The random rantings of a concerned programmer.

Archive for July, 2011

Invoking mount(2) in FreeBSD 8.x

July 17th, 2011 | Category: Random

So I’m still writing Go bindings for a lot of common FreeBSD functionality. Yesterday I implemented a means to list all mounted filesystems, so today I’m writing the bindings to mount(2) to mount/umount them.

If you look at the man page for mount, you’ll see that the function signature looks like this:

    int mount(const char *type, const char *dir, int flags, void *data);

The void* should scare you.

I haven’t been able to dig up any information about what the fuck should be passed to it (granted, I haven’t looked very hard because, judging from the contents of src/sbin/mount_*/*.c in the FreeBSD sources, it’s been entirely superseded by nmount.

    int nmount(struct iovec *iov, u_int niov, int flags);

Poking around, struct iovec (eventually included from sys/uio.h) is defined as this:

struct iovec {
    void *iov_base;
    size_t iov_len;
}

Effectively, nmount takes an array of these structs which effectively form a flattened vector of (key, value) tuples. As far as I can tell, iov_base is always a NULL-terminated char*, and iov_len should be strlen(iov_base) + 1 (for the NULL terminator).

Unfortunately, the only hints that man 2 nmount gives us is

The following options are required by all file
     systems:
	   fstype     file system type name (e.g., ``procfs'')
	   fspath     mount point pathname (e.g., ``/proc'')

     Depending on the file system type, other options may be recognized or
     required; for example, most disk-based file systems require a ``from''
     option containing the pathname of a special device in addition to the
     options listed above.

So far, the only way I’ve been able to find the actual options is to dig through mount_* sources and see what they use, but it’s pretty gross. Take, for example, the following two filesystems:

  • nullfs simply layers one vnode on top of another, effectively grafting one directory over another.
  • unionfs (roughly) does the same thing, but still lets you access the grafted-over directory in read-only mode (and can be configured to do cool shit like copy-on-write).

They’re pretty close, but let’s look at the arguments that each of them take:

nullfs
  • fstype: “nullfs”
  • fspath: Path to the directory to graft over.
  • target: Path of the directory that’s being grafted onto another.

IMHO, "target" should be "from", bikesheds, et. al.

unionfs
  • fstype: “unionfs”
  • fspath: Path to the directory where the unionfs will be mounted.
  • from: Same as “target”, above.
  • below: Makes “fspath” writable, “from” read-only (swaps default behavior)
  • errmsg: …I have no fucking idea, a char[255] which presumably is used as a buffer instead of errno?
  • …anything else passed as -oyour=mom passed to mount_unionfs?!

Maybe this is more a gripe that unionfs seems to be very shitty. And maybe I just haven’t found a nice magical table of options that every filesystem takes. But FFFFFF SERIOUSLY >:(

3 comments

getmntinfo(2) from Go — a foray into cgo

July 16th, 2011 | Category: Random

Go is a fun esoteric language that strives for system-level usage. Currently in all real operating systems, C is the dominant systems language and as such, all the functionality for interfacing with core features are exposed as raw C APIs. Go provides a C FFI layer called cgo, which handles all the preprocessing and linking magic in the background. Unfortunately, there’s little-to-no documentation available for cgo, just a couple of toy examples in Go’s misc/cgo directory (there’s actually a shitton of production examples in the Go package sources though — fucking everything uses cgo).

So, what I want to do is expose getmntinfo, which simply lists the metadata for all mounted filesystems. In C, this is pretty trivial:

#include <sys/param.h>
#include <sys/ucred.h>
#include <sys/mount.h>

#include <stdio.h>

int main() {
	struct statfs *bufs;
	int i = getmntinfo(&bufs, 0);
	int j = 0;

	for (j = 0; j < i; ++j) {
		struct statfs fs = bufs[j];
		printf("[%s] %s -> %s\n", fs.f_fstypename, 
			fs.f_mntfromname, fs.f_mntonname);
	}

	return 0;
}

This, however, presents a variety of problems for the Go implementation –

  1. We don’t really know how many struct statfs we’re getting back.
  2. The memory allocated is actually allocated statically; we just get an opaque pointer back to an in-library address.
  3. The fields of struct statfs are char[N]s rather than char*s.

