[PATCH 2/8] mm: alloc_contig_freed_pages() added

[dave@linux.vnet.ibm.com (Dave Hansen)]

On Wed, 2011-09-21 at 15:17 +0200, Michal Nazarewicz wrote:
> > This 'struct page *'++ stuff is OK, but only for small, aligned areas.
> > For at least some of the sparsemem modes (non-VMEMMAP), you could walk
> > off of the end of the section_mem_map[] when you cross a MAX_ORDER
> > boundary.  I'd feel a little bit more comfortable if pfn_to_page() was
> > being done each time, or only occasionally when you cross a section
> > boundary.
> 
> I'm fine with that.  I've used pointer arithmetic for performance reasons
> but if that may potentially lead to bugs then obviously pfn_to_page()  
> should be used

pfn_to_page() on x86 these days is usually:

	#define __pfn_to_page(pfn)      (vmemmap + (pfn))

Even for the non-vmemmap sparsemem it stays pretty quick because the
section array is in cache as you run through the loop.

There are ways to _minimize_ the number of pfn_to_page() calls by
checking when you cross a section boundary, or even at a
MAX_ORDER_NR_PAGES boundary.  But, I don't think it's worth the trouble.

-- Dave