2 * ACPI 3.0 based NUMA setup
3 * Copyright 2004 Andi Kleen, SuSE Labs.
5 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
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.
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 <linux/bootmem.h>
20 #include <asm/proto.h>
24 int acpi_numa __initdata;
26 #if (defined(CONFIG_ACPI_HOTPLUG_MEMORY) || \
27 defined(CONFIG_ACPI_HOTPLUG_MEMORY_MODULE)) \
28 && !defined(CONFIG_MEMORY_HOTPLUG)
29 #define RESERVE_HOTADD 1
32 static struct acpi_table_slit *acpi_slit;
34 static nodemask_t nodes_parsed __initdata;
35 static struct bootnode nodes[MAX_NUMNODES] __initdata;
36 static struct bootnode nodes_add[MAX_NUMNODES] __initdata;
37 static int found_add_area __initdata;
38 int hotadd_percent __initdata = 0;
39 #ifndef RESERVE_HOTADD
40 #define hotadd_percent 0 /* Ignore all settings */
43 /* Too small nodes confuse the VM badly. Usually they result
45 #define NODE_MIN_SIZE (4*1024*1024)
47 static __init int setup_node(int pxm)
49 return acpi_map_pxm_to_node(pxm);
52 static __init int conflicting_nodes(unsigned long start, unsigned long end)
55 for_each_node_mask(i, nodes_parsed) {
56 struct bootnode *nd = &nodes[i];
57 if (nd->start == nd->end)
59 if (nd->end > start && nd->start < end)
61 if (nd->end == end && nd->start == start)
67 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
69 struct bootnode *nd = &nodes[i];
74 if (nd->start < start) {
76 if (nd->end < nd->start)
81 if (nd->start > nd->end)
86 static __init void bad_srat(void)
89 printk(KERN_ERR "SRAT: SRAT not used.\n");
92 for (i = 0; i < MAX_LOCAL_APIC; i++)
93 apicid_to_node[i] = NUMA_NO_NODE;
94 for (i = 0; i < MAX_NUMNODES; i++)
95 nodes_add[i].start = nodes[i].end = 0;
98 static __init inline int srat_disabled(void)
100 return numa_off || acpi_numa < 0;
104 * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
105 * up the NUMA heuristics which wants the local node to have a smaller
106 * distance than the others.
107 * Do some quick checks here and only use the SLIT if it passes.
109 static __init int slit_valid(struct acpi_table_slit *slit)
112 int d = slit->localities;
113 for (i = 0; i < d; i++) {
114 for (j = 0; j < d; j++) {
115 u8 val = slit->entry[d*i + j];
119 } else if (val <= 10)
126 /* Callback for SLIT parsing */
127 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
129 if (!slit_valid(slit)) {
130 printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
136 /* Callback for Proximity Domain -> LAPIC mapping */
138 acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa)
143 if (pa->header.length != sizeof(struct acpi_table_processor_affinity)) {
147 if (pa->flags.enabled == 0)
149 pxm = pa->proximity_domain;
150 node = setup_node(pxm);
152 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
156 apicid_to_node[pa->apic_id] = node;
158 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
159 pxm, pa->apic_id, node);
162 #ifdef RESERVE_HOTADD
164 * Protect against too large hotadd areas that would fill up memory.
166 static int hotadd_enough_memory(struct bootnode *nd)
168 static unsigned long allocated;
169 static unsigned long last_area_end;
170 unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
171 long mem = pages * sizeof(struct page);
173 unsigned long allowed;
174 unsigned long oldpages = pages;
178 allowed = (end_pfn - e820_hole_size(0, end_pfn)) * PAGE_SIZE;
179 allowed = (allowed / 100) * hotadd_percent;
180 if (allocated + mem > allowed) {
182 /* Give them at least part of their hotadd memory upto hotadd_percent
183 It would be better to spread the limit out
184 over multiple hotplug areas, but that is too complicated
186 if (allocated >= allowed)
188 range = allowed - allocated;
189 pages = (range / PAGE_SIZE);
190 mem = pages * sizeof(struct page);
191 nd->end = nd->start + range;
193 /* Not completely fool proof, but a good sanity check */
194 addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
197 if (pages != oldpages)
198 printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
199 pages << PAGE_SHIFT);
200 last_area_end = addr + mem;
206 * It is fine to add this area to the nodes data it will be used later
207 * This code supports one contigious hot add area per node.
209 static int reserve_hotadd(int node, unsigned long start, unsigned long end)
211 unsigned long s_pfn = start >> PAGE_SHIFT;
212 unsigned long e_pfn = end >> PAGE_SHIFT;
214 struct bootnode *nd = &nodes_add[node];
216 /* I had some trouble with strange memory hotadd regions breaking
217 the boot. Be very strict here and reject anything unexpected.
