Andrew, this builds on my first patch eliminating the page_table_lock
during page table cleanup on exit.  I took a good hard look at
clear_page_tables, and realized that it's using up a lot of time to run.
It walks through a lot of empty slots looking for pte pages to free.

I came to the realization that if we could just keep a count of mapped
pages and swap entries, we'd know right away if a pte page is freeable.
This patch tracks the count for pte pages and removes them as soon as
they're unused, eliminating the need for clear_page_tables entirely.

Doing this gained another 5% in my fork/exit timing tests, so the combined
patch gives me a 10% improvement in fork/exit.

Tracking the reference counts was the last straw in overloading struct page
with pte page info, so I created a 'struct ptpage' to use when the struct
page describes a page table page.  It's a bit of a hack, but I think in the
long run will make it more understandable.

Dave McCracken

======================================================================
Dave McCracken          IBM Linux Base Kernel Team      1-512-838-3059
dmccr@us.ibm.com                                        T/L   678-3059