[deliverable/linux.git] / arch / x86_64 / mm / srat.c

/*
 * ACPI 3.0 based NUMA setup
 * Copyright 2004 Andi Kleen, SuSE Labs.
 *
 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
 *
 * Called from acpi_numa_init while reading the SRAT and SLIT tables.
 * Assumes all memory regions belonging to a single proximity domain
 * are in one chunk. Holes between them will be included in the node.
 */

#include <linux/kernel.h>
#include <linux/acpi.h>
#include <linux/mmzone.h>
#include <linux/bitmap.h>
#include <linux/module.h>
#include <linux/topology.h>
#include <asm/proto.h>
#include <asm/numa.h>
#include <asm/e820.h>

static struct acpi_table_slit *acpi_slit;

static nodemask_t nodes_parsed __initdata;
static nodemask_t nodes_found __initdata;
static struct node nodes[MAX_NUMNODES] __initdata;
static u8 pxm2node[256] = { [0 ... 255] = 0xff };

/* Too small nodes confuse the VM badly. Usually they result
   from BIOS bugs. */
#define NODE_MIN_SIZE (4*1024*1024)

static int node_to_pxm(int n);

int pxm_to_node(int pxm)
{
	if ((unsigned)pxm >= 256)
		return -1;
	/* Extend 0xff to (int)-1 */
	return (signed char)pxm2node[pxm];
}

static __init int setup_node(int pxm)
{
	unsigned node = pxm2node[pxm];
	if (node == 0xff) {
		if (nodes_weight(nodes_found) >= MAX_NUMNODES)
			return -1;
		node = first_unset_node(nodes_found); 
		node_set(node, nodes_found);
		pxm2node[pxm] = node;
	}
	return pxm2node[pxm];
}

static __init int conflicting_nodes(unsigned long start, unsigned long end)
{
	int i;
	for_each_node_mask(i, nodes_parsed) {
		struct node *nd = &nodes[i];
		if (nd->start == nd->end)
			continue;
		if (nd->end > start && nd->start < end)
			return i;
		if (nd->end == end && nd->start == start)
			return i;
	}
	return -1;
}

static __init void cutoff_node(int i, unsigned long start, unsigned long end)
{
	struct node *nd = &nodes[i];
	if (nd->start < start) {
		nd->start = start;
		if (nd->end < nd->start)
			nd->start = nd->end;
	}
	if (nd->end > end) {
		nd->end = end;
		if (nd->start > nd->end)
			nd->start = nd->end;
	}
}

static __init void bad_srat(void)
{
	int i;
	printk(KERN_ERR "SRAT: SRAT not used.\n");
	acpi_numa = -1;
	for (i = 0; i < MAX_LOCAL_APIC; i++)
		apicid_to_node[i] = NUMA_NO_NODE;
}

static __init inline int srat_disabled(void)
{
	return numa_off || acpi_numa < 0;
}

/*
 * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
 * up the NUMA heuristics which wants the local node to have a smaller
 * distance than the others.
 * Do some quick checks here and only use the SLIT if it passes.
 */
static __init int slit_valid(struct acpi_table_slit *slit)
{
	int i, j;
	int d = slit->localities;
	for (i = 0; i < d; i++) {
		for (j = 0; j < d; j++)  {
			u8 val = slit->entry[d*i + j];
			if (i == j) {
				if (val != 10)
					return 0;
			} else if (val <= 10)
				return 0;
		}
	}
	return 1;
}

/* Callback for SLIT parsing */
void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
{
	if (!slit_valid(slit)) {
		printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
		return;
	}
	acpi_slit = slit;
}

/* Callback for Proximity Domain -> LAPIC mapping */
void __init
acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa)
{
	int pxm, node;
	if (srat_disabled() || pa->flags.enabled == 0)
		return;
	pxm = pa->proximity_domain;
	node = setup_node(pxm);
	if (node < 0) {
		printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
		bad_srat();
		return;
	}
	apicid_to_node[pa->apic_id] = node;
	acpi_numa = 1;
	printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
	       pxm, pa->apic_id, node);
}

