Re: [Patch v3 1/2] dt-bindings: make sid and broadcast reg optional

From: Thierry Reding
Date: Thu Apr 25 2024 - 11:04:44 EST

On Thu Apr 25, 2024 at 11:45 AM CEST, Krzysztof Kozlowski wrote:
> On 25/04/2024 11:39, Thierry Reding wrote:
> > On Thu Apr 25, 2024 at 9:52 AM CEST, Krzysztof Kozlowski wrote:
> >> On 24/04/2024 19:04, Thierry Reding wrote:
> >>> On Wed Apr 24, 2024 at 6:26 PM CEST, Thierry Reding wrote:
> >>>> On Mon Apr 22, 2024 at 9:02 AM CEST, Krzysztof Kozlowski wrote:
> >>>>> On 12/04/2024 15:05, Sumit Gupta wrote:
> >>>>>> MC SID and Broadbast channel register access is restricted for Guest VM.
> >>>>>
> >>>>> Broadcast
> >>>>>
> >>>>>> Make both the regions as optional for SoC's from Tegra186 onwards.
> >>>>>
> >>>>> onward?
> >>>>>
> >>>>>> Tegra MC driver will skip access to the restricted registers from Guest
> >>>>>> if the respective regions are not present in the memory-controller node
> >>>>>> of Guest DT.
> >>>>>>
> >>>>>> Suggested-by: Thierry Reding <treding@xxxxxxxxxx>
> >>>>>> Signed-off-by: Sumit Gupta <sumitg@xxxxxxxxxx>
> >>>>>> ---
> >>>>>> .../nvidia,tegra186-mc.yaml | 95 ++++++++++---------
> >>>>>> 1 file changed, 49 insertions(+), 46 deletions(-)
> >>>>>>
> >>>>>> diff --git a/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml b/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>>>>> index 935d63d181d9..e0bd013ecca3 100644
> >>>>>> --- a/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>>>>> +++ b/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>>>>> @@ -34,11 +34,11 @@ properties:
> >>>>>> - nvidia,tegra234-mc
> >>>>>>
> >>>>>> reg:
> >>>>>> - minItems: 6
> >>>>>> + minItems: 4
> >>>>>> maxItems: 18
> >>>>>>
> >>>>>> reg-names:
> >>>>>> - minItems: 6
> >>>>>> + minItems: 4
> >>>>>> maxItems: 18
> >>>>>>
> >>>>>> interrupts:
> >>>>>> @@ -151,12 +151,13 @@ allOf:
> >>>>>>
> >>>>>> reg-names:
> >>>>>> items:
> >>>>>> - - const: sid
> >>>>>> - - const: broadcast
> >>>>>> - - const: ch0
> >>>>>> - - const: ch1
> >>>>>> - - const: ch2
> >>>>>> - - const: ch3
> >>>>>> + enum:
> >>>>>> + - sid
> >>>>>> + - broadcast
> >>>>>> + - ch0
> >>>>>> + - ch1
> >>>>>> + - ch2
> >>>>>> + - ch3
> >>>>>
> >>>>> I understand why sid and broadcast are becoming optional, but why order
> >>>>> of the rest is now fully flexible?
> >>>>
> >>>> The reason why the order of the rest doesn't matter is because we have
> >>>> both reg and reg-names properties and so the order in which they appear
> >>>> in the list doesn't matter. The only thing that matters is that the
> >>>> entries of the reg and reg-names properties match.
> >>>>
> >>>>> This does not even make sid/broadcast optional, but ch0!
> >>>>
> >>>> Yeah, this ends up making all entries optional, which isn't what we
> >>>> want. I don't know of a way to accurately express this in json-schema,
> >>>> though. Do you?
> >>>>
> >>>> If not, then maybe we need to resort to something like this and also
> >>>> mention explicitly in some comment that it is sid and broadcast that are
> >>>> optional.
> >>>
> >>> Actually, here's another variant that is a bit closer to what we want:
> >>>
> >>> --- >8 ---
> >>> diff --git a/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml b/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>> index 935d63d181d9..86f1475926e4 100644
> >>> --- a/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>> +++ b/Documentation/devicetree/bindings/memory-controllers/nvidia,tegra186-mc.yaml
> >>> @@ -34,11 +34,11 @@ properties:
> >>> - nvidia,tegra234-mc
> >>>
> >>> reg:
> >>> - minItems: 6
