Message-ID: <CAKddAkA6ghz3LmVwkdCJ8n+JeF-67pZi07NuPWfron46NLpKoQ@mail.gmail.com>
Date: Thu, 23 May 2024 11:26:38 +0800
From: Nick Hu <nick.hu@...ive.com>
To: Anup Patel <apatel@...tanamicro.com>
Cc: Anup Patel <anup@...infault.org>, Ulf Hansson <ulf.hansson@...aro.org>, palmer@...belt.com, 
	rafael@...nel.org, daniel.lezcano@...aro.org, paul.walmsley@...ive.com, 
	linux-pm@...r.kernel.org, linux-riscv@...ts.infradead.org, 
	linux-kernel@...r.kernel.org, zong.li@...ive.com, 
	Cyan Yang <cyan.yang@...ive.com>
Subject: Re: [PATCH] cpuidle: riscv-sbi: Add cluster_pm_enter()/exit()

Hi Anup,

On Fri, May 17, 2024 at 12:39 PM Anup Patel <apatel@...tanamicro.com> wrote:
>
> On Thu, May 16, 2024 at 9:40 AM Nick Hu <nick.hu@...ive.com> wrote:
> >
> > Hi Anup
> >
> > On Wed, May 15, 2024 at 9:46 PM Anup Patel <anup@...infault.org> wrote:
> > >
> > > Hi Nick,
> > >
> > > On Wed, May 15, 2024 at 5:45 PM Nick Hu <nick.hu@...ive.com> wrote:
> > > >
> > > > Hi Anup,
> > > >
> > > > Thank you for your guidance.
> > > > After enabling the debug message, we found a way to solve the problem
> > > > by the following steps:
> > > > 1. Add a compatible string in 'power-domains' node otherwise it won't
> > > > be the supplier of the consumers. (See of_link_to_phandle())
> > >
> > > Hmm, requiring a compatible string is odd. Where should we document
> > > this compatible string ?
> > >
> > Sorry, this is my fault. I didn't include some updates in
> > of_link_to_phandle(). This led some misunderstandings here.
> > You are right, we don't need it.
> > The supplier will be linked to the CLUSTER_PD node.
> >
> > > > 2. Move the 'power-domains' node outside the 'cpus' node otherwise it
> > > > won't be added to the device hierarchy by device_add().
> > > > 3. Update the cpuidle-riscv-sbi driver to get the pds_node from
> > > > '/power-domains'.
> > >
> > > By adding a compatible string and moving the "power-domains" node
> > > outside, you are simply forcing the OF framework to populate devices.
> > >
> > > How about manually creating platform_device for each power-domain
> > > DT node using of_platform_device_create() in sbi_pd_init() ?
> > >
> > Thanks for the suggestion! We have test the solution and it could work.
> > We was wondering if it's feasible for us to relocate the
> > 'power-domains' node outside of the /cpus? The CLUSTER_PD might
> > encompass not only the CPUs but also other components within the
> > cluster.
>
> The cpuidle-riscv-sbi driver expects "power-domains" DT node
> under "/cpus" DT node because this driver only deals with power
> domains related to CPU cluster or CPU cache-hierarchy. It does
> make sense to define L2/L3 power domains under
> "/cpus/power-domain" since these are related to CPUs.
>
> Moving the CPU "power-domains" DT node directly under "/" or
> somewhere else would mean that it covers system-wide power
> domains which is not true.
>
> I suggest we continue using "/cpus/power-domains" DT node
> only for power domains related to CPU clusters or CPU
> cache-hierarchy.
>
> For system wide power domains of SoC devices, we can either:
> 1) Use device power domains through the SBI MPXY extension
>     via different driver
> 2) Use a platform specific driver
>
Thank you for your valuable feedback. We will continue to use the
"/cpus/power-domains".

> >
> > We also look at cpuidle_psci_domain driver and it seems Arm doesn't
> > create the devices for each subnode of psci domain.
> > Is there any reason that they don't need it?
>
> Existing ARM DTS files under arch/arm64/boot/dts, use device
> power domains through SCMI (or platform specific mechanism)
> which are already populated as devices by Linux DD framework.
>
> Regards,
> Anup
>
Thank you for the explanation!

