>>   How about loading netbsd.bin at 0xa0200000, and go
>> 0xa0200000?
>> That's what we do with redboot.
>
> Still the same. I got an error message about "data
> abort" for u-boot.

Are you sure how address space is prepared at the moment
when u-boot runs go command?  ARM requires to enable
MMU to make VIVT cache work and it's unrealistic to leave
MMU turned off for real usage.  This makes somehow
confusing and aweful consequence to switch address space
which MMU provides.  It's apparent to address space will
differ in two different OS (u-boot and NetBSD) and "run
in dark before dawn" strategy is unavoidable.  I think ARM
camp is suffered from the cold fact and it's the reason why
address space arrangements are quite inconsistent across
various product implementations.

I emphasis again check the way address space is crafted.
Note; Linux has a bad habit to use a naked executable to 
start OS.  It's wrong not to use ELF header to go. 

Toru Nishimura/ALKYL Technology