Re: nonlinear swapping w/o pte_chains [Re: VMA_MERGING_FIXUP and patch]

[Andrea Arcangeli <andrea@suse.de>]

On Wed, Mar 24, 2004 at 12:18:12PM +0000, Hugh Dickins wrote:
> This subtlety in try_to_unmap_nonlinear_pte:
> 
> 	page_map_lock(page);
> 	/* check that we're not in between set_pte and page_add_rmap */
> 	if (page_mapped(page)) {
> 		unmap_pte_page(page, vma, address + offset, ptep);
> 
> Harmless, but isn't our acquisition of the page_table_lock guaranteeing
> that it cannot be in between set_pte and page_add_rmap?

I find that fragile, see the way I implemented do_anonymous_page, other
places always do page_add_rmap under the page_table_lock, but there's no
reason to require that, the swapout code already checks explicitly for
page_mapped after taking the page_map_lock, it has to do that anyways,
so I find it nicer to do it like the above and in do_anonymous_page.