/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
void __init
acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma)
{
	struct node *nd;
	unsigned long start, end;
	int node, pxm;
	int i;

	if (srat_disabled() || ma->flags.enabled == 0)
		return;
	pxm = ma->proximity_domain;
	node = setup_node(pxm);
	if (node < 0) {
		printk(KERN_ERR "SRAT: Too many proximity domains.\n");
		bad_srat();
		return;
	}
	start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32);
	end = start + (ma->length_lo | ((u64)ma->length_hi << 32));
	/* It is fine to add this area to the nodes data it will be used later*/
	if (ma->flags.hot_pluggable == 1)
		printk(KERN_INFO "SRAT: hot plug zone found %lx - %lx \n",
				start, end);
	i = conflicting_nodes(start, end);
	if (i == node) {
		printk(KERN_WARNING
		"SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
			pxm, start, end, nodes[i].start, nodes[i].end);
	} else if (i >= 0) {
		printk(KERN_ERR
		       "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
		       pxm, start, end, node_to_pxm(i),
			nodes[i].start, nodes[i].end);
		bad_srat();
		return;
	}
	nd = &nodes[node];
	if (!node_test_and_set(node, nodes_parsed)) {
		nd->start = start;
		nd->end = end;
	} else {
		if (start < nd->start)
			nd->start = start;
		if (nd->end < end)
			nd->end = end;
	}
	printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
	       nd->start, nd->end);
}

/* Sanity check to catch more bad SRATs (they are amazingly common).
   Make sure the PXMs cover all memory. */
static int nodes_cover_memory(void)
{
	int i;
	unsigned long pxmram, e820ram;

	pxmram = 0;
	for_each_node_mask(i, nodes_parsed) {
		unsigned long s = nodes[i].start >> PAGE_SHIFT;
		unsigned long e = nodes[i].end >> PAGE_SHIFT;
		pxmram += e - s;
		pxmram -= e820_hole_size(s, e);
	}

	e820ram = end_pfn - e820_hole_size(0, end_pfn);
	if (pxmram < e820ram) {
		printk(KERN_ERR
	"SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
			(pxmram << PAGE_SHIFT) >> 20,
			(e820ram << PAGE_SHIFT) >> 20);
		return 0;
	}
	return 1;
}

static void unparse_node(int node)
{
	int i;
	node_clear(node, nodes_parsed);
	for (i = 0; i < MAX_LOCAL_APIC; i++) {
		if (apicid_to_node[i] == node)
			apicid_to_node[i] = NUMA_NO_NODE;
	}
}

void __init acpi_numa_arch_fixup(void) {}

/* Use the information discovered above to actually set up the nodes. */
int __init acpi_scan_nodes(unsigned long start, unsigned long end)
{
	int i;

	/* First clean up the node list */
	for (i = 0; i < MAX_NUMNODES; i++) {
		cutoff_node(i, start, end);
		if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE)
			unparse_node(i);
	}

	if (acpi_numa <= 0)
		return -1;

	if (!nodes_cover_memory()) {
		bad_srat();
		return -1;
	}

	memnode_shift = compute_hash_shift(nodes, nodes_weight(nodes_parsed));
	if (memnode_shift < 0) {
		printk(KERN_ERR
		     "SRAT: No NUMA node hash function found. Contact maintainer\n");
		bad_srat();
		return -1;
	}

	/* Finally register nodes */
	for_each_node_mask(i, nodes_parsed)
		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)
			continue;
		if (!node_isset(cpu_to_node[i], nodes_parsed))
			numa_set_node(i, NUMA_NO_NODE);
	}
	numa_init_array();
	return 0;
}

static int node_to_pxm(int n)
{
       int i;
       if (pxm2node[n] == n)
               return n;
       for (i = 0; i < 256; i++)
               if (pxm2node[i] == n)
                       return i;
       return 0;
}

int __node_distance(int a, int b)
{
	int index;

	if (!acpi_slit)
		return a == b ? 10 : 20;
	index = acpi_slit->localities * node_to_pxm(a);
	return acpi_slit->entry[index + node_to_pxm(b)];
}

