Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754766AbbFTKBc (ORCPT ); Sat, 20 Jun 2015 06:01:32 -0400 Received: from mail-pa0-f43.google.com ([209.85.220.43]:34545 "EHLO mail-pa0-f43.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752030AbbFTKBV (ORCPT ); Sat, 20 Jun 2015 06:01:21 -0400 Message-ID: <558539E8.5050101@samsung.com> Date: Sat, 20 Jun 2015 19:01:12 +0900 From: Krzysztof Kozlowski User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.7.0 MIME-Version: 1.0 To: Michael Turquette , Bartlomiej Zolnierkiewicz CC: linux-arm-kernel@lists.infradead.org, Lukasz Majewski , Kevin Hilman , Heiko Stuebner , linux-pm@vger.kernel.org, Viresh Kumar , Stephen Boyd , Tomasz Figa , linux-kernel@vger.kernel.org, Chanwoo Choi , Thomas Abraham , Kukjin Kim , Sylwester Nawrocki , Javier Martinez Canillas , linux-samsung-soc@vger.kernel.org Subject: Re: [PATCH 1/6] clk: add CLK_RECALC_NEW_RATES clock flag for Exynos cpu clock support References: <'@samsung.com> <20150618195846.9112.7144@quantum> <15113828.ygxT0eaek9@amdc1976> <3554100.H9OH7Yrz4m@amdc1976> <20150619145357.9112.63998@quantum> In-Reply-To: <20150619145357.9112.63998@quantum> Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 9525 Lines: 192 W dniu 19.06.2015 o 23:53, Michael Turquette pisze: > Quoting Bartlomiej Zolnierkiewicz (2015-06-19 05:35:23) >> On Friday, June 19, 2015 01:19:06 PM Bartlomiej Zolnierkiewicz wrote: >>> >>> Hi, >>> >>> On Thursday, June 18, 2015 12:58:46 PM Michael Turquette wrote: >>>> Quoting Sylwester Nawrocki (2015-05-13 07:13:13) >>>>> On 03/04/15 18:43, Bartlomiej Zolnierkiewicz wrote: >>>>>> This flag is needed to fix the issue with wrong dividers being setup >>>>>> by Common Clock Framework when using the new Exynos cpu clock support. >>>>>> >>>>>> The issue happens because clk_core_set_rate_nolock() calls >>>>>> clk_calc_new_rates(clk, rate) before both pre/post clock notifiers have >>>>>> a chance to run. In case of Exynos cpu clock support pre/post clock >>>>>> notifiers are registered for mout_apll clock which is a parent of armclk >>>>>> cpu clock and dividers are modified in both pre and post clock notifier. >>>>>> This results in wrong dividers values being later programmed by >>>>>> clk_change_rate(top). To workaround the problem CLK_RECALC_NEW_RATES >>>>>> flag is added and it is set for mout_apll clock later so the correct >>>>>> divider values are re-calculated after both pre and post clock notifiers >>>>>> had run. >>>>>> >>>>>> For example when using "performance" governor on Exynos4210 Origen board >>>>>> the cpufreq-dt driver requests to change the frequency from 1000MHz to >>>>>> 1200MHz and after the change state of the relevant clocks is following: >>>>>> >>>>>> Without use of CLK_GET_RATE_NOCACHE flag: >>>>>> >>>>>> fout_apll rate: 1200000000 >>>>>> fout_apll_div_2 rate: 600000000 >>>>>> mout_clkout_cpu rate: 600000000 >>>>>> div_clkout_cpu rate: 600000000 >>>>>> clkout_cpu rate: 600000000 >>>>>> mout_apll rate: 1200000000 >>>>>> armclk rate: 1200000000 >>>>>> mout_hpm rate: 1200000000 >>>>>> div_copy rate: 300000000 >>>>>> div_hpm rate: 300000000 >>>>>> mout_core rate: 1200000000 >>>>>> div_core rate: 1200000000 >>>>>> div_core2 rate: 1200000000 >>>>>> arm_clk_div_2 rate: 600000000 >>>>>> div_corem0 rate: 300000000 >>>>>> div_corem1 rate: 150000000 >>>>>> div_periph rate: 300000000 >>>>>> div_atb rate: 300000000 >>>>>> div_pclk_dbg rate: 150000000 >>>>>> sclk_apll rate: 1200000000 >>>>>> sclk_apll_div_2 rate: 600000000 >>>>>> >>>>>> >>>>>> With use of CLK_GET_RATE_NOCACHE flag: >>>>>> >>>>>> fout_apll rate: 1200000000 >>>>>> fout_apll_div_2 rate: 600000000 >>>>>> mout_clkout_cpu rate: 600000000 >>>>>> div_clkout_cpu rate: 600000000 >>>>>> clkout_cpu rate: 600000000 >>>>>> mout_apll rate: 1200000000 >>>>>> armclk rate: 1200000000 >>>>>> mout_hpm rate: 1200000000 >>>>>> div_copy rate: 200000000 >>>>>> div_hpm rate: 200000000 >>>>>> mout_core rate: 1200000000 >>>>>> div_core rate: 1200000000 >>>>>> div_core2 rate: 1200000000 >>>>>> arm_clk_div_2 rate: 600000000 >>>>>> div_corem0 rate: 300000000 >>>>>> div_corem1 rate: 150000000 >>>>>> div_periph rate: 300000000 >>>>>> div_atb rate: 240000000 >>>>>> div_pclk_dbg rate: 120000000 >>>>>> sclk_apll rate: 150000000 >>>>>> sclk_apll_div_2 rate: 75000000 >>>>>> >>>>>> Without this change cpufreq-dt driver showed ~10 mA larger energy >>>>>> consumption when compared to cpufreq-exynos one when "performance" >>>>>> cpufreq governor was used on Exynos4210 SoC based Origen board. >>>>>> >>>>>> This issue was probably meant to be workarounded by use of >>>>>> CLK_GET_RATE_NOCACHE and CLK_DIVIDER_READ_ONLY clock flags in >>>>>> the original Exynos cpu clock patchset (in "[PATCH v12 