On Mon, 30 Dec 2013, Grygorii Strashko wrote:

quoted

quoted

diff --git a/mm/memblock.c b/mm/memblock.c
index 71b11d9..6af873a 100644
--- a/mm/memblock.c
+++ b/mm/memblock.c

@@ -707,11 +707,9 @@ void __init_memblock __next_free_mem_range(u64 *idx,

int nid,
  	struct memblock_type *rsv = &memblock.reserved;
  	int mi = *idx & 0xffffffff;
  	int ri = *idx >> 32;
-	bool check_node = (nid != NUMA_NO_NODE) && (nid != MAX_NUMNODES);

-	if (nid == MAX_NUMNODES)
-		pr_warn_once("%s: Usage of MAX_NUMNODES is depricated. Use
NUMA_NO_NODE instead\n",
-			     __func__);
+	if (WARN_ONCE(nid == MAX_NUMNODES, "Usage of MAX_NUMNODES is
deprecated. Use NUMA_NO_NODE instead\n"))
+		nid = NUMA_NO_NODE;

  	for ( ; mi < mem->cnt; mi++) {
  		struct memblock_region *m = &mem->regions[mi];

Um, why do this at runtime?  This is only used for
for_each_free_mem_range(), which is used rarely in x86 and memblock-only
code.  I'm struggling to understand why we can't deterministically fix the
callers if this condition is possible.

Unfortunately, It's not so simple as from first look :(
We've modified __next_free_mem_range_x() functions which are part of
Memblock APIs (like memblock_alloc_xxx()) and Nobootmem APIs.
These APIs are used as directly as indirectly (as part of callbacks from other
MM modules like Sparse), as result, it's not trivial to identify all places
where MAX_NUMNODES will be used as input parameter.

These functions are only used for for_each_free_mem_range() and 
for_each_free_mem_range_reverse().  I can very easily find which callers 
are passing MAX_NUMNODES deterministically.

NACK to doing this at runtime.