> I don't walk the thread descriptors.  I don't write into them.  I move
> entire double-linked lists with a dozen or so instructions.  Regardless
> of how many threads were active in the parent.
However this would cause a lot of copy-on-write faults on thread stacks
when other thread resume execution.

How about adding a MAP_DONTCOPY flag to mmap, using it for the thread
stacks and then adding yet another flag and pointer to the clone
syscall, pointing to a userspace array of addresses and flags, allowing
to specify whether vmas should be copied, ignored (or maybe shared, as a
future extension) so that userspace could specify that the current
thread stack should be copied anyway?