[RFC 00/19] Add interconnect and devfreq support for i.MX8MQ
Martin Kepplinger
martin.kepplinger at puri.sm
Tue Oct 5 06:34:19 PDT 2021
Am Donnerstag, dem 30.09.2021 um 17:22 +0300 schrieb Abel Vesa:
> On 21-09-30 10:03:46, Martin Kepplinger wrote:
> > Am Mittwoch, dem 29.09.2021 um 14:44 +0300 schrieb Abel Vesa:
> > > On 21-09-24 12:20:26, Martin Kepplinger wrote:
> > > > hi Abel,
> > > >
> > > > thank you for the update (this is actually v2 of this RFC
> > > > right?)!
> > > >
> > > > all in all this runs fine on the imx8mq (Librem 5 and devkit) I
> > > > use. For all
> > > > the pl301 nodes I'm not yet sure what I can actually test /
> > > > switch
> > > > frequencies.
> > > >
> > >
> > > You can start by looking into each of the following:
> > >
> > > $ ls -1d /sys/devices/platform/soc at 0/*/devfreq/*/trans_stat
> > >
> > > and look if the transitions happen when a specific driver that is
> > > a
> > > icc user suspends.
> > >
> > > You can also look at:
> > >
> > > /sys/kernel/debug/interconnect/interconnect_summary
> > >
> > > and:
> > >
> > > /sys/kernel/debug/interconnect/interconnect_graph
> > >
> > > > But I still have one problem: lcdif/mxfb already has the
> > > > interconnect dram
> > > > DT property and I use the following call to request bandwidth:
> > > > https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsource.puri.sm%2Fmartin.kepplinger%2Flinux-next%2F-%2Fcommit%2Fd690e4c021293f938eb2253607f92f5a64f15688&data=04%7C01%7Cabel.vesa%40nxp.com%7C7fab8aca3a5f43d56f5608d983e8da67%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C1%7C637685858400552603%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=%2FzyEQdOLU8jQuUpqJ74GTWyfrDvavz%2BxZAgv1tcIu9Y%3D&reserved=0
> > > > (mainlining this is on our todo list).
> > > >
> > > > With your patchset, I get:
> > > >
> > > > [ 0.792960] genirq: Flags mismatch irq 30. 00000004 (mxsfb-
> > > > drm)
> > > > vs. 00000004 (mxsfb-drm)
> > > > [ 0.801143] mxsfb 30320000.lcd-controller: Failed to install
> > > > IRQ
> > > > handler
> > > > [ 0.808058] mxsfb: probe of 30320000.lcd-controller failed
> > > > with
> > > > error -16
> > > >
> > > > so the main devfreq user (mxsfb) is not there :) why?
> > > >
> > >
> > > OK, I admit, this patchset doesn't provide support for all the
> > > icc
> > > consumer drivers.
> > > But that should come at a later stage. I only provided example
> > > like
> > > fec and usdhc, to show
> > > how it all fits together.
> > >
> > > > and when I remove the interconnect property from the lcdif DT
> > > > node,
> > > > mxsfb
> > > > probes again, but of course it doesn't lower dram freq as
> > > > needed.
> > > >
> > > > Do I do the icc calls wrong in mxsfb despite it working without
> > > > your
> > > > patchset, or may there be something wrong on your side that
> > > > breaks
> > > > the mxsfb IRQ?
> > > >
> > >
> > > Do you have the following changes into your tree?
