We hold i_sem during the various sync() operations to prevent livelocks:
if another thread is dirtying the file, a sync() may never return.

Or at least, that used to be true when we were using the per-address_space
page lists.  Since writeback has used radix tree traversal it is not possible
to livelock the sync() operations, because they only visit each page a single
time.

sync_page_range() (used by O_SYNC writes) has not been holding i_sem for quite
some time, for the above reasons.

The patch converts fsync(), fdatasync() and msync() to also not hold i_sem
during the radix-tree-based writeback.

Now, we _do_ still need to hold i_sem across the file->f_op->fsync() call,
because that is still based on a list_head walk, and is still livelockable.

But in the case of msync() I deliberately left i_sem untaken.  This is because
we're currently deadlockable in msync, because mmap_sem is already held, and
mmap_sem nexts inside i_sem, due to direct-io.c.

And yes, the ranking of down_read() veruss down() does matter:

	Task A			Task B		Task C

	down_read(rwsem)
				down(sem)
						down_write(rwsem)
	down(sem)
				down_read(rwsem)


C's down_write() will cause B's down_read to block.  B holds `sem', so A will
never release `rwsem'.

So the patch fixes a hard-to-hit triple-task deadlock, but adds a possible
livelock in msync().  It is possible to fix sys_msync() so that it takes i_sem
outside i_mmap_sem.  Later.
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>