From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754782AbcHSBLk (ORCPT ); Thu, 18 Aug 2016 21:11:40 -0400 Received: from mx1.redhat.com ([209.132.183.28]:36324 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754409AbcHSBLi (ORCPT ); Thu, 18 Aug 2016 21:11:38 -0400 Message-ID: <1471543363.2581.30.camel@redhat.com> Subject: Re: [PATCH] usercopy: Skip multi-page bounds checking on SLOB From: Rik van Riel To: Linus Torvalds Cc: Kees Cook , Laura Abbott , Linux Kernel Mailing List , linux-mm , kernel test robot Date: Thu, 18 Aug 2016 14:02:43 -0400 In-Reply-To: References: <20160817222921.GA25148@www.outflux.net> <1471530118.2581.13.camel@redhat.com> Organization: Red Hat, Inc. Content-Type: text/plain; charset="UTF-8" Mime-Version: 1.0 Content-Transfer-Encoding: 8bit X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.31]); Thu, 18 Aug 2016 18:02:46 +0000 (UTC) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Thu, 2016-08-18 at 10:42 -0700, Linus Torvalds wrote: > On Thu, Aug 18, 2016 at 7:21 AM, Rik van Riel > wrote: > > > > One big question I have for Linus is, do we want > > to allow code that does a higher order allocation, > > and then frees part of it in smaller orders, or > > individual pages, and keeps using the remainder? > > Yes. We've even had people do that, afaik. IOW, if you know you're > going to allocate 16 pages, you can try to do an order-4 allocation > and just use the 16 pages directly (but still as individual pages), > and avoid extra allocation costs (and to perhaps get better access > patterns if the allocation succeeds etc etc). > > That sounds odd, but it actually makes sense when you have the order- > 4 > allocation as a optimistic path (and fall back to doing smaller > orders > when a big-order allocation fails). To make that *purely* just an > optimization, you need to let the user then treat that order-4 > allocation as individual pages, and free them one by one etc. > > So I'm not sure anybody actually does that, but the buddy allocator > was partly designed for that case. That makes sense.  With that in mind, it would probably be better to just drop all of the multi-page bounds checking from the usercopy code, not conditionally on SLOB. Alternatively, we could turn the __GFP_COMP flag into its negative, and set it only on the code paths that do what Linus describes (if anyone does it). A WARN_ON_ONCE in the page freeing code could catch these cases, and point people at exactly what to do if they trigger the warning. I am unclear no how to exclude legitimate usercopies that are larger than PAGE_SIZE from triggering warnings/errors, if we cannot identify every buffer where larger copies are legitimately going. Having people rewrite their usercopy code into loops that automatically avoids triggering page crossing or >PAGE_SIZE checks would be counterproductive, since that might just opens up new attack surface.