> > >
> > > diff --git a/arch/arm64/boot/dts/freescale/imx8mq.dtsi
> > > b/arch/arm64/boot/dts/freescale/imx8mq.dtsi
> > > index 00dd8e39a595..c43a84622af5
> > > 100644
> > >
> > >
> > > ---
> > > a/arch/arm64/boot/dts/freescale/imx8mq.dtsi
> > >
> > >
> > > +++
> > > b/arch/arm64/boot/dts/freescale/imx8mq.dtsi
> > >
> > >
> > > @@ -524,7 +524,7 @@ lcdif: lcd-controller at 30320000
> > > {
> > > <&clk
> > > IMX8MQ_VIDEO_PLL1>,
> > > <&clk
> > > IMX8MQ_VIDEO_PLL1_OUT>;
> > > assigned-clock-rates = <0>, <0>,
> > > <0>,
> > > <594000000>;
> > > - interconnects = <&noc
> > > IMX8MQ_ICM_LCDIF &noc IMX8MQ_ICS_DRAM>;
> > > + interconnects = <&icc
> > > IMX8MQ_ICM_LCDIF &icc IMX8MQ_ICS_DRAM>;
> > > interconnect-names =
> > > "dram";
> > > status =
> > > "disabled";
> > >
> > >
> > >
> > >
> > >
> > > @@ -1117,7 +1117,7 @@ mipi_csi1: csi at 30a70000
> > > {
> > >
> > > <&src
> > > IMX8MQ_RESET_MIPI_CSI1_PHY_REF_RESET>,
> > > <&src
> > > IMX8MQ_RESET_MIPI_CSI1_ESC_RESET>;
> > > fsl,mipi-phy-gpr = <&iomuxc_gpr
> > > 0x88>;
> > > - interconnects = <&noc
> > > IMX8MQ_ICM_CSI1
> > > &noc IMX8MQ_ICS_DRAM>;
> > > + interconnects = <&icc
> > > IMX8MQ_ICM_CSI1
> > > &icc IMX8MQ_ICS_DRAM>;
> > > interconnect-names =
> > > "dram";
> > > status =
> > > "disabled";
> > >
> > >
> > >
> > >
> > >
> > > @@ -1169,7 +1169,7 @@ mipi_csi2: csi at 30b60000
> > > {
> > >
> > > <&src
> > > IMX8MQ_RESET_MIPI_CSI2_PHY_REF_RESET>,
> > > <&src
> > > IMX8MQ_RESET_MIPI_CSI2_ESC_RESET>;
> > > fsl,mipi-phy-gpr = <&iomuxc_gpr
> > > 0xa4>;
> > > - interconnects = <&noc
> > > IMX8MQ_ICM_CSI2
> > > &noc IMX8MQ_ICS_DRAM>;
> > > + interconnects = <&icc
> > > IMX8MQ_ICM_CSI2
> > > &icc IMX8MQ_ICS_DRAM>;
> > > interconnect-names =
> > > "dram";
> > > status =
> > > "disabled";
> > >
> > >
> > >
> > > I forgot to update these in the current version of the patchset.
> > > Will
> > > do in the next version.
> > >
> > > Also, would help a lot if you could give me a link to a tree
> > > you're
> > > testing with.
> > > That way I can look exactly at what's going on.
> > >
> > >
> >
> >
> > thanks Abel, with the above fix of existing interconnects
> > properties my
> > system runs as expected and here's the output of
> >
> > for each in `ls -1d /sys/devices/platform/soc at 0/*/devfreq/*`; do
> > echo
> > $each; cat $each/trans_stat; done
> >
> > for mxsfb requesting (max) bandwith (display on):
> >
> > /sys/devices/platform/soc at 0/32700000.noc/devfreq/32700000.noc
> > From : To
> > : 133333333 400000000 800000000 time(ms)
> > 133333333: 0 1 0 624
> > 400000000: 0 0 1 28
> > * 800000000: 1 0 0 30624
> > Total transition : 3
> > /sys/devices/platform/soc at 0/3d400000.memory-
> > controller/devfreq/3d400000.memory-controller
> > From : To
> > : 25000000 100000000 800000000 time(ms)
> > 25000000: 0 0 1 620
> > 100000000: 0 0 0 0
> > * 800000000: 1 0 0 30652
> > Total transition : 2
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 0/devfreq/soc at 0:pl301 at 0
> > From : To
> > : 25000000 133333333 333333333 time(ms)
> > 25000000: 0 0 1 616
> > 133333333: 0 0 0 0
> > * 333333333: 1 0 0 30668
> > Total transition : 2
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 1/devfreq/soc at 0:pl301 at 1
> > From : To
> > : 25000000 266666666 time(ms)
> > * 25000000: 0 0 31284
> > 266666666: 0 0 0
> > Total transition : 0
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 2/devfreq/soc at 0:pl301 at 2
> > From : To
> > : 25000000 800000000 time(ms)
> > * 25000000: 0 0 31288
> > 800000000: 1 0 0
> > Total transition : 1
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 3/devfreq/soc at 0:pl301 at 3