EXPORT_SYMBOL(__node_distance);
Commit	Line	Data
1da177e4 LT	1	/*
	2	* ACPI 3.0 based NUMA setup
	3	* Copyright 2004 Andi Kleen, SuSE Labs.
	4	*
	5	* Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
	6	*
	7	* Called from acpi_numa_init while reading the SRAT and SLIT tables.
	8	* Assumes all memory regions belonging to a single proximity domain
	9	* are in one chunk. Holes between them will be included in the node.
	10	*/
	11
	12	#include <linux/kernel.h>
	13	#include <linux/acpi.h>
	14	#include <linux/mmzone.h>
	15	#include <linux/bitmap.h>
	16	#include <linux/module.h>
	17	#include <linux/topology.h>
	18	#include <asm/proto.h>
	19	#include <asm/numa.h>
8a6fdd3e	20	#include <asm/e820.h>
1da177e4 LT	21
	22	static struct acpi_table_slit *acpi_slit;
	23
	24	static nodemask_t nodes_parsed __initdata;
	25	static nodemask_t nodes_found __initdata;
	26	static struct node nodes[MAX_NUMNODES] __initdata;
e4e94072	27	static u8 pxm2node[256] = { [0 ... 255] = 0xff };
1da177e4	28
9391a3f9 AK	29	/* Too small nodes confuse the VM badly. Usually they result
	30	from BIOS bugs. */
	31	#define NODE_MIN_SIZE (410241024)
	32
05d1fa4b AK	33	static int node_to_pxm(int n);
05d1fa4b AK	34
69e1a33f AK	35	int pxm_to_node(int pxm)
	36	{
	37	if ((unsigned)pxm >= 256)
e4e94072 AK	38	return -1;
	39	/* Extend 0xff to (int)-1 */
	40	return (signed char)pxm2node[pxm];
69e1a33f AK	41	}
69e1a33f AK	42
1da177e4 LT	43	static __init int setup_node(int pxm)
	44	{
	45	unsigned node = pxm2node[pxm];
	46	if (node == 0xff) {
	47	if (nodes_weight(nodes_found) >= MAX_NUMNODES)
	48	return -1;
	49	node = first_unset_node(nodes_found);
	50	node_set(node, nodes_found);
	51	pxm2node[pxm] = node;
	52	}
	53	return pxm2node[pxm];
	54	}
	55
	56	static __init int conflicting_nodes(unsigned long start, unsigned long end)
	57	{
	58	int i;
4b6a455c	59	for_each_node_mask(i, nodes_parsed) {
1da177e4 LT	60	struct node *nd = &nodes[i];
	61	if (nd->start == nd->end)
	62	continue;
	63	if (nd->end > start && nd->start < end)
05d1fa4b	64	return i;
1da177e4	65	if (nd->end == end && nd->start == start)
05d1fa4b	66	return i;
1da177e4 LT	67	}
	68	return -1;
	69	}
	70
	71	static __init void cutoff_node(int i, unsigned long start, unsigned long end)
	72	{
	73	struct node *nd = &nodes[i];
	74	if (nd->start < start) {
	75	nd->start = start;
	76	if (nd->end < nd->start)
	77	nd->start = nd->end;
	78	}
	79	if (nd->end > end) {
1da177e4 LT	80	nd->end = end;
	81	if (nd->start > nd->end)
	82	nd->start = nd->end;
	83	}
	84	}
	85
	86	static __init void bad_srat(void)
	87	{
2bce2b54	88	int i;
1da177e4 LT	89	printk(KERN_ERR "SRAT: SRAT not used.\n");
1da177e4 LT	90	acpi_numa = -1;
2bce2b54 AK	91	for (i = 0; i < MAX_LOCAL_APIC; i++)
2bce2b54 AK	92	apicid_to_node[i] = NUMA_NO_NODE;
1da177e4 LT	93	}
	94
	95	static __init inline int srat_disabled(void)
	96	{
	97	return numa_off \|\| acpi_numa < 0;
	98	}
	99
1584b89c AK	100	/*
	101	* A lot of BIOS fill in 10 (= no distance) everywhere. This messes
