Hi Petr et al.

I coded up something that remotely looks like the
discussed patch.
I have deviated from your proposals at some important
points:
1. I am not allocating the ring1 stack separately,
I am allocating it on a ring0 stack. Much simpler
code, although non-reentrant/preempt-unsafe.
2. I am disabling the interrupts after all. That's
because of the preempt-unsafeness. I pass up the
IOPL=1 when necessary, to avoid problems.
But I guess also with your technique the interrupts
had to be disabled, unless the ring1 stack is
per-thread.
3. I am using LAR. Do you really think it can be
slower than the whole thing of locating LDT?

Let me know if I did something stupid.
The patch is attached.
I tested (pretty much) everything in it, except
probably the "popl %esp" restartability. But that
one looks fairly simple and should work.

Does this patch look good?