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

[Andrea Arcangeli <andrea@suse.de>]

On Tue, Mar 23, 2004 at 10:44:59PM +0100, Andrea Arcangeli wrote:
> +#ifdef __HAVE_ARCH_PAGE_TEST_AND_CLEAR_DIRTY
> +			get_page(page);
> +#endif
> +			unmap_pte_page(page, vma, address, ptep);
						  ^^^^^^^ + offset

I added some stuff in the shm threaded test program to verify
correctness after swapin and it showed userspace mm corruption on the
nonlinear pages, the above + offset is the fix.

I also merged anobjrmap-6 from Hugh so ppc and ppc64 compiles fine (plus
some other arch).

Also the race in do_anonymous_page couldn't trigger, the callers of
try_to_unmap always first check if the page is mapped after taking the
page_map_lock, so I'm backing it out.

The real race triggering such bug is that I'm not taking the
page_map_lock for all the pages before unmapping them so the
--page->mapcount aren't atomic while the page is unmapped by different
address spaces at the same time from multiple cpus.

It'll all be fixed in the next update.