[PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[travis]

Fixup change NR_CPUS patchset by rebasing on 2.6.24-rc8-mm1 (from 2.6.24-rc6-mm1)
and adding last minute changes suggested by reviews.

Based on 2.6.24-rc8-mm1

Signed-off-by: Mike Travis <travis@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
---

-- 

[PATCH 1/3] x86: Change size of node ids from u8 to u16 fixup

[travis]

[-- Attachment #1: big_nodeids-fixup --]
[-- Type: text/plain, Size: 1898 bytes --]

Change the size of node ids from 8 bits to 16 bits to
accomodate more than 256 nodes.

Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Mike Travis <travis@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
---
Fixup:

Size of memnode.embedded_map needs to be changed to
accomodate 16-bit node ids as suggested by Eric.

V2->V3:
    - changed memnode.embedded_map from [64-16] to [64-8]
      (and size comment to 128 bytes)

V1->V2:
    - changed pxm_to_node_map to u16
    - changed memnode map entries to u16
---
 arch/x86/mm/numa_64.c       |    2 +-
 drivers/acpi/numa.c         |    2 +-
 include/asm-x86/mmzone_64.h |    6 +++---
 3 files changed, 5 insertions(+), 5 deletions(-)

--- a/arch/x86/mm/numa_64.c
+++ b/arch/x86/mm/numa_64.c
@@ -88,7 +88,7 @@ static int __init allocate_cachealigned_
 	unsigned long pad, pad_addr;
 
 	memnodemap = memnode.embedded_map;
-	if (memnodemapsize <= 48)
+	if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
 		return 0;
 
 	pad = L1_CACHE_BYTES - 1;
--- a/drivers/acpi/numa.c
+++ b/drivers/acpi/numa.c
@@ -38,7 +38,7 @@ ACPI_MODULE_NAME("numa");
 static nodemask_t nodes_found_map = NODE_MASK_NONE;
 
 /* maps to convert between proximity domain and logical node ID */
-static int pxm_to_node_map[MAX_PXM_DOMAINS]
+static u16 pxm_to_node_map[MAX_PXM_DOMAINS]
 				= { [0 ... MAX_PXM_DOMAINS - 1] = NID_INVAL };
 static int node_to_pxm_map[MAX_NUMNODES]
 				= { [0 ... MAX_NUMNODES - 1] = PXM_INVAL };
--- a/include/asm-x86/mmzone_64.h
+++ b/include/asm-x86/mmzone_64.h
@@ -15,9 +15,9 @@
 struct memnode {
 	int shift;
 	unsigned int mapsize;
-	u8 *map;
-	u8 embedded_map[64-16];
-} ____cacheline_aligned; /* total size = 64 bytes */
+	u16 *map;
+	u16 embedded_map[64-8];
+} ____cacheline_aligned; /* total size = 128 bytes */
 extern struct memnode memnode;
 #define memnode_shift memnode.shift
 #define memnodemap memnode.map

-- 

[PATCH 2/3] x86: Change NR_CPUS arrays in numa_64 fixup

[travis]

[-- Attachment #1: NR_CPUS-arrays-in-numa_64-fixup --]
[-- Type: text/plain, Size: 7574 bytes --]

Change the following static arrays sized by NR_CPUS to
per_cpu data variables:

	char cpu_to_node_map[NR_CPUS];


Signed-off-by: Mike Travis <travis@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
---
fixup:

    Split cpu_to_node function into "early" and "late" versions
    so that x86_cpu_to_node_map_{init,early_ptr} are not EXPORT'ed.
    This also involves setting up the percpu maps as early as possible.

