I had an observation; it seems that your support for native C++ is a limiting factor to the Hurd. One thing to remember is that I do not have any experience in Operating Systems. Second thing, forgive me if this is not the right address for this subject.
My understanding is that one thing that hurts a multi-server microkernel is the context switch. The need to store registers, stack points, copy message buffers etc. Ideally a call of any sort would be just that a function call. It seems to me that C++ prevents this, its ability to reference any memory in its memory space, and read or write that memory makes a simple function call to the kernel impossible. The basic scenario is that program A calls kernel nonblocking function B. Then right after invoking B program A modifies the data sent as parameters to B. At this point function B has no way to guarantee the correctness of the parameters passed to it. It must make a copy of those parameters so that program A can not modify the parameters after the function has validated them. My general design question, what if you moved C++ to emulation? The native language would then be a programming language that did not have random access to the entire memory space. I'll use Java as an example. In this scenario we'll have the same program A and function B except this time we'll name the parameter C. Program A would create C, initialize C and call B with C. Once B is called it locks C. This makes it so that C can no longer be changed. Now B does not have to make a copy of C but just needs to verify it. Seems faster. Also it seems at this point you would not need a separate memory space for each program. At this point a program only has access to the objects it created or that were handed to it via reference. So if one program never hands away the references to its objects then those objects are in effect in their own memory space immune to the actions of any other program. Since there is only one memory domain the need for a context switch goes away. In order to call through to the kernel there is no need to context switch to get into the kernel's memory space. All the program would need is a reference to an object defining the kernel. That object would manage the references that make up the kernel. In reality the kernel would be made up of any number of user level servers (wouldn't actually be kernel level anything?). I could go on with some other stuff, but I am not sure if I am crazy or if this is even the right audience. Mike Power
