From mboxrd@z Thu Jan 1 00:00:00 1970 From: skannan@codeaurora.org Subject: Re: [PATCH v3 1/2] PM / devfreq: Generic CPU frequency to device frequency mapping governor Date: Mon, 06 Aug 2018 22:49:48 -0700 Message-ID: <35ff3e303f5de9980979b49c89f9a687@codeaurora.org> References: <1533171465-25508-1-git-send-email-skannan@codeaurora.org> <20180802095608epcms1p33fb061543efc9ceb3ec12d5567ceffbc@epcms1p3> <45c77ee02536e237c054399cad4c7669@codeaurora.org> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <45c77ee02536e237c054399cad4c7669@codeaurora.org> Sender: linux-kernel-owner@vger.kernel.org To: myungjoo.ham@samsung.com Cc: Kyungmin Park , Chanwoo Choi , Rob Herring , Mark Rutland , georgi.djakov@linaro.org, vincent.guittot@linaro.org, daidavid1@codeaurora.org, bjorn.andersson@linaro.org, linux-pm@vger.kernel.org, devicetree@vger.kernel.org, linux-kernel@vger.kernel.org List-Id: linux-pm@vger.kernel.org On 2018-08-02 14:00, skannan@codeaurora.org wrote: > On 2018-08-02 02:56, MyungJoo Ham wrote: >>> Many CPU architectures have caches that can scale independent of the >>> CPUs. >>> Frequency scaling of the caches is necessary to make sure the cache >>> is not >>> a performance bottleneck that leads to poor performance and power. >>> The same >>> idea applies for RAM/DDR. >>> >>> To achieve this, this patch adds a generic devfreq governor that >>> takes the >>> current frequency of each CPU frequency domain and then adjusts the >>> frequency of the cache (or any devfreq device) based on the frequency >>> of >>> the CPUs. It listens to CPU frequency transition notifiers to keep >>> itself >>> up to date on the current CPU frequency. >>> >>> To decide the frequency of the device, the governor does one of the >>> following: >> >> This exactly has the same purpose with "passive" governor except for >> the >> single part: passive governor depends on another devfreq driver, not >> cpufreq driver. >> >> If both governors have many features in common, can you merge them >> into one? >> Passive governor also has "get_target_freq", which allows driver >> authors >> to define the mapping. > > Thanks for the review and pointing me to the passive governor. I agree > that at a high level they are both doing the same. I can certainly > stuff this CPU freq to dev freq mapping into passive governor, but I > think it'll just make one huge set of code that's harder to understand > and maintain because it trying to do different things under the hood. > > There are also a bunch of complexities and optimizations that come > with the cpufreq-map governor that don't fit with the passive > governor. > > 1. It's not one CPU who's frequency we have to listen to. There are > multiple CPUs/policies we have to aggregate across. > 2. We have to deal with big vs little having different needs/mappings. > 3. Since it's always just CPUfreq, I can optimize the handling in the > transition notifiers. If I have 4 different devices that are scaled > based on CPU freq, I still use only 1 transition notifier. It becomes > a bit harder to do with the passive governor. > 4. It requires that the device explicitly support the passive governor > and pick a mapping function. With cpufreq-map governor, the device > drivers don't need to make any changes. Whoever is making a > device/board can choose what devices to scale up base on CPU freq > based on their board and their needs. Even an end user can say, scale > the GPU based on my CPU based on interpolation if they choose to. > 5. If a device has some other use for the private data, it can't work > with passive governor, but can work with cpufreq-map governor. > 6. I also want to improve cpufreq-map governor to handle hotplug > correctly in later patches (needs more discussion) and that'll add > more complexity. > > I think for these reasons we shouldn't combine these two governors. > Let me know what you think. Friendly reminder. Thanks, Saravana