V2->V3:
    - add early_cpu_to_node function to keep cpu_to_node efficient
    - move and rename smp_set_apicids() to setup_percpu_maps()
    - call setup_percpu_maps() as early as possible

V1->V2:
    - Removed extraneous casts
    - Fix !NUMA builds with '#ifdef CONFIG_NUMA"
---
 arch/x86/kernel/setup64.c    |   37 +++++++++++++++++++++++++++++++++++--
 arch/x86/kernel/smpboot_64.c |   38 ++------------------------------------
 arch/x86/mm/numa_64.c        |    2 --
 arch/x86/mm/srat_64.c        |    5 +++--
 include/asm-x86/numa_64.h    |    7 -------
 include/asm-x86/topology.h   |   10 +++++++++-
 mm/page_alloc.c              |    2 +-
 7 files changed, 50 insertions(+), 51 deletions(-)

--- a/arch/x86/kernel/setup64.c
+++ b/arch/x86/kernel/setup64.c
@@ -84,6 +84,36 @@ static int __init nonx32_setup(char *str
 __setup("noexec32=", nonx32_setup);
 
 /*
+ * Copy data used in early init routines from the initial arrays to the
+ * per cpu data areas.  These arrays then become expendable and the *_ptrs
+ * are zeroed indicating that the static arrays are gone.
+ */
+void __init setup_percpu_maps(void)
+{
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		if (per_cpu_offset(cpu)) {
+			per_cpu(x86_cpu_to_apicid, cpu) =
+						x86_cpu_to_apicid_init[cpu];
+#ifdef CONFIG_NUMA
+			per_cpu(x86_cpu_to_node_map, cpu) =
+						x86_cpu_to_node_map_init[cpu];
+#endif
+		}
+		else
+			printk(KERN_NOTICE "per_cpu_offset zero for cpu %d\n",
+									cpu);
+	}
+
+	/* indicate the early static arrays are gone */
+	x86_cpu_to_apicid_early_ptr = NULL;
+#ifdef CONFIG_NUMA
+	x86_cpu_to_node_map_early_ptr = NULL;
+#endif
+}
+
+/*
  * Great future plan:
  * Declare PDA itself and support (irqstack,tss,pgd) as per cpu data.
  * Always point %gs to its beginning
@@ -104,18 +134,21 @@ void __init setup_per_cpu_areas(void)
 	for_each_cpu_mask (i, cpu_possible_map) {
 		char *ptr;
 
-		if (!NODE_DATA(cpu_to_node(i))) {
+		if (!NODE_DATA(early_cpu_to_node(i))) {
 			printk("cpu with no node %d, num_online_nodes %d\n",
 			       i, num_online_nodes());
 			ptr = alloc_bootmem_pages(size);
 		} else { 
-			ptr = alloc_bootmem_pages_node(NODE_DATA(cpu_to_node(i)), size);
+			ptr = alloc_bootmem_pages_node(NODE_DATA(early_cpu_to_node(i)), size);
 		}
 		if (!ptr)
 			panic("Cannot allocate cpu data for CPU %d\n", i);
 		cpu_pda(i)->data_offset = ptr - __per_cpu_start;
 		memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
 	}
+
+	/* setup percpu data maps early */
+	setup_percpu_maps();
 } 
 
 void pda_init(int cpu)
--- a/arch/x86/kernel/smpboot_64.c
+++ b/arch/x86/kernel/smpboot_64.c
@@ -702,7 +702,7 @@ do_rest:
 	if (boot_error) {
 		cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
 		clear_bit(cpu, (unsigned long *)&cpu_initialized); /* was set by cpu_init() */
-		clear_node_cpumask(cpu); /* was set by numa_add_cpu */
+		clear_bit(cpu, (unsigned long *)&node_to_cpumask_map[cpu_to_node(cpu)]);
 		cpu_clear(cpu, cpu_present_map);
 		cpu_clear(cpu, cpu_possible_map);
 		per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
@@ -851,39 +851,6 @@ static int __init smp_sanity_check(unsig
 	return 0;
 }
 