Regards,
Nick


> >
> > > >
> > > > So the DTS will be like:
> > > > cpus {
> > > >     ...
> > > >      domain-idle-states {
> > > >            CLUSTER_SLEEP:cluster-sleep {
> > > >                         compatible = "domain-idle-state";
> > > >                         ...
> > > >             }
> > > >      }
> > > > }
> > > > power-domains {
> > > >             compatible = "riscv,sbi-power-domains"
> > > >             ...
> > > >             ...
> > > >             CLUSTER_PD: clusterpd {
> > > >                     domain-idle-states = <&CLUSTER_SLEEP>;
> > > >             };
> > > > }
> > > > soc {
> > > >       deviceA@...{
> > > >              ...
> > > >              power-domains = <&CLUSTER_PD>;
> > > >              ...
> > > >       }
> > > > }
> > >
> > > Regards,
> > > Anup
> > >
> > > >
> > > > Regards,
> > > > Nick
> > > >
> > > > On Tue, May 14, 2024 at 10:54 PM Anup Patel <anup@...infault.org> wrote:
> > > > >
> > > > > On Tue, May 14, 2024 at 7:53 PM Anup Patel <anup@...infault.org> wrote:
> > > > > >
> > > > > > Hi Nick,
> > > > > >
> > > > > > On Tue, May 14, 2024 at 3:20 PM Nick Hu <nick.hu@...ivecom> wrote:
> > > > > > >
> > > > > > > Hi Ulf,
> > > > > > >
> > > > > > > Thank you for your valuable suggestion.
> > > > > > > I sincerely apologize for the delay in responding to your message We
> > > > > > > have diligently worked on experimenting with the suggestion you
> > > > > > > provided.
> > > > > > >
> > > > > > > As per your recommendation, we have incorporated the "power-domains=<>
> > > > > > > property" into the consumer's node, resulting in modifications to the
> > > > > > > DTS as illustrated below:
> > > > > > >
> > > > > > > cpus {
> > > > > > >     ...
> > > > > > >      domain-idle-states {
> > > > > > >            CLUSTER_SLEEP:cluster-sleep {
> > > > > > >                         compatible = "domain-idle-state";
> > > > > > >                         ...
> > > > > > >             }
> > > > > > >      }
> > > > > > >      power-domains {
> > > > > > >             ...
> > > > > > >             ...
> > > > > > >             CLUSTER_PD: clusterpd {
> > > > > > >                     domain-idle-states = <&CLUSTER_SLEEP>;
> > > > > > >             };
> > > > > > >      }
> > > > > > > }
> > > > > > > soc {
> > > > > > >       deviceA@...{
> > > > > > >              ...
> > > > > > >              power-domains = <&CLUSTER_PD>;
> > > > > > >              ...
> > > > > > >       }
> > > > > > > }
> > > > > > >
> > > > > > > However, this adjustment has led to an issue where the probe for
> > > > > > > 'deviceA' is deferred by 'device_links_check_suppliers()' within
> > > > > > > 'really_probe()'. In an attempt to mitigate this issue, we
> > > > > > > experimented with a workaround by adding the attribute
> > > > > > > "status="disabled"" to the 'CLUSTER_PD' node. This action aimed to
> > > > > > > prevent the creation of a device link between 'deviceA' and
> > > > > > > 'CLUSTER_PD'. Nevertheless, we remain uncertain about the
> > > > > > > appropriateness of this solution.
> > > > > > >
> > > > > > > Do you have suggestions on how to effectively address this issue?
> > > > > >
> > > > > > I totally missed this email since I was not CC'ed sorry about that. Please
> > > > > > use get_maintainers.pl when sending patches.
> > > > >
> > > > > I stand corrected. This patch had landed in the "spam" folder. I don't know why.
> > > > >
> > > > > Regards,
> > > > > Anup
> > > > >
> > > > > >
> > > > > > The genpd_add_provider() (called by of_genpd_add_provider_simple())
> > > > > > does mark the power-domain DT node as initialized (fwnode_dev_initialized())
> > > > > > so after the cpuidle-riscv-sbi driver is probed the 'deviceA' dependency is
> > > > > > resolved and 'deviceA' should be probed unless there are other unmet
> > > > > > dependencies.
> > > > > >
> > > > > > Try adding "#define DEBUG" before all includes in drivers/core/basec
> > > > > > and add "loglevel=8" in kernel parameters, this will print producer-consumer
> > > > > > linkage of all devices.
> > > > > >
> > > > > > Marking the power-domain DT node as "disabled" is certainly not the
> > > > > > right way.
> > > > > >
> > > > > > Regards,
> > > > > > Anup
> > > > > >
> > > > > > >
> > > > > > > Regards,
> > > > > > > Nick
> > > > > > >
> > > > > > > On Tue, Apr 30, 2024 at 4:13 PM Ulf Hansson <ulf.hansson@...aro.org> wrote:
> > > > > > > >
> > > > > > > > On Mon, 29 Apr 2024 at 18:26, Nick Hu <nick.hu@...ive.com> wrote:
> > > > > > > > >
