Re: [RFC v2 0/6] Managing cluser-level c-states with generic power domains

[Marc Titinger <mtitinger@baylibre.com>]

On 19/10/2015 22:58, Lina Iyer wrote:
> Hi Marc,
>
> I am trying to apply this on top of Axel's patches on linux-next (after
> fixing issues I saw with his v9), and running to issues applying your
> patches. Could you rebase on top of his v10 (he said he would send to
> the ML soon) ?
>

Hi Lina,

I want to replay this with Juno this afternoon first, I'll post ASAP.

Also, based on Kevin's comment I was wondering if I should drop this 
path already and try the other way as discussed (hook l2 devices to 
runtime-pm, through the CPU device), but I still need to think about 
this first.

Cheers,
Marc.



> Thanks,
> Lina
>
> On Tue, Oct 06 2015 at 08:27 -0600, Marc Titinger wrote:
>> v2:
>> - rebase on Lina Iyer's latest series
>> - remove unnecessary dependency on perf-state patches from Axel Haslam
>>
>> -----------------------
>>
>> Summary
>>
>> 1) DESCRIPTION
>> 2) DEPENDENCIES
>> 3) URL
>> ------------------------
>>
>>
>> 1) DESCRIPTION
>>
>>
>>     This patch set's underlying idea is that cluster-level c-states
>> can be managed
>> by the power domain, building upon Lina Iyers recent work on
>> CPU-domain, and Axel Haslam's
>> genpd multiple states. The power domain may contain CPU devices and
>> non-CPU devices.
>>
>> Non-CPU Devices may expose latency constraints by registering
>> intermediate power-states upon
>> probing, for instance shallower states than the deepest cluster-off
>> state. The generic
>> power domain governor may chose a device retention state in place of
>> the cluster-sleep
>> state demanded by the menu governor, and call the platform specific
>> handling to enter/leave
>> that retention state.
>>
>>
>> power-states
>> -----------
>>
>>
>> The proposed way how cluster-level c-states are declared as manageable
>> by the
>> power domain, rather than through the cpuidle-ops, relies on the
>> introduction of
>> "power-states", consistent with c-states. Here is an example of the DT
>> bindings,
>> the c-state CLUSTER_SLEEP_0 is exposed as a power-state in the
>> compatible property:
>>
>> juno.dts:           idle-states {
>>                        entry-method = "arm,psci";
>>
>>                        CPU_SLEEP_0: cpu-sleep-0 {
>>                                compatible = "arm,idle-state";
>>                                arm,psci-suspend-param = <0x0010000>;
>>                                local-timer-stop;
>>                                entry-latency-us = <100>;
>>                                exit-latency-us = <250>;
>>                                min-residency-us = <2000>;
>>                        };
>>
>>                        CLUSTER_SLEEP_0: cluster-sleep-0 {
>>                                compatible = "arm,power-state";
>>                                arm,psci-suspend-param = <0x1010000>;
>>                                local-timer-stop;
>>                                entry-latency-us = <800>;
>>                                exit-latency-us = <700>;
>>                                min-residency-us = <2500>;
>>                        };
>>         }
>>
>> This will tell cpuidle runtime_put/get the CPU devices for this
>> c-state. Eventually, the
>> actual platform handlers may be called from the genpd platform ops (in
>> place of cpuidle_ops).
>>
>> "drivers/cpuidle/cpuidle-arm.c":
>>
>> static const struct of_device_id arm_idle_state_match[] __initconst = {
>>        {.compatible = "arm,idle-state",
>>         .data = arm_enter_idle_state},
>>        {.compatible = "arm,power-state",
>>         .data = arm_enter_power_state},
>> };
>>
>>
>> In case of a power-state, arm_enter_power_state will only call
>> pm_runtime_put/get_sync
>> The power doamin will handle the power off, currently this patch set
>> lacks the final
>> call to the psci interface to have a fully fonctionnal setup
>> (and there are some genpd_lock'ing issues if put/get actually suspend
>> the CPU device.)
>>
>> Ultimately, we would like the Power Domain's simple governor to being
>> able to chose
>> the cluster power-state based on the c-states defered to it
>> (power-states) and constraints
>> added by the devices. Consequently, we need to "soak" those
>> power-states into the
>> power-domain intermediate states from Axel. Since power-states are
>> declared and handled
>> the same manner than c-states (idle-states in DT), these patches add a
>> soaking used when
>> attaching to a genpd, where power-states are parsed from the DT into
>> the genpd states:
>>
>>
>> "drivers/base/power/domain.c":
>>
>> static const struct of_device_id power_state_match[] = {
>>        {.compatible = "arm,power-state",
>>         },
>> };
>>
>> int of_genpd_device_parse_states(struct device_node *np,
>>                                 struct generic_pm_domain *genpd)
>>
>> debugfs addition
>> ---------------
>>
>> To easy debug, this patch set adds a seq-file names "states" to the
>> pm_genpd debugfs:
>>
>>    cat /sys/kernel/debug/pm_genpd/*
>>
>>      Domain             State name        Enter (ns) / Exit (ns)
>>    -------------------------------------------------------------
>>    a53_pd               cluster-sleep-0      1500000 / 800000
>>    a57_pd               cluster-sleep-0      1500000 / 800000
>>
>> And also a seq-file "timings", to help visualize the constrains of the
>> non-CPU
>> devices in a cluster PD.
>>
>>    Domain Devices, Timings in ns
>>                       Stop/Start Save/Restore, Effective
>> ----------------------------------------------------  ---
>> a57_pd
>>    /cpus/cpu@0          800   /740    1320  /1720  ,0 (cached stop)
>>    /cpus/cpu@1          800   /740    1420  /1780  ,0 (cached stop)
