth each pattern has accounted for recently (via a rolling average). If a URL matches a pattern that has been exceeding its specified limit, then the data returned is actually slowed down, with pauses between each block. If that's not possible (e.g. for CGI programs), then the server returns a special code saying 'try again later'.

Multihoming means using one machine to serve multiple hostnames. For instance, if you're an internet provider and you want to let all of your customers have customized web addresses, you might have www.joe.acme.com, www.jane.acme.com, and your own www.acme.com, all running on the same physical hardware. This feature is also known as "virtual hosts". There are three steps to setting this up.

One, make DNS entries for all of the hostnames. The current way to do this, allowed by HTTP/1.1, is to use CNAME aliases, like so:

  www.acme.com IN A 192.100.66.1
  www.joe.acme.com IN CNAME www.acme.com
  www.jane.acme.com IN CNAME www.acme.com

  www.acme.com IN A 192.100.66.1
  www.joe.acme.com IN A 192.100.66.200
  www.jane.acme.com IN A 192.100.66.201

Step two. If you're using the modern CNAME method of multihoming, then you can skip this step. Otherwise, using the older multiple-IP-address method you must set up IP aliases or multiple interfaces for the extra addresses. You can use ifconfig(8)'s alias command to tell the machine to answer to all of the different IP addresses. Example:

  ifconfig le0 www.acme.com
  ifconfig le0 www.joe.acme.com alias
  ifconfig le0 www.jane.acme.com alias

Third and last, you must set up thttpd to handle the multiple hosts. The easiest way is with the -v flag, or the ALWAYS_VHOST config.h option. This works with either CNAME multihosting or multiple-IP multihosting. What it does is send each incoming request to a subdirectory based on the hostname it's intended for. All you have to do in order to set things up is to create those subdirectories in the directory where thttpd will run. With the example above, you'd do like so:

  mkdir www.acme.com www.joe.acme.com www.jane.acme.com

  ln -s www.acme.com 192.100.66.1
  ln -s www.joe.acme.com 192.100.66.200
  ln -s www.jane.acme.com 192.100.66.201

There's an optional alternate step three if you're using multiple-IP multihosting: run a separate thttpd process for each hostname, using the -h flag to specify which one is which. This gives you more flexibility, since you can run each of these processes in separate directories, with different throttle files, etc. Example:

  thttpd -r -d /usr/www -h www.acme.com
  thttpd -r -d /usr/www/joe -u joe -h www.joe.acme.com
  thttpd -r -d /usr/www/jane -u jane -h www.jane.acme.com

thttpd lets you define your own custom error pages for the various HTTP errors. There's a separate file for each error number, all stored in one special directory. The directory name is "errors", at the top of the web directory tree. The error files should be named "errNNN.html", where NNN is the error number. So for example, to make a custom error page for the authentication failure error, which is number 401, you would put your HTML into the file "errors/err401.html". If no custom error file is found for a given error number, then the usual built-in error page is generated.

If you're using the virtual hosts option, you can also have different custom error pages for each different virtual host. In this case you put another "errors" directory in the top of that virtual host's web tree. thttpd will look first in the virtual host errors directory, and then in the server-wide errors directory, and if neither of those has an appropriate error file then it will generate the built-in error.

Sometimes another site on the net will embed your image files in their HTML files, which basically means they're stealing your bandwidth. You can prevent them from doing this by using non-local referer filtering. With this option, certain files can only be fetched via a local referer. The files have to be referenced by a local web page. If a web page on some other site references the files, that fetch will be blocked. There are three config-file variables for this feature:

urlpat

A wildcard pattern for the URLs that should require a local referer. This is typically just image files, sound files, and so on. For example:

  urlpat=**.jpg|**.gif|**.au|**.wav

For most sites, that one setting is all you need to enable referer filtering.

noemptyreferers

By default, requests with no referer at all, or a null referer, or a referer with no apparent hostname, are allowed. With this variable set, such requests are disallowed.

localpat

A wildcard pattern that specifies the local host or hosts. This is used to determine if the host in the referer is local or not. If not specified it defaults to the actual local hostname.

thttpd is very picky about symbolic links. Before delivering any file, it first checks each element in the path to see if it's a symbolic link, and expands them all out to get the final actual filename. Along the way it checks for things like links with ".." that go above the server's directory, and absolute symlinks (ones that start with a /). These are prohibited as security holes, so the server returns an error page for them. This means you can't set up your web directory with a bunch of symlinks pointing to individual users' home web directories. Instead you do it the other way around - the user web directories are real subdirs of the main web directory, and in each user's home dir there's a symlink pointing to their actual web dir.

