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 static struct acpi_table_slit *acpi_slit;
28 static nodemask_t nodes_parsed __initdata;
29 static struct bootnode nodes[MAX_NUMNODES] __initdata;
30 static struct bootnode nodes_add[MAX_NUMNODES];
31 static int found_add_area __initdata;
32 int hotadd_percent __initdata = 0;
34 /* Too small nodes confuse the VM badly. Usually they result
36 #define NODE_MIN_SIZE (4*1024*1024)
38 static __init int setup_node(int pxm)
40 return acpi_map_pxm_to_node(pxm);
43 static __init int conflicting_nodes(unsigned long start, unsigned long end)
46 for_each_node_mask(i, nodes_parsed) {
47 struct bootnode *nd = &nodes[i];
48 if (nd->start == nd->end)
50 if (nd->end > start && nd->start < end)
52 if (nd->end == end && nd->start == start)
58 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
60 struct bootnode *nd = &nodes[i];
65 if (nd->start < start) {
67 if (nd->end < nd->start)
72 if (nd->start > nd->end)
77 static __init void bad_srat(void)
80 printk(KERN_ERR "SRAT: SRAT not used.\n");
83 for (i = 0; i < MAX_LOCAL_APIC; i++)
84 apicid_to_node[i] = NUMA_NO_NODE;
85 for (i = 0; i < MAX_NUMNODES; i++)
86 nodes_add[i].start = nodes[i].end = 0;
87 remove_all_active_ranges();
90 static __init inline int srat_disabled(void)
92 return numa_off || acpi_numa < 0;
96 * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
97 * up the NUMA heuristics which wants the local node to have a smaller
98 * distance than the others.
99 * Do some quick checks here and only use the SLIT if it passes.
101 static __init int slit_valid(struct acpi_table_slit *slit)
104 int d = slit->locality_count;
105 for (i = 0; i < d; i++) {
106 for (j = 0; j < d; j++) {
107 u8 val = slit->entry[d*i + j];
109 if (val != LOCAL_DISTANCE)
111 } else if (val <= LOCAL_DISTANCE)
118 /* Callback for SLIT parsing */
119 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
121 if (!slit_valid(slit)) {
122 printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
128 /* Callback for Proximity Domain -> LAPIC mapping */
130 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
135 if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
139 if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
141 pxm = pa->proximity_domain_lo;
142 node = setup_node(pxm);
144 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
148 apicid_to_node[pa->apic_id] = node;
150 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
151 pxm, pa->apic_id, node);
154 #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE
156 * Protect against too large hotadd areas that would fill up memory.
158 static int hotadd_enough_memory(struct bootnode *nd)
160 static unsigned long allocated;
161 static unsigned long last_area_end;
162 unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
163 long mem = pages * sizeof(struct page);
165 unsigned long allowed;
166 unsigned long oldpages = pages;
170 allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
171 allowed = (allowed / 100) * hotadd_percent;
172 if (allocated + mem > allowed) {
174 /* Give them at least part of their hotadd memory upto hotadd_percent
175 It would be better to spread the limit out
176 over multiple hotplug areas, but that is too complicated
178 if (allocated >= allowed)
180 range = allowed - allocated;
181 pages = (range / PAGE_SIZE);
182 mem = pages * sizeof(struct page);
183 nd->end = nd->start + range;
185 /* Not completely fool proof, but a good sanity check */
186 addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
189 if (pages != oldpages)
190 printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
191 pages << PAGE_SHIFT);
192 last_area_end = addr + mem;
197 static int update_end_of_memory(unsigned long end)
200 if ((end >> PAGE_SHIFT) > end_pfn)
201 end_pfn = end >> PAGE_SHIFT;
205 static inline int save_add_info(void)
207 return hotadd_percent > 0;
210 int update_end_of_memory(unsigned long end) {return -1;}
211 static int hotadd_enough_memory(struct bootnode *nd) {return 1;}
212 #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE
213 static inline int save_add_info(void) {return 1;}
215 static inline int save_add_info(void) {return 0;}
219 * Update nodes_add and decide if to include add are in the zone.
