Skip to contentSkip to conversation

treewide, mm: initialize node_possible_map like the other node maps

PATCH In review Sang-Heon Jeon 2 participants 8 messages Jul 1 – Jul 2

Subsystems (6)
lore
Share & tools
lore
#0 lore raw

cover treewide, mm: initialize node_possible_map like the other node maps

While reviewing an earlier series [1], Andrew asked why we carefully initialize node_possible_map at compile time, then zero it within __init code anyway.

node_possible_map, aliased by node_states[N_POSSIBLE], is initialized to NODE_MASK_ALL at compile time, unlike the other entries, which are initialized with only node 0 set.

Architectures that use numa_memblks set node_possible_map from numa_nodes_parsed, so the compile-time value is meaningless for them.

However, a few architectures that do not use numa_memblks do not set node_possible_map on their own. Once these architectures set node_possible_map to match their own topology, the NODE_MASK_ALL initialization can be removed.

Patches 1-2 handle sparc64 and sh, which do not set node_possible_map themselves. Both set node_online_map from their topology, so set node_possible_map from node_online_map.

Patch 3 handles powerpc, which sets node_possible_map by intersecting it with node_online_map. Nothing sets node_possible_map before this, so it is still NODE_MASK_ALL and the intersection is just node_online_map. Once patch 4 changes that initial value, the intersection would no longer be node_online_map, so assign node_online_map directly.

Patch 4 initializes node_possible_map with only node 0 set, like the other node_states[] entries. NODE_MASK_ALL is then no longer used, so remove it and its only helper NODE_MASK_LAST_WORD.

[1] https://lore.kernel.org/all/[email protected]/

Sang-Heon Jeon (4):
sparc64: set node_possible_map in bootmem_init_numa()
sh: set node_possible_map in do_init_bootmem() powerpc/numa: set node_possible_map from node_online_map mm/page_alloc: initialize node_possible_map like the other node maps

 arch/powerpc/mm/numa.c   |  2 +-
 arch/sh/mm/init.c        |  2 ++
 arch/sparc/mm/init_64.c  |  2 ++
 include/linux/nodemask.h | 20 --------------------
 mm/page_alloc.c          |  2 +-
 5 files changed, 6 insertions(+), 22 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH 1/4 sparc64: set node_possible_map in bootmem_init_numa()

Unlike other NUMA architectures, sparc64 never sets node_possible_map. bootmem_init_numa() marks all nodes online and node_online_map does not change afterwards, so set node_possible_map from node_online_map.

---
 arch/sparc/mm/init_64.c | 2 ++
 1 file changed, 2 insertions(+)
arch/sparc/mm/init_64.c +2 -0
index 103db4683b16..a749f63568b7 100644
@@ -1559,6 +1559,8 @@ static int __init bootmem_init_numa(void)
 		else
 			err = numa_parse_sun4u();
 	}
+	if (!err)
+		node_possible_map = node_online_map;
 	return err;
 }
 
Signature · 1 line
2.43.0
#2 lore raw

PATCH 2/4 sh: set node_possible_map in do_init_bootmem()

In reply to Sang-Heon Jeon

Unlike other NUMA architectures, sh never sets node_possible_map. do_init_bootmem() marks node 0 online and plat_mem_setup() marks the rest online, after which node_online_map does not change, so set node_possible_map from node_online_map.

---
 arch/sh/mm/init.c | 2 ++
 1 file changed, 2 insertions(+)
arch/sh/mm/init.c +2 -0
index 110308bdef01..13954d7bbdae 100644
@@ -227,6 +227,8 @@ static void __init do_init_bootmem(void)
 	node_set_online(0);
 
 	plat_mem_setup();
+
+	node_possible_map = node_online_map;
 }
 
 static void __init early_reserve_mem(void)
Signature · 1 line
2.43.0
#3 lore raw

PATCH 3/4 powerpc/numa: set node_possible_map from node_online_map

In reply to Sang-Heon Jeon

mem_topology_setup() intersects node_possible_map with node_online_map. Nothing sets node_possible_map before this, so it is NODE_MASK_ALL and the result is just node_online_map.

In preparation for changing node_possible_map's initial value, mem_topology_setup() no longer depends on it.

No functional change.

