From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1758759Ab3K1Oix (ORCPT ); Thu, 28 Nov 2013 09:38:53 -0500 Received: from mail-qe0-f48.google.com ([209.85.128.48]:53157 "EHLO mail-qe0-f48.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751712Ab3K1Oiw (ORCPT ); Thu, 28 Nov 2013 09:38:52 -0500 Date: Thu, 28 Nov 2013 09:38:48 -0500 From: Tejun Heo To: Peter Zijlstra Cc: Oleg Nesterov , zhang.yi20@zte.com.cn, lkml , Tetsuo Handa , Ingo Molnar Subject: Re: [PATCH]: exec: avoid propagating PF_NO_SETAFFINITY into userspace child Message-ID: <20131128143848.GD3925@htj.dyndns.org> References: <20131126180420.GA18172@redhat.com> <20131127183117.GB13098@mtj.dyndns.org> <20131128091358.GH10022@twins.programming.kicks-ass.net> <20131128114542.GA3826@redhat.com> <20131128121748.GN10022@twins.programming.kicks-ass.net> <20131128133152.GA821@redhat.com> <20131128133947.GR10022@twins.programming.kicks-ass.net> <20131128141329.GB3925@htj.dyndns.org> <20131128143145.GT10022@twins.programming.kicks-ass.net> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20131128143145.GT10022@twins.programming.kicks-ass.net> User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hey, Peter. On Thu, Nov 28, 2013 at 03:31:45PM +0100, Peter Zijlstra wrote: > But the WQ_UNBOUND thingies should be just that and should thus not have > the NO_SETAFFINITY flag set because there is no valid reason to have it > set. > > Regardless of whether the threads are shared between unbound workqueues > or not. Hah? No, we do not want to allow userland to be able to set affinities on any workqueue workers, period. That's just inviting people to do weirdest things and then reporting things like "crypt jobs on some of our 500 machines end up stuck on a single cpu once in a while" which will eventually be tracked down to some weird shell script setting affinity on workers doing something else. We really want to insulate workers and pool operation from userland. e.g. unbound workqueues now default to per-NUMA affinity unless explicitly told not to, which leads to better overall behavior for most workloads. We do wanna keep those details from userland so that they can be improved in the future too. Thanks. -- tejun