fs: sync_sb_inodes fix
Nick Piggin [Tue, 6 Jan 2009 22:40:25 +0000 (14:40 -0800)]
Fix data integrity semantics required by sys_sync, by iterating over all
inodes and waiting for any writeback pages after the initial writeout.
Comments explain the exact problem.

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

fs/fs-writeback.c

index d99601a..a9ee474 100644 (file)
@@ -440,6 +440,7 @@ void generic_sync_sb_inodes(struct super_block *sb,
                                struct writeback_control *wbc)
 {
        const unsigned long start = jiffies;    /* livelock avoidance */
+       int sync = wbc->sync_mode == WB_SYNC_ALL;
 
        spin_lock(&inode_lock);
        if (!wbc->for_kupdate || list_empty(&sb->s_io))
@@ -516,7 +517,49 @@ void generic_sync_sb_inodes(struct super_block *sb,
                if (!list_empty(&sb->s_more_io))
                        wbc->more_io = 1;
        }
-       spin_unlock(&inode_lock);
+
+       if (sync) {
+               struct inode *inode, *old_inode = NULL;
+
+               /*
+                * Data integrity sync. Must wait for all pages under writeback,
+                * because there may have been pages dirtied before our sync
+                * call, but which had writeout started before we write it out.
+                * In which case, the inode may not be on the dirty list, but
+                * we still have to wait for that writeout.
+                */
+               list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
+                       struct address_space *mapping;
+
+                       if (inode->i_state & (I_FREEING|I_WILL_FREE))
+                               continue;
+                       mapping = inode->i_mapping;
+                       if (mapping->nrpages == 0)
+                               continue;
+                       __iget(inode);
+                       spin_unlock(&inode_lock);
+                       /*
+                        * We hold a reference to 'inode' so it couldn't have
+                        * been removed from s_inodes list while we dropped the
+                        * inode_lock.  We cannot iput the inode now as we can
+                        * be holding the last reference and we cannot iput it
+                        * under inode_lock. So we keep the reference and iput
+                        * it later.
+                        */
+                       iput(old_inode);
+                       old_inode = inode;
+
+                       filemap_fdatawait(mapping);
+
+                       cond_resched();
+
+                       spin_lock(&inode_lock);
+               }
+               spin_unlock(&inode_lock);
+               iput(old_inode);
+       } else
+               spin_unlock(&inode_lock);
+
        return;         /* Leave any unwritten inodes on s_io */
 }
 EXPORT_SYMBOL_GPL(generic_sync_sb_inodes);
@@ -596,13 +639,16 @@ void sync_inodes_sb(struct super_block *sb, int wait)
                .range_start    = 0,
                .range_end      = LLONG_MAX,
        };
-       unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
-       unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
 
-       wbc.nr_to_write = nr_dirty + nr_unstable +
-                       (inodes_stat.nr_inodes - inodes_stat.nr_unused) +
-                       nr_dirty + nr_unstable;
-       wbc.nr_to_write += wbc.nr_to_write / 2;         /* Bit more for luck */
+       if (!wait) {
+               unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY);
+               unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS);
+
+               wbc.nr_to_write = nr_dirty + nr_unstable +
+                       (inodes_stat.nr_inodes - inodes_stat.nr_unused);
+       } else
+               wbc.nr_to_write = LONG_MAX; /* doesn't actually matter */
+
        sync_sb_inodes(sb, &wbc);
 }