Re: lockless pagecache Cassini regression

[Nick Piggin <npiggin@suse.de>]

On Fri, Jan 04, 2008 at 03:58:31AM -0800, David Miller wrote:
> From: Nick Piggin <npiggin@suse.de>
> Date: Fri, 4 Jan 2008 12:33:52 +0100
> 
> > Just for interest, the lockless pagecache actually makes
> > page->_count unstable for all pages that _have ever_ been pagecache
> > pages (since the last quiescent rcu state, anyway). Basically, it
> > looks up and takes a ref on the struct page without ever having a
> > prior pin or reference on that page. It can do this because it knows
> > the struct page won't actually get freed. After taking the ref, it
> > rechecks that it has got the right page...
> 
> Ok, I understand the needs now.
> 
> I think the way the drivers/net/niu.c driver handles things
> would work better.  It only performs get_page(), atomic
> increments on compound_head(page)->_count, and __free_page().
> Is that all legal with the lockless pagecache?

Yes, you can use the regular refcounting and freeing operations. If the
lockless pagecache has got a speculative reference on the page after the
driver drops all "real" references, it will take care of freeing the page.

 
> If so I can likely rework the Cassini page management to
> behave similarly.

If you got the chance, that would be very nice.

Thanks,
Nick