On Wed, Mar 23, 2022 at 09:06:18PM +0100, Max Krummenacher wrote: > Am Mittwoch, den 23.03.2022, 16:58 +0100 schrieb Maxime Ripard: > > On Wed, Mar 23, 2022 at 09:42:11AM +0100, Max Krummenacher wrote: > > > Am Freitag, den 18.03.2022, 17:53 +0000 schrieb Dave Stevenson: > > > > On Fri, 18 Mar 2022 at 17:16, Maxime Ripard wrote: > > > > > On Fri, Mar 18, 2022 at 05:05:11PM +0000, Dave Stevenson wrote: > > > > > > On Fri, 18 Mar 2022 at 16:35, Maxime Ripard wrote: > > > > > > > On Mon, Mar 07, 2022 at 04:26:56PM +0100, Max Krummenacher wrote: > > > > > > > > On Wed, Mar 2, 2022 at 5:22 PM Marek Vasut wrote: > > > > > > > > > On 3/2/22 15:21, Maxime Ripard wrote: > > > > > > > > > > On Wed, Feb 23, 2022 at 04:25:19PM +0100, Max Krummenacher wrote: > > > > > > > > > > > The goal here is to set the element bus_format in the struct > > > > > > > > > > > panel_desc. This is an enum with the possible values defined in > > > > > > > > > > > include/uapi/linux/media-bus-format.h. > > > > > > > > > > > > > > > > > > > > > > The enum values are not constructed in a way that you could calculate > > > > > > > > > > > the value from color channel width/shift/mapping/whatever. You rather > > > > > > > > > > > would have to check if the combination of color channel > > > > > > > > > > > width/shift/mapping/whatever maps to an existing value and otherwise > > > > > > > > > > > EINVAL out. > > > > > > > > > > > > > > > > > > > > > > I don't see the value in having yet another way of how this > > > > > > > > > > > information can be specified and then having to write a more > > > > > > > > > > > complicated parser which maps the dt data to bus_format. > > > > > > > > > > > > > > > > > > > > Generally speaking, sending an RFC without explicitly stating what you > > > > > > > > > > want a comment on isn't very efficient. > > > > > > > > > > > > > > > > > > Isn't that what RFC stands for -- Request For Comment ? > > > > > > > > > > > > > > > > I hoped that the link to the original discussion was enough. > > > > > > > > > > > > > > > > panel-simple used to have a finite number of hardcoded panels selected > > > > > > > > by their compatible. > > > > > > > > The following patchsets added a compatible 'panel-dpi' which should > > > > > > > > allow to specify the panel in the device tree with timing etc. > > > > > > > > > > > > > > > > https://patchwork.kernel.org/project/dri-devel/patch/20200216181513.28109-6-sam@xxxxxxxxxxxx/ > > > > > > > > In the same release cycle part of it got reverted: > > > > > > > > > > > > > > > > https://patchwork.kernel.org/project/dri-devel/patch/20200314153047.2486-3-sam@xxxxxxxxxxxx/ > > > > > > > > With this it is no longer possible to set bus_format. > > > > > > > > > > > > > > > > The explanation what makes the use of a property "data-mapping" not a > > > > > > > > suitable way in that revert > > > > > > > > is a bit vague. > > > > > > > > > > > > > > Indeed, but I can only guess. BGR666 in itself doesn't mean much for > > > > > > > example. Chances are the DPI interface will use a 24 bit bus, so where > > > > > > > is the padding? > > > > > > > > > > > > > > I think that's what Sam and Laurent were talking about: there wasn't > > > > > > > enough information encoded in that property to properly describe the > > > > > > > format, hence the revert. > > > > > > I agree that the strings used to set "data-mapping" weren't self explaining. > > > However, as there was a > > > clear 1:1 relation to the bus_format value the meaning > > > wasn't ambiguous at all. > > > > > > > > > MEDIA_BUS_FMT_RGB666_1X18 defines an 18bit bus, therefore there is no > > > > > > padding. "bgr666" was selecting that media bus code (I won't ask about > > > > > > the rgb/bgr swap). > > > > > > > > > > > > If there is padding on a 24 bit bus, then you'd use (for example) > > > > > > MEDIA_BUS_FMT_RGB666_1X24_CPADHI to denote that the top 2 bits of each > > > > > > colour are the padding. Define and use a PADLO variant if the padding > > > > > > is the low bits. > > > > > > > > > > Yeah, that's kind of my point actually :) > > > > > > > > Ah, OK :) > > > > > > > > > Just having a rgb666 string won't allow to differentiate between > > > > > MEDIA_BUS_FMT_RGB666_1X18 and MEDIA_BUS_FMT_RGB666_1X24_CPADHI: both are > > > > > RGB666 formats. Or we could say that it's MEDIA_BUS_FMT_RGB666_1X18 and > > > > > then when we'll need MEDIA_BUS_FMT_RGB666_1X24_CPADHI we'll add a new > > > > > string but that usually leads to inconsistent or weird names, so this > > > > > isn't ideal. > > > > > > We're on the