6/6] clk: >>>>>> samsung: remove unused clock aliases and update clock flags" patch) >>>>>> but usage of these flags is not sufficient to fix the issue observed. >>>>>> >>>>>> Cc: Thomas Abraham >>>>>> Cc: Tomasz Figa >>>>>> Cc: Mike Turquette >>>>>> Cc: Javier Martinez Canillas >>>>>> Signed-off-by: Bartlomiej Zolnierkiewicz >>>>>> --- >>>>>> drivers/clk/clk.c | 3 +++ >>>>>> include/linux/clk-provider.h | 1 + >>>>>> 2 files changed, 4 insertions(+) >>>>>> >>>>>> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c >>>>>> index f85c8e2..97cc73e 100644 >>>>>> --- a/drivers/clk/clk.c >>>>>> +++ b/drivers/clk/clk.c >>>>>> @@ -1771,6 +1771,9 @@ static void clk_change_rate(struct clk_core *clk) >>>>>> if (clk->notifier_count && old_rate != clk->rate) >>>>>> __clk_notify(clk, POST_RATE_CHANGE, old_rate, clk->rate); >>>>>> >>>>>> + if (clk->flags & CLK_RECALC_NEW_RATES) >>>>>> + (void)clk_calc_new_rates(clk, clk->new_rate); >>>>>> + >>>>>> /* >>>>>> * Use safe iteration, as change_rate can actually swap parents >>>>>> * for certain clock types. >>>>>> diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h >>>>>> index 28abf1b..8d1aebe 100644 >>>>>> --- a/include/linux/clk-provider.h >>>>>> +++ b/include/linux/clk-provider.h >>>>>> @@ -31,6 +31,7 @@ >>>>>> #define CLK_GET_RATE_NOCACHE BIT(6) /* do not use the cached clk rate */ >>>>>> #define CLK_SET_RATE_NO_REPARENT BIT(7) /* don't re-parent on rate change */ >>>>>> #define CLK_GET_ACCURACY_NOCACHE BIT(8) /* do not use the cached clk accuracy */ >>>>>> +#define CLK_RECALC_NEW_RATES BIT(9) /* recalc rates after notifications */ >>>>> >>>>> Mike, Stephen, what do you think about this? I'm rather resistant to >>>>> this new flag approach, it looks like a hack. I don't seem to have better >>>>> ideas to address the missing rate recalculation issue though. >>>> >>>> I also do not like it. The root of the problem is the use of clk >>>> notifiers to change clk rates. This is also a hack and it points towards >>>> some missing infrastructure in the clock framework. >>> >>> I'm very surprised by this. Clock changes using clock notifiers in >>> Thomas' original patchset were Acked by you: >>> >>> "[PATCH v12 1/6] clk: samsung: add infrastructure to register cpu clocks" >>> https://www.marc.info/?l=linux-arm-kernel&m=141657613203808&w=2 >>> >>> I've only fixed issues present within the original code (this 4 lines >>> workaround/hack change to clock subsystem is a result of this), I have >>> not changed it fundamentally. >> >> Moreover similar changes for rockchip SoCs (which were explicitly based >> on Thomas' patches as noted in the code!) have already been merged in >> 3.18: >> >> http://lkml.iu.edu/hypermail/linux/kernel/1410.1/02644.html >> >> and are available in commit f6fba5f6967dbc062a7c138d67e2314220f5dd04 >> ("clk: rockchip: add new clock-type for the cpuclk"). >> >> I understand that my findings have uncovered some clock subsystem >> deficiencies which resulted in afterthought about fundamental design >> of cpu clocks but I have a feeling that our patches are now being >> unjustly punished for making these issues public. >> >> I agree that current patches are not perfect (especially this patch) >> but they are good enough IMO. Please also understand that there were >> some serious work put into validating and reviewing them. > > You know what? You're right. > > I don't really like this cpu-clock code (and similarly I don't like > Tegra's EMC code which requires access to clk_hw_reparent). But I slept > on this issue overnight and it doesn't seem right for me to hold back > these patches when the better solution is currently vaporware (I have > some code but it's far from ready). > > It occurs to me that the best decision I can take is to merge it now and > then force you guys to switch over when the new infrastructure is > available. That is more reasonable than delaying the patches getting > pulled. > > So how to merge it? Viresh has given his Ack and is OK for the cpufreq > changes to go through another tree. I can take the whole thing with > Kukjin's Ack on the ARM dts patch, or we can set up an immutable > branch. I already acked the DTS [0] patch and I'm fine with it. However there would be some conflicts if this went through tree different than Samsung. There are many Exynos4 DTS changes in the queue for 4.2. Are there any objections for picking the DTS patch to Samsung tree? Best regards, Krzysztof [0] https://lkml.org/lkml/2015/5/7/999 > > If you really want to make my life easy you can send a pull request for > these patches (and the other Exynos cpufreq/cpu-clock series). > -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in Please read the FAQ at http://www.tux.org/lkml/