Message-ID: <7hmsilqrw3.fsf@baylibre.com>
Date: Wed, 30 Oct 2024 13:04:28 -0700
From: Kevin Hilman <khilman@...libre.com>
To: Tomi Valkeinen <tomi.valkeinen@...asonboard.com>, Nishanth Menon
 <nm@...com>, Tero Kristo <kristo@...nel.org>, Santosh Shilimkar
 <ssantosh@...nel.org>, Ulf Hansson <ulf.hansson@...aro.org>
Cc: linux-arm-kernel@...ts.infradead.org, linux-pm@...r.kernel.org,
 linux-kernel@...r.kernel.org, vishalm@...com, sebin.francis@...com,
 d-gole@...com, Devarsh Thakkar <devarsht@...com>, Vignesh Raghavendra
 <vigneshr@...com>, Tomi Valkeinen <tomi.valkeinen@...asonboard.com>
Subject: Re: [PATCH RFC] pmdomain: ti-sci: Set PD on/off state according to
 the HW state

Tomi Valkeinen <tomi.valkeinen@...asonboard.com> writes:

> At the moment the driver sets the power state of all the PDs it creates
> to off, regardless of the actual HW state. This has two drawbacks:
>
> 1) The kernel cannot disable unused PDs automatically for power saving,
>    as it thinks they are off already
>
> 2) A more specific case (but perhaps applicable to other scenarios
>    also): bootloader enabled splash-screen cannot be kept on the screen.
>
> The issue in 2) is that the driver framework automatically enables the
> device's PD before calling probe() and disables it after the probe().
> This means that when the display subsystem (DSS) driver probes, but e.g.
> fails due to deferred probing, the DSS PD gets turned off and the driver
> cannot do anything to affect that.
>
> Solving the 2) requires more changes to actually keep the PD on during
> the boot, but a prerequisite for it is to have the correct power state
> for the PD.
>
> The downside with this patch is that it takes time to call the 'is_on'
> op, and we need to call it for each PD. In my tests with AM62 SK, using
> defconfig, I see an increase from ~3.5ms to ~7ms. However, the added
> feature is valuable, so in my opinion it's worth it.
>
> The performance could probably be improved with a new firmware API which
> returns the power states of all the PDs.

Agreed.  I think we have to pay this performance price for correctness,
and we can optimizie it later with improvements to the SCI firmware and
a new API.

> There's also a related HW issue at play here: if the DSS IP is enabled
> and active, and its PD is turned off without first disabling the DSS
> display outputs, the DSS IP will hang and causes the kernel to halt if
> and when the DSS driver accesses the DSS registers the next time.

Ouch.

> With the current upstream kernel, with this patch applied, this means
> that if the bootloader enables the display, and the DSS driver is
> compiled as a module, the kernel will at some point disable unused PDs,
> including the DSS PD. When the DSS module is later loaded, it will hang
> the kernel.
>
> The same issue is already there, even without this patch, as the DSS
> driver may hit deferred probing, which causes the PD to be turned off,
> and leading to kernel halt when the DSS driver is probed again. This
> issue has been made quite rare with some arrangements in the DSS
> driver's probe, but it's still there.
>
> So, because of the DSS hang issues, I think this patch is still an RFC.

Like you said, I think that DSS hang is an issue independently of this
patch, so it shouldn't hold this up IMO.

> Hopefully we can sort out all the issues, but this patch (or similar)
> will be part of the solution so I'd like to get some acks/nacks/comments
> for this. Also, this change might have side effects to other devices
> too, if the drivers expect the PD to be on, so testing is needed.
>
> Signed-off-by: Tomi Valkeinen <tomi.valkeinen@...asonboard.com>

We already discussed this a bit off-list, but for the record, I agree
with the approach.

I also tested it on k3-am62a7-sk where I've been doing the other TI SCI
pmdomain work and everything still working fine.

Reviewed-by: Kevin Hilman <khilman@...libre.com>
Tested-by: Kevin Hilman <khilman@...libre.com>

Kevin

Powered by blists - more mailing lists