Hi Krzysztof, Arend,
On Wed, Aug 21, 2024 at 4:46 PM Krzysztof Kozlowski <krzk@xxxxxxxxxx> wrote:
>
> On 20/08/2024 21:29, Arend van Spriel wrote:
> > On August 20, 2024 5:51:03 PM Krzysztof Kozlowski <krzk@xxxxxxxxxx> wrote:
> >
> >> On 20/08/2024 17:36, Krzysztof Kozlowski wrote:
> >>> On 20/08/2024 14:50, Arend van Spriel wrote:
> >>>> On 8/20/2024 1:39 PM, Krzysztof Kozlowski wrote:
> >>>>> On 20/08/2024 13:27, Krzysztof Kozlowski wrote:
> >>>>>> On Tue, Aug 20, 2024 at 12:12:15PM +0200, Arend van Spriel wrote:
> >>>>>>> When extending the bindings for Apple PCIe devices the compatible property
> >>>>>>> specification was changed. However, it was changed such that for these
> >>>>>>> devices it was no longer necessary to have "brcm,bcm4329-fmac" listed as
> >>>>>>> string in the compatible list property as it was before that extension.
> >>>>>>
> >>>>>> Apart that this was never tested... That statement is not true. Look at
> >>>>>> "fixed" commit - it is not doing like that at all.
> >>>>>>
> >>>>>> I don't understand the reasoning.
> >>>>>>
> >>>>>>> This patch restores that constraint.
> >>>>>>>
> >>>>>>> Fixes: e2e37224e8b3 ("dt-bindings: net: bcm4329-fmac: Add Apple properties
> >>>>>>> & chips")
> >>>>>>> Signed-off-by: Arend van Spriel <arend.vanspriel@xxxxxxxxxxxx>
> >>>>>>> ---
> >>>>>>> .../net/wireless/brcm,bcm4329-fmac.yaml | 19 ++++++++++---------
> >>>>>>> 1 file changed, 10 insertions(+), 9 deletions(-)
> >>>>>>>
> >>>>>>> diff --git
> >>>>>>> a/Documentation/devicetree/bindings/net/wireless/brcm,bcm4329-fmac.yaml
> >>>>>>> b/Documentation/devicetree/bindings/net/wireless/brcm,bcm4329-fmac.yaml
> >>>>>>> index e564f20d8f41..47f90446322f 100644
> >>>>>>> --- a/Documentation/devicetree/bindings/net/wireless/brcm,bcm4329-fmac.yaml
> >>>>>>> +++ b/Documentation/devicetree/bindings/net/wireless/brcm,bcm4329-fmac.yaml
> >>>>>>> @@ -7,7 +7,7 @@ $schema: http://devicetree.org/meta-schemas/core.yaml#
> >>>>>>> title: Broadcom BCM4329 family fullmac wireless SDIO/PCIE devices
> >>>>>>>
> >>>>>>> maintainers:
> >>>>>>> - - Arend van Spriel <arend@xxxxxxxxxxxx>
> >>>>>>> + - Arend van Spriel <arend.vanspriel@xxxxxxxxxxxx>
> >>>>>>>
> >>>>>>> description:
> >>>>>>> The Broadcom Single chip MAC part for the BCM4329 family and
> >>>>>>> @@ -27,7 +27,6 @@ properties:
> >>>>>>> - brcm,bcm4341b0-fmac
> >>>>>>> - brcm,bcm4341b4-fmac
> >>>>>>> - brcm,bcm4341b5-fmac
> >>>>>>> - - brcm,bcm4329-fmac
> >>>>>>> - brcm,bcm4330-fmac
> >>>>>>> - brcm,bcm4334-fmac
> >>>>>>> - brcm,bcm43340-fmac
> >>>>>>> @@ -46,13 +45,15 @@ properties:
> >>>>>>> - cypress,cyw43012-fmac
> >>>>>>> - infineon,cyw43439-fmac
> >>>>>>> - const: brcm,bcm4329-fmac
> >>>>>>> - - enum:
> >>>>>>> - - brcm,bcm4329-fmac
> >>>>>>> - - pci14e4,43dc # BCM4355
> >>>>>>> - - pci14e4,4464 # BCM4364
> >>>>>>> - - pci14e4,4488 # BCM4377
> >>>>>>> - - pci14e4,4425 # BCM4378
> >>>>>>> - - pci14e4,4433 # BCM4387
> >>>>>>> + - items:
> >>>>>>> + - enum:
> >>>>>>> + - pci14e4,43dc # BCM4355
> >>>>>>> + - pci14e4,4464 # BCM4364
> >>>>>>> + - pci14e4,4488 # BCM4377
> >>>>>>> + - pci14e4,4425 # BCM4378
> >>>>>>> + - pci14e4,4433 # BCM4387
> >>>>>>> + - const: brcm,bcm4329-fmac
> >>>>>>> + - const: brcm,bcm4329-fmac
> >>>>>>
> >>>>>> And this does not make sense... You claim that some constrained was
> >>>>>> droppped and you re-add it, but in fact you still add the same code as
> >>>>>> it was before.
