On Wednesday 06 February 2002 5:09 am, Jason R Thorpe wrote:
> On Tue, Feb 05, 2002 at 07:44:27PM -0800, Jason R Thorpe wrote:
>  > This says to me that the recursive PT mapping is somehow screwed up.
>  >
>  > Now, for recursive PT mappings to work properly, both the L1 and L2
>  > descriptors must have the same format.  This is not generally the case
>  > with ARM PTEs.
>
> Yah, and after discussing this w/ Bill Sommerfeld, he set me straight
> and I now have my head wrapped around the non-recursive PT mapping used
> by the ARM pmap (sigh, I have been working on this code too long; my
> eyes are starting to glaze over).
>
> That still doesn't change my puzzlement with what's going on with the
> mdpage change...

I've not had time to look in depth, so this could be completly clutching at 
straws.   but there's a difference in the way vm_pages are allocated:
	freepages = 0;
	for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
		freepages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
[snip]
	pagecount = ((freepages + 1) << PAGE_SHIFT) /
	    (PAGE_SIZE + sizeof(struct vm_page));

If I read the above it doesn't actually give a vm_page for every page of 
memory, compare that with how the pv_heads were allocated:
	{
	       	int npages, lcv;
		vsize_t s;

		npages = 0;
		for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
		npages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
		s = (vsize_t) (sizeof(struct pv_head) * npages +
				sizeof(char) * npages);
		s = round_page(s); /* round up */
		boot_head = (char *)uvm_pageboot_alloc(s);
		bzero((char *)boot_head, s);
		if (boot_head == 0)
			panic("pmap_init: unable to allocate pv_heads");
	}


Note that it doesn't do the strange bit of maths.  it even uses pv_heads on 
its own memory.

Quick check with some maths:

say we've got 8192 pages, that's about 32MB.  In the old style you got 8192 
pv_heads, with the new style you get 8168 vm_pages, because I guess we don't 
need vm_pages to manage where the vm_pages were allocated?  but I'm wondering 
if we do need pv_heads for them?

Note I'm quite possibly wrong though. I'll try to find time this evening to 
play with the code.

Chris