On Thu, 2004-08-05 at 14:13, Andrew Morton wrote:
> Dave Hansen <haveblue@us.ibm.com> wrote:
> >
> > When using CONFIG_NONLINEAR, a zone's mem_map isn't contiguous, and
> > isn't allocated in the same place.  This means that nonlinear doesn't
> > really have a mem_map[] to pass into free_area_init_node() or 
> > memmap_init_zone() which makes any sense.  
> 
> argh, sorry.  It's this patch which I dropped due to psychedelic screen
> syndrome.  The "break out zone free list initialization" patch is innocent, and
> was included in rc3-mm1.

OK, here's a replacement for the bad one.  I could provide an
incremental instead if you prefer.

Turns out I was initializing the mem_map too late.  Not only did
pfn_to_page() use the pgdats, but it also used mem_map.  Not
surprisingly, mem_map == NULL did not make things happy.

My earlier testing didn't catch it because I was always using either
DISCONTIGMEM or NONLINEAR (even on plain SMP because I'm weird), both of
which skip the code that I screwed up.  

-- Dave