---
 arch/powerpc/mm/numa.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/mm/numa.c +1 -1
index f4cf3ae036de..2fdecae90a01 100644
@@ -1179,7 +1179,7 @@ void __init mem_topology_setup(void)
 	 * that we expect to make use of for this platform's affinity
 	 * calculations.
 	 */
-	nodes_and(node_possible_map, node_possible_map, node_online_map);
+	node_possible_map = node_online_map;
 
 	find_possible_nodes();
 
Signature · 1 line
2.43.0
maintainer
#4 lore raw

Re: [PATCH 3/4] powerpc/numa: set node_possible_map from node_online_map

Quoted 19 lines · Sang-Heon Jeon

mem_topology_setup() intersects node_possible_map with node_online_map. Nothing sets node_possible_map before this, so it is NODE_MASK_ALL and the result is just node_online_map.

In preparation for changing node_possible_map's initial value, mem_topology_setup() no longer depends on it.

No functional change.

---
 arch/powerpc/mm/numa.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/mm/numa.c +0 -0
index f4cf3ae036de..2fdecae90a01 100644

arch/powerpc/mm/numa.c void __init mem_topology_setup(void) · line 1184

Quoted 7 lines
arch/powerpc/mm/numa.c +1 -1
@@ -1179,7 +1179,7 @@ void __init mem_topology_setup(void)
 	 * that we expect to make use of for this platform's affinity
 	 * calculations.
 	 */
-	nodes_and(node_possible_map, node_possible_map, node_online_map);
+	node_possible_map = node_online_map;
 
 	find_possible_nodes();

And in find_possible_nodes():

         for (i = 0; i < max_nodes; i++) {
                 if (!node_possible(i))
                         node_set(i, node_possible_map);
         }

So, assuming the code reaches to that point, you unconditionally online all the offlined nodes.

This doesn't look correct at the first glance, and definitely it's not a "No functional change" type of things.

#5 lore raw

Re: [PATCH 3/4] powerpc/numa: set node_possible_map from node_online_map

Hi Yury,

Quoted 31 lines · Yury Norov
Quoted 19 lines · Sang-Heon Jeon

mem_topology_setup() intersects node_possible_map with node_online_map. Nothing sets node_possible_map before this, so it is NODE_MASK_ALL and the result is just node_online_map.

In preparation for changing node_possible_map's initial value, mem_topology_setup() no longer depends on it.

No functional change.

---
 arch/powerpc/mm/numa.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/mm/numa.c +0 -0
index f4cf3ae036de..2fdecae90a01 100644

arch/powerpc/mm/numa.c void __init mem_topology_setup(void) · line 1184

Quoted 7 lines
arch/powerpc/mm/numa.c +1 -1
@@ -1179,7 +1179,7 @@ void __init mem_topology_setup(void)
       * that we expect to make use of for this platform's affinity
       * calculations.
       */
-     nodes_and(node_possible_map, node_possible_map, node_online_map);
+     node_possible_map = node_online_map;
 
      find_possible_nodes();

And in find_possible_nodes():

Quoted 7 lines
         for (i = 0; i < max_nodes; i++) {
                 if (!node_possible(i))
                         node_set(i, node_possible_map);
         }

So, assuming the code reaches to that point, you unconditionally online all the offlined nodes.

This loop only sets bits in node_possible_map, not in node_online_map, so I can't find how it onlines the offlined nodes unexpectedly. Please correct me if I'm wrong or missing something.

Or are you talking about the case where max_nodes exceeds MAX_NUMNODES, which was also noticed by sashiko? I think it is a pre-existing issue, and this patch does not affect it.

Quoted 2 lines

This doesn't look correct at the first glance, and definitely it's not a "No functional change" type of things.

Best Regards,
Sang-Heon Jeon

maintainer
#6 lore raw

Re: [PATCH 3/4] powerpc/numa: set node_possible_map from node_online_map

Quoted 42 lines · Sang-Heon Jeon

Hi Yury,

Quoted 31 lines · Yury Norov
Quoted 19 lines · Sang-Heon Jeon

mem_topology_setup() intersects node_possible_map with node_online_map. Nothing sets node_possible_map before this, so it is NODE_MASK_ALL and the result is just node_online_map.

In preparation for changing node_possible_map's initial value, mem_topology_setup() no longer depends on it.

No functional change.

