On Tue, May 30, 2023 at 10:53:31AM +0200, Krzysztof Kozlowski wrote: > On 22/05/2023 21:15, Kevin Hilman wrote: > > Krzysztof Kozlowski <krzysztof.kozlowski@xxxxxxxxxx> writes: > > > >> On 16/05/2023 23:31, Kevin Hilman wrote: > >> > >>>> Third is to use versioned IP blocks. > >>>> > >>>> The second case also would work, if it is applicable to you (you really > >>>> have fallback matching all devices). Third solution depends on your > >>>> versioning and Rob expressed dislike about it many times. > >>>> > >>>> We had many discussions on mailing lists, thus simplifying the review - > >>>> I recommend the first choice. For a better recommendation you should say > >>>> a bit more about the block in different SoCs. > >>> > >>> I'll try to say a bit more about the PHY block, but in fact, it's not > >>> just about differences between SoCs. On the same SoC, 2 different PHYs > >>> may have different features/capabilities. > >>> > >>> For example, on MT8365, There are 2 PHYs: CSI0 and CSI1. CSI0 can > >>> function as a C-PHY or a D-PHY, but CSI1 can only function as D-PHY > >>> (used as the example in the binding patch[1].) On another related SoC, > >>> there are 3 PHYs, where CSI0 is C-D but CSI1 & CSI2 are only D. > >>> > >>> So that's why it seems (at least to me) that while we need SoC > >>> compatible, it's not enough. We also need properties to describe > >>> PHY-specific features (e.g. C-D PHY) > >> > >> I recall the same or very similar case... It bugs me now, but > >> unfortunately I cannot find it. > >> > >>> > >>> Of course, we could rely only on SoC-specific compatibles describe this. > >>> But then driver will need an SoC-specific table with the number of PHYs > >>> and per-PHY features for each SoC encoded in the driver. Since the > >>> driver otherwise doesn't (and shouldn't, IMHO) need to know how many > >>> PHYs are on each SoC, I suggested to Julien that perhaps the additional > >>> propery was the better solution. > >> > >> Phys were modeled as separate device instances, so you would need > >> difference in compatible to figure out which phy is it. > >> > >> Other way could be to create device for all phys and use phy-cells=1. > >> Whether it makes sense, depends on the actual datasheet - maybe the > >> split phy per device is artificial? There is one PHY block with two > >> address ranges for each PHY - CSI0 and CSI1 - but it is actually one > >> block? You should carefully check this because once design is chosen, > >> you won't be able to go back to other and it might be a problem (e.g. > >> there is some top-level block for powering on all CSI instances). > > > > We're pretty sure these are multiple instances of the IP block as they > > can operate completely independently. > > > >>> > >>> To me it seems redundant to have the driver encode PHYs-per-SoC info, > >>> when the per-SoC DT is going to have the same info, so my suggestion was > >>> to simplify the driver and have this kind of hardware description in the > >>> DT, and keep the driver simple, but we are definitely open to learning > >>> the "right way" of doing this. > >> > >> The property then is reasonable. It should not be bool, though, because > >> it does not scale. There can be next block which supports only D-PHY on > >> CSI0 and C-PHY on CSI1? Maybe some enum or list, depending on possible > >> configurations. > > > > OK, looks like include/dt-bindings/phy/phy.y already has > > > > #define PHY_TYPE_DPHY 10 > > #define PHY_TYPE_CPHY 11 > > > > we'll add a PHY_TYPE_CDPHY and use that. Sound reasonable? > > Yes. Currently it is usually used as phy-cells argument (after the phy > number/lane/ID), but cdns,phy-type and intel,phy-mode use it directly as > property in provider. In both cases they have a bit different meaning > than yours. You want to list all supported modes or narrow/restrict > them. Maybe hisilicon,fixed-mode fits your purpose? > Hi Krzysztof, Thanks for the suggestion, using something like hisilicon,fixed-node looks like a good fit. With mediatek,fixed-node, by default CSI node will be considered as CD-PHY capable (unless the fixed-mode property is set.) so I won't need anymore the new define PHY_TYPE_CDPHY introduced in v3. Also introducing mediatek,fixed-mode suggests that PHYs not declaring the fixed mode property support mode selection, so I suggest to add a phy argument (#phy-cells = <1>) to select the mode (D or C mode). Exactly what is done by the hsilicon driver. How does that sound to you? Best, Julien > Best regards, > Krzysztof >
