From mboxrd@z Thu Jan 1 00:00:00 1970 From: Dmitry Osipenko Subject: Re: [PATCH V5 11/18] clk: tegra210: Add support for Tegra210 clocks Date: Fri, 19 Jul 2019 01:52:59 +0300 Message-ID: <20190719015259.30941f3c@dimatab> References: <351a07d4-ba90-4793-129b-b1a733f95531@nvidia.com> <9271ae75-5663-e26e-df26-57cba94dab75@nvidia.com> <7ae3df9a-c0e9-cf71-8e90-4284db8df82f@nvidia.com> <46b55527-da5d-c0b7-1c14-43b5c6d49dfa@nvidia.com> <2de9a608-cf38-f56c-b192-7ffed65092f8@nvidia.com> <5eedd224-77b0-1fc9-4e5e-d884b41a64ed@nvidia.com> <89f23878-d4b2-2305-03e5-8a3e781c2b02@gmail.com> <20190718194222.GH12715@pdeschrijver-desktop.Nvidia.com> <056496ed-9abf-6907-c61c-a99ccf23b834@gmail.com> <1c85cb35-ce7c-1dd1-f637-0c91b2b36db3@nvidia.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: <1c85cb35-ce7c-1dd1-f637-0c91b2b36db3@nvidia.com> Sender: linux-kernel-owner@vger.kernel.org To: Sowjanya Komatineni Cc: Peter De Schrijver , sboyd@kernel.org, Michael Turquette , Joseph Lo , thierry.reding@gmail.com, jonathanh@nvidia.com, tglx@linutronix.de, jason@lakedaemon.net, marc.zyngier@arm.com, linus.walleij@linaro.org, stefan@agner.ch, mark.rutland@arm.com, pgaikwad@nvidia.com, linux-clk@vger.kernel.org, linux-gpio@vger.kernel.org, jckuo@nvidia.com, talho@nvidia.com, linux-tegra@vger.kernel.org, linux-kernel@vger.kernel.org, mperttunen@nvidia.com, spatra@nvidia.com, robh+dt@kernel.org, devicetree@vger.kernel.org List-Id: devicetree@vger.kernel.org =D0=92 Thu, 18 Jul 2019 13:36:35 -0700 Sowjanya Komatineni =D0=BF=D0=B8=D1=88=D0=B5=D1=82: > On 7/18/19 1:26 PM, Dmitry Osipenko wrote: > > 18.07.2019 22:42, Peter De Schrijver =D0=BF=D0=B8=D1=88=D0=B5=D1=82: =20 > >> On Thu, Jul 18, 2019 at 02:44:56AM +0300, Dmitry Osipenko wrote: =20 > >>>> dependencies I am referring are dfll_ref, dfll_soc, and DVFS > >>>> peripheral clocks which need to be restored prior to DFLL > >>>> reinit. =20 > >>> Okay, but that shouldn't be a problem if clock dependencies are > >>> set up properly. > >>> =20 > >>>>>> reverse list order during restore might not work as all other > >>>>>> clocks are in proper order no with any ref clocks for plls > >>>>>> getting restored prior to their clients =20 > >>>>> Why? The ref clocks should be registered first and be the roots > >>>>> for PLLs and the rest. If it's not currently the case, then > >>>>> this need to be fixed. You need to ensure that each clock is > >>>>> modeled properly. If some child clock really depends on > >>>>> multiple parents, then the parents need to in the correct order > >>>>> or CCF need to be taught about such multi-dependencies. > >>>>> > >>>>> If some required feature is missed, then you have to implement > >>>>> it properly and for all, that's how things are done in > >>>>> upstream. Sometimes it's quite a lot of extra work that > >>>>> everyone are benefiting from in the end. > >>>>> > >>>>> [snip] =20 > >>>> Yes, we should register ref/parents before their clients. > >>>> > >>>> cclk_g clk is registered last after all pll and peripheral > >>>> clocks are registers during clock init. > >>>> > >>>> dfllCPU_out clk is registered later during dfll-fcpu driver > >>>> probe and gets added to the clock list. > >>>> > >>>> Probably the issue seems to be not linking dfll_ref and dfll_soc > >>>> dependencies for dfllCPU_out thru clock list. > >>>> > >>>> clk-dfll driver during dfll_init_clks gets ref_clk and soc_clk > >>>> reference thru DT. =20 > >> The dfll does not have any parents. It has some clocks which are > >> needed for the logic part of the dfll to function, but there's no > >> parent clock as such unlike for peripheral clocks or PLLs where > >> the parent is at least used as a reference. The I2C controller of > >> the DFLL shares the lines with a normal I2C controller using some > >> arbitration logic. That logic only works if the clock for the > >> normal I2C controller is enabled. So you need probably 3 clocks > >> enabled to initialize the dfll in that case. I don't think it > >> makes sense to add complicated logic to the clock core to deal > >> with this rather strange case. To me it makes more sense to use > >> pmops and open code the sequence there. =20 > > It looks to me that dfllCPU is a PLL and dfll_ref is its reference > > parent, while dfll_soc clocks the logic that dynamically > > reconfigures dfllCPU in background. I see that PLLP is defined as a > > parent for dfll_ref and dfll_soc in the code. Hence seems dfll_ref > > should be set as a parent for dfllCPU, no? =20 >=20 > dfll_soc will not be restored by the time dfllCPU resume happens > after dfll_ref. >=20 > without dfll_soc, dfllCPU cannot be resumed either. So if we decide > to use parent we should use dfll_soc. Okay, my point is that the parents should be properly specified any ways. > > Either way is good to me, given that DFLL will be disabled during > > suspend. Resetting DFLL on DFLL's driver resume using PM ops should > > be good. And then it also will be better to error out if DFLL is > > active during suspend on the DFLL's driver suspend. =20 >=20 > Doing in dfll-fcpu pm_ops is much better as it happens right after > all clocks are restored and unlike other clock enables, dfll need > dfll controller programming as well and is actually registered in > dfll-fcpu driver. >=20 > With this, below is the sequence: >=20 > CPUFreq suspend switches CPU to PLLP and disables dfll >=20 > Will add dfll_suspend/resume in dfll-fcpu driver and in dfll suspend=20 > will check for dfll active and will error out suspend. >=20 > dfll resume does dfll reinit. >=20 > CPUFreq resume enables dfll and switches CPU to dfll. >=20 >=20 > Will go with doing in dfll-fcpu pm_ops rather than parenting > dfllCPU_OUT... >=20 Sounds good.