Linus Torvalds (torvalds@osdl.org) on 5/15/06 wrote:
>Shapiro (and to some degree Tanenbaum) also makes the
>mistake of equating different address spaces with the
>notion of modularity. They have nothing to do with each
>other.

This is the most sensible thing about this "micro-vs-mono" debate that anyone has said so far. However, it's important not to ignore the protection aspect of having separate address spaces, so I'd say something more like (modularity + isolation != address spaces).

> You can be modular without using hardware to enforce
>it

Yes! And you can have isolation without the hardware assist too -- for example, Singularity.

> and you can generate a horribly messy system where
>two processes are intimately aware of how each other works
>even if they are separated by MMU boundaries.

Yes again -- for example, MINIX prior to version 3 (I haven't looked at MINIX 3 yet, so I can't comment on it).

>As I already alluded to (and expanded on in my reply to
>myself), I actually think using "designer languages" may
>be a much better solution to the modularity problem than
>the microkernel approach.

By "designer languages", are you making a veiled reference to Spec#? ;-)

I agree that the language level seems to be the most sensible place to solve the problem of performance versus modularity and isolation. The time for this approach has obviously not come yet, since there is a huge amount of existing code out there that would take decades to rewrite in a "designer language", but I can see such an idea taking off eventually.

However, I disagree that hardware-enforced protection boundaries are a necessary part of the definition of "microkernel". As you said before, modularity and separate address spaces are orthogonal concepts. Also, separate address spaces are not a precondition for process isolation, as systems like Singularity demonstrate.

I think it should be obvious that modularity is the key to good software engineering, and isolation is necessary for reliability, so perhaps the "hardware-isolated ukernel" is just an ugly intermediate step on the way to a much better idea.

[Disclaimer: I am not an OS researcher. I work on commercial database software for a living, not OSes. I'm not sure how this affects the dimensions of the grain of salt with which you take my opinions. :-)]