[PATCH] can_share_swap_page: use page_mapcount
authorHugh Dickins <hugh@veritas.com>
Wed, 22 Jun 2005 00:15:12 +0000 (17:15 -0700)
committerLinus Torvalds <torvalds@ppc970.osdl.org>
Wed, 22 Jun 2005 01:46:21 +0000 (18:46 -0700)
commitc475a8ab625d567eacf5e30ec35d6d8704558062
tree0971bef7b876f1b3eb160621fc2b61cb5313827b
parentd296e9cd02c92e576ecce5344026a4df4353cdb2
[PATCH] can_share_swap_page: use page_mapcount

Remember that ironic get_user_pages race?  when the raised page_count on a
page swapped out led do_wp_page to decide that it had to copy on write, so
substituted a different page into userspace.  2.6.7 onwards have Andrea's
solution, where try_to_unmap_one backs out if it finds page_count raised.

Which works, but is unsatisfying (rmap.c has no other page_count heuristics),
and was found a few months ago to hang an intensive page migration test.  A
year ago I was hesitant to engage page_mapcount, now it seems the right fix.

So remove the page_count hack from try_to_unmap_one; and use activate_page in
unuse_mm when dropping lock, to replace its secondary effect of helping
swapoff to make progress in that case.

Simplify can_share_swap_page (now called only on anonymous pages) to check
page_mapcount + page_swapcount == 1: still needs the page lock to stabilize
their (pessimistic) sum, but does not need swapper_space.tree_lock for that.

In do_swap_page, move swap_free and unlock_page below page_add_anon_rmap, to
keep sum on the high side, and correct when can_share_swap_page called.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/memory.c
mm/rmap.c
mm/swapfile.c