On Tue, 6 Mar 2007, Mark Gross wrote:
> Let me give your idea a spin and get back to you.
>
Something like the following might be a little better.
[ You might consider adding this as a configuration option such
as CONFIG_NODE_POWER_MANAGEMENT so that power_managed_node(nid)
always returns 0 when this isn't defined in arch/x86_64/Kconfig. ]
David
---
arch/x86_64/mm/numa.c | 59 ++++++++++++++++++++++++++++++++++++++--
arch/x86_64/mm/srat.c | 7 +++-
include/asm-x86_64/numa.h | 2 +
include/asm-x86_64/topology.h | 7 +++++
include/linux/topology.h | 18 ++++++++++++
mm/bootmem.c | 34 +++++++++++++----------
mm/mempolicy.c | 9 +++++-
mm/page_alloc.c | 19 ++++++++++++-
mm/slab.c | 7 +++++
mm/sparse.c | 8 +++++
10 files changed, 146 insertions(+), 24 deletions(-)
diff --git a/arch/x86_64/mm/numa.c b/arch/x86_64/mm/numa.c
--- a/arch/x86_64/mm/numa.c
+++ b/arch/x86_64/mm/numa.c
@@ -11,6 +11,7 @@
#include <linux/ctype.h>
#include <linux/module.h>
#include <linux/nodemask.h>
+#include <linux/acpi.h>
#include <asm/e820.h>
#include <asm/proto.h>
@@ -159,8 +160,23 @@ static void * __init
early_node_mem(int nodeid, unsigned long start, unsigned long end,
unsigned long size)
{
- unsigned long mem = find_e820_area(start, end, size);
+ unsigned long mem;
void *ptr;
+
+ /*
+ * If this is a power-managed node, we need to allocate this memory
+ * elsewhere so we remap it, if possible.
+ */
+ if (power_managed_node(nodeid)) {
+ int new_node = node_remap(nodeid, nodes_parsed, non_pm_nodes);
+ if (nodeid != new_node) {
+ start = NODE_DATA(new_node)->node_start_pfn;
+ end = start + NODE_DATA(new_node)->node_spanned_pages;
+ nodeid = new_node;
+ }
+ }
+ mem = find_e820_area(start, end, size);
+
if (mem != -1L)
return __va(mem);
ptr = __alloc_bootmem_nopanic(size,
@@ -180,6 +196,7 @@ void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long en
unsigned long nodedata_phys;
void *bootmap;
const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
+ int reserve_nid = nodeid;
start = round_up(start, ZONE_ALIGN);
@@ -218,6 +235,13 @@ void __init setup_node_bootmem(int nodeid, unsigned long start, unsigned long en
free_bootmem_with_active_regions(nodeid, end);
+ /*
+ * Make sure we reserve bootmem on a node that is not under power
+ * management.
+ */
+ if (power_managed_node(nodeid))
+ reserve_nid = node_remap(nodeid, nodes_parsed, non_pm_nodes);
+
reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size);
reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, bootmap_pages<<PAGE_SHIFT);
#ifdef CONFIG_ACPI_NUMA
@@ -242,6 +266,9 @@ void __init setup_node_zones(int nodeid)
memmapsize = sizeof(struct page) * (end_pfn-start_pfn);
limit = end_pfn << PAGE_SHIFT;
#ifdef CONFIG_FLAT_NODE_MEM_MAP
+ if (power_managed_node(nodeid))
+ nodeid = node_remap(nodeid, nodes_parsed, non_pm_nodes);
+
NODE_DATA(nodeid)->node_mem_map =
__alloc_bootmem_core(NODE_DATA(nodeid)->bdata,
memmapsize, SMP_CACHE_BYTES,
@@ -255,7 +282,7 @@ void __init setup_node_zones(int nodeid)
void __init numa_init_array(void)
{
- int rr, i;
+ int rr, i, nodeid;
/* There are unfortunately some poorly designed mainboards around
that only connect memory to a single CPU. This breaks the 1:1 cpu->node
mapping. To avoid this fill in the mapping for all possible
@@ -265,7 +292,11 @@ void __init numa_init_array(void)
for (i = 0; i < NR_CPUS; i++) {
if (cpu_to_node[i] != NUMA_NO_NODE)
continue;
- numa_set_node(i, rr);
+ if (power_managed_node(rr))
+ nodeid = node_remap(rr, nodes_parsed, non_pm_nodes);
+ else
+ nodeid = rr;
+ numa_set_node(i, nodeid);
rr = next_node(rr, node_online_map);
if (rr == MAX_NUMNODES)
rr = first_node(node_online_map);
@@ -681,3 +712,25 @@ int pfn_valid(unsigned long pfn)
}
EXPORT_SYMBOL(pfn_valid);
#endif
+
+int __power_managed_node(int nid)
+{
+ return !node_isset(node_to_pxm(nid), non_pm_nodes);
+}
+
+int __power_managed_memory_present(void)
+{
+ return !nodes_full(non_pm_nodes);
+}
+
+int __find_closest_non_pm_node(int nid)
+{
+ int node;
+ nodemask_t new_nodes;
+
+ nodes_and(new_nodes, non_pm_nodes, node_online_map);
+ node = next_node(nid, non_pm_nodes);
