Adding System Calls (an OpenBSD Example)

BSD Forums writes "Kernel programming sometimes feels like a dark art where application programmers should never venture, but sometimes it's the right way to solve a problem. (Oh, and it's also very interesting.) One of the easiest places to start is by adding a new system call to a kernel. Kevin Lo explains how and why, with the OpenBSD kernel in this OnLamp article."

Comments

other examples

[fred ugly]

A simple Google search brings up plenty of info on how to do this in Linux as well...

Re:other examples

[smittyoneeach]

A useful first post. An hour later, it's still lonely. Why?

Re:other examples

[fred ugly]

yeah, i know... and i didn't even think to mention that it was a FP. oh well, my karma shouldn't suffer terribly.

Interesting, however....

[Anonymous Coward]

I think my last attempt at 'hello-world' may have introduced a local root hole. Kernel hacking is probably knowledge I shouldn't have ... I'm sure I'm not alone =)

Sometimes this is more trouble then it seems.

[John Sokol]

I have done this several times in the past and what happens is you then have an application trapped on a non-standard system. Any system that it runs that application will need your kernal patch with the new calls.

This gets even worse when the OS keeps upgrading and you are forced to migrate your changes up to be able to use current hardware. (this is need when all of the supported hardware is no longer available new)

In practice it turn into a major undertaking everytime a new OS release came out.

It's is a cool thing to have fun with. But think twice before you base a product or application on a kernel change. (unless you can get the main development tree to adopt it.) Or have the resources to maintain your own OS development team.

Adding new syscontrols and sockopts are also great fun.

At one point we had a versions of FreeBSD that could run DES encrypted Binaries, access the hard driver serial numbers and Mount a CD from HTTP or FTP connection, transmit Datagrams masquerading as TCP connections and be able to process Router Alert packets.

John

Re:Sometimes this is more trouble then it seems.

[SuperFrink]

An excellent point.

I wrote patches to the linux kernel and apache to allow all PHP and CGI requests to run as the owner of the file being requested. Actually the UID changes for each request, even static pages. So long as a stat() can be done it should work. This means the files don't have to be readable by the webserver.

The system stores the user ID on a stack and returns to the webserver user ID when the request has been served so no forking or set-uid files are needed. (Yes root needs to configure permissions at boot time or at least before switching users will work.)

I was explaining the system to someone at work (where we look after several webservers used by different people). When asked weather I intended to install this on the work servers. My reply was no, if it were my company I would definitely consider it but I did not want to saddle the next admin after me with the job of keeping a custom kernel patch in line with future releases.

That said I say if you think you might be interested in operating systems then doing some reading and eventually playing with kernel code can be a valuable experience. (I also wrote the first result from the above google search. The second link is the more current version.)

-Chad

PS: the DES binaries sound like a neat project. for that mater so does mounting over http.

Syscalls

[wazlaf]

Awesome, we'll get even more syscalls! Linus is going to love this..

Re:Syscalls

[Homology]

Awesome, we'll get even more syscalls! Linus is going to love this..

The article is about syscalls in the OpenBSD kernel. I don't think that Linus is very interested in recieving patches for new syscalls in a BSD kernel.

Eternal pet peeve

[Emil Brink]

Now, I know the code in the article is only sample code, and as such need not be perfect as long as it gets its point across. I'd say it does that just fine, it was an interesting read. But I can't help feel a bit saddened by code such as this snippet:

size_t len;
char buffer[1024]; /* must bounds check all user values */
if (SCARG(uap, len) > (size_t)1024)
return(EINVAL);

Here, the author needlessly repeats the "1024" constant, which introduces a fine opportunity to make an error by only changing the buffer definition and not the constant, later on. In my world, cases such as these is where you really want to use the sizeof operator, and simply write that check like so:

if (SCARG(uap, len) > sizeof buffer)
return EINVAL;

Also note that this gets rid of the cast, since the sizeof creates a value of type size_t. There are no parens involved, since neither sizeof nor return are function calls. This seems to be a matter of personal style, though. Of course, sizeof is properly used in the copying call, so I guess this is just a minor point anyway. I also suspect there's an off-by-one error lurking in there, but am not familiar enough with the OpenBSD copying semantics to be sure.

Re:Eternal pet peeve

[Viol8]

" There are no parens involved, since neither sizeof nor return are function calls. This seems to be a matter of personal style, though"

Its not just a case of personal taste , using sizeof with parenthesis will only work for types that only have a 1 token declaration

Eg: sizeof int

works fine , whereas

sizeof int * sizeof struct foo

will give an error with most compilers. So in this sense the sizeof operator is not treated
syntatically in an equivalent way to the return operator. Just another of C's little foibles.

Re:Eternal pet peeve

[Viol8]

Missed out a newline

sizeof int *
sizeof struct foo

should have been seperate.

Re:Eternal pet peeve

[Emil Brink]

I wasn't being complete since I wanted to, er, not sound too dorky. The sizeof operator takes a single argument, that's true. If the argument is to be a type name, such as int or your struct foo, then that type name needs to be a cast expression, which as we all know is simply the type name in parens. Using a type name without parens is, as far as I know, a syntax error.