numa.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825
  1. /* Common code for 32 and 64-bit NUMA */
  2. #include <linux/kernel.h>
  3. #include <linux/mm.h>
  4. #include <linux/string.h>
  5. #include <linux/init.h>
  6. #include <linux/bootmem.h>
  7. #include <linux/memblock.h>
  8. #include <linux/mmzone.h>
  9. #include <linux/ctype.h>
  10. #include <linux/module.h>
  11. #include <linux/nodemask.h>
  12. #include <linux/sched.h>
  13. #include <linux/topology.h>
  14. #include <asm/e820.h>
  15. #include <asm/proto.h>
  16. #include <asm/dma.h>
  17. #include <asm/acpi.h>
  18. #include <asm/amd_nb.h>
  19. #include "numa_internal.h"
  20. int __initdata numa_off;
  21. nodemask_t numa_nodes_parsed __initdata;
  22. struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  23. EXPORT_SYMBOL(node_data);
  24. static struct numa_meminfo numa_meminfo
  25. #ifndef CONFIG_MEMORY_HOTPLUG
  26. __initdata
  27. #endif
  28. ;
  29. static int numa_distance_cnt;
  30. static u8 *numa_distance;
  31. static __init int numa_setup(char *opt)
  32. {
  33. if (!opt)
  34. return -EINVAL;
  35. if (!strncmp(opt, "off", 3))
  36. numa_off = 1;
  37. #ifdef CONFIG_NUMA_EMU
  38. if (!strncmp(opt, "fake=", 5))
  39. numa_emu_cmdline(opt + 5);
  40. #endif
  41. #ifdef CONFIG_ACPI_NUMA
  42. if (!strncmp(opt, "noacpi", 6))
  43. acpi_numa = -1;
  44. #endif
  45. return 0;
  46. }
  47. early_param("numa", numa_setup);
  48. /*
  49. * apicid, cpu, node mappings
  50. */
  51. s16 __apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  52. [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  53. };
  54. int __cpuinit numa_cpu_node(int cpu)
  55. {
  56. int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
  57. if (apicid != BAD_APICID)
  58. return __apicid_to_node[apicid];
  59. return NUMA_NO_NODE;
  60. }
  61. cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
  62. EXPORT_SYMBOL(node_to_cpumask_map);
  63. /*
  64. * Map cpu index to node index
  65. */
  66. DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  67. EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  68. void __cpuinit numa_set_node(int cpu, int node)
  69. {
  70. int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
  71. /* early setting, no percpu area yet */
  72. if (cpu_to_node_map) {
  73. cpu_to_node_map[cpu] = node;
  74. return;
  75. }
  76. #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  77. if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
  78. printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
  79. dump_stack();
  80. return;
  81. }
  82. #endif
  83. per_cpu(x86_cpu_to_node_map, cpu) = node;
  84. if (node != NUMA_NO_NODE)
  85. set_cpu_numa_node(cpu, node);
  86. }
  87. void __cpuinit numa_clear_node(int cpu)
  88. {
  89. numa_set_node(cpu, NUMA_NO_NODE);
  90. }
  91. /*
  92. * Allocate node_to_cpumask_map based on number of available nodes
  93. * Requires node_possible_map to be valid.
  94. *
  95. * Note: cpumask_of_node() is not valid until after this is done.
