current -mm tree includes node-hotplug codes.
But by following reason , ia64's node-hotplug doesn't work well now.
Following patch will fix it. I'd like to post this patch against next -mm.
Feedbacks are welcome.
1. empty-node-fix : avoid creating empty node
SRAT's enable bit just shows 'you can read this entry'. But the kernel know
this and checks each entries are vaild or not later.
But pxm_bit/node_online_mask is not treated as they should be.
The kernel creates empty node, which has no cpu, no memory.
Becasue of the empty node, node-hot-added will not create new NODE_DATA at
hotadd event. It's already created at boot time as empty node.
I'm now thinking of allocate NODE_DATA on local (hot-added) node. So,
avoiding to allocate empty NODE_DATA (allocated on off-node) is necessary.
My concern is whether there is a nice way to detect I/O only node at boot
time or not. (if we need it) If someone shows it, I'll add it to my patch.
2. cpu-to-node fix: fix cpu-to-node mapping at cpu hotplug
cpu hotplug on NUMA has to map cpu to its node. From its comment in the code,
it expects the container hotplug will map pxm to correct node.
But the container hotplug itself doesn't it now and acpi_map_pxm_to_node()
is introduced.
We also need to update node_to_cpu_mask[] and cpu_to_node_map[].
BTW, our team's node-hotplug considers (cpu + memory) hotplug by ACPI's container.
Does anyone has plan of cpu-only-node-hotplug or I/O-only-node-hotplug ?
If someone has, I'll develop memory-less-node hotplug, which just allocates
NODE_DATA of hot-added node.
-Kame
Remove empty node -- a node which containes no cpu, no memory (and no I/O).
When empty node is onlined, it allocates NODE_DATA(). This causes
for_each_online_node() walks through unused NODE_DATA.
Because an empty node has no memory, its NODE_DATA is allocated off-node.
Now, Node-hot-add is introduced to -mm. It can alloc NODE_DATA dynamically.
But if empty node exists, node-hotplug cannot allocate new NODE_DATA in local
memory on-node(*)
I think it's good chance to remove empty node, which came from mishandling of
pxm in SRAT.
TBD: I/O only node detections scheme should be fixed. Does anyone have a
suggestion ?
(*) Allocating NODE_DATA in local memory at node-hotplug is on my TBD list.
not posted yet.
Signed-Off-By: KAMEZAWA Hiroyuki <[email protected]>
Index: linux-2.6.17-rc4-mm3/arch/ia64/kernel/setup.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/kernel/setup.c 2006-05-25 18:48:15.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/kernel/setup.c 2006-05-25 18:50:20.000000000 +0900
@@ -418,7 +418,7 @@
if (early_console_setup(*cmdline_p) == 0)
mark_bsp_online();
-
+ reserve_memory();
#ifdef CONFIG_ACPI
/* Initialize the ACPI boot-time table parser */
acpi_table_init();
Index: linux-2.6.17-rc4-mm3/arch/ia64/mm/contig.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/mm/contig.c 2006-05-25 18:48:15.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/mm/contig.c 2006-05-25 18:49:24.000000000 +0900
@@ -146,8 +146,6 @@
{
unsigned long bootmap_size;
- reserve_memory();
-
/* first find highest page frame number */
max_pfn = 0;
efi_memmap_walk(find_max_pfn, &max_pfn);
Index: linux-2.6.17-rc4-mm3/arch/ia64/mm/discontig.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/mm/discontig.c 2006-05-25 18:48:15.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/mm/discontig.c 2006-05-25 18:49:40.000000000 +0900
@@ -443,8 +443,6 @@
