Re: [PATCH 0/4] cpufreq: use generic cpufreq drivers for Exynos5250 platform

From: Javier Martinez Canillas
Date: Fri May 15 2015 - 13:48:17 EST


Hello Bartlomiej,

On 04/22/2015 07:37 PM, Bartlomiej Zolnierkiewicz wrote:
>
> On Tuesday, April 21, 2015 04:45:56 PM Kevin Hilman wrote:
>> Bartlomiej Zolnierkiewicz <b.zolnierkie@xxxxxxxxxxx> writes:
>>
>> > On Monday, April 20, 2015 02:07:33 PM Kevin Hilman wrote:
>> >> Bartlomiej Zolnierkiewicz <b.zolnierkie@xxxxxxxxxxx> writes:
>> >>
>> >> > Hi,
>> >> >
>> >> > This patch series removes the use of Exynos5250 specific support
>> >> > from exynos-cpufreq driver and enables the use of cpufreq-dt driver
>> >> > for this platform. The exynos-cpufreq driver itself is also removed
>> >> > as it is no longer used/needed after Exynos5250 support removal.
>> >> >
>> >> > This patch series has been tested on Exynos5250 based Arndale board.
>> >> >
>> >> > Depends on:
>> >> > - next-20150330 branch of linux-next kernel tree
>> >> > - "[PATCH 0/6] cpufreq: use generic cpufreq drivers for Exynos4210
>> >> > platform" [1]
>> >> > - "[PATCH 0/6] cpufreq: use generic cpufreq drivers for Exynos4x12
>> >> > platform" [2]
>> >> > - "[PATCH] cpufreq: exynos: remove dead ->need_apll_change method" [3]
>> >>
>> >> Any chance you could prepare a branch with all the dependencies for easy
>> >> testing?
>> >
>> > All cpufreq changes with needed dependencies are now availble in
>> >
>> > https://github.com/bzolnier/linux.git
>> >
>> > repository and the branch is
>> >
>> > next-20150330-generic-cpufreq-exynos5420-5800-v2
>>
>> Great, thanks.
>>
>> >> Also, The previous version from Thomas was v12, and this one is neither
>> >> versioned nor has any reference to what may have changed since that
>> >
>> > Please note that Thomas' patchset was split on separate parts (this is
>> > part #3) and heavily modified so the previous versioning was dropped.
>> >
>> > The cover letter of part #1 ("[PATCH 0/6] cpufreq: use generic cpufreq
>> > drivers for Exynos4210 platform") contains detailed changelog on what has
>> > been changed since Thomas' original v12 patch series. Individual Thomas'
>> > patches which were modified by me also contain such information.
>> >
>> > Part #2 ("[PATCH 0/6] cpufreq: use generic cpufreq drivers for Exynos4x12
>> > platform") was entirely new code when compared to Thomas' v12 patchset so
>> > its cover letter doesn't contain such detailed changelog as part #1.
>> >
>> > The newly posted part #4 ("[PATCH 0/8] cpufreq: add generic cpufreq driver
>> > support for Exynos5250/5800 platforms" https://lkml.org/lkml/2015/4/21/314)
>> > also contains the detailed changelog.
>> >
>> > However for part #3 (this one, "[PATCH 0/4] cpufreq: use generic cpufreq
>> > drivers for Exynos5250 platform") such summary changelog got missed for
>> > some reason. Here it is:
>> > - split Exynos5250 support from the original patch
>> > - moved E5250_CPU_DIV[0,1]() macros to clk-exynos5250.c
>> > - added CPU regulator supply property for Google Spring board
>> > - removed exynos-cpufreq driver entirely as it is no longer used/needed
>>
>> Great, thanks for clarifying.
>>
>> >> version. Also, on v12, I had several comments[1] and wonder if they've
>> >> been addressed.
>> >
>> > All issues previously reported should have been fixed. If you still see
>> > some problems please let me know.
>> >
>> > [ I see now that exynos5420-arndale-octa.dts, exynos5420-peach-pit.dts,
>> > exynos5420-smdk5420.dts and exynos5800-peach-pi.dts should also have
>> > been updated to contain CPU cluster regulator supply properties or else
>> > if the default vdd_arm/vdd_kfc regulator state is set to too low value
>> > there may be problems with stability when switching to higher than
>> > default frequencies. I have posted v2 version of patch #2/8 of part #4
>> > and pushed v2 combined branch on github. Sorry for the inconvenience. ]
>>
>> I've now tested your v2 branch with the bL switcher disabled, CPUidle
>> enabled and CPUfreq enabled.
>>
>> With the default governor set to performance, it fails to boot. The last
>> kernel messages on the console are:
>
> [ Small explanation for people not following the discussion from
> the start:
>
> This testing is relevant to part #4 of the rework: "[PATCH 0/8]
> cpufreq: add generic cpufreq driver support for Exynos5250/5800
> platforms" (https://lkml.org/lkml/2015/4/21/314";), not this one
> which is part #3 and has no known issues. ]
>

I know that Exynos5420/5422/5800 is related to part #4 and not #3 but I
wanted to answer in this thread since here is where Kevin reported the
issue. I tried your next-20150330-generic-cpufreq-exynos5420-5800-v2-debug
branch with exynos_defconfig plus CONFIG_BL_SWITCHER disabled and:

CONFIG_ARM_BIG_LITTLE_CPUFREQ=y
CONFIG_ARM_DT_BL_CPUFREQ=y

By default CONFIG_CPU_FREQ_GOV_PERFORMANCE=y but with that option it fails
to boot as well on my Exynos5420 Peach Pit so seems to be exactly what Kevin
reported on the Exynos5800 Peach Pi Chromebook.

