Re: [PATCH] Factor in buddy allocator alignment requirements in node memory alignment

["Martin J. Bligh" <mbligh@mbligh.org>]

--On Monday, May 16, 2005 12:43:17 -0700 christoph <christoph@scalex86.org> wrote:

> On Mon, 16 May 2005, Dave Hansen wrote:
> 
>> On Mon, 2005-05-16 at 12:05 -0700, christoph wrote:
>> > Memory for nodes on i386 is currently aligned on 2 MB boundaries.
>> > However, the buddy allocator needs pages to be aligned on
>> > PAGE_SIZE << MAX_ORDER which is 8MB if MAX_ORDER = 11.
>> 
>> Why do you need this?  Are you planning on allowing NUMA KVA remap pages
>> to be handed over to the buddy allocator?  That would be a major
>> departure from what we do now, and I'd be very interested in seeing how
>> that is implemented before a infrastructure for it goes in.
> 
> Because the buddy allocator is complaining about wrongly allocated zones!
> 
> in page_alloc.c:
> 
> static void __init free_area_init_core(struct pglist_data *pgdat,
>                 unsigned long *zones_size, unsigned long *zholes_size)
> {
> ...
> 
>   const unsigned long zone_required_alignment = 1UL << (MAX_ORDER-1);
> 
> ...
> 
>              if ((zone_start_pfn) & (zone_required_alignment-1))
>                         printk(KERN_CRIT "BUG: wrong zone alignment, it will crash\n");

IIRC, we decided that warning was worthless ... and I *think* Andy fixed
it to do non-aligned zones, though it might have been someone else. Andy,
can you remember what you did to fix this up? 

M.