{
int node;
- reserve_memory();
-
if (num_online_nodes() == 0) {
printk(KERN_ERR "node info missing!\n");
node_set_online(0);
Index: linux-2.6.17-rc4-mm3/arch/ia64/kernel/acpi.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/kernel/acpi.c 2006-05-25 18:48:15.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/kernel/acpi.c 2006-05-26 16:38:35.000000000 +0900
@@ -515,6 +515,43 @@
num_node_memblks++;
}
+/* online node if node has valid memory */
+static
+int find_valid_memory_range(unsigned long start, unsigned long end, void *arg)
+{
+ int i;
+ struct node_memblk_s *p;
+ start = __pa(start);
+ end = __pa(end);
+ for (i = 0; i < num_node_memblks; ++i) {
+ p = &node_memblk[i];
+ if (end < p->start_paddr)
+ continue;
+ if (p->start_paddr + p->size <= start)
+ continue;
+ node_set_online(p->nid);
+ }
+ return 0;
+}
+
+static void
+acpi_online_node_fixup(void)
+{
+ int i, cpu;
+ /* online node if a node has available cpus */
+ for (i = 0; i < srat_num_cpus; ++i)
+ for (cpu = 0; cpu < available_cpus; ++cpu)
+ if (smp_boot_data.cpu_phys_id[cpu] ==
+ node_cpuid[i].phys_id) {
+ node_set_online(node_cpuid[i].nid);
+ break;
+ }
+ /* memory */
+ efi_memmap_walk(find_valid_memory_range, NULL);
+
+ /* TBD: check I/O devices which have valid nid. and online it*/
+}
+
void __init acpi_numa_arch_fixup(void)
{
int i, j, node_from, node_to;
@@ -526,22 +563,28 @@
return;
}
- /*
- * MCD - This can probably be dropped now. No need for pxm ID to node ID
- * mapping with sparse node numbering iff MAX_PXM_DOMAINS <= MAX_NUMNODES.
- */
nodes_clear(node_online_map);
+ /* MAP pxm to nid */
for (i = 0; i < MAX_PXM_DOMAINS; i++) {
if (pxm_bit_test(i)) {
- int nid = acpi_map_pxm_to_node(i);
- node_set_online(nid);
+ /* this makes pxm <-> nid mapping */
+ acpi_map_pxm_to_node(i);
}
}
+ /* convert pxm information to nid information */
- /* set logical node id in memory chunk structure */
for (i = 0; i < num_node_memblks; i++)
node_memblk[i].nid = pxm_to_node(node_memblk[i].nid);
+ for (i = 0; i < srat_num_cpus; i++)
+ node_cpuid[i].nid = pxm_to_node(node_cpuid[i].nid);
+
+ /*
+ * confirm node is online or not.
+ * onlined node will have their own NODE_DATA
+ */
+ acpi_online_node_fixup();
+
/* assign memory bank numbers for each chunk on each node */
for_each_online_node(i) {
int bank;
@@ -552,9 +595,6 @@
node_memblk[j].bank = bank++;
}
- /* set logical node id in cpu structure */
- for (i = 0; i < srat_num_cpus; i++)
- node_cpuid[i].nid = pxm_to_node(node_cpuid[i].nid);
printk(KERN_INFO "Number of logical nodes in system = %d\n",
num_online_nodes());
At node hotplug, cpu can be added before memory depends on evaluation order of
firmware(ACPI) information.
Current ia64's cpu hotplug make an assumption at binding cpu to node.
/*
* Assuming that the container driver would have set the proximity
* domain and would have initialized pxm_to_node(pxm_id) && pxm_flag
*/
If nid is invalid here, cpu is bound to node 0.
So, all cpus on the new node goes to node 0 if cpu is evaluated before memory.
We have node hotplug in -mm now. The container doesn't fixes pxm<->nid
conversion but acpi_map_pxm_to_nid() does it. cpu hotplug should call
acpi_map_pxm_to_nid() to map cpu to new nid. This patch makes cpu hotplug
to call acpi_map_pxm_to_nid().
This fix will map cpus to the correct node.
As a side effect, this shows another problem. node_to_cpu_mask[] should be
updated correctly. This patch also fixes it.
