Linus Torvalds wrote:
> Is this only ever used for kernel register state?
>
> Because in the _general_ case, the code should likely be something like
>
> 	if (user_mode_vm(regs)) {
> 		gdb_regs[GDB_SS] = regs->ss;
> 		gdb_regs[GDB_SP] = regs->sp;
> 	} else {
> 		gdb_regs[GDB_SS] = __KERNEL_DS;
> 		gdb_regs[GDB_SP] = (unsigned long)&regs->sp
> 	}
>   

You have discovered a long standing corner case.  The only way you can
end up with user_mode_vm() being true is the death by NMI watch dog or
from the IPI to bring the non master kgdb cpus into debugger. By default
the kgdb exception handler is not going to end up with that condition,
because there is a check against regs in the kgdb notify handler for the
non NMI/IPI events.

Thank you for the suggestion.   I went ahead and tested it out to
confirm the behavior, as well as to run the standard set of kgdb
regression tests.

The corner case is fixed in the attached patch, and I updated for_linus
branch with just this patch, if you would prefer to pull it.

 
git://git.kernel.org/pub/scm/linux/kernel/git/jwessel/linux-2.6-kgdb.git
for_linus

Thanks,
Jason.