VFS/fs_stack: use locking around i_size_write in 32-bit systems

LTP's iogen01 doio tests hang nicely on 32-bit SMP when /tmp is a unionfs
mount of a tmpfs.  See the comment on i_size_write in linux/fs.h: it needs
to be locked, otherwise i_size_read can spin forever waiting for a lost
seqcount update.

Most filesystems are already holding i_mutex for this, but unionfs calls
fsstack_copy_inode_size from many places, not necessarily holding i_mutex.
Use the low-level i_lock within fsstack_copy_inode_size when 32-bit SMP.

Checked the entire unionfs code to ensure this is the right fix for
i_size_write().

Also compared to what other file systems do when they have to handle inodes,
esp. not their own inodes (e.g., network file systems have to access the
exported file system's inodes).  Found out that most such file systems not just
don't lock around i_size_write, but they don't even use i_size_read or
i_size_write to access the inode's size.

CC: Mike Halcrow <mhalcrow@us.ibm.com>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Erez Zadok <ezk@cs.sunysb.edu>

diff --git a/fs/stack.c b/fs/stack.c
index 7913fe53a3c9317723696352c9f16bb45397c616..4336f2b32dfb8182b4d19082c262b450e9bcfc87 100644 (file)
--- a/fs/stack.c
+++ b/fs/stack.c
@@ -21,8 +21,14 @@
  */
 void fsstack_copy_inode_size(struct inode *dst, const struct inode *src)
 {
+#if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
+       spin_lock(&dst->i_lock);
+#endif
        i_size_write(dst, i_size_read(src));
        dst->i_blocks = src->i_blocks;
+#if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
+       spin_unlock(&dst->i_lock);
+#endif
 }
 EXPORT_SYMBOL_GPL(fsstack_copy_inode_size);