Re: [PATCH 2/4] dt-bindings: hwinfo: ti,k3-socinfo: Add nvmem-cells

From: Markus Schneider-Pargmann
Date: Mon Mar 18 2024 - 12:10:30 EST

On Tue, Mar 05, 2024 at 11:41:31AM -0600, Andrew Davis wrote:
> On 3/5/24 11:01 AM, Krzysztof Kozlowski wrote:
> > On 05/03/2024 15:42, Andrew Davis wrote:
> > > On 3/5/24 8:11 AM, Krzysztof Kozlowski wrote:
> > > > On 05/03/2024 12:17, Markus Schneider-Pargmann wrote:
> > > > > Hi Krzysztof,
> > > > >
> > > > > On Tue, Mar 05, 2024 at 08:43:03AM +0100, Krzysztof Kozlowski wrote:
> > > > > > On 04/03/2024 11:36, Markus Schneider-Pargmann wrote:
> > > > > > > Hi,
> > > > > > >
> > > > > > > On Sat, Feb 17, 2024 at 03:25:30PM +0100, Krzysztof Kozlowski wrote:
> > > > > > > > On 14/02/2024 10:31, Markus Schneider-Pargmann wrote:
> > > > > > > > > Hi Rob,
> > > > > > > > >
> > > > > > > > > On Tue, Feb 06, 2024 at 06:43:05PM +0000, Rob Herring wrote:
> > > > > > > > > > On Tue, Feb 06, 2024 at 03:37:09PM +0100, Markus Schneider-Pargmann wrote:
> > > > > > > > > > > The information k3-socinfo requires is stored in an efuse area. This
> > > > > > > > > > > area is required by other devices/drivers as well, so using nvmem-cells
> > > > > > > > > > > can be a cleaner way to describe which information are used.
> > > > > > > > > > >
> > > > > > > > > > > If nvmem-cells are supplied, the address range is not required.
> > > > > > > > > > > Cells chipvariant, chippartno and chipmanufacturer are introduced to
> > > > > > > > > > > cover all required information.
> > > > > > > > > > >
> > > > > > > > > > > Signed-off-by: Markus Schneider-Pargmann <msp@xxxxxxxxxxxx>
> > > > > > > > > > > Reviewed-by: Andrew Davis <afd@xxxxxx>
> > > > > > > > > > > ---
> > > > > > > > > > > .../bindings/hwinfo/ti,k3-socinfo.yaml | 23 ++++++++++++++++++-
> > > > > > > > > > > 1 file changed, 22 insertions(+), 1 deletion(-)
> > > > > > > > > > >
> > > > > > > > > > > diff --git a/Documentation/devicetree/bindings/hwinfo/ti,k3-socinfo.yaml b/Documentation/devicetree/bindings/hwinfo/ti,k3-socinfo.yaml
> > > > > > > > > > > index dada28b47ea0..f085b7275b7d 100644
> > > > > > > > > > > --- a/Documentation/devicetree/bindings/hwinfo/ti,k3-socinfo.yaml
> > > > > > > > > > > +++ b/Documentation/devicetree/bindings/hwinfo/ti,k3-socinfo.yaml
> > > > > > > > > > > @@ -26,9 +26,24 @@ properties:
> > > > > > > > > > > reg:
> > > > > > > > > > > maxItems: 1
> > > > > > > > > > > + nvmem-cells:
> > > > > > > > > > > + $ref: /schemas/types.yaml#/definitions/phandle-array
> > > > > > > > > > > +
> > > > > > > > > > > + nvmem-cell-names:
> > > > > > > > > > > + items:
> > > > > > > > > > > + - const: chipvariant
> > > > > > > > > > > + - const: chippartno
> > > > > > > > > > > + - const: chipmanufacturer
> > > > > > > > > > > +
> > > > > > > > > > > required:
> > > > > > > > > > > - compatible
> > > > > > > > > > > - - reg
> > > > > > > > > > > +
> > > > > > > > > > > +oneOf:
> > > > > > > > > > > + - required:
> > > > > > > > > > > + - reg
> > > > > > > > > > > + - required:
> > > > > > > > > > > + - nvmem-cells
> > > > > > > > > > > + - nvmem-cell-names
> > > > > > > > > > > additionalProperties: false
> > > > > > > > > > > @@ -38,3 +53,9 @@ examples:
> > > > > > > > > > > compatible = "ti,am654-chipid";
> > > > > > > > > > > reg = <0x43000014 0x4>;
> > > > > > > > > > > };
> > > > > > > > > > > + - |
> > > > > > > > > > > + chipid: chipid@14 {
> > > > > > > > > > > + compatible = "ti,am654-chipid";
> > > > > > > > > >
> > > > > > > > > > This isn't compatible if you have a completely different way to access
> > > > > > > > > > it.
> > > > > > > > >
> > > > > > > > > Thanks, it is not entirely clear to me how I could go forward with this?
> > > > > > > > > Are you suggesting to use a different compatible? Or is it something
> > > > > > > > > else I could do to proceed with this conversion?
> > > > > > > >
> > > > > > > > What you claim now, is that you have one device with entirely different
> > > > > > > > interfaces and programming model. So either this is not the same device