-/*
- * Copy data used in early init routines from the initial arrays to the
- * per cpu data areas.  These arrays then become expendable and the
- * *_ptrs are zeroed indicating that the static arrays are gone.
- */
-void __init smp_set_apicids(void)
-{
-	int cpu;
-
-	for_each_possible_cpu(cpu) {
-		if (per_cpu_offset(cpu)) {
-			per_cpu(x86_cpu_to_apicid, cpu) =
-						x86_cpu_to_apicid_init[cpu];
-#ifdef CONFIG_NUMA
-			per_cpu(x86_cpu_to_node_map, cpu) =
-						x86_cpu_to_node_map_init[cpu];
-#endif
-			per_cpu(x86_bios_cpu_apicid, cpu) =
-						x86_bios_cpu_apicid_init[cpu];
-		}
-		else
-			printk(KERN_NOTICE "per_cpu_offset zero for cpu %d\n",
-									cpu);
-	}
-
-	/* indicate the early static arrays are gone */
-	x86_cpu_to_apicid_early_ptr = NULL;
-#ifdef CONFIG_NUMA
-	x86_cpu_to_node_map_early_ptr = NULL;
-#endif
-	x86_bios_cpu_apicid_early_ptr = NULL;
-}
-
 static void __init smp_cpu_index_default(void)
 {
 	int i;
@@ -906,7 +873,6 @@ void __init smp_prepare_cpus(unsigned in
 	smp_cpu_index_default();
 	current_cpu_data = boot_cpu_data;
 	current_thread_info()->cpu = 0;  /* needed? */
-	smp_set_apicids();
 	set_cpu_sibling_map(0);
 
 	if (smp_sanity_check(max_cpus) < 0) {
@@ -1060,7 +1026,7 @@ void remove_cpu_from_maps(void)
 	cpu_clear(cpu, cpu_callout_map);
 	cpu_clear(cpu, cpu_callin_map);
 	clear_bit(cpu, (unsigned long *)&cpu_initialized); /* was set by cpu_init() */
-	clear_node_cpumask(cpu);
+	clear_bit(cpu, (unsigned long *)&node_to_cpumask_map[cpu_to_node(cpu)]);
 }
 
 int __cpu_disable(void)
--- a/arch/x86/mm/numa_64.c
+++ b/arch/x86/mm/numa_64.c
@@ -35,8 +35,6 @@ u16 x86_cpu_to_node_map_init[NR_CPUS] = 
 	[0 ... NR_CPUS-1] = NUMA_NO_NODE
 };
 void *x86_cpu_to_node_map_early_ptr;
-EXPORT_SYMBOL(x86_cpu_to_node_map_init);
-EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr);
 DEFINE_PER_CPU(u16, x86_cpu_to_node_map) = NUMA_NO_NODE;
 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map);
 
--- a/arch/x86/mm/srat_64.c
+++ b/arch/x86/mm/srat_64.c
@@ -382,9 +382,10 @@ int __init acpi_scan_nodes(unsigned long
 			setup_node_bootmem(i, nodes[i].start, nodes[i].end);
 
 	for (i = 0; i < NR_CPUS; i++) {
-		if (cpu_to_node(i) == NUMA_NO_NODE)
+		int node = cpu_to_node(i);
+		if (node == NUMA_NO_NODE)
 			continue;
-		if (!node_isset(cpu_to_node(i), node_possible_map))
+		if (!node_isset(node, node_possible_map))
 			numa_set_node(i, NUMA_NO_NODE);
 	}
 	numa_init_array();
--- a/include/asm-x86/numa_64.h
+++ b/include/asm-x86/numa_64.h
@@ -29,15 +29,8 @@ extern void setup_node_bootmem(int nodei
 
 #ifdef CONFIG_NUMA
 extern void __init init_cpu_to_node(void);
-
-static inline void clear_node_cpumask(int cpu)
-{
-	clear_bit(cpu, (unsigned long *)&node_to_cpumask_map[cpu_to_node(cpu)]);
-}
-
 #else
 #define init_cpu_to_node() do {} while (0)
-#define clear_node_cpumask(cpu) do {} while (0)
 #endif
 
