kcgi is an open source CGI and FastCGI library for C web applications. It is minimal, secure, and auditable. To start, install the library then read the usage guide. Contact Kristaps with questions or comments. kcgi is a BSD.lv project.
First, check if kcgi isn't already a third-part port for your system, such as for OpenBSD or FreeBSD. If so, install using that system.
If not, you'll need a modern UNIX system.
To date, kcgi has been built and run on GNU/Linux machines, BSD (OpenBSD, FreeBSD), and Mac OSX (Snow Leopard, Lion) on i386 and AMD64.
It has been deployed under Apache, nginx, and OpenBSD's httpd(8)
(the latter two natively over FastCGI and via the
Begin by downloading kcgi.tgz and verify the archive with kcgi.tgz.sha512.
Once downloaded, compile the software with
make, which will automatically run a
configuration script to conditionally deploy portability glue.
Finally, install the software using
make install, optionally specifying the
PREFIX if you don't intend to use /usr/local.
If kcgi doesn't compile, please send me the config.log
file and the output of the failed compilation.
If you're running on an operating system with an unsupported sandbox, let me know and we can work
together to fit it into the configuration and portability layer.
Lastly, I'd love to compile with mingw for Microsoft machines:
please contact me if you can do the small amount of work (I think?) to port the
other non-Microsoft functions.
The kcgi manpages, starting with kcgi(3), are the canonical source of documentation. The following are introductory materials to the system.
Applications using kcgi behave just like any other application.
To compile kcgi applications, just include the kcgi.h
header file and make sure it appears in the compiler inclusion path.
(According to C99, you'll need to include stdint.h before it for the
int64_t type used for parsing integers.)
Linking is similarly normative: link to libkcgi and, if your system has
compression support, libz.
Well-deployed web servers, such as the default OpenBSD server, by
default are deployed within a chroot(2). If
this is the case, you'll need to statically link your binary.
If running within a chroot(2) and on
OpenBSD, be aware that the sandbox method requires /dev/systrace within the
By default, this file does not exist in the web server root.
Moreover, the default web server root mount-point, /var, is mounted
This complication does not exist for the other sandboxes.
FastCGI applications may either be started directly by the web server (which is popular with Apache) or
externally given a socket and kfcgi(8) (this method is normative for OpenBSD's httpd(8) and
suggested for the security precautions taken by the wrapper).
The bulk of kcgi's CGI handling lies in khttp_parse(3), which fully parses the HTTP request. Application developers must invoke this function before all others. For FastCGI, it's split between khttp_fcgi_init(3), which initialises context; and khttp_fcgi_parse(3), which receives new parsed requests. In either case, requests must be freed by an khttp_free(3).
All functions isolate the parsing and validation of untrusted network data within a sandboxed child process. Sandboxes limit the environment available to a process, so exploitable errors in the parsing process (or validation with third-party libraries) cannot touch the system environment. This parsed data is returned to the parent process over a socket. In the following, the HTTP parser and input validator manage a single HTTP request, while connection delegator accepts new HTTP requests and passes them along.
This method of sandboxing the untrusted parsing process follows OpenSSH, and requires special handling for each operating system:
- seccomp(2) (Linux)
This requires a fairly new kernel (≥Linux 3.5).
It is supplemented by
setrlimit(2)limiting. For the time being, this feature is only available for x86, x86_64, and arm architectures. If you're using another one, please send me your
uname -mand, if you know if it, the correct
- systrace(4) (OpenBSD)
This requires the existence of /dev/systrace if running in a
chroot(2), which is strongly suggested. If you're using a stock OpenBSD, make sure that the mount-point of /dev/systrace isn't mounted
- sandbox_init(3) (Apple OSX)
This uses the sandboxing profile for
pure computationas provided in Mac OS X Leopard and later. This is supplemented by resource limiting via
- capsicum(4) (FreeBSD)
Uses the capabilities facility on FreeBSD 10 and later.
This is supplemented by resource limiting with
Since validation occurs within the sandbox, special care must be taken that validation routines don't access the environment (e.g., by opening files, network connections, etc.), as the child might be abruptly killed by the sandbox facility. (Not all sandboxes do this.) If required, this kind of validation can take place after the parse validation sequence.
The connection delegator is similar, but has different sandboxing rules, as it must manage an open socket connection and respond to new requests.
kcgi is shipped with a fully automated testing framework executed with
Interfacing systems can also make use of this by working with the kcgiregress(3) library.
This framework acts as a mini-webserver, listening on a local port, translating an HTTP document into a
minimal CGI request, and passing the request to a kcgi CGI client.
For internal tests, test requests are constructed with libcurl.
The automated test framework, at the moment, only has a few tests for basic functionality and sandboxing. The binding local port is fixed, too; so if you plan on running the regression suite, you may need to tweak its access port.
Another testing framework exists for use with the American
To use this, you'll need to compile the
make afl target with your compiler of choice, e.g.,
make clean, then
make afl CC=afl-gcc.
Then run the
afl-fuzz tool on the
afl-urlencoded binaries using the test cases (and dictionaries, for the first) provided.
Security comes at a price. By design, kcgi incurs overhead in three ways: first, spawning a child to process the untrusted network data; second, enacting the sandbox framework; and third, passing parsed pairs back to the parent context.
This figure illustrates the cost of running kcgi against a baseline on OpenBSD 5.5 running nginx and slowcgi(8), i386. It shows the empirical cumulative distribution of a statisically-significant number of page requests (>1000) as measured by ab(1).
Line (1) shows a static file being served by the web server. The high speed is due to the file (most likely) being cached by the web server and/or kernel. Moving right, line (2) shows a basic CGI request producing no content. The CGI simply exits with an HTTP 200. This reflects the cost of invoking slowcgi(8). Line (3) is a simple kcgi that simply emits an HTTP 200. This instance is neither sandboxed nor does it compress output. The overhead from line (2) is due to the additional child being spawned. Line (4) shows the addition of sandboxing without compression, in this case via systrace(4). Lastly, line (5) shows both compression and sandboxing.
In this bar chart, I show the same but for relative comparison of the distribution means. The similarity between the compressed and non-compressed versions are due to the small amount of data being transmitted in the response body.