Re: [PATCH] nfs: clear_commit_release incorrectly handle truncated page

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Tue, 2010-02-02 at 18:17 +0300, Dmitry Monakhov wrote: 
> Trond Myklebust <trond.myklebust@fys.uio.no> writes:
> 
> > On Tue, 2010-02-02 at 13:36 +0300, Dmitry Monakhov wrote: 
> >> After page was truncated it lost it's mapping, this result in null
> >> pointer dereference on bdi_stat update. In fact we have to decrement
> >> bdi_stat even for truncated pages, so let's pass correct mapping in
> >> function arguments. Patch against linux-2.6
> >> ##TEST_CASE
> >> /*
> >> Tast case for bug in nfs_clear_request_commit()
> >> caused by null pointer dereference in case of truncated page.
> >> It takes less than 10 minutes to reproduce the bug.
> >
> > Something is wrong here. nfs_release_page() returns '0' if the 
> > page has an associated write request (i.e. PagePrivate is set), and so
> > both invalidate_complete_page() and invalidate_complete_page2() will
> > fail.
> >
> > So what is truncating the page?
> truncate_inode_page()
>   truncate_complete_page()
>     if (page_has_private(page))
>        do_invalidatepage()
>          ->nfs_invalidate_page()

do_invalidate_page() is called before remove_from_page_cache(), so
page->mapping should still be set.

Trond