220 * Both SPARSE and RESERVE need nodes_add infomation.
221 * This code supports one contigious hot add area per node.
223 static int reserve_hotadd(int node, unsigned long start, unsigned long end)
225 unsigned long s_pfn = start >> PAGE_SHIFT;
226 unsigned long e_pfn = end >> PAGE_SHIFT;
227 int ret = 0, changed = 0;
228 struct bootnode *nd = &nodes_add[node];
230 /* I had some trouble with strange memory hotadd regions breaking
231 the boot. Be very strict here and reject anything unexpected.
232 If you want working memory hotadd write correct SRATs.
234 The node size check is a basic sanity check to guard against
236 if ((signed long)(end - start) < NODE_MIN_SIZE) {
237 printk(KERN_ERR "SRAT: Hotplug area too small\n");
241 /* This check might be a bit too strict, but I'm keeping it for now. */
242 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
244 "SRAT: Hotplug area %lu -> %lu has existing memory\n",
249 if (!hotadd_enough_memory(&nodes_add[node])) {
250 printk(KERN_ERR "SRAT: Hotplug area too large\n");
256 if (nd->start == nd->end) {
261 if (nd->start == end) {
265 if (nd->end == start) {
270 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
273 ret = update_end_of_memory(nd->end);
276 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
280 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
282 acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
284 struct bootnode *nd, oldnode;
285 unsigned long start, end;
291 if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
295 if ((ma->flags & ACPI_SRAT_MEM_ENABLED) == 0)
298 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) && !save_add_info())
300 start = ma->base_address;
301 end = start + ma->length;
302 pxm = ma->proximity_domain;
303 node = setup_node(pxm);
305 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
309 i = conflicting_nodes(start, end);
312 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
313 pxm, start, end, nodes[i].start, nodes[i].end);
316 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
317 pxm, start, end, node_to_pxm(i),
318 nodes[i].start, nodes[i].end);
324 if (!node_test_and_set(node, nodes_parsed)) {
328 if (start < nd->start)
334 printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
336 e820_register_active_regions(node, nd->start >> PAGE_SHIFT,
337 nd->end >> PAGE_SHIFT);
338 push_node_boundaries(node, nd->start >> PAGE_SHIFT,
339 nd->end >> PAGE_SHIFT);
341 if ((ma->flags & ACPI_SRAT_MEM_HOT_PLUGGABLE) &&
342 (reserve_hotadd(node, start, end) < 0)) {
343 /* Ignore hotadd region. Undo damage */
344 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
346 if ((nd->start | nd->end) == 0)
347 node_clear(node, nodes_parsed);
351 /* Sanity check to catch more bad SRATs (they are amazingly common).
352 Make sure the PXMs cover all memory. */
353 static int __init nodes_cover_memory(const struct bootnode *nodes)
356 unsigned long pxmram, e820ram;
359 for_each_node_mask(i, nodes_parsed) {
360 unsigned long s = nodes[i].start >> PAGE_SHIFT;
361 unsigned long e = nodes[i].end >> PAGE_SHIFT;
363 pxmram -= absent_pages_in_range(s, e);
364 if ((long)pxmram < 0)
368 e820ram = end_pfn - absent_pages_in_range(0, end_pfn);
369 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
370 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
372 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
373 (pxmram << PAGE_SHIFT) >> 20,
374 (e820ram << PAGE_SHIFT) >> 20);
380 static void unparse_node(int node)
383 node_clear(node, nodes_parsed);
384 for (i = 0; i < MAX_LOCAL_APIC; i++) {
385 if (apicid_to_node[i] == node)
386 apicid_to_node[i] = NUMA_NO_NODE;
390 void __init acpi_numa_arch_fixup(void) {}
392 /* Use the information discovered above to actually set up the nodes. */
393 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
400 /* First clean up the node list */
401 for (i = 0; i < MAX_NUMNODES; i++) {
402 cutoff_node(i, start, end);
403 if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
409 if (!nodes_cover_memory(nodes)) {
414 memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES);
415 if (memnode_shift < 0) {
417 "SRAT: No NUMA node hash function found. Contact maintainer\n");
422 node_possible_map = nodes_parsed;
424 /* Finally register nodes */
425 for_each_node_mask(i, node_possible_map)
426 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
427 /* Try again in case setup_node_bootmem missed one due
428 to missing bootmem */
429 for_each_node_mask(i, node_possible_map)
431 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
433 for (i = 0; i < NR_CPUS; i++) {
434 if (cpu_to_node[i] == NUMA_NO_NODE)
436 if (!node_isset(cpu_to_node[i], node_possible_map))
437 numa_set_node(i, NUMA_NO_NODE);
443 #ifdef CONFIG_NUMA_EMU
444 static int __init find_node_by_addr(unsigned long addr)
446 int ret = NUMA_NO_NODE;
449 for_each_node_mask(i, nodes_parsed) {
451 * Find the real node that this emulated node appears on. For
452 * the sake of simplicity, we only use a real node's starting
453 * address to determine which emulated node it appears on.