 #endif
--- a/include/asm-x86/topology.h
+++ b/include/asm-x86/topology.h
@@ -38,7 +38,7 @@ extern cpumask_t node_to_cpumask_map[];
 #define NUMA_NO_NODE	((u16)(~0))
 
 /* Returns the number of the node containing CPU 'cpu' */
-static inline int cpu_to_node(int cpu)
+static inline int early_cpu_to_node(int cpu)
 {
 	u16 *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
 
@@ -50,6 +50,14 @@ static inline int cpu_to_node(int cpu)
 		return NUMA_NO_NODE;
 }
 
+static inline int cpu_to_node(int cpu)
+{
+	if(per_cpu_offset(cpu))
+		return per_cpu(x86_cpu_to_node_map, cpu);
+	else
+		return NUMA_NO_NODE;
+}
+
 /*
  * Returns the number of the node containing Node 'node'. This
  * architecture is flat, so it is a pretty simple function!
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -1783,7 +1783,7 @@ EXPORT_SYMBOL(free_pages);
 static unsigned int nr_free_zone_pages(int offset)
 {
 	/* Just pick one node, since fallback list is circular */
-	pg_data_t *pgdat = NODE_DATA(numa_node_id());
+	pg_data_t *pgdat = NODE_DATA(cpu_to_node(raw_smp_processor_id()));
 	unsigned int sum = 0;
 
 	struct zonelist *zonelist = pgdat->node_zonelists + offset;

-- 

[PATCH 3/3] x86: Change bios_cpu_apicid to percpu data variable fixup

[travis]

[-- Attachment #1: change-bios_cpu_apicid-to-percpu-fixup --]
[-- Type: text/plain, Size: 2564 bytes --]

Change static bios_cpu_apicid array to a per_cpu data variable.
This includes using a static array used during initialization
similar to the way x86_cpu_to_apicid[] is handled.

There is one early use of bios_cpu_apicid in apic_is_clustered_box().
The other reference in cpu_present_to_apicid() is called after
smp_set_apicids() has setup the percpu version of bios_cpu_apicid.


Signed-off-by: Mike Travis <travis@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
---
V1->V2:
    - Removed extraneous casts
    - Add slight optimization to apic_is_clustered_box()
      [don't reference x86_bios_cpu_apicid_early_ptr each pass.]
---
 arch/x86/kernel/apic_64.c  |    6 +++---
 arch/x86/kernel/setup64.c  |    3 +++
 arch/x86/kernel/setup_64.c |    1 +
 3 files changed, 7 insertions(+), 3 deletions(-)

--- a/arch/x86/kernel/apic_64.c
+++ b/arch/x86/kernel/apic_64.c
@@ -1191,9 +1191,9 @@ __cpuinit int apic_is_clustered_box(void
 
 	/* Problem:  Partially populated chassis may not have CPUs in some of
 	 * the APIC clusters they have been allocated.  Only present CPUs have
-	 * x86_bios_cpu_apicid entries, thus causing zeroes in the bitmap.  Since
-	 * clusters are allocated sequentially, count zeros only if they are
-	 * bounded by ones.
+	 * x86_bios_cpu_apicid entries, thus causing zeroes in the bitmap.
+	 * Since clusters are allocated sequentially, count zeros only if
+	 * they are bounded by ones.
 	 */
 	clusters = 0;
 	zeros = 0;
--- a/arch/x86/kernel/setup64.c
+++ b/arch/x86/kernel/setup64.c
@@ -96,6 +96,8 @@ void __init setup_percpu_maps(void)
 		if (per_cpu_offset(cpu)) {
 			per_cpu(x86_cpu_to_apicid, cpu) =
 						x86_cpu_to_apicid_init[cpu];
+			per_cpu(x86_bios_cpu_apicid, cpu) =
+						x86_bios_cpu_apicid_init[cpu];
 #ifdef CONFIG_NUMA
 			per_cpu(x86_cpu_to_node_map, cpu) =
 						x86_cpu_to_node_map_init[cpu];
@@ -108,6 +110,7 @@ void __init setup_percpu_maps(void)
 