---
 arch/powerpc/mm/numa.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/powerpc/mm/numa.c +0 -0
index f4cf3ae036de..2fdecae90a01 100644

arch/powerpc/mm/numa.c void __init mem_topology_setup(void) · line 1184

Quoted 7 lines
arch/powerpc/mm/numa.c +1 -1
@@ -1179,7 +1179,7 @@ void __init mem_topology_setup(void)
       * that we expect to make use of for this platform's affinity
       * calculations.
       */
-     nodes_and(node_possible_map, node_possible_map, node_online_map);
+     node_possible_map = node_online_map;
 
      find_possible_nodes();

And in find_possible_nodes():

Quoted 7 lines
         for (i = 0; i < max_nodes; i++) {
                 if (!node_possible(i))
                         node_set(i, node_possible_map);
         }

So, assuming the code reaches to that point, you unconditionally online all the offlined nodes.

Quoted 7 lines

This loop only sets bits in node_possible_map, not in node_online_map, so I can't find how it onlines the offlined nodes unexpectedly. Please correct me if I'm wrong or missing something.

Or are you talking about the case where max_nodes exceeds MAX_NUMNODES, which was also noticed by sashiko? I think it is a pre-existing issue, and this patch does not affect it.

Oops, sorry, It's my bad. They are nodemask_t, not pointers, so in "node_possible_map = node_online_map" you copy by value.

Thanks,
Yury

#7 lore raw

PATCH 4/4 mm/page_alloc: initialize node_possible_map like the other node maps

In reply to Sang-Heon Jeon

node_states[N_POSSIBLE], aliased by node_possible_map, is initialized to NODE_MASK_ALL, unlike the other entries, which are initialized with only node 0 set.

On !NUMA, MAX_NUMNODES == 1, so initializing node_possible_map to NODE_MASK_ALL is equivalent to initializing it with only node 0 set.

On NUMA, every architecture now sets node_possible_map after parsing its NUMA topology, so nothing relies on the initial value.

So, instead of initializing node_possible_map to NODE_MASK_ALL, set only node 0, like the other entries. Then NODE_MASK_ALL is no longer used, so remove it and its only helper NODE_MASK_LAST_WORD.

No functional change.

---
 include/linux/nodemask.h | 20 --------------------
 mm/page_alloc.c          |  2 +-
 2 files changed, 1 insertion(+), 21 deletions(-)
include/linux/nodemask.h +0 -20
index b842aa525546..484fe366a94c 100644
@@ -47,7 +47,6 @@
  *					MAX_NUMNODES
  *
  * nodemask_t nodemask_of_node(node)	Return nodemask with bit 'node' set
- * NODE_MASK_ALL			Initializer - all bits set
  * NODE_MASK_NONE			Initializer - no bits set
  * unsigned long *nodes_addr(mask)	Array of unsigned long's in mask
  *
@@ -296,25 +295,6 @@ static __always_inline unsigned int __first_unset_node(const nodemask_t *maskp)
 	return min(MAX_NUMNODES, find_first_zero_bit(maskp->bits, MAX_NUMNODES));
 }
 
-#define NODE_MASK_LAST_WORD BITMAP_LAST_WORD_MASK(MAX_NUMNODES)
-
-#if MAX_NUMNODES <= BITS_PER_LONG
-
-#define NODE_MASK_ALL							\
-((nodemask_t) { {							\
-	[BITS_TO_LONGS(MAX_NUMNODES)-1] = NODE_MASK_LAST_WORD		\
-} })
-
-#else
-
-#define NODE_MASK_ALL							\
-((nodemask_t) { {							\
-	[0 ... BITS_TO_LONGS(MAX_NUMNODES)-2] = ~0UL,			\
-	[BITS_TO_LONGS(MAX_NUMNODES)-1] = NODE_MASK_LAST_WORD		\
-} })
-
-#endif
-
 #define NODE_MASK_NONE							\
 ((nodemask_t) { {							\
 	[0 ... BITS_TO_LONGS(MAX_NUMNODES)-1] =  0UL			\
mm/page_alloc.c +1 -1
index 784a6e2f8ba0..c0f17dede824 100644
@@ -191,7 +191,7 @@ EXPORT_SYMBOL(latent_entropy);
  * Array of node states.
  */
 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
-	[N_POSSIBLE] = NODE_MASK_ALL,
+	[N_POSSIBLE] = { { [0] = 1UL } },
 	[N_ONLINE] = { { [0] = 1UL } },
 #ifndef CONFIG_NUMA
 	[N_NORMAL_MEMORY] = { { [0] = 1UL } },
Signature · 1 line
2.43.0