>> [...]
>> [ 3.426021] cpu cpu0: bL_cpufreq_init: CPU 0 initialized
>> [ 3.431189] cpu cpu4: bL_cpufreq_init: CPU 4 initialized
>>
>> However, with the default governor set to userspace it boots fine until
>> my userspace (ubuntu) tries to enable the ondemand governor, and then it
>> hangs.
>>
>> For it to boot, I have to disable the ondemand governor, and set the
>> default governor to userspace.

Disabling CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE and the performance
governor and enabling the userpace governor and setting it as default
makes the kernel to boot again which is the same behaviour Kevin had.

Also I can see that the CPUS have a cpufreq and the scaling driver
is the arm b.L one:

# cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_driver
arm-big-little
arm-big-little
arm-big-little
arm-big-little
arm-big-little
arm-big-little
arm-big-little
arm-big-little

>
> I've tried with ARM big.LITTLE cpuidle support enabled (I've just noticed
> that it is not turned on in exynos_defconfig) and my ODROID-XU3 board
> fails to boot. This happens even with cpufreq support disabled (hard
> lockup happens during mmc initialization which is done just after cpufreq
> initialization).
>
> Could you please check if disabling cpuidle support helps?

I don't have CPUIdle enabled since as you said is not turned on in exynos
defconfig by default so I think that is just a red herring.

>
>> As I reported earlier on Thomas' series, I suspect this is related to
>> the fact that the higher OPPs aren't really functional without voltage
>> scaling also supported.
>
> Part #4 contains voltage scaling support for arm_big_little[_dt] driver
> so this should not be a problem any longer.
>
> You may try next-20150330-generic-cpufreq-exynos5420-5800-v2-debug
> branch from my github (with cpufreq debugging printks enabled) to check
> whether the voltage scaling is indeed done on your board.
>

The last boot log shown on the serial console with your debug branch is:

[ 3.078885] cpu cpu0: _get_cluster_clk_and_freq_table: clk: ee516a80 & freq table: ee511540, cluster: 0
[ 3.088128] cpu cpu0: bL_cpufreq_init: CPU 0 physical_cluster 0
[ 3.094024] cpu cpu0: bL_cpufreq_init: CPU 1 physical_cluster 0
[ 3.099912] cpu cpu0: bL_cpufreq_init: CPU 2 physical_cluster 0
[ 3.105871] cpu cpu0: bL_cpufreq_init: CPU 3 physical_cluster 0
[ 3.111710] cpu cpu0: bL_cpufreq_init: CPU 0 initialized
[ 3.117291] cpu cpu4: _get_cluster_clk_and_freq_table: clk: ee516d80 & freq table: ee50c180, cluster: 1
[ 3.126379] cpu cpu4: bL_cpufreq_init: CPU 4 physical_cluster 1
[ 3.132276] cpu cpu4: bL_cpufreq_init: CPU 5 physical_cluster 1
[ 3.138169] cpu cpu4: bL_cpufreq_init: CPU 6 physical_cluster 1
[ 3.144069] cpu cpu4: bL_cpufreq_init: CPU 7 physical_cluster 1
[ 3.149967] cpu cpu4: bL_cpufreq_init: CPU 4 initialized
[ 3.155453] arm_big_little: bL_cpufreq_set_rate: cpu: 4, old cluster: 1, new cluster: 1, freq: 1300000
[ 3.164551] arm_big_little: bL_cpufreq_set_rate_cluster: cpu 4, cluster: 1, 600 MHz, -1 mV --> 1300 MHz, -1 mV
[ 3.174646] arm_big_little: bL_cpufreq_register: Registered platform driver: dt-bl
[ 3.182305] Unable to handle kernel paging request at virtual address 60000186

so it seems to be a bug in the code and not about OPPs not functional due
voltage scaling not working as Kevin guessed.

However, as Kevin mentioned is strange that the regulator voltages are -1
as you can see in the above boot log.

Also when setting a scaling_setspeed to one of the scaling available
frequencies for a given CPU, I see in the log that the frequency is
scaled for all the cluster and correctly reported in scaling_cur_freq
but the scaled voltage is always reported as -1 again:

[ 899.002840] arm_big_little: bL_cpufreq_set_rate_cluster: cpu 0, cluster: 0, 1800 MHz, -1 mV --> 1100 MHz, -1 mV
[ 946.153852] arm_big_little: bL_cpufreq_set_rate: cpu: 0, old cluster: 0, new cluster: 0, freq: 1200000
[ 946.161785] arm_big_little: bL_cpufreq_set_rate_cluster: cpu 0, cluster: 0, 1100 MHz, -1 mV --> 1200 MHz, -1 mV
[ 975.328748] arm_big_little: bL_cpufreq_set_rate: cpu: 4, old cluster: 1, new cluster: 1, freq: 1200000
[ 975.336663] arm_big_little: bL_cpufreq_set_rate_cluster: cpu 4, cluster: 1, 600 MHz, -1 mV --> 1200 MHz, -1 mV

>> I'm also seeing the wait_until_divider_stable errors when switching
>> between the available A7 OPPs. I'd reported this one earlier as well,
>> along with the script to reproduce it.
>
> I've tried your script and it works fine for me (I only needed to change
> cpu4 to cpu0 as on ODROID-XU3 CPUs 0,5,6,7 are A7 and 1,2,3,4 are A15).
>
> Best regards,
> --
> Bartlomiej Zolnierkiewicz
> Samsung R&D Institute Poland
> Samsung Electronics
>

Best regards,
Javier
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/