Skip to contentSkip to conversation

2/5 in support of hot-add memory x86_64 create arch_find_node

PATCH In review keith mannthey 4 participants 5 messages Jul 29, 2006 – Jul 31, 2006

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 2/5 in support of hot-add memory x86_64 create arch_find_node

With the advent of the new ACPI hot-plug memory driver and mechanism
is needed to deal with ACPI add memory events that do not contain the pxm (node) information. I do not believe that the add-event is required to contain this information so I create a arch_find_node generic layer used in the generic add_memory function.

If add_memory is called with node < 0 arch_find_node is invoked to
fine the correct node to add the memory. This created the generic construct of arch_find_node.

This was built against 2.6.18-rc2.

include/linux/memory_hotplug.h +5 -1
diff -urN orig/include/linux/memory_hotplug.h work/include/linux/memory_hotplug.h
@@ -132,7 +132,11 @@
 }
 #endif /* CONFIG_NUMA */
 #endif /* CONFIG_HAVE_ARCH_NODEDATA_EXTENSION */
-
+#ifdef CONFIG_ARCH_FIND_NODE
+	extern int arch_find_node(unsigned long, unsigned long);
+#else
+	static inline int arch_find_node(unsigned long a,  unsigned long b) {return 0;}
+#endif
 #else /* ! CONFIG_MEMORY_HOTPLUG */
 /*
  * Stub functions for when hotplug is off
mm/memory_hotplug.c +7 -2
diff -urN orig/mm/memory_hotplug.c work/mm/memory_hotplug.c
@@ -233,12 +233,17 @@
 
 
 
-int add_memory(int nid, u64 start, u64 size)
+int add_memory(int node, u64 start, u64 size)
 {
 	pg_data_t *pgdat = NULL;
 	int new_pgdat = 0;
-	int ret;
+	int ret,nid;
 
+	if (node < 0) 
+		nid = arch_find_node(start,size);
+	else
+		nid = node;
+	
 	if (!node_online(nid)) {
 		pgdat = hotadd_new_pgdat(nid, start);
 		if (!pgdat)
#1 lore raw
Quoted 9 lines · keith mannthey

With the advent of the new ACPI hot-plug memory driver and mechanism
is needed to deal with ACPI add memory events that do not contain the pxm (node) information. I do not believe that the add-event is required to contain this information so I create a arch_find_node generic layer used in the generic add_memory function.

If add_memory is called with node < 0 arch_find_node is invoked to
fine the correct node to add the memory. This created the generic construct of arch_find_node.

It would be cleaner to always call add_memory from architecture specific code instead of such ugly hooks

-Andi

#2 lore raw
Quoted 7 lines · Andi Kleen
Quoted 6 lines · keith mannthey

With the advent of the new ACPI hot-plug memory driver and mechanism
is needed to deal with ACPI add memory events that do not contain the pxm (node) information. I do not believe that the add-event is required to contain this information so I create a arch_find_node generic layer used in the generic add_memory function.

Quoted 6 lines
Quoted 3 lines

If add_memory is called with node < 0 arch_find_node is invoked to
fine the correct node to add the memory. This created the generic construct of arch_find_node.

It would be cleaner to always call add_memory from architecture specific code instead of such ugly hooks

Since 2.6.18 it goes

acpi (generic calling add_memory)
add_memory (generic)
arch_add_memory (x86_64)
__add_pages (generic)

The generic add_memory call is good as it does the dynamic pgdat allocation for new nodes. The generic add_memory call need the correct node infomation.

Do you think we should do

acpi (generic calling add_memory)
arch_pre_add_memory? (x86_64)
add_memory (generic)
arch_add_memory (x86_64)
__add_pages (generic)

?
It would just create a function hook.

Thanks,
Keith

#3 lore raw
In reply to Andi Kleen
Quoted 7 lines · On Sat, 29 Jul 2006 18:25:15 +0200 Andi Kleen <[email protected]> wrote:
Quoted 6 lines · keith mannthey

With the advent of the new ACPI hot-plug memory driver and mechanism
is needed to deal with ACPI add memory events that do not contain the pxm (node) information. I do not believe that the add-event is required to contain this information so I create a arch_find_node generic layer used in the generic add_memory function.

Quoted 7 lines
Quoted 3 lines

If add_memory is called with node < 0 arch_find_node is invoked to
fine the correct node to add the memory. This created the generic construct of arch_find_node.

It would be cleaner to always call add_memory from architecture specific code instead of such ugly hooks

Hi, Keith

I don't like insert such a check (nid < 0) in add_memory(), either. Could you add it before calling add_memory() ?
(for example, find_pxm parh in acpi's add memory code.)

-Kame

#4 lore raw
Quoted 14 lines
Quoted 11 lines · On Sat, 29 Jul 2006 18:25:15 +0200 Andi Kleen <[email protected]> wrote:
Quoted 9 lines · keith mannthey

With the advent of the new ACPI hot-plug memory driver and mechanism
is needed to deal with ACPI add memory events that do not contain the pxm (node) information. I do not believe that the add-event is required to contain this information so I create a arch_find_node generic layer used in the generic add_memory function.

If add_memory is called with node < 0 arch_find_node is invoked to
fine the correct node to add the memory. This created the generic construct of arch_find_node.

Quoted 8 lines
Quoted 3 lines

It would be cleaner to always call add_memory from architecture specific code instead of such ugly hooks

Hi, Keith

I don't like insert such a check (nid < 0) in add_memory(), either. Could you add it before calling add_memory() ?
(for example, find_pxm parh in acpi's add memory code.)

Agree.
I think that arch_find_node() should be called in acpi_get_node().

For the time being, I would like to post folloing patch for 2.6.18 to fix returning -1 if _PXM is not defined.
Because, not only your case, but also there might be a case that hot add code will be executed with NUMA kernel on NON-NUMA box.

Then, arch_find_node() should be add before this line.

Thanks.

-----

 drivers/acpi/numa.c |    4 ++++
 1 files changed, 4 insertions(+)
drivers/acpi/numa.c +4 -0
Index: pxm_find1/drivers/acpi/numa.c
===================================================================
@@ -263,6 +263,10 @@ int acpi_get_node(acpi_handle *handle)
 	if (pxm >= 0)
 		node = acpi_map_pxm_to_node(pxm);
 
+	/* _PXM might not be defined by firmware. */
+	if (node < 0)
+		node = 0;
+
 	return node;
 }
 EXPORT_SYMBOL(acpi_get_node);
Signature · 1 line
Yasunori Goto