From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754880Ab0ECRKZ (ORCPT ); Mon, 3 May 2010 13:10:25 -0400 Received: from tex.lwn.net ([70.33.254.29]:41750 "EHLO vena.lwn.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754099Ab0ECRKY (ORCPT ); Mon, 3 May 2010 13:10:24 -0400 Date: Mon, 3 May 2010 11:10:20 -0600 From: Jonathan Corbet To: Kevin Hilman Cc: mgross@linux.intel.com, aili@codeaurora.org, dwalker@codeaurora.org, tiwai@suse.de, bruce.w.allan@intel.com, davidb@quicinc.com, mcgrof@gmail.com, pavel@ucw.cz, linux-pm , lkml Subject: Re: [PATCH]PM QOS refresh against next-20100430 Message-ID: <20100503111020.2dab8003@bike.lwn.net> In-Reply-To: <87tyqo537l.fsf@deeprootsystems.com> References: <20100430212043.GA30315@linux.intel.com> <20100503103300.6330e522@bike.lwn.net> <87y6g1547o.fsf@deeprootsystems.com> <20100503104250.7605d2bc@bike.lwn.net> <87tyqo537l.fsf@deeprootsystems.com> Organization: LWN.net X-Mailer: Claws Mail 3.7.5 (GTK+ 2.20.0; x86_64-redhat-linux-gnu) Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 8bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, 03 May 2010 10:01:50 -0700 Kevin Hilman wrote: > > Well, shallow could mean that the state lacks the CPUIDLE_FLAG_DEEP > > flag; that should be relatively portable. In any case, it seems > > more so than "if I put in a 55us latency requirement, I'll stay out > > of C3". > > I guess it depends on your goal. Do you just want to stay out of C3 > on your current platform? or do you want to stay out of any low-power > state (on any platform) where you'll have a latency of > 55 usecs? My case is a camera driver; as long as latencies are on the scale of the frame rate, things won't go too wrong. 1ms latencies would not be a huge problem. What happens is that the hardware corrupts the data if it goes into C3 while video capture is happening. Comments in drivers/net/e1000e/netdev.c suggest that they were dealing with similar issues. > The former is not portable (as C-states don't have the same meanings > across arches/SoCs) where as using a real-world number in usecs will > have meaning on any platform. Understood. But the end result is that I've ended up specifying a latency requirement that is far tighter than the situation really needs in order to achieve the real goal. > My main concern is that drivers not be written with latency > constraints that assume an Intel-centric set of power states. There > are other SoCs out there with different sets of states and > corresponding latencies, so keeping things in real-world numbers > (latency, throughput, etc.) seems to me to be the only portable way. Understood. Perhaps the pm_qos interface isn't the best way to achieve the real goal here, but it's all we have at the moment. Thanks, jon