In two cases (AIO-for-direct-IO and some CDROM DMA stuff which Jens
did), we need to run set_page_dirty() in interrupt context.  After DMA
hardware has altered userspace pages for direct-IO reads.

But mapping->page_lock, mapping->private_lock and inode_lock are not
irq-safe.  And really, we don't want to convert those locks just for this
problem.

So what we do is to dirty the pages *before* starting I/O.  Then, in
interrupt completion context, check to see that they are still dirty.
If so then there is nothing to do.  But if the pages were cleaned while
the direct-IO is in progress we need to redirty them.  The code uses
schedule_work() for that.

Of course, we could use schedule_work() for all BIOs and pages.  The
speculative dirty-before-starting-IO is just an optimisation.  It's
quite unlikely that any of the pages will be cleaned during the direct
IO.


This code is basically untestable under normal circumstances, because the
relevant pages are pinned via get_user_pages().  This makes
is_page_cache_freeable() evaluate false and the VM doesn't try to write them
out anyway.  But if the pages happen to be MAP_SHARED file pages, pdflush
could clean them.  The way I tested this code was to disable the call to
bio_set_pages_dirty() in fs/direct-io.c.