>>    /D1                  660   /580    16560 /6080  ,2199420 (cached stop)
>>
>>
>> Device power-states
>> -------------------
>>
>> some devices, like L2 caches, may feature a shallower retention mode,
>> between CPU_SLEEP_0
>> and CLUSTER_SLEEP_0, in which mode the L2 memory is not powered off,
>> leading to faster
>> resume than CLUSTER_SLEEP_0.
>>
>> One way to handle device constrains and retention features in the
>> power-domain, is to
>> allow devices to register a new power-state (consistent with a c-state).
>>
>> idle-states:
>>
>>                        D1_RETENTION: d1-retention {
>>                                compatible = "arm,power-state";
>>                                /*leave the psci param, for demo/testing:
>>                                * the psci cpuidle driver will not
>> currently
>>                                * understand that a c-state shall not
>> have it's
>>                                * table entry with a firmware command.
>>                                * the actual .power_on/off would be
>> registered
>>                                * by the DECLARE macro for a given
>> domain*/
>>                                arm,psci-suspend-param = <0x1010000>;
>>                                local-timer-stop;
>>                                entry-latency-us = <800>;
>>                                exit-latency-us = <200>;
>>                                min-residency-us = <2500>;
>>                        };
>>
>>
>>        D1 {
>>                compatible = "fake,fake-driver";
>>                name = "D1";
>>                constraint = <30000>;
>>                power-domains = <&a53_pd>;
>>         power-states =<&D1_RETENTION>;
>>        };
>>
>>
>> The genpd simple governor can now upon suspend of the last-man CPU
>> chose a shallower
>> retention state than CLUSTER_SLEEP_0.
>>
>> In order to achieve this, this patch set added the power-state parsing
>> during the
>> genpd_dev_pm_attach call. Multiple genpd states are now inserted in a
>> sorted manner
>> according to their depth: see pm_genpd_insert_state in
>> "drivers/base/power/domain.c".
>>
>>
>>
>> 2) DEPENDENCIES
>>
>>     This patch set applies over linux-4.2rc5 plus the following
>> ordered dependencies:
>>
>> * Ulf Hansson:
>>
>> 6637131 New          [V4] PM / Domains: Remove intermediate states
>> from the power off sequence
>>
>> * Lina Iyer's patch series:
>>
>> 7118981 Not Applicable [v2,1/7] PM / Domains: Allocate memory outside
>> domain locks
>> 7118991 Not Applicable [v2,2/7] PM / Domains: Support IRQ safe PM domains
>> 7119001 Not Applicable [v2,3/7] drivers: cpu: Define CPU devices as
>> IRQ safe
>> 7119011 Not Applicable [v2,4/7] PM / Domains: Introduce PM domains for
>> CPUs/clusters
>> 7119021 Not Applicable [v2,5/7] ARM: cpuidle: Add runtime PM support
>> for CPU idle
>> 7119031 Not Applicable [v2,6/7] ARM64: smp: Add runtime PM support for
>> CPU hotplug
>> 7119041 Not Applicable [v2,7/7] ARM: smp: Add runtime PM support for
>> CPU hotplug
>>
>> * John Medhurst:
>>
>> 6303671 New          arm64: dts: Add idle-states for Juno
>>
>> * Axel Haslam:
>>
>> 6301741 Not Applicable [v7,1/5] PM / Domains: prepare for multiple states
>> 6301751 Not Applicable [v7,2/5] PM / Domains: core changes for
>> multiple states
>> 6301781 Not Applicable [v7,3/5] PM / Domains: make governor select
>> deepest state
>> 6301771 Not Applicable [v7,4/5] ARM: imx6: pm: declare pm domain
>> latency on power_state struct.
>> 6301761 Not Applicable [v7,5/5] PM / Domains: remove old power on/off
>> latencies.
>>
>> 2) URL
>>
>> playable from https://github.com/mtitinger/linux-pm.git
>>
>> by adding the "fake driver D1" and launching the test-dev-state.sh
>> script.
>> this will show the power domain suspending to an intermediate state,
>> based on the
>> device constraints.
>>
>>    domain                      status pstate     slaves
>>           /device                                      runtime status
>> -----------------------------------------------------------------------------------
>>
>> a53_pd                          on
>>    /devices/system/cpu/cpu0                            active
>>    /devices/system/cpu/cpu3                            suspended
>>    /devices/system/cpu/cpu4                            suspended
>>    /devices/system/cpu/cpu5                            suspended
>> a57_pd                          d1-retention
>>    /devices/system/cpu/cpu1                            suspended
>>    /devices/system/cpu/cpu2                            suspended
>>    /devices/platform/D1
>>
>> ----------------------------------------------------------------------
>>
>> Marc Titinger (6):
>>  arm64: Juno: declare generic power domains for both clusters.
>>  PM / Domains: prepare for devices that might register a power state
>>  PM / Domains: introduce power-states consistent with c-states.
>>  PM / Domains: succeed & warn when attaching non-irqsafe devices to an
>>    irq-safe domain.
>>  arm: cpuidle: let genpd handle the cluster power transition with
>>    'power-states'
>>  PM / Domains: add debugfs 'states' and 'timings' seq files
>>
>> .../devicetree/bindings/arm/idle-states.txt        |  21 +-
>> .../devicetree/bindings/power/power_domain.txt     |  29 ++
>> arch/arm64/boot/dts/arm/juno.dts                   |  25 +-
>> drivers/base/power/cpu-pd.c                        |   5 +
>> drivers/base/power/domain.c                        | 415
>> +++++++++++++++------
>> drivers/cpuidle/cpuidle-arm.c                      |  52 ++-
>> include/linux/pm_domain.h                          |  21 +-
>> 7 files changed, 437 insertions(+), 131 deletions(-)
>>
>> --
>> 1.9.1
>>