Message-ID: <20201027131919.pyllaznj7vg4o5fm@quickly>
Date:   Tue, 27 Oct 2020 08:19:19 -0500
From:   Nishanth Menon <nm@...com>
To:     Tony Lindgren <tony@...mide.com>
CC:     Peter Ujfalusi <peter.ujfalusi@...com>, <t-kristo@...com>,
        <devicetree@...r.kernel.org>, <robh+dt@...nel.org>,
        <linux-kernel@...r.kernel.org>,
        <linux-arm-kernel@...ts.infradead.org>
Subject: Re: [PATCH] arm64: dts: ti: k3-j7200-main: Add McASP nodes

Hi Tony,

   Thanks for joining the discussion.

On 15:07-20201027, Tony Lindgren wrote:
> * Nishanth Menon <nm@...com> [201026 14:58]:
> > On 13:38-20201007, Peter Ujfalusi wrote:
> > [...]
> > > >>>> +		status = "disabled";
> > > >>>
> > > >>> I see that there is inconsistent usage of "disabled" in our SoC.dts
> > > >>>
> > > >>> Our generic rule has been set them to disabled in board.dtsi
> > > >>> McASP and DSS for existing SoC dts do not follow this.. which is a tad
> > > >>> confusing.. (considering that not even all uarts come out on every board
> > > >>> and every uart needs pinmux to function..)
> > > >>
> > > >> "keep them disabled because several required properties are not present
> > > >> as they are board specific."
> > > >>
> > > >> In board file the enabled mcasp must be updated with options that is
> > > >> required for operation. Without those option the McASP can not be
> > > >> initialized.
> > > >>
> > > >> I think we have been revisiting the very same discussion every time we
> > > >> have a new SoC with McASP...
> > > >>
> > > > 
> > > > Yep.. This doe'snt really follow the rest of the SoC definition. [1]
> > > > came to mind. The McASP discussion is a variation in the debate of the
> > > > same.
> > > 
> > > Right, saying status = "okay" to a node which is missing required
> > > properties (which can only be added by boards when the McASP is
> > > connected up) does not sound a good solution.
> > > How should the SW handle that? Fail the device probe and return with
> > > -EINVAL or eat up the error and just probe with broken configuration.
> > > Since the peripheral is not used, the broken configuration will not
> > > cause much runtime errors as there will be no runtime use of the peripheral.
> > > 
> > > status of fail or fail-sss is not a good one either, their definition is:
> > > "Indicates that the device is not operational. A serious error was
> > > detected in the device, and it is unlikely to become operational without
> > > repair."
> > > 
> > > The peripheral is fine, we are just trying to enable it without
> > > providing the needed properties.
> > > 
> > > > I'd argue Serdes, or for that matter any IP that has a link to
> > > > outside-the-SoC world has the same discussion point.
> > > 
> > > status = "disabled" is still the closest thing for everything which have
> > > external dependencies. There is not much point to enable an i2c bus
> > > without making sure that the signals are actually routed to the pins
> > > where they supposed to go.
> 
> There is one reason to keep using the default status = "okay", for
> i2c Linux can properly idle the device on boot without dependencies
> to a certain boot loader version.

Yes, but in K3 architecture (and I suspect in many SCMI like
architectures involving heterogenous systems), Linux does'nt even own the
system power management of peripherals at all. In fact,
we dont even see all the SoC components. if a particular I2C is being
used by some other processor, then that node is disabled for linux.

So, I am wondering if the argument of "status = okay" helps Linux PM
idle unused devices holds good on architectures where System PM is not
owned by Linux. That (system level PM) is not really Linux's job anymore.

> > > Or from other pow: a board design is not based on what is _not_
> > > connected to outside world, but you actually _connect_ or _enable_
> > > certain peripherals to external components, connectors.
> > 
> > OK, I will buy the argument that the current status thingy is a bit
> > overloaded and does'nt imply the correct state we need it to imply with
> > "fail-sss" either - I remember an argument for "fail-incomplete", but
> > that never happened anyways.
> > 
> > Lets add this argument to the commit message and repost after testing
> > on 5.10-rc1 please?
> 
> We should use status = "disabled" when the hardware is not accessible
> to the SoC. That goes for things like secure world devices, and hardware

I'd argue 'not accessible to Linux instance' (considering how visibility
of peripherals are constrained further in VMs)

> that relies on a missing external clock for example. For other things
> we should just stick to the default which "okay" :)

ok, the debate here (and looks like the DSS series as well[1]) is
what do you do when the there are not enough properties for a valid
device node. Do you still do status="okay" and let the driver fail ?
The driver folks are arguing, they don't want to do that since it is an
incomplete description of the hardware..

[1]
https://lore.kernel.org/linux-arm-kernel/20201027104132.105485-2-tomi.valkeinen@ti.com/
-- 
Regards,
Nishanth Menon
Key (0xDDB5849D1736249D) / Fingerprint: F8A2 8693 54EB 8232 17A3  1A34 DDB5 849D 1736 249D

Powered by blists - more mailing lists