This patch is against pre7-9 (almost)
I will prepare a patch against pre8 but it might take
until Monday.

Tested on PPro 180 UP system with 96MB RAM.
(Bennos latency tests)

This shrink_mmap checks all pages at the priority level
and frees the one it can. And then it is not retried
with the same level.

It does not move pages around unnecessarily thus keeping
the list LRU (unscanned but referenced pages are not
moved up... bad)

I think the locking is SMP safe [Ingo].
  pagemap_lru_lock or page_locked

It uses the >> priority trick since that will result
in scanning double amount of pages next time.
(pages already scanned + the same no of prev. unscanned)

vmscan.c is modified to call only once per priority level.

(it has the goto sleep patch too)

/RogerL

--
Home page:
  http://www.norran.net/nra02596/