On Tue, 4 Dec 2012, Hugh Dickins wrote: > > Yes, your patch fixes it Mel, but I prefer it as below, with a couple > of mods: removing the no longer true comment, and leaving shmem_swapin() > alone with just a comment. It appears to be the job of the rather weird > mpol_cond_copy() to drop the reference on the original mempolicy, and > clear MPOL_F_SHARED so the copy won't need one (it's trying to cope with > the fact that swapin_readahead will make an unknown number of calls to > alloc_page_vma). So I'd rather not add another mpol_cond_put there, > whose cond will never be met. Hold on, ignore that patch for now, I think I had my priorities upside down: it would be better for shmem_swapin() to behave as you proposed, and we delete the mpol_cond_copy() weirdness instead. Your 00442ad04a5e changed alloc_pages_vma() to keep its refcounting in balance, so it now does not matter that swapin_readahead() makes an unknown number of calls to it: we should simply take a reference before and drop it after, just as you do in shmem_alloc_page(). I'd still like to revisit alloc_page_vma(), and its refcount manipulations do now appear redundant; but changing that is not something I want to get into in a last minute rush. But getting rid of mpol_cond_copy() should be safe and clear, I'll test that out now and reply with an updated patch (or else admit I got confused). Hugh -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>