On 04/04/23 17:24, Ulf Hansson wrote: > On Mon, 3 Apr 2023 at 19:42, Vignesh Raghavendra <vigneshr@xxxxxx> wrote: >> >> Hi, >> >> On 31/03/23 20:01, Ulf Hansson wrote: >>> On Fri, 31 Mar 2023 at 12:17, Vignesh Raghavendra <vigneshr@xxxxxx> wrote: >>>> >>>> From: Aswath Govindraju <a-govindraju@xxxxxx> >>>> >>>> Add support for suspend/resume and pm_runtime resume/suspend. >>>> >>>> Signed-off-by: Aswath Govindraju <a-govindraju@xxxxxx> >>>> Signed-off-by: Georgi Vlaev <g-vlaev@xxxxxx> >>>> Signed-off-by: Vignesh Raghavendra <vigneshr@xxxxxx> >>>> --- >>>> >>>> Since RFC v2: >>>> Address all comments around sdhci_am654_remove() >>>> Set autosuspend_delay to -1 as SDHCI will host rootfs via SD/eMMC and >>>> autosuspend can cause long latency for user interactive applications >>> >>> I am curious about the "long" latency. Do you have some data that you >>> can share for how long it takes to runtime-resume the device? >>> >> >> As part of wider teting, I discovered that on certain boards (J721e EVM >> for example), there is a latency of ~500ms for controller to start up >> (HW limitation around card detect debounce timer [0]). This results in >> momentary (but noticeable) lag when playing a video for example. > > Ohh, that kind of sucks. There is no way to workaround that I guess? Unfortunately no, > >> >> >> [0[ >> https://elixir.bootlin.com/linux/latest/source/drivers/mmc/host/sdhci_am654.c#L361 >> >> >>> Using -1 as the default value for the autosuspend_delay seems fine to >>> me, but did you consider using a longer timeout than the common 100ms? >>> Could that perhaps be a better default option? >> >> Its more of the startup delay than how long controller should remain >> active. So tuning autosuspend_delay does not seem to help. >> >>> >>> Of course, in the end this is a choice of performance in favor of >>> wasting energy. >> >> Yes, this is choice of opt-in vs opt-out. Users can tune the delay via >> sysfs as needed. >> >> I am setting it to -1 to avoid regressing any userspace apps (vs >> enabling autosuspend and then asking impacted users to set to -1) > > Alright, makes sense! > >> >> One solution is to use -1 on platforms where latency is significant >> based on compatible / SoC detection? > > That's a good idea! It seems a pity to make all SoCs suffer, if there > is only a subset that has this problem. Although, let's consider that > as an improvement that can be done on top of $subject patch. Sure, will follow up with a patch to reduce autosuspend_delay on platforms w/o above issue. > > So, applied for next, thanks! Great, thanks! -- Regards Vignesh
