On Thu, Oct 17, 2013 at 10:42:53AM +0900, Simon Horman wrote: > On Wed, Oct 16, 2013 at 11:33:43AM +0200, Laurent Dufour wrote: > > On 16/10/2013 09:28, Simon Horman wrote: > > > On Tue, Oct 15, 2013 at 07:05:28PM +0200, Laurent Dufour wrote: > > >> While reviewing fs2dt.c in the common kexec directory, in order to use it to > > >> support little endian ppc64 architecture, I found some endianess > > >> conversion's issues. > > >> > > >> In dt_reserve, dt_base is a pointer and thus should not be converted. > > >> > > >> In checkprop, values read from the device tree are big endian encoded and > > >> should be converted if CPU is running in little endian mode. > > >> > > >> In add_dyn_reconf_usable_mem_property__, fix extraneous endianess conversion > > >> of rlen which should be natively used to increase the dt pointer. > > > > > > These changes seem logical to me. > > > > > > However, I'm unsure how the code might have worked as-is. > > > In particular I'm pretty sure ARM was running in little endian > > > mode when I worked on DT support for it. Perhaps these code > > > paths were not executed. Or the errors cancelled themselves out somehow? > > > > I've to admit that I don't know how this has worked for ARM. > > I run this patch with a set of change I'm working on to support PPC64 > > little endian architecture. My tests were done with kexec-tools built in > > little endian mode, using this fs2dt.c file and I need this fix to build > > a valid device tree. > > > > The fix is dt_reserve is hit unless the device tree is large enough. I > > had to reduce INIT_TREE_WORDS to hit it. > > > > I think the ones in checkprop, add_dyn_reconf_usable_mem_property__ are > > Power specific since RTAS and 'ibm,dynamic-memory' are required to hit them. > > > > This may explain why this issue are not seen in ARM mode. > > Perhaps, its a bit of a mystery. > I think that I will apply your change as it does seem to be correct. And I have. It should appear on kernel.org shortly.
