22.7 Advanced Topics

22.7.1 How Does It Work?

FreeBSD has an abstraction called an ``execution class loader''. This is a wedge into the execve(2) system call.

What happens is that FreeBSD has a list of loaders, instead of a single loader with a fallback to the #! loader for running any shell interpreters or shell scripts.

Historically, the only loader on the Unix platform examined the magic number (generally the first 4 or 8 bytes of the file) to see if it was a binary known to the system, and if so, invoked the binary loader.

If it was not the binary type for the system, the execve(2) call returned a failure, and the shell attempted to start executing it as shell commands.

The assumption was a default of ``whatever the current shell is''.

Later, a hack was made for sh(1) to examine the first two characters, and if they were :\n, then it invoked the csh(1) shell instead (we believe SCO first made this hack).

What FreeBSD does now is go through a list of loaders, with a generic #! loader that knows about interpreters as the characters which follow to the next whitespace next to last, followed by a fallback to /bin/sh.

For the Linux ABI support, FreeBSD sees the magic number as an ELF binary (it makes no distinction between FreeBSD, Solaris, Linux, or any other OS which has an ELF image type, at this point).

The ELF loader looks for a specialized brand, which is a comment section in the ELF image, and which is not present on SVR4/Solaris ELF binaries.

For Linux binaries to function, they must be branded as type Linux; from brandelf(1):

    # brandelf -t Linux file

When this is done, the ELF loader will see the Linux brand on the file.

When the ELF loader sees the Linux brand, the loader replaces a pointer in the proc structure. All system calls are indexed through this pointer (in a traditional Unix system, this would be the sysent[] structure array, containing the system calls). In addition, the process flagged for special handling of the trap vector for the signal trampoline code, and sever other (minor) fix-ups that are handled by the Linux kernel module.

The Linux system call vector contains, among other things, a list of sysent[] entries whose addresses reside in the kernel module.

When a system call is called by the Linux binary, the trap code dereferences the system call function pointer off the proc structure, and gets the Linux, not the FreeBSD, system call entry points.

In addition, the Linux mode dynamically reroots lookups; this is, in effect, what the union option to FS mounts (not the unionfs!) does. First, an attempt is made to lookup the file in the /compat/linux/original-path directory, then only if that fails, the lookup is done in the /original-path directory. This makes sure that binaries that require other binaries can run (e.g., the Linux toolchain can all run under Linux ABI support). It also means that the Linux binaries can load and exec FreeBSD binaries, if there are no corresponding Linux binaries present, and that you could place a uname(1) command in the /compat/linux directory tree to ensure that the Linux binaries could not tell they were not running on Linux.

In effect, there is a Linux kernel in the FreeBSD kernel; the various underlying functions that implement all of the services provided by the kernel are identical to both the FreeBSD system call table entries, and the Linux system call table entries: file system operations, virtual memory operations, signal delivery, System V IPC, etc... The only difference is that FreeBSD binaries get the FreeBSD glue functions, and Linux binaries get the Linux glue functions (most older OS's only had their own glue functions: addresses of functions in a static global sysent[] structure array, instead of addresses of functions dereferenced off a dynamically initialized pointer in the proc structure of the process making the call).

Which one is the native FreeBSD ABI? It does not matter. Basically the only difference is that (currently; this could easily be changed in a future release, and probably will be after this) the FreeBSD glue functions are statically linked into the kernel, and the Linux glue functions can be statically linked, or they can be accessed via a kernel module.

Yeah, but is this really emulation? No. It is an ABI implementation, not an emulation. There is no emulator (or simulator, to cut off the next question) involved.

So why is it sometimes called ``Linux emulation''? To make it hard to sell FreeBSD! Really, it is because the historical implementation was done at a time when there was really no word other than that to describe what was going on; saying that FreeBSD ran Linux binaries was not true, if you did not compile the code in or load a module, and there needed to be a word to describe what was being loaded--hence ``the Linux emulator''.