same page that the strings that were used aren't self > > > explaining and do not follow a pattern which would make it easy to > > > extend. However that is something I addressed in my RFC proposal, not? > > > > > > > > > The string matching would need to be extended to have some string to > > > > > > select those codes ("lvds666" is a weird choice from the original > > > > > > patch). > > > > > > > > > > > > Taking those media bus codes and handling them appropriately is > > > > > > already done in vc4_dpi [1], and the vendor tree has gained > > > > > > BGR666_1X18 and BGR666_1X24_CPADHI [2] as they aren't defined in > > > > > > mainline. > > > > > > > > > > > > Now this does potentially balloon out the number of MEDIA_BUS_FMT_xxx > > > > > > defines needed, but that's the downside of having defines for all > > > > > > formats. > > > > > > > > > > > > (I will admit to having a similar change in the Pi vendor tree that > > > > > > allows the media bus code to be selected explicitly by hex value). > > > > > > > > > > I think having an integer value is indeed better: it doesn't change much > > > > > in the device tree if we're using a header, it makes the driver simpler > > > > > since we don't have to parse a string, and we can easily extend it or > > > > > rename the define, it won't change the ABI. > > > > > > Fine with me. I'm fine with integers too. Strings give the false impression that new values can be added with a lower risk of a conflict, but that just a false impression. > > > > > I'm not sure using the raw media bus format value is ideal though, since > > > > > that value could then be used by any OS, and it would effectively force > > > > > the mbus stuff down their throat. > > > > > > I disagree here, this forces us to use code to map the device tree enum > > > to the kernel enum for Linux, i.e. adds complexity and maintenance work > > > if additional bus_formats are needed. > > > Assuming there is another OS which uses the device tree it would not > > > make a difference, that OS would still need to map the device tree enum > > > to the corresponding representation in their kernel. > > > > So, you don't want to do something in Linux, but would expect someone > > else to be completely ok with that? > > Yes, sort of. > Recycling the values as used currently in the Linux kernel rather than > inventing a new numbering will make the Linux code a little easier to > write, read and maintain without any negative effect on how that other > OSs would have to map the DT representation to their internal representation. > > Would you rather have something like: > > <the common dt-bindings header file> > DT_BUS_FMT_RGB666_1X18 1 > DT_BUS_FMT_RGB888_1X24 2 > ... > > <panel-simple.c> > switch (bus-format) { > case DT_BUS_FMT_RGB666_1X18: > bus_format = MEDIA_BUS_FMT_RGB666_1X18; break; > case DT_BUS_FMT_RGB888_1X24: > bus_format = MEDIA_BUS_FMT_RGB888_1X24; break; > ... I'm having a bit of trouble providing comments on this RFC, because I don't believe I have a good enough overview of the different cases we need to support (and in particular the corner cases). Max, you mentioned an interesting one on Tegra platforms, would you maybe have a short (or long, who knows) list of use cases you need to support now, or just know about ? I think it would be easier to discuss the problem and the best solution with concrete examples. One particular thing that needs to be taken into account is that not all devices (I'm talking about both the panel side and the source side) use a data bus with contiguous bits. How to map a format to D[23:0] is one thing, but there are devices that document pins as R[7:0], G[7:0], B[7:0] (possibly with some permutations of the components). It's quite easy to map between those two representations, once a mapping is defined. I'd like these things to be considered explicitly instead of relying on an implicit shared knowledge, as in my experience implicit rules lead to one version per participant in the conversation :-) > > > I would copy the definitions of media-bus-format.h into a header in > > > include/dt-bindings similarly as it is done for > > > include/dt-bindings/display/sdtv-standards.h for TV standards. > > > > That might not be an option: that header is licensed under the GPL, > > device trees are usually licensed under GPL+MIT, and we don't have any > > requirements on the license for other projects using a DT (hence the > > dual license). > > That one I didn't consider. That would be solved by a newly invented > enum. -- Regards, Laurent Pinchart