	102	* up the NUMA heuristics which wants the local node to have a smaller
	103	* distance than the others.
	104	* Do some quick checks here and only use the SLIT if it passes.
	105	*/
	106	static __init int slit_valid(struct acpi_table_slit *slit)
	107	{
	108	int i, j;
	109	int d = slit->localities;
	110	for (i = 0; i < d; i++) {
	111	for (j = 0; j < d; j++) {
	112	u8 val = slit->entry[d*i + j];
	113	if (i == j) {
	114	if (val != 10)
	115	return 0;
	116	} else if (val <= 10)
	117	return 0;
	118	}
	119	}
	120	return 1;
	121	}
	122
1da177e4 LT	123	/* Callback for SLIT parsing */
	124	void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
	125	{
1584b89c AK	126	if (!slit_valid(slit)) {
	127	printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
	128	return;
	129	}
1da177e4 LT	130	acpi_slit = slit;
	131	}
	132
	133	/* Callback for Proximity Domain -> LAPIC mapping */
	134	void __init
	135	acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa)
	136	{
	137	int pxm, node;
	138	if (srat_disabled() \|\| pa->flags.enabled == 0)
	139	return;
	140	pxm = pa->proximity_domain;
	141	node = setup_node(pxm);
	142	if (node < 0) {
	143	printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
	144	bad_srat();
	145	return;
	146	}
0b07e984	147	apicid_to_node[pa->apic_id] = node;
1da177e4	148	acpi_numa = 1;
0b07e984 AK	149	printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
0b07e984 AK	150	pxm, pa->apic_id, node);
1da177e4 LT	151	}
	152
	153	/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
	154	void __init
	155	acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma)
	156	{
	157	struct node *nd;
	158	unsigned long start, end;
	159	int node, pxm;
	160	int i;
	161
	162	if (srat_disabled() \|\| ma->flags.enabled == 0)
	163	return;
1da177e4 LT	164	pxm = ma->proximity_domain;
	165	node = setup_node(pxm);
	166	if (node < 0) {
	167	printk(KERN_ERR "SRAT: Too many proximity domains.\n");
	168	bad_srat();
	169	return;
	170	}
	171	start = ma->base_addr_lo \| ((u64)ma->base_addr_hi << 32);
	172	end = start + (ma->length_lo \| ((u64)ma->length_hi << 32));
69cb62eb AK	173	/* It is fine to add this area to the nodes data it will be used later*/
	174	if (ma->flags.hot_pluggable == 1)
	175	printk(KERN_INFO "SRAT: hot plug zone found %lx - %lx \n",
	176	start, end);
1da177e4	177	i = conflicting_nodes(start, end);
05d1fa4b AK	178	if (i == node) {
	179	printk(KERN_WARNING
	180	"SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
	181	pxm, start, end, nodes[i].start, nodes[i].end);
	182	} else if (i >= 0) {
1da177e4	183	printk(KERN_ERR
05d1fa4b AK	184	"SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
	185	pxm, start, end, node_to_pxm(i),
	186	nodes[i].start, nodes[i].end);
1da177e4 LT	187	bad_srat();
	188	return;
	189	}
	190	nd = &nodes[node];
	191	if (!node_test_and_set(node, nodes_parsed)) {
	192	nd->start = start;
	193	nd->end = end;
	194	} else {
	195	if (start < nd->start)
	196	nd->start = start;
	197	if (nd->end < end)
