]> err.no Git - linux-2.6/blobdiff - mm/rmap.c
[PATCH] delete from_swap_cache BUG_ONs
[linux-2.6] / mm / rmap.c
index 884d6d1928bce0b438ec7d43d7b33918e690be9c..facb8cdca665cc9b80debc37c7c8ba5d6ad7d763 100644 (file)
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -34,9 +34,8 @@
  *       anon_vma->lock
  *         mm->page_table_lock
  *           zone->lru_lock (in mark_page_accessed)
- *           swap_list_lock (in swap_free etc's swap_info_get)
+ *           swap_lock (in swap_duplicate, swap_info_get)
  *             mmlist_lock (in mmput, drain_mmlist and others)
- *             swap_device_lock (in swap_duplicate, swap_info_get)
  *             mapping->private_lock (in __set_page_dirty_buffers)
  *             inode_lock (in set_page_dirty's __mark_inode_dirty)
  *               sb_lock (within inode_lock in fs/fs-writeback.c)
@@ -242,6 +241,42 @@ unsigned long page_address_in_vma(struct page *page, struct vm_area_struct *vma)
        return vma_address(page, vma);
 }
 
+/*
+ * Check that @page is mapped at @address into @mm.
+ *
+ * On success returns with mapped pte and locked mm->page_table_lock.
+ */
+pte_t *page_check_address(struct page *page, struct mm_struct *mm,
+                         unsigned long address)
+{
+       pgd_t *pgd;
+       pud_t *pud;
+       pmd_t *pmd;
+       pte_t *pte;
+
+       /*
+        * We need the page_table_lock to protect us from page faults,
+        * munmap, fork, etc...
+        */
+       spin_lock(&mm->page_table_lock);
+       pgd = pgd_offset(mm, address);
+       if (likely(pgd_present(*pgd))) {
+               pud = pud_offset(pgd, address);
+               if (likely(pud_present(*pud))) {
+                       pmd = pmd_offset(pud, address);
+                       if (likely(pmd_present(*pmd))) {
+                               pte = pte_offset_map(pmd, address);
+                               if (likely(pte_present(*pte) &&
+                                          page_to_pfn(page) == pte_pfn(*pte)))
+                                       return pte;
+                               pte_unmap(pte);
+                       }
+               }
+       }
+       spin_unlock(&mm->page_table_lock);
+       return ERR_PTR(-ENOENT);
+}
+
 /*
  * Subfunctions of page_referenced: page_referenced_one called
  * repeatedly from either page_referenced_anon or page_referenced_file.
@@ -251,9 +286,6 @@ static int page_referenced_one(struct page *page,
 {
        struct mm_struct *mm = vma->vm_mm;
        unsigned long address;
-       pgd_t *pgd;
-       pud_t *pud;
-       pmd_t *pmd;
        pte_t *pte;
        int referenced = 0;
 
@@ -263,39 +295,18 @@ static int page_referenced_one(struct page *page,
        if (address == -EFAULT)
                goto out;
 
-       spin_lock(&mm->page_table_lock);
-
-       pgd = pgd_offset(mm, address);
-       if (!pgd_present(*pgd))
-               goto out_unlock;
-
-       pud = pud_offset(pgd, address);
-       if (!pud_present(*pud))
-               goto out_unlock;
-
-       pmd = pmd_offset(pud, address);
-       if (!pmd_present(*pmd))
-               goto out_unlock;
-
-       pte = pte_offset_map(pmd, address);
-       if (!pte_present(*pte))
-               goto out_unmap;
-
-       if (page_to_pfn(page) != pte_pfn(*pte))
-               goto out_unmap;
-
-       if (ptep_clear_flush_young(vma, address, pte))
-               referenced++;
-
-       if (mm != current->mm && !ignore_token && has_swap_token(mm))
-               referenced++;
+       pte = page_check_address(page, mm, address);
+       if (!IS_ERR(pte)) {
+               if (ptep_clear_flush_young(vma, address, pte))
+                       referenced++;
 
-       (*mapcount)--;
+               if (mm != current->mm && !ignore_token && has_swap_token(mm))
+                       referenced++;
 
-out_unmap:
-       pte_unmap(pte);
-out_unlock:
-       spin_unlock(&mm->page_table_lock);
+               (*mapcount)--;
+               pte_unmap(pte);
+               spin_unlock(&mm->page_table_lock);
+       }
 out:
        return referenced;
 }
@@ -502,9 +513,6 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
 {
        struct mm_struct *mm = vma->vm_mm;
        unsigned long address;
-       pgd_t *pgd;
-       pud_t *pud;
-       pmd_t *pmd;
        pte_t *pte;
        pte_t pteval;
        int ret = SWAP_AGAIN;
@@ -515,30 +523,9 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
        if (address == -EFAULT)
                goto out;
 
-       /*
-        * We need the page_table_lock to protect us from page faults,
-        * munmap, fork, etc...
-        */
-       spin_lock(&mm->page_table_lock);
-
-       pgd = pgd_offset(mm, address);
-       if (!pgd_present(*pgd))
-               goto out_unlock;
-
-       pud = pud_offset(pgd, address);
-       if (!pud_present(*pud))
-               goto out_unlock;
-
-       pmd = pmd_offset(pud, address);
-       if (!pmd_present(*pmd))
-               goto out_unlock;
-
-       pte = pte_offset_map(pmd, address);
-       if (!pte_present(*pte))
-               goto out_unmap;
-
-       if (page_to_pfn(page) != pte_pfn(*pte))
-               goto out_unmap;
+       pte = page_check_address(page, mm, address);
+       if (IS_ERR(pte))
+               goto out;
 
        /*
         * If the page is mlock()d, we cannot swap it out.
@@ -551,27 +538,6 @@ 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);
@@ -598,13 +564,12 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
                dec_mm_counter(mm, anon_rss);
        }
 
-       inc_mm_counter(mm, rss);
+       dec_mm_counter(mm, rss);
        page_remove_rmap(page);
        page_cache_release(page);
 
 out_unmap:
        pte_unmap(pte);
-out_unlock:
        spin_unlock(&mm->page_table_lock);
 out:
        return ret;
@@ -639,7 +604,7 @@ static void try_to_unmap_cluster(unsigned long cursor,
        pgd_t *pgd;
        pud_t *pud;
        pmd_t *pmd;
-       pte_t *pte;
+       pte_t *pte, *original_pte;
        pte_t pteval;
        struct page *page;
        unsigned long address;
@@ -671,7 +636,7 @@ static void try_to_unmap_cluster(unsigned long cursor,
        if (!pmd_present(*pmd))
                goto out_unlock;
 
-       for (pte = pte_offset_map(pmd, address);
+       for (original_pte = pte = pte_offset_map(pmd, address);
                        address < end; pte++, address += PAGE_SIZE) {
 
                if (!pte_present(*pte))
@@ -707,8 +672,7 @@ static void try_to_unmap_cluster(unsigned long cursor,
                (*mapcount)--;
        }
 
-       pte_unmap(pte);
-
+       pte_unmap(original_pte);
 out_unlock:
        spin_unlock(&mm->page_table_lock);
 }
@@ -860,3 +824,4 @@ int try_to_unmap(struct page *page)
                ret = SWAP_SUCCESS;
        return ret;
 }
+