> > From : To
> > : 25000000 800000000 time(ms)
> > * 25000000: 0 0 31292
> > 800000000: 1 0 0
> > Total transition : 1
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 4/devfreq/soc at 0:pl301 at 4
> > From : To
> > : 25000000 333333333 time(ms)
> > 25000000: 0 1 648
> > * 333333333: 0 0 30652
> > Total transition : 1
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 5/devfreq/soc at 0:pl301 at 5
> > From : To
> > : 25000000 500000000 time(ms)
> > * 25000000: 0 0 31304
> > 500000000: 1 0 0
> > Total transition : 1
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 6/devfreq/soc at 0:pl301 at 6
> > From : To
> > : 25000000 500000000 time(ms)
> > * 25000000: 0 0 31308
> > 500000000: 0 0 0
> > Total transition : 0
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 7/devfreq/soc at 0:pl301 at 7
> > From : To
> > : 25000000 128000000 500000000 time(ms)
> > * 25000000: 0 0 0 31312
> > 128000000: 0 0 0 0
> > 500000000: 1 0 0 0
> > Total transition : 1
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 8/devfreq/soc at 0:pl301 at 8
> > From : To
> > : 25000000 133333333 time(ms)
> > * 25000000: 0 0 31316
> > 133333333: 0 0 0
> > Total transition : 0
> > /sys/devices/platform/soc at 0/soc at 0:pl301 at 9/devfreq/soc at 0:pl301 at 9
> > From : To
> > : 25000000 133333333 266666666 time(ms)
> > 25000000: 0 0 5 1052
> > 133333333: 0 0 0 0
> > * 266666666: 5 0 0 30268
> > Total transition : 10
> >
> >
> > but with display off (mxsfb not requesting anything), I get the
> > same
> > fast freqs for noc and memory-controller. They should use the
> > lowest
> > freqs. Only pl301 at 4 switches to 25mhz in that case. That's odd.
> >
>
> Well, have a look at:
>
> /sys/devices/platform/soc at 0/soc at 0:pl301 at 9/devfreq/soc at 0:pl301 at 9
>
> even in the output you gave here, you can see that there are 5
> transisions between 25MHz and 266MHz. BTW, that is the USDHC pl301.
>
> I'm assuming you're booting with rootfs from usdhc not through nfs,
> right? Anyway, the noc and dram clocks rate only drop when there is
> no user enabling its own icc path to the dram.
>
> Keep in mind that the benefit of this approach is not only to drop
> the
> dram clock rate, but also to drop the rates of all the bus clocks on
> whenever possible.
>
> Yes, the perfect scenario would be, from power consumption point of
> view at least,
> have dram clock rate as low as possible and as long as possible,
> which
> implicitly means there is no one requesting the higher rate.
>
> If you want to observe the transitions number change for the dram
> devfreq node as well, you can run a simple sync from userspace and
> that will
> trigger a "high rate" request for the usdhc. Note, this will only
> happen
> if there are no other users asking for the higher rate.
correct, I boot from usdhc. and when I disable interconnect in usdhc
the behaviour actually makes sense. tree:
https://source.puri.sm/martin.kepplinger/linux-next/-/commits/5.15-rc4/librem5__integration_byzantium_new_devfreq_interconnect
And I can see the system using a bit less power in the "display off"
case now (and various freqs switching to the lowest).
I didn't yet test whether the new "consumers" (for example usb)
correctly request more bandwidth now.
The only thing I see is with the "display on" case, that
"32700000.interconnect" is switched to 800mhz now, where it used 400mhz
before this patchset. I should be able to find out why though :)
so, for a proof of concept (and after what you mentioned to change for
a next revision) this looks good to me.
thanks a lot!
martin
More information about the linux-arm-kernel
mailing list