buffer_head: fix private_list handling

There are two possible races in handling of private_list in buffer cache.

1) When fsync_buffers_list() processes a private_list, it clears
   b_assoc_mapping and moves buffer to its private list.  Now
   drop_buffers() comes, sees a buffer is on list so it calls
   __remove_assoc_queue() which complains about b_assoc_mapping being
   cleared (as it cannot propagate possible IO error).  This race has been
   actually observed in the wild.

2) When fsync_buffers_list() processes a private_list,
   mark_buffer_dirty_inode() can be called on bh which is already on the
   private list of fsync_buffers_list().  As buffer is on some list (note
   that the check is performed without private_lock), it is not readded to
   the mapping's private_list and after fsync_buffers_list() finishes, we
   have a dirty buffer which should be on private_list but it isn't.  This
   race has not been reported, probably because most (but not all) callers
   of mark_buffer_dirty_inode() hold i_mutex and thus are serialized with
   fsync().

Fix these issues by not clearing b_assoc_map when fsync_buffers_list()
moves buffer to a dedicated list and by reinserting buffer in private_list
when it is found dirty after we have submitted buffer for IO.  We also
change the tests whether a buffer is on a private list from
!list_empty(&bh->b_assoc_buffers) to bh->b_assoc_map so that they are
single word reads and hence lockless checks are safe.

Signed-off-by: Jan Kara <jack@suse.cz>
Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/fs/buffer.c b/fs/buffer.c
index 6f0bddddcf4aadb69c6b1de52eedb97e5735bba6..3ebccf4aa7e3280318b4ae8fbd2b28514d929749 100644 (file)
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -678,7 +678,7 @@ void mark_buffer_dirty_inode(struct buffer_head *bh, struct inode *inode)
        } else {
                BUG_ON(mapping->assoc_mapping != buffer_mapping);
        }
-       if (list_empty(&bh->b_assoc_buffers)) {
+       if (!bh->b_assoc_map) {
                spin_lock(&buffer_mapping->private_lock);
                list_move_tail(&bh->b_assoc_buffers,
                                &mapping->private_list);
@@ -794,6 +794,7 @@ static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
 {
        struct buffer_head *bh;
        struct list_head tmp;
+       struct address_space *mapping;
        int err = 0, err2;
 
        INIT_LIST_HEAD(&tmp);
@@ -801,9 +802,14 @@ static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
        spin_lock(lock);
        while (!list_empty(list)) {
                bh = BH_ENTRY(list->next);
+               mapping = bh->b_assoc_map;
                __remove_assoc_queue(bh);
+               /* Avoid race with mark_buffer_dirty_inode() which does
+                * a lockless check and we rely on seeing the dirty bit */
+               smp_mb();
                if (buffer_dirty(bh) || buffer_locked(bh)) {
                        list_add(&bh->b_assoc_buffers, &tmp);
+                       bh->b_assoc_map = mapping;
                        if (buffer_dirty(bh)) {
                                get_bh(bh);
                                spin_unlock(lock);
@@ -822,8 +828,17 @@ static int fsync_buffers_list(spinlock_t *lock, struct list_head *list)
 
        while (!list_empty(&tmp)) {
                bh = BH_ENTRY(tmp.prev);
-               list_del_init(&bh->b_assoc_buffers);
                get_bh(bh);
+               mapping = bh->b_assoc_map;
+               __remove_assoc_queue(bh);
+               /* Avoid race with mark_buffer_dirty_inode() which does
+                * a lockless check and we rely on seeing the dirty bit */
+               smp_mb();
+               if (buffer_dirty(bh)) {
+                       list_add(&bh->b_assoc_buffers,
+                                &bh->b_assoc_map->private_list);
+                       bh->b_assoc_map = mapping;
+               }
                spin_unlock(lock);
                wait_on_buffer(bh);
                if (!buffer_uptodate(bh))
@@ -1195,7 +1210,7 @@ void __brelse(struct buffer_head * buf)
 void __bforget(struct buffer_head *bh)
 {
        clear_buffer_dirty(bh);
-       if (!list_empty(&bh->b_assoc_buffers)) {
+       if (bh->b_assoc_map) {
                struct address_space *buffer_mapping = bh->b_page->mapping;
 
                spin_lock(&buffer_mapping->private_lock);
@@ -3022,7 +3037,7 @@ drop_buffers(struct page *page, struct buffer_head **buffers_to_free)
        do {
                struct buffer_head *next = bh->b_this_page;
 
-               if (!list_empty(&bh->b_assoc_buffers))
+               if (bh->b_assoc_map)
                        __remove_assoc_queue(bh);
                bh = next;
        } while (bh != head);