Re: [PATCH RFC v1 2/8] kernel/cpu_pm: Manage runtime PM in the idle path for CPUs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 12 October 2018 at 09:43, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
> On Fri, Oct 12, 2018 at 12:08 AM Lina Iyer <ilina@xxxxxxxxxxxxxx> wrote:
>>
>> On Thu, Oct 11 2018 at 14:56 -0600, Rafael J. Wysocki wrote:
>> >On Wednesday, October 10, 2018 11:20:49 PM CEST Raju P.L.S.S.S.N wrote:
>> >> From: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
>> >>
>> >> To allow CPUs being power managed by PM domains, let's deploy support for
>> >> runtime PM for the CPU's corresponding struct device.
>> >>
>> >> More precisely, at the point when the CPU is about to enter an idle state,
>> >> decrease the runtime PM usage count for its corresponding struct device,
>> >> via calling pm_runtime_put_sync_suspend(). Then, at the point when the CPU
>> >> resumes from idle, let's increase the runtime PM usage count, via calling
>> >> pm_runtime_get_sync().
>> >>
>> >> Cc: Lina Iyer <ilina@xxxxxxxxxxxxxx>
>> >> Co-developed-by: Lina Iyer <lina.iyer@xxxxxxxxxx>
>> >> Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
>> >> Signed-off-by: Raju P.L.S.S.S.N <rplsssn@xxxxxxxxxxxxxx>
>> >> (am from https://patchwork.kernel.org/patch/10478153/)
>> >> ---
>> >>  kernel/cpu_pm.c | 11 +++++++++++
>> >>  1 file changed, 11 insertions(+)
>> >>
>> >> diff --git a/kernel/cpu_pm.c b/kernel/cpu_pm.c
>> >> index 67b02e1..492d4a8 100644
>> >> --- a/kernel/cpu_pm.c
>> >> +++ b/kernel/cpu_pm.c
>> >> @@ -16,9 +16,11 @@
>> >>   */
>> >>
>> >>  #include <linux/kernel.h>
>> >> +#include <linux/cpu.h>
>> >>  #include <linux/cpu_pm.h>
>> >>  #include <linux/module.h>
>> >>  #include <linux/notifier.h>
>> >> +#include <linux/pm_runtime.h>
>> >>  #include <linux/spinlock.h>
>> >>  #include <linux/syscore_ops.h>
>> >>
>> >> @@ -91,6 +93,7 @@ int cpu_pm_enter(void)
>> >>  {
>> >>      int nr_calls;
>> >>      int ret = 0;
>> >> +    struct device *dev = get_cpu_device(smp_processor_id());
>> >>
>> >>      ret = cpu_pm_notify(CPU_PM_ENTER, -1, &nr_calls);
>> >>      if (ret)
>> >> @@ -100,6 +103,9 @@ int cpu_pm_enter(void)
>> >>               */
>> >>              cpu_pm_notify(CPU_PM_ENTER_FAILED, nr_calls - 1, NULL);
>> >>
>> >> +    if (!ret && dev && dev->pm_domain)
>> >> +            pm_runtime_put_sync_suspend(dev);
>> >
>> >This may cause a power domain to go off, but if it goes off, then the idle
>> >governor has already selected idle states for all of the CPUs in that domain.
>> >
>> >Is there any way to ensure that turning the domain off (and later on) will
>> >no cause the target residency and exit latency expectations for those idle
>> >states to be exceeded?
>> >
>> Good point.
>>
>> The cluster states should account for that additional latency.
>
> But even then, you need to be sure that the idle governor selected
> "cluster" states for all of the CPUs in the cluster.  It might select
> WFI for one of them for reasons unrelated to the distance to the next
> timer (so to speak), for example.

The approach here is that cpu_pm_enter isn't called for WFI, hence
there is no pm_runtime_get|put() done for a CPU going to WFI. In other
words, the cluster will never be powered off/on if there is any CPU in
the cluster in WFI.

>
>> Just the CPU's power down states need not care about that.
>
> The meaning of this sentence isn't particularly clear to me. :-)
>
>> But, it would be nice if the PM domain governor could be cognizant of
>> the idle state chosen for each CPU, that way we dont configure the
>> domain to be powered off when the CPUs have just chosen to power down
>> (not chosen a cluster state). I think that is a whole different topic to
>> discuss.
>
> This needs to be sorted out before the approach becomes viable, though.
>
> Basically, the domain governor needs to track what the idle governor
> did for all of the CPUs in the domain and only let the domain go off
> if the latency matches all of the states selected by the idle
> governor.  Otherwise the idle governor's assumptions would be violated
> and it would become essentially useless overhead.

That is correct!

The reason why it works in the current approach, is because there is
only one additional CPU idle state besides WFI. Hence
pm_runtime_get|put() is correctly called, as it's done only when the
cpuidle-governor has picked the deepest idle state for the CPU.

To solve this in the long run (where CPUs have > 1 idle state besides
WFI), I think there are two options.

1)
We either have to select the idle state of the CPU (and not only the
cluster) as part of the genpd and the genpd governors. This makes
genpd in charge and can thus decide internally what idle state that
should be selected, even hierarchically. Then it becomes a matter of
how to share code and interact between cpuidle/cpuidle-governors and
genpd.

2)
Leave genpd and the genpd governor to deal with cluster idle states,
but not the CPU idle states, as is the suggested approach. Then, to
solve the problem you pointed out, we need to provide the cpuidle
driver with information about which of the available idle state(s) it
should call pm_runtime_get|put() for, as to allow cluster idle
management through genpd.

I am currently looking at option 2), as I think it requires less
changes and I guess it's better to move slowly forward.

Does it make sense?

Kind regards
Uffe



[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux