On Mon, Feb 13, 2006 at 09:35:00AM +0000, Martin Husemann wrote:
> I'm trying to get the nvidia nforce ethernet lkm working on amd64. It does
> not load, crashes the kernel instead.
> 
> On investigation I found that I can not load *any* lkm.
> 
> I get a failed assertion: end < VM_MAX_KERNEL_ADDRESS
> in uvm/uvm_km.c:479
> 
> with the backtrace:
> uvm_km_check_empty()
> uvm_map()
> uvm_km_alloc()
> lkmioctl()
> 
> 
> (this is a DEBUG, DIAGNOSTIC, LOCKDEBUG kernel, and yes, I do know that the
> pf.o would not have worked anyway in this kernel)

The vm_map that LKM space is allocated from is lkm_map. This is usually
the same as kernel_map, but on a few architectures it is different.
On amd64 it is different, since the LKM code needs to be in proximity
to the kernel code, to make the branches/jumps fit.

end < VM_MAX_KERNEL_ADDRESS is *not* true for lkm_map space.

The problem is that VM_MAP_IS_KERNEL, which the check is conditional
on, is coded as: (vm_map_pmap(map) == pmap_kernel(). This is true
for lkm_map as well, so things go wrong.

I can't think of an easy fix right away, but it should not be
too hard.

- Frank