	198	nd->end = end;
	199	}
1da177e4 LT	200	printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
	201	nd->start, nd->end);
	202	}
	203
8a6fdd3e AK	204	/* Sanity check to catch more bad SRATs (they are amazingly common).
	205	Make sure the PXMs cover all memory. */
	206	static int nodes_cover_memory(void)
	207	{
	208	int i;
	209	unsigned long pxmram, e820ram;
	210
	211	pxmram = 0;
	212	for_each_node_mask(i, nodes_parsed) {
	213	unsigned long s = nodes[i].start >> PAGE_SHIFT;
	214	unsigned long e = nodes[i].end >> PAGE_SHIFT;
	215	pxmram += e - s;
	216	pxmram -= e820_hole_size(s, e);
	217	}
	218
	219	e820ram = end_pfn - e820_hole_size(0, end_pfn);
	220	if (pxmram < e820ram) {
	221	printk(KERN_ERR
	222	"SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
	223	(pxmram << PAGE_SHIFT) >> 20,
	224	(e820ram << PAGE_SHIFT) >> 20);
	225	return 0;
	226	}
	227	return 1;
	228	}
	229
9391a3f9 AK	230	static void unparse_node(int node)
	231	{
	232	int i;
	233	node_clear(node, nodes_parsed);
	234	for (i = 0; i < MAX_LOCAL_APIC; i++) {
	235	if (apicid_to_node[i] == node)
	236	apicid_to_node[i] = NUMA_NO_NODE;
	237	}
	238	}
	239
1da177e4 LT	240	void __init acpi_numa_arch_fixup(void) {}
	241
	242	/* Use the information discovered above to actually set up the nodes. */
	243	int __init acpi_scan_nodes(unsigned long start, unsigned long end)
	244	{
	245	int i;
8a6fdd3e	246
e58e0d03	247	/* First clean up the node list */
9391a3f9	248	for (i = 0; i < MAX_NUMNODES; i++) {
e58e0d03	249	cutoff_node(i, start, end);
9391a3f9 AK	250	if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE)
9391a3f9 AK	251	unparse_node(i);
e58e0d03 AK	252	}
e58e0d03 AK	253
9391a3f9 AK	254	if (acpi_numa <= 0)
	255	return -1;
	256
8a6fdd3e AK	257	if (!nodes_cover_memory()) {
	258	bad_srat();
	259	return -1;
	260	}
	261
1da177e4 LT	262	memnode_shift = compute_hash_shift(nodes, nodes_weight(nodes_parsed));
	263	if (memnode_shift < 0) {
	264	printk(KERN_ERR
	265	"SRAT: No NUMA node hash function found. Contact maintainer\n");
	266	bad_srat();
	267	return -1;
	268	}
e58e0d03 AK	269
	270	/* Finally register nodes */
	271	for_each_node_mask(i, nodes_parsed)
1da177e4	272	setup_node_bootmem(i, nodes[i].start, nodes[i].end);
1da177e4 LT	273	for (i = 0; i < NR_CPUS; i++) {
	274	if (cpu_to_node[i] == NUMA_NO_NODE)
	275	continue;
	276	if (!node_isset(cpu_to_node[i], nodes_parsed))
69d81fcd	277	numa_set_node(i, NUMA_NO_NODE);
1da177e4 LT	278	}
	279	numa_init_array();
	280	return 0;
	281	}
	282
05d1fa4b	283	static int node_to_pxm(int n)
1da177e4 LT	284	{
	285	int i;
	286	if (pxm2node[n] == n)
	287	return n;
	288	for (i = 0; i < 256; i++)
	289	if (pxm2node[i] == n)
	290	return i;
	291	return 0;
	292	}
	293
	294	int __node_distance(int a, int b)
	295	{
	296	int index;
	297
	298	if (!acpi_slit)
	299	return a == b ? 10 : 20;
	300	index = acpi_slit->localities * node_to_pxm(a);
	301	return acpi_slit->entry[index + node_to_pxm(b)];
	302	}
	303
	304	EXPORT_SYMBOL(__node_distance);