On Wed, 28 Apr 2010 03:05:43 +0200 Andrea Arcangeli <aarcange@xxxxxxxxxx> wrote: > On Wed, Apr 28, 2010 at 09:39:48AM +0900, KAMEZAWA Hiroyuki wrote: > > Seems nice. > > What did you mean with objrmap inconsistency? I think this is single > threaded there, userland didn't run yet and I don't think page faults > could run. Maybe it's safer to add a VM_BUG_ON(vma->anon_vma) just > before vma->anon_vma = anon_vma to be sure nothing run in between. > I mean following relationship. vma_address(vma, page1) <-> address <-> pte <-> page2 If page1 == page2, objrmap is consistent. If page1 != page2, objrmap is inconsistent. > > I'll test this but I think we need to take care of do_mremap(), too. > > And it's more complicated.... > > do_mremap has to be safe by: > > 1) adjusting page->index atomically with the pte updates inside pt > lock (while it moves from one pte to another) > I reviewed do_mremap again and am thinking it's safe. new_vma = copy_vma(); .... ----------(*) move_ptes(). munmap unnecessary range. At (*), if page1 != page2, rmap_walk will not run correctly. But it seems copy_vma() keeps page1==page2 As I reported, when there is a problem, vma_address() returns an address but that address doesn't contain migration_pte. BTW, page->index is not updated, we just keep [start_address, pgoff] to be sane value. Thanks, -Kame -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>