Re: MM patches against 2.5.31

[Christian Ehrhardt]

On Mon, Aug 26, 2002 at 12:24:50PM -0700, Andrew Morton wrote:
> The flaw is in doing the put_page_testzero() outside of any locking
> which would prevent other CPUs from finding and "rescuing" the zero-recount
> page.
> 
> CPUA:
> 	if (put_page_testzero()) {
> 		/* Here's the window */
> 		spin_lock(lru_lock);
> 		list_del(page->lru);
> 
> CPUB:
> 
> 	spin_lock(lru_lock);
> 	page = list_entry(lru);
> 	page_cache_get(page);	/* If this goes from 0->1, we die */
> 	...
> 	page_cache_release(page);	/* double free */

So what we want CPUB do instead is

	spin_lock(lru_lock);
	page = list_entry(lru)

	START ATOMIC 
		page_cache_get(page);
		res = (page_count (page) == 1)
	END ATOMIC

	if (res) {
		atomic_dec (&page->count);
		continue;  /* with next page */
	}
	...
	page_cache_release (page);

I.e. we want to detect _atomically_ that we just raised the page count
from zero to one. My patch actually has a solution that implements the
needed atomic operation above by means of the atomic functions that we
currently have on all archs (it's called get_page_testzero and
should probably called get_page_testone).
The more I think about this the more I think this is the way to go.

      regards   Christian

-- 
THAT'S ALL FOLKS!