+ if (node == MAX_NUMNODES)
+ node = first_node(non_pm_nodes);
+ return node;
+}
diff --git a/arch/x86_64/mm/srat.c b/arch/x86_64/mm/srat.c
--- a/arch/x86_64/mm/srat.c
+++ b/arch/x86_64/mm/srat.c
@@ -25,10 +25,11 @@ int acpi_numa __initdata;
static struct acpi_table_slit *acpi_slit;
-static nodemask_t nodes_parsed __initdata;
+nodemask_t nodes_parsed __initdata;
static struct bootnode nodes_add[MAX_NUMNODES];
static int found_add_area __initdata;
int hotadd_percent __initdata = 0;
+nodemask_t non_pm_nodes __read_mostly = NODE_MASK_ALL;
/* Too small nodes confuse the VM badly. Usually they result
from BIOS bugs. */
@@ -298,8 +299,10 @@ acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
return;
start = ma->base_address;
end = start + ma->length;
- pxm = ma->proximity_domain;
+ pxm = ma->proximity_domain & 0x0000ffff;
node = setup_node(pxm);
+ if (ma->proximity_domain & (1 << 31))
+ node_clear(node, non_pm_nodes);
if (node < 0) {
printk(KERN_ERR "SRAT: Too many proximity domains.\n");
bad_srat();
diff --git a/include/asm-x86_64/numa.h b/include/asm-x86_64/numa.h
--- a/include/asm-x86_64/numa.h
+++ b/include/asm-x86_64/numa.h
@@ -18,6 +18,8 @@ extern int numa_off;
extern void numa_set_node(int cpu, int node);
extern void srat_reserve_add_area(int nodeid);
extern int hotadd_percent;
+extern nodemask_t nodes_parsed __initdata;
+extern nodemask_t non_pm_nodes;
extern unsigned char apicid_to_node[256];
#ifdef CONFIG_NUMA
diff --git a/include/asm-x86_64/topology.h b/include/asm-x86_64/topology.h
--- a/include/asm-x86_64/topology.h
+++ b/include/asm-x86_64/topology.h
@@ -18,6 +18,13 @@ extern int __node_distance(int, int);
/* #else fallback version */
#endif
+extern int __power_managed_memory_present(void);
+extern int __power_managed_node(int);
+extern int __find_closest_non_pm_node(int);
+#define power_managed_memory_present() __power_managed_memory_present()
+#define power_managed_node(nid) __power_managed_node(nid)
+#define find_closest_non_pm_node(nid) __find_closest_non_pm_node(nid)
+
#define cpu_to_node(cpu) (cpu_to_node[cpu])
#define parent_node(node) (node)
#define node_to_first_cpu(node) (first_cpu(node_to_cpumask[node]))
diff --git a/include/linux/topology.h b/include/linux/topology.h
--- a/include/linux/topology.h
+++ b/include/linux/topology.h
@@ -67,6 +67,24 @@
#ifndef PENALTY_FOR_NODE_WITH_CPUS
#define PENALTY_FOR_NODE_WITH_CPUS (1)
#endif
+#ifndef power_managed_memory_present
+static inline int power_managed_memory_present(void)
+{
+ return 0;
+}
+#endif
+#ifndef power_managed_node
+static inline int power_managed_node(int nid)
+{
+ return 0;
+}
+#endif
+#ifndef find_closest_non_pm_node
+static inline int find_closest_non_pm_node(int nid)
+{
+ return nid;
+}
+#endif
/*
* Below are the 3 major initializers used in building sched_domains:
diff --git a/mm/bootmem.c b/mm/bootmem.c
--- a/mm/bootmem.c
+++ b/mm/bootmem.c
@@ -417,14 +417,16 @@ unsigned long __init free_all_bootmem(void)
void * __init __alloc_bootmem_nopanic(unsigned long size, unsigned long align,
unsigned long goal)
{
- bootmem_data_t *bdata;
void *ptr;
-
- list_for_each_entry(bdata, &bdata_list, list) {
- ptr = __alloc_bootmem_core(bdata, size, align, goal, 0);
- if (ptr)
- return ptr;
- }
+ int i;
+
+ for_each_online_node(i)
+ if (!power_managed_node(i)) {
+ ptr = __alloc_bootmem_core(NODE_DATA(i)->bdata,
+ size, align, goal, 0);
+ if (ptr)
+ return ptr;
+ }
return NULL;
}
@@ -463,15 +465,17 @@ void * __init __alloc_bootmem_node(pg_data_t *pgdat, unsigned long size,
void * __init __alloc_bootmem_low(unsigned long size, unsigned long align,
unsigned long goal)
{
- bootmem_data_t *bdata;
void *ptr;
-
- list_for_each_entry(bdata, &bdata_list, list) {
- ptr = __alloc_bootmem_core(bdata, size, align, goal,
- ARCH_LOW_ADDRESS_LIMIT);
- if (ptr)
- return ptr;
- }
+ int i;
+
+ for_each_online_node(i)
+ if (!power_managed_node(i)) {
+ ptr = __alloc_bootmem_core(NODE_DATA(i)->bdata,
+ size, align, goal,
+ ARCH_LOW_ADDRESS_LIMIT);
+ if (ptr)
+ return ptr;
+ }
/*
* Whoops, we cannot satisfy the allocation request.
