Re: Note describing poor dcache utilization under high memory pressure

[Linus Torvalds]

On Tue, 29 Jan 2002, Oliver Xymoron wrote:
>
> The "detached mm" approach should be sufficiently parallel to the
> read-only page directory entries that the two can use almost the same
> framework.

Yes. I suspect that it can be trivially hidden in just two
architecture-specific functions, ie something like "detach_pgd(pgd)" and
"attach_pgd_entry(mm, address)".

>	 The downside is faults on reads in the detached case, but that
> shouldn't be significantly worse than the original copy, thanks to the
> large fanout.

Right. We'd get a few "unnecessary" page faults, but they should be on the
order of 0.1% of the necessary ones. In fact, with pre-faulting in
sys_fork(), I wouldn't be surprised if the common case is to not have any
directory-related page faults at all.

		Linus