From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754065AbbDYMvB (ORCPT ); Sat, 25 Apr 2015 08:51:01 -0400 Received: from shelob.surriel.com ([74.92.59.67]:51093 "EHLO shelob.surriel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753154AbbDYMu4 (ORCPT ); Sat, 25 Apr 2015 08:50:56 -0400 Message-ID: <553B8DA9.3060600@surriel.com> Date: Sat, 25 Apr 2015 08:50:49 -0400 From: Rik van Riel User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.4.0 MIME-Version: 1.0 To: Heiko Carstens CC: linux-kernel@vger.kernel.org, Andy Lutomirsky , Frederic Weisbecker , Peter Zijlstra , williams@redhat.com Subject: Re: [PATCH v2] context_tracking: remove local_irq_save from __acct_update_integrals References: <20150424111653.2a87a103@annuminas.surriel.com> <20150425094346.GA5897@osiris> In-Reply-To: <20150425094346.GA5897@osiris> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 04/25/2015 05:43 AM, Heiko Carstens wrote: > On Fri, Apr 24, 2015 at 11:16:53AM -0400, Rik van Riel wrote: >> V2: introduce signed_cputime_t to deal with 64 bit cputime_t on >> 32 bit architectures, and use READ_ONCE to ensure the value >> is always read atomically (Heiko Karstens) > > Erm, that's not what I said ;) > READ_ONCE() only fixes the isssue that with your previous code the > compiler was free to generate code that accesses the memory value > several times. Ah indeed, you are right. >> - local_irq_save(flags); >> time = stime + utime; >> - dtime = time - tsk->acct_timexpd; >> + dtime = time - READ_ONCE(tsk->acct_timexpd); >> + /* >> + * This code is called both from irq context and from >> + * task context. There is a race where irq context advances >> + * tsk->acct_timexpd to a value larger than time, creating >> + * a negative value. In that case, the irq has already >> + * updated the statistics. >> + */ >> + if (unlikely((signed_cputime_t)dtime <= 0)) >> + return; >> + > > ...the READ_ONCE() doesn't give you any guarantees about reading > tsk->acct_timexpd in an atomic way. > Well, actually you don't need atomic semantics, but only to make sure that > the read access happens with a single instruction, since you want to protect > against interrupts. > But still: if the size of acct_timexpd is 64 bit READ_ONCE() may still result > in two instructions on 32 bit architectures. > (or isn't there currently no 32 bit architecture with 64 bit cputime_t left?) Even if there is (maybe some ARM system?), can we even guarantee that a single instruction to read 64 bits exists on such a system? -- All rights reversed.