> > > > > > > > or you just wrote bindings to whatever you have in driver.
> > > > > > > >
> > > > > > > > Nothing in commit msg explains this.
> > > > > > > >
> > > > > > > > What you should do? Depends. If you just write bindings for driver, then
> > > > > > > > stop. It's a NAK. Instead write bindings for hardware.
> > > > > > > >
> > > > > > > > If the first choice, just the hardware is somehow like this, then
> > > > > > > > explain in commit msg and device description, how this device can be
> > > > > > > > connected over other bus, not MMIO. You can draw some schematics in
> > > > > > > > commit msg explaining architecture etc.
> > > > > > >
> > > > > > > Sorry the information provided in the commit message is not very clear.
> > > > > > >
> > > > > > > The basic access to the registes is still MMIO. nvmem is used to have a
> > > > > > > better abstraction and cleaner description of the hardware.
> > > > > > >
> > > > > > > Currently most of the data is exported using the parent syscon device.
> > > > > > > The relevant data is read-only and contained in a single register with
> > > > > > > offset 0x14:
> > > > > > > - Chip variant
> > > > > > > - Chip part number
> > > > > > > - Chip manufacturer
> > > > > > >
> > > > > > > There are more read-only registers in this section of address space.
> > > > > > > These are relevant to other components as they define the operating
> > > > > > > points for example. For the OPP table relevant are chip variant and chip
> > > > > > > speed (which is in a different register).
> > > > > > >
> > > > > > > Instead of devices refering to this whole register range of 0x20000 in
> > > > > >
> > > > > > Whaaaaat?
> > > > > >
> > > > > > > size, I would like to introduce this nvmem abstraction in between that
> > > > > > > describes the information and can directly be referenced by the devices
> > > > > > > that depend on it. In this case the above mentioned register with offset
> > > > > > > 0x14 is instead described as nvmem-layout like this:
> > > > > > >
> > > > > > > nvmem-layout {
> > > > > > > compatible = "fixed-layout";
> > > > > > > #address-cells = <1>;
> > > > > > > #size-cells = <1>;
> > > > > > >
> > > > > > > chip_manufacturer: jtagidmfg@14 {
> > > > > > > reg = <0x14 0x2>;
> > > > > > > bits = <1 11>;
> > > > > > > };
> > > > > > >
> > > > > > > chip_partno: jtagidpartno@15 {
> > > > > > > reg = <0x15 0x3>;
> > > > > > > bits = <4 16>;
> > > > > > > };
> > > > > > >
> > > > > > > chip_variant: jtagidvariant@17 {
> > > > > > > reg = <0x17 0x1>;
> > > > > > > bits = <4 4>;
> > > > > > > };
> > > > > > >
> > > > > > > chip_speed: jtaguseridspeed@18 {
> > > > > > > reg = <0x18 0x4>;
> > > > > > > bits = <6 5>;
> > > > > > > };
> > > > > > >
> > > > > > > The underlying registers are still the same but they are not hidden
> > > > > > > by the syscon phandles anymore.
> > > > > > >
> > > > > > > The device that consumes this data would now use
> > > > > > >
> > > > > > > nvmem-cells = <&chip_variant>, <&chip_speed>;
> > > > > > > nvmem-cell-names = "chipvariant", "chipspeed";
> > > > > > >
> > > > > > > instead of
> > > > > > >
> > > > > > > syscon = <&wkup_conf>;
> > > > > >
> > > > > > syscon allows you this as well - via phandle arguments.
> > > > > >
> > > > > > nvmem is for non-volatile memory, like OCOTP and eFUSE. This is not for
> > > > > > accessing regular MMIO registers of system-controller, regardless
> > > > > > whether they are read-only or not (regmap handles this nicely, BTW).
> > > > > > Although probably Apple efuses and few others can confuse here. It still
> > > > > > looks like you convert regular system-controller block into nvmem,
> > > > > > because you prefer that Linux driver abstraction...
> > > > >
> > > > > The above mentioned data is set in the factory. There is other
> > > > > non-volatile data, like device feature registers, in the same address
> > > > > region, as well as OTP data like MAC and USB IDs. But it is not a pure
> > > > > non-volatile memory region. The data is copied into these registers by
