From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756371Ab2EBJJ4 (ORCPT ); Wed, 2 May 2012 05:09:56 -0400 Received: from mail-ob0-f174.google.com ([209.85.214.174]:49402 "EHLO mail-ob0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756098Ab2EBJJy convert rfc822-to-8bit (ORCPT ); Wed, 2 May 2012 05:09:54 -0400 MIME-Version: 1.0 In-Reply-To: <20120502081705.GB16976@quack.suse.cz> References: <1335778207-6511-1-git-send-email-jack@suse.cz> <20120502081705.GB16976@quack.suse.cz> Date: Wed, 2 May 2012 19:09:54 +1000 Message-ID: Subject: Re: [PATCH] Describe race of direct read and fork for unaligned buffers From: Nick Piggin To: Jan Kara Cc: Jeff Moyer , KOSAKI Motohiro , Michael Kerrisk , LKML , linux-man@vger.kernel.org, linux-mm@kvack.org, mgorman@suse.de, Andrea Arcangeli , Woodman Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8BIT Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 2 May 2012 18:17, Jan Kara wrote: > On Wed 02-05-12 01:50:46, Nick Piggin wrote: >> KOSAKI-san is correct, I think. >> >> The race is something like this: >> >> DIO-read >>     page = get_user_pages() >>                                                         fork() >>                                                             COW(page) >>                                                          touch(page) >>     DMA(page) >>     page_cache_release(page); >> >> So whether parent or child touches the page, determines who gets the >> actual DMA target, and who gets the copy. >  OK, this is roughly what I understood from original threads as well. So > if our buffer is page aligned and its size is page aligned, you would hit > the corruption only if you do modify the buffer while IO to / from that buffer > is in progress. And that would seem like a really bad programming practice > anyway. So I still believe that having everything page size aligned will > effectively remove the problem although I agree it does not aim at the core > of it. I see what you mean. I'm not sure, though. For most apps it's bad practice I think. If you get into realm of sophisticated, performance critical IO/storage managers, it would not surprise me if such concurrent buffer modifications could be allowed. We allow exactly such a thing in our pagecache layer. Although probably those would be using shared mmaps for their buffer cache. I think it is safest to make a default policy of asking for IOs against private cow-able mappings to be quiesced before fork, so there are no surprises or reliance on COW details in the mm. Do you think?