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -1609,8 +1609,13 @@ void __init numa_policy_init(void)
/* Set interleaving policy for system init. This way not all
the data structures allocated at system boot end up in node zero. */
- if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
- printk("numa_policy_init: interleaving failed\n");
+ if (power_managed_memory_present()) {
+ if (do_set_mempolicy(MPOL_DEFAULT, &node_online_map))
+ printk("numa_policy_init: default failed\n");
+ } else {
+ if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
+ printk("numa_policy_init: interleaving failed\n");
+ }
}
/* Reset policy of current process to default */
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2599,6 +2599,14 @@ int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
* sizeof(wait_queue_head_t);
if (system_state == SYSTEM_BOOTING) {
+ struct pglist_data *alloc_pgdat = pgdat;
+
+ if (power_managed_node(pgdat->node_id)) {
+ int nid;
+ nid = find_closest_non_pm_node(pgdat->node_id);
+ alloc_pgdat = NODE_DATA(nid);
+ }
+
zone->wait_table = (wait_queue_head_t *)
alloc_bootmem_node(pgdat, alloc_size);
} else {
@@ -3203,6 +3211,7 @@ static void __init alloc_node_mem_map(struct pglist_data *pgdat)
if (!pgdat->node_mem_map) {
unsigned long size, start, end;
struct page *map;
+ struct pglist_data *alloc_pgdat = pgdat;
/*
* The zone's endpoints aren't required to be MAX_ORDER
@@ -3214,8 +3223,14 @@ static void __init alloc_node_mem_map(struct pglist_data *pgdat)
end = ALIGN(end, MAX_ORDER_NR_PAGES);
size = (end - start) * sizeof(struct page);
map = alloc_remap(pgdat->node_id, size);
- if (!map)
- map = alloc_bootmem_node(pgdat, size);
+ if (!map) {
+ if (power_managed_node(pgdat->node_id)) {
+ int nid;
+ nid = find_closest_non_pm_node(pgdat->node_id);
+ alloc_pgdat = NODE_DATA(nid);
+ }
+ map = alloc_bootmem_node(alloc_pgdat, size);
+ }
pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
printk(KERN_DEBUG
"Node %d memmap at 0x%p size %lu first pfn 0x%p\n",
diff --git a/mm/slab.c b/mm/slab.c
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -3399,6 +3399,10 @@ __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
if (unlikely(nodeid == -1))
nodeid = numa_node_id();
+ /* We cannot allocate objects to nodes subject to power management */
+ if (power_managed_node(nodeid))
+ nodeid = find_closest_non_pm_node(nodeid);
+
if (unlikely(!cachep->nodelists[nodeid])) {
/* Node not bootstrapped yet */
ptr = fallback_alloc(cachep, flags);
@@ -3672,6 +3676,9 @@ out:
#ifdef CONFIG_NUMA
void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
{
+ /* We cannot allocate objects to nodes subject to power management */
+ if (power_managed_node(nodeid))
+ nodeid = find_closest_non_pm_node(nodeid);
return __cache_alloc_node(cachep, flags, nodeid,
__builtin_return_address(0));
}
diff --git a/mm/sparse.c b/mm/sparse.c
--- a/mm/sparse.c
+++ b/mm/sparse.c
@@ -50,6 +50,10 @@ static struct mem_section *sparse_index_alloc(int nid)
unsigned long array_size = SECTIONS_PER_ROOT *
sizeof(struct mem_section);
+ /* The node we allocate on must not be subject to power management */
+ if (power_managed_node(nid))
+ nid = find_closest_non_pm_node(nid);
+
if (slab_is_available())
section = kmalloc_node(array_size, GFP_KERNEL, nid);
else
@@ -215,6 +219,10 @@ static struct page *sparse_early_mem_map_alloc(unsigned long pnum)
struct mem_section *ms = __nr_to_section(pnum);
int nid = sparse_early_nid(ms);
+ /* The node we allocate on must not be subject to power management */
+ if (power_managed_node(nid))
+ nid = find_closest_non_pm_node(nid);
+
map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
if (map)
return map;