Signed-Off-By: KAMEZAWA Hiroyuki <[email protected]>
arch/ia64/kernel/acpi.c | 10 +++++-----
arch/ia64/kernel/numa.c | 15 ++++++++++++---
include/asm-ia64/topology.h | 1 +
3 files changed, 18 insertions(+), 8 deletions(-)
Index: linux-2.6.17-rc4-mm3/arch/ia64/kernel/numa.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/kernel/numa.c 2006-05-26 16:37:50.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/kernel/numa.c 2006-05-26 17:08:14.000000000 +0900
@@ -30,6 +30,17 @@
cpumask_t node_to_cpu_mask[MAX_NUMNODES] __cacheline_aligned;
+/* called by cpu hotplug. */
+void __cpuinit arch_update_cpu_to_node(int cpu, int newnode)
+{
+ int oldnode = cpu_to_node(cpu);
+ cpu_to_node_map[cpu] = (newnode >= 0)? newnode : 0;
+ cpu_clear(cpu, node_to_cpu_mask[oldnode]);
+ if (newnode >= 0)
+ cpu_set(cpu, node_to_cpu_mask[newnode]);
+}
+
+
/**
* build_cpu_to_node_map - setup cpu to node and node to cpumask arrays
*
@@ -50,8 +61,6 @@
node = node_cpuid[i].nid;
break;
}
- cpu_to_node_map[cpu] = (node >= 0) ? node : 0;
- if (node >= 0)
- cpu_set(cpu, node_to_cpu_mask[node]);
+ arch_update_cpu_to_node(cpu, node);
}
}
Index: linux-2.6.17-rc4-mm3/arch/ia64/kernel/acpi.c
===================================================================
--- linux-2.6.17-rc4-mm3.orig/arch/ia64/kernel/acpi.c 2006-05-26 16:38:35.000000000 +0900
+++ linux-2.6.17-rc4-mm3/arch/ia64/kernel/acpi.c 2006-05-26 17:05:35.000000000 +0900
@@ -812,16 +812,16 @@
{
#ifdef CONFIG_ACPI_NUMA
int pxm_id;
+ int nid;
pxm_id = acpi_get_pxm(handle);
+ nid = acpi_map_pxm_to_node(pxm_id);
- /*
- * Assuming that the container driver would have set the proximity
- * domain and would have initialized pxm_to_node(pxm_id) && pxm_flag
- */
- node_cpuid[cpu].nid = (pxm_id < 0) ? 0 : pxm_to_node(pxm_id);
+ node_cpuid[cpu].nid = nid;
node_cpuid[cpu].phys_id = physid;
+
+ arch_update_cpu_to_node(cpu, nid);
#endif
return (0);
}
Index: linux-2.6.17-rc4-mm3/include/asm-ia64/topology.h
===================================================================
--- linux-2.6.17-rc4-mm3.orig/include/asm-ia64/topology.h 2006-05-26 16:37:50.000000000 +0900
+++ linux-2.6.17-rc4-mm3/include/asm-ia64/topology.h 2006-05-26 17:05:35.000000000 +0900
@@ -54,6 +54,7 @@
*/
#define pcibus_to_node(bus) PCI_CONTROLLER(bus)->node
+void arch_update_cpu_to_node(int cpu, int nid);
void build_cpu_to_node_map(void);
#define SD_CPU_INIT (struct sched_domain) { \
On Fri, 26 May 2006 17:56:22 +0900
KAMEZAWA Hiroyuki <[email protected]> wrote:
> current -mm tree includes node-hotplug codes.
>
> But by following reason , ia64's node-hotplug doesn't work well now.
>
> Following patch will fix it. I'd like to post this patch against next -mm.
> Feedbacks are welcome.
>
> 1. empty-node-fix : avoid creating empty node
> SRAT's enable bit just shows 'you can read this entry'. But the kernel know
^^^^
And
-Kame
> 1. empty-node-fix : avoid creating empty node
> SRAT's enable bit just shows 'you can read this entry'. But the kernel know
> this and checks each entries are vaild or not later.
>
> But pxm_bit/node_online_mask is not treated as they should be.
> The kernel creates empty node, which has no cpu, no memory.
I would like to mention about background of this more.
I thought if enable bit of each SRAT entry is on, then its entry's
object is usable for OS.
However, SRAT specification says only
"If clear, the OSPM ignores the contents of the Processor Local
APIC/SAPIC (or Memory) Affinity Structure."
So, our firmware team (or Micro $oft) interprets this
"If enable bit is on, then this entry is just readable by OS.
The object of its entry MIGHT NOT EXIST. This entry can be used for
reserve resource for memory/cpu which can be hot-add later."
They implemented it.
I really really hate this. :-(
But, indeed, ACPI spec. says just IGNORE if clear. They are correct.
Current linux code checks memory and cpu existence by other ways.
But, PXM remains even if they don't exist. The first patch is to remove it.
Bye.
--
Yasunori Goto