mm: m(un)lock avoid ZERO_PAGE
Hugh Dickins [Tue, 22 Sep 2009 00:03:32 +0000 (17:03 -0700)]
I'm still reluctant to clutter __get_user_pages() with another flag, just
to avoid touching ZERO_PAGE count in mlock(); though we can add that later
if it shows up as an issue in practice.

But when mlocking, we can test page->mapping slightly earlier, to avoid
the potentially bouncy rescheduling of lock_page on ZERO_PAGE - mlock
didn't lock_page in olden ZERO_PAGE days, so we might have regressed.

And when munlocking, it turns out that FOLL_DUMP coincidentally does
what's needed to avoid all updates to ZERO_PAGE, so use that here also.
Plus add comment suggested by KAMEZAWA Hiroyuki.

Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: Rik van Riel <riel@redhat.com>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Nick Piggin <npiggin@suse.de>
Acked-by: Mel Gorman <mel@csn.ul.ie>
Cc: Minchan Kim <minchan.kim@gmail.com>
Reviewed-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/mlock.c

index 22041aa..bd6f0e4 100644 (file)
@@ -198,17 +198,26 @@ static long __mlock_vma_pages_range(struct vm_area_struct *vma,
                for (i = 0; i < ret; i++) {
                        struct page *page = pages[i];
 
-                       lock_page(page);
-                       /*
-                        * Because we lock page here and migration is blocked
-                        * by the elevated reference, we need only check for
-                        * file-cache page truncation.  This page->mapping
-                        * check also neatly skips over the ZERO_PAGE(),
-                        * though if that's common we'd prefer not to lock it.
-                        */
-                       if (page->mapping)
-                               mlock_vma_page(page);
-                       unlock_page(page);
+                       if (page->mapping) {
+                               /*
+                                * That preliminary check is mainly to avoid
+                                * the pointless overhead of lock_page on the
+                                * ZERO_PAGE: which might bounce very badly if
+                                * there is contention.  However, we're still
+                                * dirtying its cacheline with get/put_page:
+                                * we'll add another __get_user_pages flag to
+                                * avoid it if that case turns out to matter.
+                                */
+                               lock_page(page);
+                               /*
+                                * Because we lock page here and migration is
+                                * blocked by the elevated reference, we need
+                                * only check for file-cache page truncation.
+                                */
+                               if (page->mapping)
+                                       mlock_vma_page(page);
+                               unlock_page(page);
+                       }
                        put_page(page); /* ref from get_user_pages() */
                }
 
@@ -309,9 +318,23 @@ void munlock_vma_pages_range(struct vm_area_struct *vma,
        vma->vm_flags &= ~VM_LOCKED;
 
        for (addr = start; addr < end; addr += PAGE_SIZE) {
-               struct page *page = follow_page(vma, addr, FOLL_GET);
-               if (page) {
+               struct page *page;
+               /*
+                * Although FOLL_DUMP is intended for get_dump_page(),
+                * it just so happens that its special treatment of the
+                * ZERO_PAGE (returning an error instead of doing get_page)
+                * suits munlock very well (and if somehow an abnormal page
+                * has sneaked into the range, we won't oops here: great).
+                */
+               page = follow_page(vma, addr, FOLL_GET | FOLL_DUMP);
+               if (page && !IS_ERR(page)) {
                        lock_page(page);
+                       /*
+                        * Like in __mlock_vma_pages_range(),
+                        * because we lock page here and migration is
+                        * blocked by the elevated reference, we need
+                        * only check for file-cache page truncation.
+                        */
                        if (page->mapping)
                                munlock_vma_page(page);
                        unlock_page(page);