> > > > > the ROM at boot.
> > > >
> > > > Still entire block is MMIO IP in your SoC, not a efuse/OTP hardware.
> > > > nvmem is not for regular MMIO registers which are sometimes R, sometimes RW.
> > >
> > > Most eFuse/OTP hardware is accessed via MMIO, not sure what that changes.
> >
> > Just check exiting NVMEM drivers, except Apple I think most if not all
> > are not syscon blocks.
> >
>
> We don't want it to be a syscon block either. Syscon is just another Linux
> interface for accessing MMIO areas that found its way in to DT. NVMEM
> is another way, which as a DT construct is more "correct" as the area
> we are describing here *is* a non-volatile memory. Not a "syscon"?? whatever
> that is.
>
> > Following such approach, each hardware block, even USB or PCI, which
> > exposes a read-only register with some fused value, e.g. version, should
> > be nvmem?
> >
>
> If those fused values are grouped into a region then yes, why not. Wouldn't
> that be more correct to describe them as they actually are instead of
> hiding them behind a "syscon" block?
>
> > >
> > > This "block" is a whole bunch of smaller logical chunks of registers,
> > > some are actually mapped to eFuses like our MAC addresses. Regions
> > > like factory fused MAC addresses are exactly what nvmem does well[0].
> > >
> > > Yes, we *could* just have this whole area be one massive blanked syscon
> > > region that every driver just manually pokes into with syscon phandles
> > > everywhere. But that is hacky and hides details, it is not how DT normally
> > > looks. We would like to correctly model our device now with nodes for each
> > > "reg" region. We took the syscon shortcut before, and we want to correct
> > > that mistake.
> >
> > Wait, you now mix up hardware description with Linux interface.
> > Describing each register as nvmem field is not a better way of
> > describing hardware. It is unnecessarily too granular and results in
> > huge and unmaintainable DTS. It is however convenient because it is nice
> > API for other devices.
>
> It is not convenient. How we have it currently as a blanket syscon node
> that each driver can simply poke whatever address it wants is much easier.
> We are now trying to do the more difficult (but more correct) thing here by
> modeling our non-volatile memory areas as they are as nvmem nodes.
>
> > But claiming that MMIO register block is better
> > represented as nvmem is not correct. It is still MMIO block with
> > registers, like everywhere else in every other device.
> >
>
> Everything is MMIO on these SoCs, we don't have any sideband band
> IO ports. Following that to its logical conclusion we should just
> make the entire memory space reg = <0x0 0xffffffff> one big syscon node
> then have all other driver phandle into that for their MMIO access.
>
> > >
> > > So what are our options? Is the objection here that this is a new nvmem
> > > way of modeling this region changes the compatible "ti,am654-chipid"? If
> > > so then would you be open to us adding a new compatible that uses the
> > > nvmem nodes? We could then convert over one by one and keeping full
> > > backwards compatibility while we do it.
> >
> > Switching from MMIO to nvmem for chipid is a different interface, so as
> > Rob pointed out devices are not really compatible. You claim that
> > devices are compatible, because there is *NO REAL NVMEM* but MMIO
> > wrapped in nvmem. So do you see the logic here?
> >
>
> If the interface changing means it is not compatible then that is
> fine, we would use a different compatible string to identify the
> interface to be used.

So what is the acceptable approach here?

Best,
Markus

>
> This is not uncommon, the example that comes to my mind is "gpio-leds".
> We used to just have named GPIO pins for our LEDs, now we can
> put "gpio-leds" on top to better represent what is going on in HW.
> Even though physically nothing changed, we just now have a better way
> to model that HW in DT.
>
> Andrew
>
> > Best regards,
> > Krzysztof
> >
> >