Re: [PATCH 2/4] AFS: Add a function to excise a rejected write from the pagecache

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 24 May 2007 23:34:33 +0100
David Howells <dhowells@redhat.com> wrote:

> Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > So my reason for asking the above is to try to find a way to make all these
> > new PG-error games just go away.
> 
> Yeah.  However, there needs to be something to cover the gap between releasing
> PG_writeback and getting PG_lock.  They have to be done in that order to avoid
> deadlocking against truncate and other stuff, but that leaves a window in which
> the page appears to be in a good state - one in which prepare_write() or
> page_mkwrite() can potentially leak through.

hm.  I don't see why that race window would be a problem in practice: the
page-exciser does a lock_page();wait_on_page_writeback() as normal, then
proceeds with its business?

But given that this doesn't work right for some reason, can we use PG_error
and then handle that appropriately in the filesystem's ->prepare_write() and
->page_mkwrite()?

> Nick Piggin talked about using an extra lock, but as far as I can tell, that
> just compounds the deadlock problems.
> 
> I suppose I could leave something in page->private that indicated that the
> page was defunct, but that'd have to be done by the filesystem, probably
> before calling cancel_rejected_write().

Well, using PG_error is OK and appropriate for that if it's localised to
the fs.  But I'd be a bit worried about requiring that the VFS maintain
some special protocol for it, partly because it would be such a
rarely-tested thing.