On Thu, Dec 6, 2018 at 7:32 AM Andreas Färber <afaerber@xxxxxxx> wrote: > > Am 05.12.18 um 05:17 schrieb Rob Herring: > > On Tue, Dec 4, 2018 at 7:22 PM Andreas Färber <afaerber@xxxxxxx> wrote: > >> > >> Rob, > >> > >> Am 04.12.18 um 19:36 schrieb Rob Herring: > >>> I've put together a script to move the dts files and update the > >>> makefiles. It doesn't handle files not following a common prefix which > >>> isn't many and some includes within the dts files will need some fixups > >>> by hand. > >>> > >>> MAINTAINERS will also need updating. > >>> > >>> A few questions: > >>> > >>> Do we want to move absolutely everything to subdirs? > >> > >> This refactoring is a terrible idea! > > > > How do you really feel? > > > >> While it would've been nice to have more structure from the start, > >> bootloaders like U-Boot expect a flat structure for arm .dtb files now. > >> If you start installing them into subdirs instead, they won't find the > >> files anymore under the hardcoded name. > >> > >> Doing this only for new platforms would be much less invasive and allow > >> to prepare bootloaders accordingly. > > > > That was my suggestion where this started for the new RDA platform. > > Olof preferred to move everything and that's my desire too. > > > >> Alternatively, white-list which ones > >> are safe to move around. > > > > I'd prefer to know which ones the distros don't want moved. That > > should be easier to figure out. We also need that anyways in context > > of what platforms we care about compatibility. > > > > Another option is dtbs_install target could flatten the installed > > dtbs. That is the only part the distros should depend on. > > I'd be okay with distinguishing source vs. install location. Due to the > issue I mention below (and more) we can't use install_dtbs for openSUSE > and had to reimplement it, which we'd need to (and can) adjust. What would be needed for dtbs_install to work? arm64 needs to support a flat install? If it doesn't work for Debian or openSUSE, I'm not sure why we have it. So I'd like to make it work. > >> But don't just script a refactoring because it > >> looks nicer in the source tree, without testing what side effects this > >> can have for board/distro users of the compiled files in practice. > >> We already had that discussion for arm64 because Debian chose to ignore > >> the kernel-installed subdirectories and installed .dtb files into a flat > >> directory, which collided with openSUSE sticking to the kernel choice. > > > > So everyone already deals with subdirs or not with arm and arm64 > > already, seems like they can deal with this. I will raise the topic on > > the cross-distro list though. > > Sounds like you're twisting words... The keyword was "hardcoded" paths - > one way or another (not "and") depending on the kernel choices being > flat for arm, vendor subdir for arm64. > > >> This topic becomes even more important with EBBR: There is neither a > >> mechanism in place to sync .dts files into U-Boot or EDK2 source trees, > >> nor are capsule updates implemented in U-Boot for easily deploying such > >> bootloaders with new .dts sources or paths yet. > > > > EBBR actually says firmware (including dtbs) goes in directories named > > by vendor. > > Fine, but unrelated. If the distros want dtbs in a flat dir and EBBR says otherwise, then it is related. > >> And I can assure you > >> that just getting users to dd the right bootloader can be difficult... > >> Since DT forward and backward compatibility is often being neglected, > >> for example with optional properties or renamed compatibles that break > >> booting with previous drivers, new kernel versions often need updated > >> Device Trees to make use of new/enhanced drivers. Therefore it is > >> unfortunately often enough a necessity to load newer kernel-based .dtb > >> files matching the kernel (as opposed to the dream of kernel-independent > >> hardware descriptions) when working with the latest -rc or -next kernels > >> at least. For examples of DTs needing updates, look no further than > >> Linaro's 96boards - in case of hikey960/EDK2 GRUB is another layer where > >> .dtb paths may be hardcoded, ditto for arm; and Armada was an example > >> where the upstream bindings for the network IP changed incompatibly. > > > > Compatibility is an issue, yes, but that really has nothing to do with this. > > > >> DT overlays are another topic that is not making any progress upstream > >> according to the ELCE BoF, so beyond the Raspberry Pi the only known > >> working way to apply them is to write a U-Boot boot.scr script, which > >> can either reuse $fdtcontroladdr DT or use the filename $fdtfile or > >> hardcode one, the latter two of which would break with your renaming. > > > > DT overlays also have nothing to do with this as there aren't any in > > the kernel. I'm not inclined to take any either with a flat tree. > > We're already at 1800+ files. > > Read again: a) Breaking DT changes and b) the desire to use Overlays > instead of replacing the bootloaders for each change are _reasons_ why > people depend on .dtb filenames from the kernel source tree for their > boot flow today. Nothing to do with downstream .dtbo files. > > For example, remember when I reported that the kernel didn't compile DTs > with -@? No reaction except for Frank asking to be CC'ed - was it ever > fixed??? Do EDK2's or U-Boot's built-in DTs compile with -@ today? IIRC, Frank objected to changing this globally because it will bloat all dtbs. And then no one did the work to make it a per dtb option. Maybe that was the same issue in another thread. > Raspberry Pi overlays in U-Boot work because it switched to passing the > DT through from the proprietary firmware. > Point being that while it would be nice to get a current, compatible DT > via UEFI tables and ignore .dtb filenames outside of a few bootloaders, > in reality we're not quite there yet for all platforms. > > I see no problem (except for naming choices) moving new targets like RDA > to subfolders because we can then hardcode it the new way; I also assume > deeply embedded targets like stm32f4 or targets with no mainline > bootloaders yet like owl-s500 could be refactored. > I do see problems refactoring widely used SBC targets like sunXi though. > > Regards, > Andreas > > -- > SUSE Linux GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany > GF: Felix Imendörffer, Jane Smithard, Graham Norton > HRB 21284 (AG Nürnberg)
