Re: [patch 1/8] fix race in clear_page_dirty_for_io()

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

(cc's reinstated)

On Wed, 07 Mar 2007 09:09:50 +0100 Miklos Szeredi <miklos@szeredi.hu> wrote:

> There's a race in clear_page_dirty_for_io() that allows a page to have
> cleared PG_dirty, while being mapped read-write into the page table(s).

I assume you refer to this:

		 * FIXME! We still have a race here: if somebody
		 * adds the page back to the page tables in
		 * between the "page_mkclean()" and the "TestClearPageDirty()",
		 * we might have it mapped without the dirty bit set.
		 */
		if (page_mkclean(page))
			set_page_dirty(page);
		if (TestClearPageDirty(page)) {
			dec_zone_page_state(page, NR_FILE_DIRTY);
			return 1;
		}

I guess the comment actually refers to a writefault after the
set_page_dirty() and before the TestClearPageDirty().  The fault handler
will run set_page_dirty() and will return to userspace to rerun the write. 
The page then gets set pte-dirty but this thread of control will now make
the page !PageDirty() and will write it out.

With Nick's proposed lock-the-page-in-pagefaults patches, we have
lock_page() synchronisation between pagefaults and
clear_page_dirty_for_io() which I think will fix this.