[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)
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

index 1c0a3db78a052e2ad0e166b9b0c1af44109623da..da91b7bf998605677a87d4d0500dde2acf11aa28 100644 (file)
@@ -1686,10 +1686,6 @@ static int do_swap_page(struct mm_struct * mm,
        }
 
        /* The page isn't present yet, go ahead with the fault. */
-               
-       swap_free(entry);
-       if (vm_swap_full())
-               remove_exclusive_swap_page(page);
 
        inc_mm_counter(mm, rss);
        pte = mk_pte(page, vma->vm_page_prot);
@@ -1697,12 +1693,16 @@ static int do_swap_page(struct mm_struct * mm,
                pte = maybe_mkwrite(pte_mkdirty(pte), vma);
                write_access = 0;
        }
-       unlock_page(page);
 
        flush_icache_page(vma, page);
        set_pte_at(mm, address, page_table, pte);
        page_add_anon_rmap(page, vma, address);
 
+       swap_free(entry);
+       if (vm_swap_full())
+               remove_exclusive_swap_page(page);
+       unlock_page(page);
+
        if (write_access) {
                if (do_wp_page(mm, vma, address,
                                page_table, pmd, pte) == VM_FAULT_OOM)
index 9827409eb7c7ea4026e27b34fdaceacd909cabc3..89770bd25f31be71877813a292793fcc62e01116 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -539,27 +539,6 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
                goto out_unmap;
        }
 
-       /*
-        * Don't pull an anonymous page out from under get_user_pages.
-        * GUP carefully breaks COW and raises page count (while holding
-        * page_table_lock, as we have here) to make sure that the page
-        * cannot be freed.  If we unmap that page here, a user write
-        * access to the virtual address will bring back the page, but
-        * its raised count will (ironically) be taken to mean it's not
-        * an exclusive swap page, do_wp_page will replace it by a copy
-        * page, and the user never get to see the data GUP was holding
-        * the original page for.
-        *
-        * This test is also useful for when swapoff (unuse_process) has
-        * to drop page lock: its reference to the page stops existing
-        * ptes from being unmapped, so swapoff can make progress.
-        */
-       if (PageSwapCache(page) &&
-           page_count(page) != page_mapcount(page) + 2) {
-               ret = SWAP_FAIL;
-               goto out_unmap;
-       }
-
        /* Nuke the page table entry. */
        flush_cache_page(vma, address, page_to_pfn(page));
        pteval = ptep_clear_flush(vma, address, pte);
index da48405cd9a3199d9088d1221864fa77a76a108d..60cd24a55204efc5c84956443d02f9806bd66a9d 100644 (file)
@@ -276,61 +276,37 @@ void swap_free(swp_entry_t entry)
 }
 
 /*
- * Check if we're the only user of a swap page,
- * when the page is locked.
+ * How many references to page are currently swapped out?
  */
-static int exclusive_swap_page(struct page *page)
+static inline int page_swapcount(struct page *page)
 {
-       int retval = 0;
-       struct swap_info_struct * p;
+       int count = 0;
+       struct swap_info_struct *p;
        swp_entry_t entry;
 
        entry.val = page->private;
        p = swap_info_get(entry);
        if (p) {
-               /* Is the only swap cache user the cache itself? */
-               if (p->swap_map[swp_offset(entry)] == 1) {
-                       /* Recheck the page count with the swapcache lock held.. */
-                       write_lock_irq(&swapper_space.tree_lock);
-                       if (page_count(page) == 2)
-                               retval = 1;
-                       write_unlock_irq(&swapper_space.tree_lock);
-               }
+               /* Subtract the 1 for the swap cache itself */
+               count = p->swap_map[swp_offset(entry)] - 1;
                swap_info_put(p);
        }
-       return retval;
+       return count;
 }
 
 /*
  * We can use this swap cache entry directly
  * if there are no other references to it.
- *
- * Here "exclusive_swap_page()" does the real
- * work, but we opportunistically check whether
- * we need to get all the locks first..
  */
 int can_share_swap_page(struct page *page)
 {
-       int retval = 0;
+       int count;
 
-       if (!PageLocked(page))
-               BUG();
-       switch (page_count(page)) {
-       case 3:
-               if (!PagePrivate(page))
-                       break;
-               /* Fallthrough */
-       case 2:
-               if (!PageSwapCache(page))
-                       break;
-               retval = exclusive_swap_page(page);
-               break;
-       case 1:
-               if (PageReserved(page))
-                       break;
-               retval = 1;
-       }
-       return retval;
+       BUG_ON(!PageLocked(page));
+       count = page_mapcount(page);
+       if (count <= 1 && PageSwapCache(page))
+               count += page_swapcount(page);
+       return count == 1;
 }
 
 /*
@@ -529,9 +505,10 @@ static int unuse_mm(struct mm_struct *mm,
 
        if (!down_read_trylock(&mm->mmap_sem)) {
                /*
-                * Our reference to the page stops try_to_unmap_one from
-                * unmapping its ptes, so swapoff can make progress.
+                * Activate page so shrink_cache is unlikely to unmap its
+                * ptes while lock is dropped, so swapoff can make progress.
                 */
+               activate_page(page);
                unlock_page(page);
                down_read(&mm->mmap_sem);
                lock_page(page);