Slashdot Mirror
Could Linux Become A Microkernel?
Kris Warkentin writes: "This question is not entirely intended to start a debate about the pros and cons of microkernels vs. monolithic ones. What I would really like to know, however, is how _feasible_ it would be to convert the Linux kernel to a microkernel. I was looking at how the QNX kernel offers only core services like threading, IPC, process creation, memory management, initial interrupt handling, etc. Everything else functions as a process within its own memory space. Linux can be configured so that it is much like this with other things (filesystems, etc.) compiled as modules. The key difference is that all the modules are operating in kernel space. So, the question is, how difficult do you think it would be to devise a communication protocol to let modules function outside of kernel space and merely talk to the kernel? What would be the cost and benefits? Would it be possible to have both types in the same source tree? (say, as a compile option)"
No. In NT 3.5.x, just as in 4.0 and, as far as I know, 5.0^H^H^HW2K, drivers for devices such as disk and network controllers, and network protocol implementations run in kernel mode in a fashion similar to the way they function in various UNIX systems.
Yes, rendering was done in 3.x by sending messages to the Win32 subsystem process...
...but if that's sufficient to make NT a microkernel, then, well, err, umm, Linux - or {Free,Net,Open}BSD, or Solaris, or HP-UX, or AIX, or Digital UNIX, or... - are also microkernels if they're running X; in systems running X, the rendering code runs in "a process within its own memory space", i.e. the X server, in user mode.
BTW, "the GUI" runs, in part, in user mode even on NT 4 - the low-level rendering is done in kernel drivers, but the toolkit - the equivalent of Motif or GTK+ or Qt or... - lives, as far as I know, in user32.dll, which is a library that calls routines in gdi32.dll to get stuff rendered.
user32.dll is, as far as I know, just user-mode library code, as is gdi32.dll; on 3.x, gdi32.dll sent messages to the Win32 subsystem process, and, in 4.x and later, it goes through the kernel driver in at least some cases. (The fact that it's a shared library means that binaries built for 3.x should just continue to work - the ABI for drawing stuff on the screen is, in effect, a bunch of "call routine XXX in this library, with these arguments" items, and the way routine XXX accomplishes that can change from release to release without affecting programs that don't go around the back of the library.)
(user32.dll probably roughly corresponds to your toolkit library or libraries in X, and gdi32.dll probably rougly corresponds to Xlib, although there may be differences.)
Others have mentioned MkLinux, which is a version of Linux which runs on top of the Mach microkernel. By modern standards, Mach isn't so "micro". On my Hurd partition, the gnumach executable weighs in at 726kb compressed, and about 1.6Mb uncompressed. Compare with ntoskrnl.exe, which is 907kb on NT 4.0 enterprise server. Both of these are comparable with the size of an average linux or BSD monolithic kernel, which sit around the megabyte mark uncompressed.
The QNX kernel, on the other hand, is something like 8kb in size, which fits in the cache of a 486. Even the BeOS kernel is only something like 78kb compressed. Not that size is the only concern (so my wife keeps telling me), but in general, the less code that runs in the kernel, the easier it is to say something about how secure it is. Also the easier it is to change things while the system is running.
I hate to sound like Andrew Tanenbaum, but MkLinux and the Hurd are now obsolete too. Mach belongs to the old school of microkernels which were popular 10-15 years ago, but with the benefit of hindsight, we know better. Nowadays, for example, we know that you don't even need to do VM swapping inside the kernel.
There are some projects of note which may result in a product which is cleaner and better designed than Linux. Here are some suggestions:
sub f{($f)=@_;print"$f(q{$f});";}f(q{sub f{($f)=@_;print"$f(q{$f});";}f});