mm: protect set_page_dirty() from ongoing truncation
authorJohannes Weiner <hannes@cmpxchg.org>
Thu, 8 Jan 2015 22:32:18 +0000 (14:32 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 8 Jan 2015 23:10:51 +0000 (15:10 -0800)
Tejun, while reviewing the code, spotted the following race condition
between the dirtying and truncation of a page:

__set_page_dirty_nobuffers()       __delete_from_page_cache()
  if (TestSetPageDirty(page))
                                     page->mapping = NULL
     if (PageDirty())
       dec_zone_page_state(page, NR_FILE_DIRTY);
       dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
    if (page->mapping)
      account_page_dirtied(page)
        __inc_zone_page_state(page, NR_FILE_DIRTY);
__inc_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);

which results in an imbalance of NR_FILE_DIRTY and BDI_RECLAIMABLE.

Dirtiers usually lock out truncation, either by holding the page lock
directly, or in case of zap_pte_range(), by pinning the mapcount with
the page table lock held.  The notable exception to this rule, though,
is do_wp_page(), for which this race exists.  However, do_wp_page()
already waits for a locked page to unlock before setting the dirty bit,
in order to prevent a race where clear_page_dirty() misses the page bit
in the presence of dirty ptes.  Upgrade that wait to a fully locked
set_page_dirty() to also cover the situation explained above.

Afterwards, the code in set_page_dirty() dealing with a truncation race
is no longer needed.  Remove it.

Reported-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/writeback.h
mm/memory.c
mm/page-writeback.c

index a219be961c0a2cb7ab21ce08353bdb1c968d2b92..00048339c23e4f252ee6a4b15cd38b49b8032de4 100644 (file)
@@ -177,7 +177,6 @@ int write_cache_pages(struct address_space *mapping,
                      struct writeback_control *wbc, writepage_t writepage,
                      void *data);
 int do_writepages(struct address_space *mapping, struct writeback_control *wbc);
-void set_page_dirty_balance(struct page *page);
 void writeback_set_ratelimit(void);
 void tag_pages_for_writeback(struct address_space *mapping,
                             pgoff_t start, pgoff_t end);
index d7e497e98f463f4ded7ce384513be439f190bde0..c6565f00fb38bdc89e0c0ae061786af4bc9c1667 100644 (file)
@@ -2137,17 +2137,24 @@ reuse:
                if (!dirty_page)
                        return ret;
 
-               /*
-                * Yes, Virginia, this is actually required to prevent a race
-                * with clear_page_dirty_for_io() from clearing the page dirty
-                * bit after it clear all dirty ptes, but before a racing
-                * do_wp_page installs a dirty pte.
-                *
-                * do_shared_fault is protected similarly.
-                */
                if (!page_mkwrite) {
-                       wait_on_page_locked(dirty_page);
-                       set_page_dirty_balance(dirty_page);
+                       struct address_space *mapping;
+                       int dirtied;
+
+                       lock_page(dirty_page);
+                       dirtied = set_page_dirty(dirty_page);
+                       VM_BUG_ON_PAGE(PageAnon(dirty_page), dirty_page);
+                       mapping = dirty_page->mapping;
+                       unlock_page(dirty_page);
+
+                       if (dirtied && mapping) {
+                               /*
+                                * Some device drivers do not set page.mapping
+                                * but still dirty their pages
+                                */
+                               balance_dirty_pages_ratelimited(mapping);
+                       }
+
                        /* file_update_time outside page_lock */
                        if (vma->vm_file)
                                file_update_time(vma->vm_file);
index d5d81f5384d16f09076fdb4dc06cdf45f6d238ef..6f4335238e33311de251a647fe725d06d5897060 100644 (file)
@@ -1541,16 +1541,6 @@ pause:
                bdi_start_background_writeback(bdi);
 }
 
-void set_page_dirty_balance(struct page *page)
-{
-       if (set_page_dirty(page)) {
-               struct address_space *mapping = page_mapping(page);
-
-               if (mapping)
-                       balance_dirty_pages_ratelimited(mapping);
-       }
-}
-
 static DEFINE_PER_CPU(int, bdp_ratelimits);
 
 /*
@@ -2123,32 +2113,25 @@ EXPORT_SYMBOL(account_page_dirtied);
  * page dirty in that case, but not all the buffers.  This is a "bottom-up"
  * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
  *
- * Most callers have locked the page, which pins the address_space in memory.
- * But zap_pte_range() does not lock the page, however in that case the
- * mapping is pinned by the vma's ->vm_file reference.
- *
- * We take care to handle the case where the page was truncated from the
- * mapping by re-checking page_mapping() inside tree_lock.
+ * The caller must ensure this doesn't race with truncation.  Most will simply
+ * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
+ * the pte lock held, which also locks out truncation.
  */
 int __set_page_dirty_nobuffers(struct page *page)
 {
        if (!TestSetPageDirty(page)) {
                struct address_space *mapping = page_mapping(page);
-               struct address_space *mapping2;
                unsigned long flags;
 
                if (!mapping)
                        return 1;
 
                spin_lock_irqsave(&mapping->tree_lock, flags);
-               mapping2 = page_mapping(page);
-               if (mapping2) { /* Race with truncate? */
-                       BUG_ON(mapping2 != mapping);
-                       WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
-                       account_page_dirtied(page, mapping);
-                       radix_tree_tag_set(&mapping->page_tree,
-                               page_index(page), PAGECACHE_TAG_DIRTY);
-               }
+               BUG_ON(page_mapping(page) != mapping);
+               WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
+               account_page_dirtied(page, mapping);
+               radix_tree_tag_set(&mapping->page_tree, page_index(page),
+                                  PAGECACHE_TAG_DIRTY);
                spin_unlock_irqrestore(&mapping->tree_lock, flags);
                if (mapping->host) {
                        /* !PageAnon && !swapper_space */
@@ -2305,12 +2288,10 @@ int clear_page_dirty_for_io(struct page *page)
                /*
                 * We carefully synchronise fault handlers against
                 * installing a dirty pte and marking the page dirty
-                * at this point. We do this by having them hold the
-                * page lock at some point after installing their
-                * pte, but before marking the page dirty.
-                * Pages are always locked coming in here, so we get
-                * the desired exclusion. See mm/memory.c:do_wp_page()
-                * for more comments.
+                * at this point.  We do this by having them hold the
+                * page lock while dirtying the page, and pages are
+                * always locked coming in here, so we get the desired
+                * exclusion.
                 */
                if (TestClearPageDirty(page)) {
                        dec_zone_page_state(page, NR_FILE_DIRTY);