On 2020-08-25 12:50, Stephen Boyd wrote:
Quoting Bjorn Andersson (2020-08-24 09:42:12)
On Fri 21 Aug 14:41 PDT 2020, Stephen Boyd wrote:
> Quoting Sibi Sankar (2020-08-21 13:49:20)
> > Add GENPD_FLAG_NO_SUSPEND/RESUME flags to instruct genpd to keep the
> > status of the PM domain unaltered during suspend/resume respectively.
> > The flags are aimed at power domains coupled to co-processors which
> > enter low-power modes independent to that of the application processor.
> >
> > Specifically the flags are to be used by the power domains exposed
> > by the AOSS QMP driver linked to modem, adsp, cdsp remoteprocs. These
> > power domains are used to notify the Always on Subsystem (AOSS) that
> > a particular co-processor is up. AOSS uses this information to wait
> > for the co-processors to suspend before starting its sleep sequence.
> > The application processor powers off these power domains only if the
> > co-processor has crashed or powered off and remains unaltered during
> > system suspend/resume.
>
> Why are these power domains instead of some QMP message sent during
> remote proc power up?
The understanding I gained as I researched this, was that with this
property enabled resources related to the particular subsystem will be
kept enabled when the apss enters some power save mode. So my
interpretation was that it does "keep something powered".
It looks like it tells AOSS that the processor is booted and to start
considering these processors in the SoC wide system suspend sequence.
Otherwise I guess the RPMh buckets associated with these remoteprocs
don't count in the aggregation and sleep/wake sequences that AOSS runs
through when putting the SoC into low power mode. I'm not sure it
actually "keeps something powered" so much as it lets something be
powered off. Sibi?
That is just a part of equation i.e
AOSS doesn't enter sleep until the
remote processors enter RPMh assisted
sleep. This also implies that if the
respective remote processor has to come
out of low power states it will need
to wait for AOSS to come out of sleep.
So clearly remote processors are dependent
on certain resources to be enabled by
the AOSS but the resources may not be
restricted to just corners.
Another question, why can't the processors tell AOSS themselves about
their boot state? I guess because they may crash or be powered down and
then AOSS wouldn't know? Fair enough I guess, but I don't think this is
mentioned anywhere.
> If this has been discussed before feel free to
> disregard and please link to prior mailing list discussions.
>
There where some discussions related to the "QDSS clk" in that series,
but I don't remember getting any feedback on modelling these things as
power-domains.
> I find it odd that this is modeled as a power domain instead of some
> Qualcomm specific message that the remoteproc driver sends to AOSS. Is
> there some sort of benefit the driver gets from using the power domain
> APIs for this vs. using a custom API?
We need to send "up" and "down" notifications and this needs to happen
at the same time as other standard resources are enabled/disabled.
Further more, at the time the all resources handled by the downstream
driver was either power-domains (per above understanding) or clocks,
so
it made sense to me not to spin up a custom API.
So the benefit is not spinning up a custom API? I'm not Ulf, but it
looks like this is hard to rationalize about as a power domain. It
doesn't have any benefit to model it this way besides to make it
possible to turn on with other power domains.
This modem remoteproc drivers isn't SoC agnostic anyway, it relies on
SMEM APIs, so standing up another small qmp_remoteproc_booted() and
qmp_remoteproc_shutdown() API would avoid adding a genpd flag here that
probably will never be used outside of this corner-case. There is also
some get/put EPROBE_DEFER sort of logic to implement, but otherwise it
would be possible to do this outside of power domains, and that seems
better given that this isn't really a power domain to start with.
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project.
