On Thu, Jun 21, 2012 at 1:17 PM, Tejun Heo [off-list ref] wrote:

Hello, Yinghai.

On Tue, Jun 19, 2012 at 07:57:45PM -0700, Yinghai Lu wrote:

quoted

if it is that case, that change could fix other problem problem too.

--- during the one free reserved.regions could double the array.

Yeah, that sounds much more attractive to me too.  Some comments on
the patch tho.

quoted

 /**
  * memblock_double_array - double the size of the memblock regions array
  * @type: memblock type of the regions array being doubled

@@ -216,7 +204,7 @@ static int __init_memblock memblock_doub

      /* Calculate new doubled size */
      old_size = type->max * sizeof(struct memblock_region);
-     new_size = old_size << 1;
+     new_size = PAGE_ALIGN(old_size << 1);

We definintely can use some comments explaining why we want page
alignment.  It's kinda subtle.

yes.

This is a bit confusing here because old_size is the proper size
without padding while new_size is page aligned size with possible
padding.  Maybe discerning {old|new}_alloc_size is clearer?  Also, I
think adding @new_cnt variable which is calculated together would make
the code easier to follow.  So, sth like,

       /* explain why page aligning is necessary */
       old_size = type->max * sizeof(struct memblock_region);
       old_alloc_size = PAGE_ALIGN(old_size);

       new_max = type->max << 1;
       new_size = new_max * sizeof(struct memblock_region);
       new_alloc_size = PAGE_ALIGN(new_size);

and use alloc_sizes for alloc/frees and sizes for everything else.

ok, will add new_alloc_size, old_alloc_size.

quoted

 unsigned long __init free_low_memory_core_early(int nodeid)
 {
      unsigned long count = 0;
-     phys_addr_t start, end;
+     phys_addr_t start, end, size;
      u64 i;

-     /* free reserved array temporarily so that it's treated as free area */
-     memblock_free_reserved_regions();
+     for_each_free_mem_range(i, MAX_NUMNODES, &start, &end, NULL)
+             count += __free_memory_core(start, end);

-     for_each_free_mem_range(i, MAX_NUMNODES, &start, &end, NULL) {
-             unsigned long start_pfn = PFN_UP(start);
-             unsigned long end_pfn = min_t(unsigned long,
-                                           PFN_DOWN(end), max_low_pfn);
-             if (start_pfn < end_pfn) {
-                     __free_pages_memory(start_pfn, end_pfn);
-                     count += end_pfn - start_pfn;
-             }
-     }
+     /* free range that is used for reserved array if we allocate it */
+     size = get_allocated_memblock_reserved_regions_info(&start);
+     if (size)
+             count += __free_memory_core(start, start + size);

I'm afraid this is too early.  We don't want the region to be unmapped
yet.  This should only happen after all memblock usages are finished
which I don't think is the case yet.

No, it is not early. at that time memblock usage is done.

Also I tested one system with huge memory, duplicated the problem on
KVM that Sasha met.
my patch fixes the problem.

please check attached patch.

Also I add another patch to double check if there is any reference
with reserved.region.
so far there is no reference found.

Thanks

Yinghai