On Jun 28, 2011, at 4:39 PM, Benjamin Herrenschmidt wrote:

> On Tue, 2011-06-28 at 14:54 -0500, Becky Bruce wrote:
>> struct page *alloc_huge_page_node(struct hstate *h, int nid);
>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>> index 6402458..2db81ea 100644
>> --- a/mm/hugetlb.c
>> +++ b/mm/hugetlb.c
>> @@ -1105,8 +1105,14 @@ static void __init
>> gather_bootmem_prealloc(void)
>>        struct huge_bootmem_page *m;
>> 
>>        list_for_each_entry(m, &huge_boot_pages, list) {
>> -               struct page *page = virt_to_page(m);
>>                struct hstate *h = m->hstate;
>> +#ifdef CONFIG_HIGHMEM
>> +               struct page *page = pfn_to_page(m->phys >>
>> PAGE_SHIFT);
>> +               free_bootmem_late((unsigned long)m,
>> +                                 sizeof(struct huge_bootmem_page));
>> +#else
>> +               struct page *page = virt_to_page(m);
>> +#endif
>>                __ClearPageReserved(page);
> 
> Why do you add free_bootmem_late() in the highmem case and not the
> normal case ?

Because there was no bootmem allocation in the normal case - the non-highmem version stores data structure in the huge page itself.  This is perfectly fine as long as you have a mapping.  Since this isn't true for HIGHMEM pages, I allocate bootmem to store the early data structure that stores information about the hugepage (this happens in arch-specific code in alloc_bootmem_huge_page).

-Becky