455 if (addr >= nodes[i].start && addr < nodes[i].end) {
464 * In NUMA emulation, we need to setup proximity domain (_PXM) to node ID
465 * mappings that respect the real ACPI topology but reflect our emulated
466 * environment. For each emulated node, we find which real node it appears on
467 * and create PXM to NID mappings for those fake nodes which mirror that
468 * locality. SLIT will now represent the correct distances between emulated
469 * nodes as a result of the real topology.
471 void __init acpi_fake_nodes(const struct bootnode *fake_nodes, int num_nodes)
474 int fake_node_to_pxm_map[MAX_NUMNODES] = {
475 [0 ... MAX_NUMNODES-1] = PXM_INVAL
477 unsigned char fake_apicid_to_node[MAX_LOCAL_APIC] = {
478 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
481 printk(KERN_INFO "Faking PXM affinity for fake nodes on real "
483 for (i = 0; i < num_nodes; i++) {
486 nid = find_node_by_addr(fake_nodes[i].start);
487 if (nid == NUMA_NO_NODE)
489 pxm = node_to_pxm(nid);
490 if (pxm == PXM_INVAL)
492 fake_node_to_pxm_map[i] = pxm;
494 * For each apicid_to_node mapping that exists for this real
495 * node, it must now point to the fake node ID.
497 for (j = 0; j < MAX_LOCAL_APIC; j++)
498 if (apicid_to_node[j] == nid)
499 fake_apicid_to_node[j] = i;
501 for (i = 0; i < num_nodes; i++)
502 __acpi_map_pxm_to_node(fake_node_to_pxm_map[i], i);
503 memcpy(apicid_to_node, fake_apicid_to_node, sizeof(apicid_to_node));
505 nodes_clear(nodes_parsed);
506 for (i = 0; i < num_nodes; i++)
507 if (fake_nodes[i].start != fake_nodes[i].end)
508 node_set(i, nodes_parsed);
509 WARN_ON(!nodes_cover_memory(fake_nodes));
512 static int null_slit_node_compare(int a, int b)
514 return node_to_pxm(a) == node_to_pxm(b);
517 static int null_slit_node_compare(int a, int b)
521 #endif /* CONFIG_NUMA_EMU */
523 void __init srat_reserve_add_area(int nodeid)
525 if (found_add_area && nodes_add[nodeid].end) {
528 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
529 "for node %d at %Lx-%Lx\n",
530 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
531 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
533 total_mb *= sizeof(struct page);
535 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
536 "pre-allocated memory.\n", (unsigned long long)total_mb);
537 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
538 nodes_add[nodeid].end - nodes_add[nodeid].start);
542 int __node_distance(int a, int b)
547 return null_slit_node_compare(a, b) ? LOCAL_DISTANCE :
549 index = acpi_slit->locality_count * node_to_pxm(a);
550 return acpi_slit->entry[index + node_to_pxm(b)];
553 EXPORT_SYMBOL(__node_distance);
555 int memory_add_physaddr_to_nid(u64 start)
560 if (nodes_add[i].start <= start && nodes_add[i].end > start)
565 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);