 	/* indicate the early static arrays are gone */
 	x86_cpu_to_apicid_early_ptr = NULL;
+	x86_bios_cpu_apicid_early_ptr = NULL;
 #ifdef CONFIG_NUMA
 	x86_cpu_to_node_map_early_ptr = NULL;
 #endif
--- a/arch/x86/kernel/setup_64.c
+++ b/arch/x86/kernel/setup_64.c
@@ -390,6 +390,7 @@ void __init setup_arch(char **cmdline_p)
 #ifdef CONFIG_SMP
 	/* setup to use the early static init tables during kernel startup */
 	x86_cpu_to_apicid_early_ptr = (void *)&x86_cpu_to_apicid_init;
+	x86_bios_cpu_apicid_early_ptr = (void *)&x86_bios_cpu_apicid_init;
 #ifdef CONFIG_NUMA
 	x86_cpu_to_node_map_early_ptr = (void *)&x86_cpu_to_node_map_init;
 #endif

-- 

Re: [PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[Mike Travis]

Hi Andrew,

My automatic scripts accidentally sent this mail prematurely.  Please hold
off applying yet.

Thanks,
Mike

travis@sgi.com wrote:
> Fixup change NR_CPUS patchset by rebasing on 2.6.24-rc8-mm1 (from 2.6.24-rc6-mm1)
> and adding last minute changes suggested by reviews.
> 
> Based on 2.6.24-rc8-mm1
> 
> Signed-off-by: Mike Travis <travis@sgi.com>
> Reviewed-by: Christoph Lameter <clameter@sgi.com>
> ---
> 

Re: [PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[Ingo Molnar]

* Mike Travis <travis@sgi.com> wrote:

> Hi Andrew,
> 
> My automatic scripts accidentally sent this mail prematurely.  Please 
> hold off applying yet.

I've picked it up for x86.git and i'll keep testing it (the patches seem 
straightforward) and will report any problems with the bite-head-off 
option unset.

[ The 32-bit NUMA compile issue is orthogonal to these patches - it's 
  due to the lack of 32-bit NUMA support in your changes :) That needs 
  fixing before this could go into v2.6.25. ]

	Ingo

Re: [PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[Mike Travis]

Ingo Molnar wrote:
> * Mike Travis <travis@sgi.com> wrote:
> 
>> Hi Andrew,
>>
>> My automatic scripts accidentally sent this mail prematurely.  Please 
>> hold off applying yet.
> 
> I've picked it up for x86.git and i'll keep testing it (the patches seem 
> straightforward) and will report any problems with the bite-head-off 
> option unset.
> 
> [ The 32-bit NUMA compile issue is orthogonal to these patches - it's 
>   due to the lack of 32-bit NUMA support in your changes :) That needs 
>   fixing before this could go into v2.6.25. ]
> 
> 	Ingo

I hadn't considered doing 32-bit NUMA changes as I didn't know if the
NR_CPUS count would really be increased for the 32-bit architecture.
I have been trying though not to break it. ;-)

Thanks,
Mike

Re: [PATCH 0/3] x86: Reduce memory and intra-node effects with large count NR_CPUs fixup

[Christoph Lameter]

On Fri, 18 Jan 2008, Mike Travis wrote:

 > I hadn't considered doing 32-bit NUMA changes as I didn't know if the
> NR_CPUS count would really be increased for the 32-bit architecture.
> I have been trying though not to break it. ;-)

32bit NUMA is tricky because ZONE_NORMAL memory is only available on node 
0. There have been thorny difficult to debug issues in the past...