"Ah, but in kernel land we have preemption and reentrency (just to prove it, even with a fully frozen desktop, network and disc drive still work). So Acorn manage to make kernel level things with PMT, but not user applications ? They manage to make RISC iX with PMt but not ROS ?"
You're confusing modules (which work in the background, and are not PMTed as such - they're event driven CMTed) and processes. Much of RISC OS's kernel is not re-entrent at all. FileCore is one big irritating example, as are lots of other bits.
If, for example, you're in an OS_GBPB call to read in 2GB of data say, into one block (doesn't matter if this is actually possible or not, it's just an example.) pretty much the whole system stops until it's done. Under a pre-emptable, re-enterant kernel, other processes could still run happily, and use OS_GBPB themselves (perhaps even to read data from the same file.). It appears that you're confused on what kernel pre-emption is.
"Ah no, you cannot make a predictable system with PMT. At the best you'll make a soft or a hard RT system... That's not 100% predicatable. Why do you think Montavista tried to put different schedulers inside Linux kernel?"
Errr, I'm quite sure you're mad. Real Time operating systems are *all* about being predictable. If you think they're not predictable, that QNX installation at your local nuclear power plant, or WindRiver VxWorks in that cruise missile are in deep shit.
If what you mean is you can't predict what process is running at any given point, then that's true. It also doesn't matter. What predictability means in real-time scheduler terms is "when I wiggle this interrupt line into the CPU, this process will be swapped in within a certian number of microseconds to deal with it" *that's* predictability, and what is easily done with PMT, and a complete are to do with CMT.
Montavista specialise in embedding Linux. I certianly hope most people have learnt by now that Linux is utterly unsuitable for hard real time. But that's not the same as embedding.