On Thu, 21 Nov 2013 18:49:08 +0100, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote: > On Thu, Nov 21, 2013 at 4:53 PM, Grant Likely <grant.likely@xxxxxxxxxxxx> wrote: > >> My changes don't change the current behavior much: currently > >> early_init_dt_scan() is already called with &__dtb_start in several places. > >> If this is broken, it's already broken. > > > > Yes, but it is called on platforms that already make that assumption > > that __dtb_start must be copied and therefore call > > unflatten_and_copy_devicetree(). This change makes *all* platforms do > > that. That breaks arm, arm64, c6x, microblaze, some mips platforms, and > > powerpc! > > arm, arm64, and powerpc don't have builtin DTBs, hence no change. > microblaze and c6x already did "early_init_devtree(&__dtb_start)" before. > > That leaves mips, which has DT handling in too many variants, which I > didn't all verify. > > >> > memory. The dtb section can also potentially contain multiple .dtb > >> > blobs. In the use case that you care about you are probably only > >> > >> Multiple dtb blobs are currently handled in platform-specific code, which > >> passes the right dtb to early_init_dt_scan(). > > > > The problem is that it makes the default dt completely random because > > the generic code still deferrences __dtb_start. > > Only if no DT was passed by the bootloader. And only if there really > is something at __dtb_start. Before it would fail if no DTB was passed > by the bootloader. > > >> > thinking about one, but it is entirely possible for device drivers to > >> > have a dtb linked in which may break this if it gets linked in a > >> > different order. The specific example I'm thinking about is I want to > >> > have the DT selftest code load an overlay to get testcase data from a > >> > dtb blob. > >> > > >> > The other concern I have here is that I don't really want this to be the > >> > default on a lot of platforms. ARM and PowerPC for instance should only > >> > get the default dtb from the boot wrapper. It needs to be configurable > >> > in some way. > >> > >> On ARM and PowerPC, the section is empty, hence &__dbt_start == > >> &__dtb_end. > > > > There is no guarantee that that will always be so. The patch makes some > > poor assumtions, but they shouldn't be difficult to fix. > > OK, how to proceed? Add a global flag, bool __init of_dtb_needs_copy, and make it default to false. On platforms that need it, change it to true. Make unflatten_device_tree() check the flag and copy the tree if necessary. Make early_init_dt_scan() set the flag if it choses to use a default built-in tree. Then get rid of unflatten_and_copy_devicetree() because it has become redundant. Finally, make an explicit backup __dtb pointer. Don't assume that __dtb_start == __dtb_end is the correct test. We never want to have that behavour on arm/powerpc/microblaze, but that doesn't mean there won't be compiled in dtbs for other reasons. Create a __dtb_default symbol require platforms to do something specific (ie. select a config symbol) to set it to __dtb_start. It's not perfect, but at least it acknowledges the problem and can be refined later. g. -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