> >>> + minItems: 4
> >>> maxItems: 18
> >>>
> >>> reg-names:
> >>> - minItems: 6
> >>> + minItems: 4
> >>> maxItems: 18
> >>>
> >>> interrupts:
> >>> @@ -146,17 +146,21 @@ allOf:
> >>> then:
> >>> properties:
> >>> reg:
> >>> + minItems: 4
> >>> maxItems: 6
> >>> description: 5 memory controller channels and 1 for stream-id registers
> >>>
> >>> reg-names:
> >>> - items:
> >>> - - const: sid
> >>> - - const: broadcast
> >>> - - const: ch0
> >>> - - const: ch1
> >>> - - const: ch2
> >>> - - const: ch3
> >>> + anyOf:
> >>> + - items:
> >>> + enum: [ sid, broadcast, ch0, ch1, ch2, ch3 ]
> >>> + uniqueItems: true
> >>> + minItems: 6
> >>> +
> >>> + - items:
> >>> + enum: [ ch0, ch1, ch2, ch3 ]
> >>> + uniqueItems: true
> >>> + minItems: 4
> >>>
> >>> - if:
> >>> properties:
> >>> @@ -165,29 +169,22 @@ allOf:
> >>> then:
> >>> properties:
> >>> reg:
> >>> - minItems: 18
> >>> + minItems: 16
> >>> description: 17 memory controller channels and 1 for stream-id registers
> >>>
> >>> reg-names:
> >>> - items:
> >>> - - const: sid
> >>> - - const: broadcast
> >>> - - const: ch0
> >>> - - const: ch1
> >>> - - const: ch2
> >>> - - const: ch3
> >>> - - const: ch4
> >>> - - const: ch5
> >>> - - const: ch6
> >>> - - const: ch7
> >>> - - const: ch8
> >>> - - const: ch9
> >>> - - const: ch10
> >>> - - const: ch11
> >>> - - const: ch12
> >>> - - const: ch13
> >>> - - const: ch14
> >>> - - const: ch15
> >>> + anyOf:
> >>> + - items:
> >>> + enum: [ sid, broadcast, ch0, ch1, ch2, ch3, ch4, ch5, ch6, ch7,
> >>> + ch8, ch9, ch10, ch11, ch12, ch13, ch14, ch15 ]
> >>> + minItems: 18
> >>> + uniqueItems: true
> >>> +
> >>> + - items:
> >>> + enum: [ ch0, ch1, ch2, ch3, ch4, ch5, ch6, ch7, ch8, ch9, ch10,
> >>> + ch11, ch12, ch13, ch14, ch15 ]
> >>> + minItems: 16
> >>> + uniqueItems: true
> >>
> >> No, because order is strict.
> >
> > Why? I realize that prior to this the order was indeed strict and it's
>
> That's the policy for entire Devicetree. I said why in other email:
> because any bindings consumer can take it via indices.
>
> > common to have these listed in strict order in the DTS files. However,
> > this is an arbitrary restriction that was introduced in the patch that
> > added reg-names. However, */*-names properties have always assumed the
> > ordering to be non-strict because each entry from the * property gets
> > matched up with the corresponding entry in the *-names property, so the
> > ordering is completely irrelevant.
>
> This was raised so many times... reg-names is just a helper. It does not
> change the fact that order should be strict and if binding defined the
> order, it is an ABI.

Sorry, but that's not how we've dealt with this in the past. Even though
this was now ten or more years ago, I distinctly recall that when we
started adding these *-names properties and at the time it was very much
implied that the order didn't matter.

The only use-case that I know of where order was always meant to matter
is backwards-compatibility for devices that used to have a single entry
(hence drivers couldn't rely on *-names to resolve the index) and then
had additional entries added. The *-names entry for that previously
single entry would now obviously have to always be first in the list to
preserve backwards-compatibility.

Besides, if reg-names was really only a helper, then it would also be
completely redundant. Many device tree bindings have *-names properties
marked as "required" precisely because of the role that they serve.

Thierry

Attachment: signature.asc
Description: PGP signature