> > > > > > > > > On Tue, Apr 30, 2024 at 12:22 AM Nick Hu <nick.hu@...ive.com> wrote:
> > > > > > > > > >
> > > > > > > > > > Hi Ulf
> > > > > > > > > >
> > > > > > > > > > On Mon, Apr 29, 2024 at 10:32 PM Ulf Hansson <ulf.hansson@...aro.org> wrote:
> > > > > > > > > > >
> > > > > > > > > > > On Mon, 26 Feb 2024 at 07:51, Nick Hu <nick.hu@...ivecom> wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > When the cpus in the same cluster are all in the idle state, the kernel
> > > > > > > > > > > > might put the cluster into a deeper low power state Call the
> > > > > > > > > > > > cluster_pm_enter() before entering the low power state and call the
> > > > > > > > > > > > cluster_pm_exit() after the cluster woken up.
> > > > > > > > > > > >
> > > > > > > > > > > > Signed-off-by: Nick Hu <nick.hu@...ive.com>
> > > > > > > > > > >
> > > > > > > > > > > I was not cced this patch, but noticed that this patch got queued up
> > > > > > > > > > > recently. Sorry for not noticing earlier.
> > > > > > > > > > >
> > > > > > > > > > > If not too late, can you please drop/revert it? We should really move
> > > > > > > > > > > away from the CPU cluster notifiers. See more information below.
> > > > > > > > > > >
> > > > > > > > > > > > ---
> > > > > > > > > > > >  drivers/cpuidle/cpuidle-riscv-sbi.c | 24 ++++++++++++++++++++++--
> > > > > > > > > > > >  1 file changed, 22 insertions(+), 2 deletions(-)
> > > > > > > > > > > >
> > > > > > > > > > > > diff --git a/drivers/cpuidle/cpuidle-riscv-sbi.c b/drivers/cpuidle/cpuidle-riscv-sbi.c
> > > > > > > > > > > > index e8094fc92491..298dc76a00cf 100644
> > > > > > > > > > > > --- a/drivers/cpuidle/cpuidle-riscv-sbi.c
> > > > > > > > > > > > +++ b/drivers/cpuidle/cpuidle-riscv-sbi.c
> > > > > > > > > > > > @@ -394,6 +394,7 @@ static int sbi_cpuidle_pd_power_off(struct generic_pm_domain *pd)
> > > > > > > > > > > >  {
> > > > > > > > > > > >         struct genpd_power_state *state = &pd->states[pd->state_idx];
> > > > > > > > > > > >         u32 *pd_state;
> > > > > > > > > > > > +       int ret;
> > > > > > > > > > > >
> > > > > > > > > > > >         if (!state->data)
> > > > > > > > > > > >                 return 0;
> > > > > > > > > > > > @@ -401,6 +402,10 @@ static int sbi_cpuidle_pd_power_off(struct generic_pm_domain *pd)
> > > > > > > > > > > >         if (!sbi_cpuidle_pd_allow_domain_state)
> > > > > > > > > > > >                 return -EBUSY;
> > > > > > > > > > > >
> > > > > > > > > > > > +       ret = cpu_cluster_pm_enter();
> > > > > > > > > > > > +       if (ret)
> > > > > > > > > > > > +               return ret;
> > > > > > > > > > >
> > > > > > > > > > > Rather than using the CPU cluster notifiers, consumers of the genpd
> > > > > > > > > > > can register themselves to receive genpd on/off notifiers.
> > > > > > > > > > >
> > > > > > > > > > > In other words, none of this should be needed, right?
> > > > > > > > > > >
> > > > > > > > > > Thanks for the feedback!
> > > > > > > > > > Maybe I miss something, I'm wondering about a case like below:
> > > > > > > > > > If we have a shared L2 cache controller inside the cpu cluster power
> > > > > > > > > > domain and we add this controller to be a consumer of the power
> > > > > > > > > > domain, Shouldn't the genpd invoke the domain idle only after the
> > > > > > > > > > shared L2 cache controller is suspended?
> > > > > > > > > > Is there a way that we can put the L2 cache down while all cpus in the
> > > > > > > > > > same cluster are idle?
> > > > > > > > > > > [...]
> > > > > > > > > Sorry, I made some mistake in my second question.
> > > > > > > > > Update the question here:
> > > > > > > > > Is there a way that we can save the L2 cache states while all cpus in the
> > > > > > > > > same cluster are idle and the cluster could be powered down?
> > > > > > > >
> > > > > > > > If the L2 cache is a consumer of the cluster, the consumer driver for
> > > > > > > > the L2 cache should register for genpd on/off notifiers.
> > > > > > > >
> > > > > > > > The device representing the L2 cache needs to be enabled for runtime
> > > > > > > > PM, to be taken into account correctly by the cluster genpd In this
> > > > > > > > case, the device should most likely remain runtime suspended, but
> > > > > > > > instead rely on the genpd on/off notifiers to understand when
> > > > > > > > save/restore of the cache states should be done.
> > > > > > > >
> > > > > > > > Kind regards
> > > > > > > > Uffe
> >

Powered by blists - more mailing lists