Thankfully, calling getmntinfo is pretty trivial –

func GetMntInfo() []MntInfo {
	var tmp *C.struct_statfs;
	i := int(C.getmntinfo(&tmp, 0))

It’s pretty close to the C version — we just allocate a pointer, and pass a pointer to it in. getmntinfo sets the value of the pointer to an internal array of struct statfs‘s and lets us go along our merry way. Naturally, we want to marshal it to the appropriate Go types.


	info := make([]MntInfo, i)
	for j, _ := range(info) {

So we create an array to marshal values into and begin to iterate through it.

This is where it gets nasty. All we have right now is an opaque pointer to a struct statfs — in C we’d just use pointer arithmetic to get the other entries in the array. Go, fortunately, explicitly disallows pointer arithmetic. I’m not sure what the appropriate method to get values out of it is. First, I tried something like


foo := (*[]MntInfo)(unsafe.Pointer(tmp))
item := (*foo)[j]

But that seems to cause a panic (no idea why). I got tired of dicking with it and threw in the cards, simply exposing the following C function in the cgo header –


struct statfs* offset(struct statfs *v, int i) {
	return v + i;
}

With that, there’s no need to dick with much of anything, so we can get the current struct statfs of the iteration pass via


		s := C.offset(tmp, C.int(j))

Finally, the char[16] values need to be marshaled out. Unfortunately, the C.GoString marshaling function only takes a char* and it’s too damn stubborn to take an implicitly-convertible type (noting that X* != X[]). The other beef is that cgo’s type system processes a char[] strangely as a []_C_char_type, so we can index it perfectly fine (but not implicitly coerce it into a pointer).

So we juggle some types and shit all over unsafe.Pointer and make it do what we want –

		info[j].FsType = C.GoString((*C.char)
			(unsafe.Pointer(&s.f_fstypename[0])))
		info[j].MntFrom = C.GoString((*C.char)
			(unsafe.Pointer(&s.f_mntfromname[0])))
		info[j].MntOn = C.GoString((*C.char)
			(unsafe.Pointer(&s.f_mntonname[0])))
	}

	return info
}

And, after several hours of not finding any fucking documentation and screaming at the fucking monitor the damn thing finally works. I’m completely glossing over the terrible shitty build system they’ve got set up (it basically only provides functionality to INSTALL to built cgo packages — I haven’t found a way to actually build and link them otherwise) — will probably have to read through all the fucking makefiles that do evil shit.

At some point just doing everything in C is easier, I suspect :|


will post full code listing in a sec

4 comments

Calling a templated member function of a typedef’d template class

July 11th, 2011 | Category: Random

C++ is insane.

Assume you have a templated Object:

template  
struct Object {
	template  void func(){};
};

And you want to wrap up the instance in a Proxy object:

template 
struct Proxy {
	typedef Object WrappedType;
	WrappedType obj;

	static void Func() {
		Proxy *self = new Proxy;
		self->obj.func();
	}
};

Pretty straightforward, but when you actually try to invoke Proxy::Func on an arbitrary T using g++

struct Foo {};

int main() {
	Proxy::Func();
	return 0;
}

g++ shits itself completely:

$ g++ test1.cpp 
test1.cpp: In static member function ‘static void Proxy::Func()’:
test1.cpp:13: error: ‘Foo’ was not declared in this scope
test1.cpp:13: error: expected primary-expression before ‘)’ token
$ g++ --version
i686-apple-darwin10-g++-4.2.1 (GCC) 4.2.1 (Apple Inc. build 5666) (dot 3)

Fucking fantastic.

Some tinkering reveals that the compiler is getting confused as to what the fuck obj.func is somewhere. The following implementation of Func works fine (but defeats the point of using templates) --

	static void Func() {
		Proxy *self = new Proxy;
		Object bar = self->obj;
		bar.func();
	}

I searched for awhile and turned up jack diddly squat, then a co-worker informed me the fix is to use the following:

	static void Func() {
		Proxy *self = new Proxy;
		self->obj.template func();
	}

I don't know what the fuck this instance.template function<..>() bullshit is, but apparently MSVC implicitly puts it in there for you. I've certainly never seen it before and it's completely orthogonal to any fix I would have assumed.

tl;dr C++ is a clusterfuck.


EDIT: A stack overflow post which contains a reference to the C++03 standard (14.2/4) in the answers. fml.

4 comments