Re: invalidate_inode_pages in 2.5.32/3

[Daniel Phillips <phillips@arcor.de>]

On Friday 13 September 2002 00:30, Rik van Riel wrote:
> On Thu, 12 Sep 2002, Andrew Morton wrote:
> 
> > Well the lazy invalidation would be OK - defer that to the next
> > userspace access,
> 
> I think I have an idea on how to do that, here's some pseudocode:
> 
> invalidate_page(struct page * page) {
> 	SetPageInvalidated(page);
> 	rmap_lock(page);
> 	for_each_pte(pte, page) {
> 		make pte PROT_NONE;
> 		flush TLBs for this virtual address;
> 	}
> 	rmap_unlock(page);
> }
> 
> And in the page fault path:
> 
> if (pte_protection(pte) == PROT_NONE && PageInvalidated(pte_page_pte)) {
> 	clear_pte(ptep);
> 	page_cache_release(page);
> 	mm->rss--;
> }
> 
> What do you think, is this simple enough that it would work ? ;)

Too simple to work, unfortunately.  We have to at least 1) lock the
page and 2) remove it from the page cache.  Can we remove a page from
the page cache while it still has pte references?  I suppose we can,
it turns into an anonymous page.  But isn't that the reason we didn't
do it in invalidate_inode_pages in the first place?  However, since
we now, in addition, mark the page invalidated, it doesn't matter any
more what kind of page it is, so I suppose that's ok.

And we need a locked_page_cache_release->free_locked_page.

-- 
Daniel