ext3_block_truncate_page() is calling grab_cache_page() inside a JBD
transaction.  This is wrong, because transactions nest inside lock_page().

The deadlock is against shrink_list->ext3_journalled_writepage->journal_start.

This was not noticed before because we never used to journal writepage() data
in journalled-data mode.  And because the deadlock against
generic_file_write() is covered up by i_sem.

Rework things so that we lock the page prior to starting a transaction.