218 If you want working memory hotadd write correct SRATs.
220 The node size check is a basic sanity check to guard against
222 if ((signed long)(end - start) < NODE_MIN_SIZE) {
223 printk(KERN_ERR "SRAT: Hotplug area too small\n");
227 /* This check might be a bit too strict, but I'm keeping it for now. */
228 if (e820_hole_size(s_pfn, e_pfn) != e_pfn - s_pfn) {
229 printk(KERN_ERR "SRAT: Hotplug area has existing memory\n");
233 if (!hotadd_enough_memory(&nodes_add[node])) {
234 printk(KERN_ERR "SRAT: Hotplug area too large\n");
241 if (nd->start == nd->end) {
246 if (nd->start == end) {
250 if (nd->end == start) {
255 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
258 if ((nd->end >> PAGE_SHIFT) > end_pfn)
259 end_pfn = nd->end >> PAGE_SHIFT;
262 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
267 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
269 acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma)
271 struct bootnode *nd, oldnode;
272 unsigned long start, end;
278 if (ma->header.length != sizeof(struct acpi_table_memory_affinity)) {
282 if (ma->flags.enabled == 0)
284 if (ma->flags.hot_pluggable && hotadd_percent == 0)
286 start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32);
287 end = start + (ma->length_lo | ((u64)ma->length_hi << 32));
288 pxm = ma->proximity_domain;
289 node = setup_node(pxm);
291 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
295 i = conflicting_nodes(start, end);
298 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
299 pxm, start, end, nodes[i].start, nodes[i].end);
302 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
303 pxm, start, end, node_to_pxm(i),
304 nodes[i].start, nodes[i].end);
310 if (!node_test_and_set(node, nodes_parsed)) {
314 if (start < nd->start)
320 printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
323 #ifdef RESERVE_HOTADD
324 if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) {
325 /* Ignore hotadd region. Undo damage */
326 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
328 if ((nd->start | nd->end) == 0)
329 node_clear(node, nodes_parsed);
334 /* Sanity check to catch more bad SRATs (they are amazingly common).
335 Make sure the PXMs cover all memory. */
336 static int nodes_cover_memory(void)
339 unsigned long pxmram, e820ram;
342 for_each_node_mask(i, nodes_parsed) {
343 unsigned long s = nodes[i].start >> PAGE_SHIFT;
344 unsigned long e = nodes[i].end >> PAGE_SHIFT;
346 pxmram -= e820_hole_size(s, e);
347 pxmram -= nodes_add[i].end - nodes_add[i].start;
348 if ((long)pxmram < 0)
352 e820ram = end_pfn - e820_hole_size(0, end_pfn);
353 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
354 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
356 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
357 (pxmram << PAGE_SHIFT) >> 20,
358 (e820ram << PAGE_SHIFT) >> 20);
364 static void unparse_node(int node)
367 node_clear(node, nodes_parsed);
368 for (i = 0; i < MAX_LOCAL_APIC; i++) {
369 if (apicid_to_node[i] == node)
370 apicid_to_node[i] = NUMA_NO_NODE;
374 void __init acpi_numa_arch_fixup(void) {}
376 /* Use the information discovered above to actually set up the nodes. */
377 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
381 /* First clean up the node list */
382 for (i = 0; i < MAX_NUMNODES; i++) {
383 cutoff_node(i, start, end);
384 if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
393 if (!nodes_cover_memory()) {
398 memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES);
399 if (memnode_shift < 0) {
401 "SRAT: No NUMA node hash function found. Contact maintainer\n");
406 /* Finally register nodes */
407 for_each_node_mask(i, nodes_parsed)
408 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
409 /* Try again in case setup_node_bootmem missed one due
410 to missing bootmem */
411 for_each_node_mask(i, nodes_parsed)
413 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
415 for (i = 0; i < NR_CPUS; i++) {
416 if (cpu_to_node[i] == NUMA_NO_NODE)
418 if (!node_isset(cpu_to_node[i], nodes_parsed))
419 numa_set_node(i, NUMA_NO_NODE);
425 void __init srat_reserve_add_area(int nodeid)
427 if (found_add_area && nodes_add[nodeid].end) {
430 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
431 "for node %d at %Lx-%Lx\n",
432 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
433 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
435 total_mb *= sizeof(struct page);
437 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
438 "pre-allocated memory.\n", (unsigned long long)total_mb);
439 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
440 nodes_add[nodeid].end - nodes_add[nodeid].start);
444 int __node_distance(int a, int b)
449 return a == b ? 10 : 20;
450 index = acpi_slit->localities * node_to_pxm(a);
451 return acpi_slit->entry[index + node_to_pxm(b)];
454 EXPORT_SYMBOL(__node_distance);