The CGI pattern is also affected - it gets matched against the fully-expanded filename. So, if you have a single CGI directory but then put a symbolic link in it pointing somewhere else, that won't work. The CGI program will be treated as a regular file and returned to the client, instead of getting run. This could be confusing.

thttpd is also picky about file permissions. It wants data files (HTML, images) to be world readable. Readable by the group that the thttpd process runs as is not enough - thttpd checks explicitly for the world-readable bit. This is so that no one ever gets surprised by a file that's not set world-readable and yet somehow is readable by the HTTP server and therefore the *whole* world.

The same logic applies to directories. As with the standard Unix "ls" program, thttpd will only let you look at the contents of a directory if its read bit is on; but as with data files, this must be the world-read bit, not just the group-read bit.

thttpd also wants the execute bit to be *off* for data files. A file that is marked executable but doesn't match the CGI pattern might be a script or program that got accidentally left in the wrong directory. Allowing people to fetch the contents of the file might be a security breach, so this is prohibited. Of course if an executable file *does* match the CGI pattern, then it just gets run as a CGI.

In summary, data files should be mode 644 (rw-r--r--), directories should be 755 (rwxr-xr-x) if you want to allow indexing and 711 (rwx--x--x) to disallow it, and CGI programs should be mode 755 (rwxr-xr-x) or 711 (rwx--x--x).

thttpd does all of its logging via syslog(3). The facility it uses is configurable. Aside from error messages, there are only a few log entry types of interest, all fairly similar to CERN Common Log Format:

  Aug  6 15:40:34 acme thttpd[583]: 165.113.207.103 - - "GET /file" 200 357
  Aug  6 15:40:43 acme thttpd[583]: 165.113.207.103 - - "HEAD /file" 200 0
  Aug  6 15:41:16 acme thttpd[583]: referer http://www.acme.com/ -> /dir
  Aug  6 15:41:16 acme thttpd[583]: user-agent Mozilla/1.1N

Relevant config.h option: LOG_FACILITY.

If you'd rather log directly to a file, you can use the -l command-line flag. But note that error messages still go to syslog.

thttpd handles a couple of signals, which you can send via the standard Unix kill(1) command:

INT,TERM

These signals tell thttpd to shut down immediately. Any requests in progress get aborted.

USR1

This signal tells thttpd to shut down as soon as it's done servicing all current requests. In addition, the network socket it uses to accept new connections gets closed immediately, which means a fresh thttpd can be started up immediately.

HUP

This signal tells thttpd to close and re-open its (non-syslog) log file, for instance if you rotated the logs and want thttpd to start using the new one. However, this feature isn't actually that useful at the moment. The problem is that thttpd will generally be started as root, so that it can bind to port 80; then it gives up the root uid as soon as it can, for security reasons. But if you later send it a HUP, it will try to re-open the log file without root access and will generally fail. Also, if you're running inside a chroot tree, as you should be, the log file won't even be accessible. Currently the best alternative for log rotation is to send a USR1 signal, shutting down thttpd altogether, and then restart it.

thtpasswd(1), syslogtocern(8)

Many thanks to contributors, reviewers, testers: John LoVerso, Jordan Hayes, Chris Torek, Jim Thompson, Barton Schaffer, Geoff Adams, Dan Kegel, John Hascall, Bennett Todd, KIKUCHI Takahiro, Catalin Ionescu. Special thanks to Craig Leres for substantial debugging and development, and for not complaining about my coding style very much.

Copyright © 1995,1998,1999,2000 by Jef Poskanzer <jef@acme.com>. All rights reserved.