> >>>>>>
> >>>>>> NAK.
> >>>>>
> >>>>> Ah, the last "const" actually makes sense, I missed that.
> >>>>>
> >>>>> Commit still however lacks rationale why these devices are compatible.
> >>>>> Plus existing rationale that e2e37224e8b3 changed something is entirely
> >>>>> WRONG. It changed nothing. ZERO. It only added new devices, which was
> >>>>> claimed are not compatible with brcm,bcm4329-fmac.
> >>>>
> >>>> So is that claim true? What does it mean that these new devices are not
> >>>> compatible. If they are they should be in a separate binding or the
> >>>
> >>> Whether binding is separate or not, is just way of organizing things.
> >>>
> >>>> applicable properties for these devices should be made conditional.
> >>>
> >>> Could be if they are not applicable.
> >>>
> >>>>
> >>>>> Now if you claim that original commit which said "these devices are not
> >>>>> compatible with brcm,bcm4329-fmac", then please provide arguments, not
> >>>>> just say "other commit did something". It did nothing...
> >>>>
> >>>> Not entirely true. Indeed new devices were added for which no
> >>>> "brcm,bcm4329-fmac" string is required in the compatible property. Also
> >>>> the commit added new properties for these new devices. Now in my opinion
> >>>> a driver should not use these properties without a "compatible" check.
> >>>> Hope we can agree to that. However, the driver patch for supporting the
> >>>
> >>> Sorry, I don't follow. Why the driver would need to check for compatible?
> >>>
> >>>> binding change does no such thing. So if we leave the binding as it
> >>>> currently is the driver will have to check if compatible has any of the
> >>>> listed PCI IDs before processing the properties. As all properties old
> >>>
> >>> Why driver needs to check it? Are these properties not valid?
> >
> > How would the driver know other than the compatible property? The node
> > with properties is delivered by the bus driver. If that comes with
> > guarantees about validity than that's great.
>
> I still do not follow what is the problem being addressed by driver
> needing to check.
>
> >
> >>>
> >>>
> >>>> and new are marked as optional I can not come up with an argument that
> >>>> these new devices are *not* compatible with brcm,bcm4329-fmac.
> >>>
> >>> Compatibility is expressed by implementing same programming interfasce
> >>> (or its subset) thus being able to bind via fallback and correctly
> >>> operate in given SW.
> >
> > This exactly what I mean to say (and apparently fail to do so ;-s ).
> >
> >>> I don't know whether that's the case here, so rephrasing my earlier
> >>> comments - the commit msg should focus on this aspect and tell that
> >>> devices are fully compatible, thus they should use fallback.
> >>>
> >>> Quick look at drivers told me that not - they are not compatible...
> >
> > Okay. That puts use in different corner of the arena. Can you elaborate
> > how you come to that assessment? Is that based on the fact that some of
> > the properties are SDIO-only?
>
> The simplest: because they do not use the same match/bind code. Plus PCI
> devices never used half of brcmf_of_probe(). Although that's more of a
> reason these are significantly different.
>
> >
> >>>
> >> Another thing is that calling SDIO and PCI devices compatible is quite a
> >> stretch... Clearly hardware-wise they are very different and Linux does
> >> not use the same interfaces to match/bind them.
> >
> > These are wifi devices which hardware-wise are 95% the same. If you find
> > the block diagram with IP cores for these devices (enough google results
> > to find some) you can see they sometimes even have both PCI and SDIO
> > block on-chip although only one is used so they can be considered 100%
> > the same. In both cases the bus driver will attach the DT node to the
> > binding device.
>
> I understand they are similar, but it does not matter if that is 95% or
> even 99% if the interface is different. Linux cannot use these devices
> through the same interface. However if you claim it can, then please
> write appropriate commit msg.
>
> My entire objection hare started not because I believe these are not
> compatible (although based on different buses I believe they are not
> compatible), but because the argument was about that other commit. That
> argument is not correct to make the change. Correct argument to make the
> change could be: These devices are compatible, because of foo and bar.
A way to put it might be something like:
Almost all Broadcom wireless chipsets, some Broadcom ethernet chipsets
and most Broadcom-derived Infineon and Cypress wireless chipsets share
the same rough layout:
System bus -> bus glue logic -> internal bus -> RF chipset(s)
Previous generations had separate drivers for the bus glue logic and
then probed the internal bus for the RF chipset(s), however in the
current generation that separation is mostly just a historical
artifact and abstraction layer.
The firmwares for all WiFi chipsets in this generation also provide
fundamentally identical interfaces with any differences either probed
at runtime or derived from data in the firmware. As such, once the
glue logic is abstracted away, all chipsets of this generation are
fundamentally identical from a software perspective. Therefore the
only data needed to use one of these chipsets is the type of bus to
select the glue logic driver and the firmware to load which is derived
from the vendor and product IDs.
The only real hardware differences between different chipsets are that
some have external interrupt lines or additional clocks and I'm not
sure if hooking those up is actually mandatory.
Thanks,
Julian Calaby
--
Julian Calaby
Email: julian.calaby@xxxxxxxxx
Profile: http://www.google.com/profiles/julian.calaby/
