Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752435AbdHRRz1 (ORCPT + 2 others); Fri, 18 Aug 2017 13:55:27 -0400 Received: from mga14.intel.com ([192.55.52.115]:24480 "EHLO mga14.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752249AbdHRRzZ (ORCPT ); Fri, 18 Aug 2017 13:55:25 -0400 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.41,393,1498546800"; d="scan'208";a="139221206" Date: Fri, 18 Aug 2017 23:23:02 +0530 From: Rajneesh Bhardwaj To: Andy Shevchenko Cc: "Peter Zijlstra (Intel)" , Platform Driver , "dvhart@infradead.org" , Andy Shevchenko , "linux-kernel@vger.kernel.org" , Vishwanath Somayaji , dbasehore@chromium.org, rjw@rjwysocki.net, rajatja@google.com Subject: Re: your mail Message-ID: <20170818175301.GA12806@raj-desk2.iind.intel.com> References: <20170818174214.GA12519@raj-desk2.iind.intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20170818174214.GA12519@raj-desk2.iind.intel.com> User-Agent: Mutt/1.5.24 (2015-08-30) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Return-Path: On Fri, Aug 18, 2017 at 11:12:14PM +0530, Rajneesh Bhardwaj wrote: > Bcc: > Subject: Re: [PATCH] platform/x86: intel_pmc_core: Add Package C-states > residency info > Reply-To: > In-Reply-To: > Please ignore my previous email without subject. It was sent by mistake. > On Fri, Aug 18, 2017 at 08:17:32PM +0300, Andy Shevchenko wrote: > > +PeterZ (since I mentioned his name) > > > > On Fri, Aug 18, 2017 at 5:58 PM, Rajneesh Bhardwaj > > wrote: > > > On Fri, Aug 18, 2017 at 03:57:34PM +0300, Andy Shevchenko wrote: > > >> On Fri, Aug 18, 2017 at 3:37 PM, Rajneesh Bhardwaj > > >> wrote: > > >> > This patch introduces a new debugfs entry to read current Package C-state > > >> > residency values and, one new kernel API to read the Package C-10 residency > > >> > counter. > > >> > > > >> > Package C-state residency MSRs provide useful debug information about system > > >> > idle states. In idle states system must enter deeper Package C-states. > > > > >> Why this patch is needed? > > > > > > Andy, I'll try to give some background for this. > > > > > > This is needed to enhance the S0ix failure debug capabilities from within > > > the kernel. On ChromeOS we have S0ix failsafe kernel framework that is used > > > to validate S0ix and report the blockers in case of a failure. > > > https://patchwork.kernel.org/patch/9148999/ > > > > (It's not part of upstream) > > Sorry i sent an older link. There are fresh attempts to get this into > mainline kernel and looks like there is a traction for it. > https://patchwork.kernel.org/patch/9831229/ > > Package C-state (PC10) validation is discussed there. > > > > > > So far only intel_pmc_slp_s0_counter_read is called by this framework to > > > check whether the previous attempt to enter S0ix was success or not. > > > > I harder see even a single user of that API in current kernel. It > > should be unexported and removed I think. > > > > > Having > > > another PC10 counter related exported function enhances the S0ix debug since > > > PC10 state is a prerequisite to enter S0ix. > > > > > >> See, we have turbostat and cpupower user space tools which do this > > >> without any additional code to be written in kernel. What prevents > > >> your user space application do the same? > > >> > > >> Moreover, we have events for cstate, I assume perf or something alike > > >> can monitor those counters as well. > > > > > > You're right, perhaps the debugfs is redundant when we have those user space > > > tools but such tools are not available readily for all platforms/distros. > > > Interfaces like /dev/cpu/*/msr that turbostat uses are not available on all > > > the platforms. > > > PMC driver is a debug driver so i thought its better to show Package C-state > > > related info for low power debug here. > > > > > >> > > >> Sorry, NAK. > > > > > > This patch has two parts i.e. exported PC10 API and the debugfs. Based on > > > the above explanation, if the patch is not good as is, please let me know if > > > i should drop the debugfs part and respin a v2 with just the exported API or > > > drop this totally. > > > > > > Thanks for the feedback and thanks for taking time to review! > > > > Reading above makes me think that entire design of this is misguided. > > Since the most of values are counters they better to be accessed in a > > way how perf does. > > > > In case you need *in-kernel* facility, do some APIs (if it's not done > > yet) for events drivers first. > > cstate event driver is already in upstream. > > > > Sorry, NAK for entire patch until it would be blessed by people like Peter Z. > > > > -- > > With Best Regards, > > Andy Shevchenko > > -- > Best Regards, > Rajneesh -- Best Regards, Rajneesh