💾 Archived View for perso.pw › blog › articles › vger-security.gmi captured on 2023-05-24 at 18:19:16. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2023-01-29)
-=-=-=-=-=-=-
I would like to share about Vger internals in regards to how the security was thought to protect vger users and host systems.
I claim about security in Vger as its main feature, I even wrote Vger to have a secure gemini server that I can trust. Why so? It's written in C and I'm a beginner developer in this language, this looks like a scam.
I chose to follow the best practice I'm aware of from the very first line. My goal is to be sure Vger can't be used to exfiltrate data from the host on which it runs or to allow it to run arbirary command. While I may have missed corner case in which it could crash, I think a crash is the worse that can happen with Vger.
Vger doesn't have to manage connections or TLS, this was a lot of code already removed by this design choice. There are better tools which are exactly made for this purpose, so it's time to reuse other people good work.
Vger is run by inetd daemon, allowing to choose the user running vger. Using a dedicated user is always a good idea to prevent any harm in case of issue, but it's really not sufficient to protect vger to behave badly.
Another kind of security benefit is that vger runtime isn't looping like a daemon awaiting new connections. Vger accept a request, read a file if exist and gives its result and terminates. This is less error prone because no variable can be reused or tricked after a loop that could leave the code in an inconsistent or vulnerable state.
A critical vger feature is the ability to chroot into a directory, meaning the directory is now seen as the root of the file system (/var/gemini would be seen as /) and prevent vger to escape it. In addition to the chroot feature, the feature allow vger to drop to an unprivileged user.
/*
* use chroot() if a user is specified requires root user to be
* running the program to run chroot() and then drop privileges
*/
if (strlen(user) > 0) {
/* is root? */
if (getuid() != 0) {
syslog(LOG_DAEMON, "chroot requires program to be run as root");
errx(1, "chroot requires root user");
}
/* search user uid from name */
if ((pw = getpwnam(user)) == NULL) {
syslog(LOG_DAEMON, "the user %s can't be found on the system", user);
err(1, "finding user");
}
/* chroot worked? */
if (chroot(path) != 0) {
syslog(LOG_DAEMON, "the chroot_dir %s can't be used for chroot", path);
err(1, "chroot");
}
chrooted = 1;
if (chdir("/") == -1) {
syslog(LOG_DAEMON, "failed to chdir(\"/\")");
err(1, "chdir");
}
/* drop privileges */
if (setgroups(1, &pw->pw_gid) ||
setresgid(pw->pw_gid, pw->pw_gid, pw->pw_gid) ||
setresuid(pw->pw_uid, pw->pw_uid, pw->pw_uid)) {
syslog(LOG_DAEMON, "dropping privileges to user %s (uid=%i) failed",
user, pw->pw_uid);
err(1, "Can't drop privileges");
}
}
Vger only requires standard C includes, this avoid leaving trust to dozens of developers using fragile or barely tested code.
In addition to all the previous security practices, OpenBSD is offering a few functions to help restricting a lot what Vger can do.
The first function is pledge, allowing to restrict the system calls that can happen within the code itself. The current syscalls allowed in vger are related to the categories "rpath" and "stdio", basically standard input/output and reading files/directories only. This mean after pledge() is called, if any syscall not in those two categories is used, vger will be killed and a pledge error will be reported in the logs.
The second function is unveil, which will basically restrict access to the filesystem to anything but what you list, with the permission. Currently, vger only allows file access in read-only mode in the base directory used to serve files.
Here is an extract of the code relative to the OpenBSD specific code. With unveil available everywhere chroot wouldn't be required.
#ifdef __OpenBSD__
/*
* prevent access to files other than the one in path
*/
if (chrooted) {
eunveil("/", "r");
} else {
eunveil(path, "r");
}
/*
* prevent system calls other parsing queryfor fread file and
* write to stdio
*/
if (pledge("stdio rpath", NULL) == -1) {
syslog(LOG_DAEMON, "pledge call failed");
err(1, "pledge");
}
#endif
I made my best to use the least code possible before reducing Vger capabilities. Only the code managing the parameters is done before activating chroot and/or unveil/pledge.
int
main(int argc, char **argv)
{
char request [GEMINI_REQUEST_MAX] = {'\0'};
char hostname [GEMINI_REQUEST_MAX] = {'\0'};
char uri [PATH_MAX] = {'\0'};
char user [_SC_LOGIN_NAME_MAX] = "";
int virtualhost = 0;
int option = 0;
char *pos = NULL;
while ((option = getopt(argc, argv, ":d:l:m:u:vi")) != -1) {
switch (option) {
case 'd':
estrlcpy(chroot_dir, optarg, sizeof(chroot_dir));
break;
case 'l':
estrlcpy(lang, "lang=", sizeof(lang));
estrlcat(lang, optarg, sizeof(lang));
break;
case 'm':
estrlcpy(default_mime, optarg, sizeof(default_mime));
break;
case 'u':
estrlcpy(user, optarg, sizeof(user));
break;
case 'v':
virtualhost = 1;
break;
case 'i':
doautoidx = 1;
break;
}
}
/*
* do chroot if a user is supplied run pledge/unveil if OpenBSD
*/
drop_privileges(user, chroot_dir);
Unix is made of small component that can work together as small bricks to build something more complex. Vger is based on this idea by delegating the listening daemon handling incoming requests to another software (let's say relayd or haproxy). And then, what's left from the gemini specs once you delegate TLS is to take account of a request and return some content, which is well suited for a program accepting a request on its standard input and giving the result on standard ouput. Inetd is a key here to make such a program compatible with a daemon like relayd or haproxy. When a connection is made into the TLS listening daemon, a local port will trigger inetd that will run the command, passing the network content to the binary into its stdin.
CGI support was added in order to allow Vger to make dynamic content instead of serving only static files. It has a fine grained control, you can allow only one file to be executable as a CGI or a whole directory of files. When serving a CGI, vger forks, a pipe is opened between the two processes and a process is using execlp to run the cgi and transmit its output to vger.
From the beginning, I wrote a set of tests to be sure that once a kind of request or a use case work I can easily check I won't break it. This isn't about security but about reliability. When I push a new version on the git repository, I am absolutely confident it will work for the users. It was also an invaluable help for writing Vger.
As vger is a simple binary that accept data in stdin and output data on stdout, it is simple to write tests like this. The following example will run vger with a request, as the content is local and within the git repository, the output is predictable and known.
printf "gemini://host.name/autoidx/\r\n" | vger -d var/gemini/
From here, it's possible to build an automatic test by checking the checksum of the output to the checksum of the known correct output. Of course, when you make a new use case, this requires manually generating the checksum to use it as a comparison later.
OUT=$(printf "gemini://host.name/autoidx/\r\n" | ../vger -d var/gemini/ -i | md5) if ! [ $OUT = "770a987b8f5cf7169e6bc3c6563e1570" ] then echo "error" exit 1 fi
At this time, vger as 19 use case in its test suite.
By using the program `entr` and a Makefile to manage the build process, it was very easy to trigger the testing process while working on the source code, allowing me to check the test suite only by saving my current changes. Anytime a .c file is modified, entr will trigger a make test command that will be displayed in a dedicated terminal.
ls *.c | entr make test
Realtime integration tests? :)
By using best practices, reducing the amount of code and using only system libraries, I am quite confident about Vger good security. The only real issue could be to have too many connections leading to a quite high load due to inetd spawning new processes and doing a denial of services. This could be avoided by throttling simultaneous connection in the TLS daemon.
If you want to contribute, please do, and if you find a security issue please contact me, I'll be glad to examine the issue.