mm: munlock use mapcount to avoid terrible overhead

A process spent 30 minutes exiting, just munlocking the pages of a large
anonymous area that had been alternately mprotected into page-sized vmas:
for every single page there's an anon_vma walk through all the other
little vmas to find the right one.

A general fix to that would be a lot more complicated (use prio_tree on
anon_vma?), but there's one very simple thing we can do to speed up the
common case: if a page to be munlocked is mapped only once, then it is our
vma that it is mapped into, and there's no need whatever to walk through
all the others.

Okay, there is a very remote race in munlock_vma_pages_range(), if between
its follow_page() and lock_page(), another process were to munlock the
same page, then page reclaim remove it from our vma, then another process
mlock it again.  We would find it with page_mapcount 1, yet it's still
mlocked in another process.  But never mind, that's much less likely than
the down_read_trylock() failure which munlocking already tolerates (in
try_to_unmap_one()): in due course page reclaim will discover and move the
page to unevictable instead.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Michel Lespinasse <walken@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

diff --git a/mm/mlock.c b/mm/mlock.c
index 70d2b988c1265cc53846475fae5520824ea3f46b..360a7ba2c952c0a5d2254c4648a34ffd779c60bc 100644 (file)
--- a/mm/mlock.c
+++ b/mm/mlock.c
@@ -110,7 +110,10 @@ void munlock_vma_page(struct page *page)
        if (TestClearPageMlocked(page)) {
                dec_zone_page_state(page, NR_MLOCK);
                if (!isolate_lru_page(page)) {
-                       int ret = try_to_munlock(page);
+                       int ret = SWAP_AGAIN;
+
+                       if (page_mapcount(page) > 1)
+                               ret = try_to_munlock(page);
                        /*
                         * did try_to_unlock() succeed or punt?
                         */