[PATCH 2/4] unionfs: 32-bit needs lock for i_size

[Hugh Dickins <hugh@veritas.com>]

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.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
---
This works for me, to help get through LTP, but is not enough: what about
those places which use i_size_write in common code called by unionfs?
They ought to participate in this i_locking too.  Needs an audit to
enumerate such places and decide what course to take.

Bigger question: is unionfs calling down to filesystems without holding
their i_mutex, where some may be relying on their i_mutex to be held?

 fs/stack.c |    6 ++++++
 1 file changed, 6 insertions(+)

--- 2.6.24-rc5-mm1/fs/stack.c	2007-12-05 10:38:38.000000000 +0000
+++ unionfs2/fs/stack.c	2007-12-05 16:50:15.000000000 +0000
@@ -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((struct inode *)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);