  96. * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
  97. */
  98. void __init setup_node_to_cpumask_map(void)
  99. {
  100. unsigned int node, num = 0;
  101. /* setup nr_node_ids if not done yet */
  102. if (nr_node_ids == MAX_NUMNODES) {
  103. for_each_node_mask(node, node_possible_map)
  104. num = node;
  105. nr_node_ids = num + 1;
  106. }
  107. /* allocate the map */
  108. for (node = 0; node < nr_node_ids; node++)
  109. alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
  110. /* cpumask_of_node() will now work */
  111. pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
  112. }
  113. static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
  114. struct numa_meminfo *mi)
  115. {
  116. /* ignore zero length blks */
  117. if (start == end)
  118. return 0;
  119. /* whine about and ignore invalid blks */
  120. if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
  121. pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
  122. nid, start, end);
  123. return 0;
  124. }
  125. if (mi->nr_blks >= NR_NODE_MEMBLKS) {
  126. pr_err("NUMA: too many memblk ranges\n");
  127. return -EINVAL;
  128. }
  129. mi->blk[mi->nr_blks].start = start;
  130. mi->blk[mi->nr_blks].end = end;
  131. mi->blk[mi->nr_blks].nid = nid;
  132. mi->nr_blks++;
  133. return 0;
  134. }
  135. /**
  136. * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
  137. * @idx: Index of memblk to remove
  138. * @mi: numa_meminfo to remove memblk from
  139. *
  140. * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
  141. * decrementing @mi->nr_blks.
  142. */
  143. void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
  144. {
  145. mi->nr_blks--;
  146. memmove(&mi->blk[idx], &mi->blk[idx + 1],
  147. (mi->nr_blks - idx) * sizeof(mi->blk[0]));
  148. }
  149. /**
  150. * numa_add_memblk - Add one numa_memblk to numa_meminfo
  151. * @nid: NUMA node ID of the new memblk
  152. * @start: Start address of the new memblk
  153. * @end: End address of the new memblk
  154. *
  155. * Add a new memblk to the default numa_meminfo.
  156. *
  157. * RETURNS:
  158. * 0 on success, -errno on failure.
  159. */
  160. int __init numa_add_memblk(int nid, u64 start, u64 end)
  161. {
  162. return numa_add_memblk_to(nid, start, end, &numa_meminfo);
  163. }
  164. /* Initialize NODE_DATA for a node on the local memory */
  165. static void __init setup_node_data(int nid, u64 start, u64 end)
  166. {
  167. const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
  168. u64 nd_pa;
  169. void *nd;
  170. int tnid;
  171. /*
  172. * Don't confuse VM with a node that doesn't have the
  173. * minimum amount of memory:
  174. */
  175. if (end && (end - start) < NODE_MIN_SIZE)
  176. return;
  177. start = roundup(start, ZONE_ALIGN);
  178. printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n",
  179. nid, start, end);
  180. /*
  181. * Allocate node data. Try node-local memory and then any node.
  182. * Never allocate in DMA zone.
  183. */
  184. nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid);
  185. if (!nd_pa) {
  186. pr_err("Cannot find %zu bytes in node %d\n",
  187. nd_size, nid);
  188. return;
  189. }
  190. nd = __va(nd_pa);
  191. /* report and initialize */
  192. printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n",
  193. nd_pa, nd_pa + nd_size - 1);
  194. tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
  195. if (tnid != nid)
  196. printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
  197. node_data[nid] = nd;
  198. memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
  199. NODE_DATA(nid)->node_id = nid;
  200. NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
  201. NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
  202. node_set_online(nid);
  203. }
  204. /**
  205. * numa_cleanup_meminfo - Cleanup a numa_meminfo
  206. * @mi: numa_meminfo to clean up
  207. *
  208. * Sanitize @mi by merging and removing unncessary memblks. Also check for
  209. * conflicts and clear unused memblks.
  210. *
  211. * RETURNS:
  212. * 0 on success, -errno on failure.
  213. */
  214. int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
  215. {
  216. const u64 low = 0;
  217. const u64 high = PFN_PHYS(max_pfn);
  218. int i, j, k;
  219. /* first, trim all entries */
  220. for (i = 0; i < mi->nr_blks; i++) {
  221. struct numa_memblk *bi = &mi->blk[i];
  222. /* make sure all blocks are inside the limits */
  223. bi->start = max(bi->start, low);
  224. bi->end = min(bi->end, high);
  225. /* and there's no empty block */
  226. if (bi->start >= bi->end)
  227. numa_remove_memblk_from(i--, mi);
  228. }
  229. /* merge neighboring / overlapping entries */
  230. for (i = 0; i < mi->nr_blks; i++) {
  231. struct numa_memblk *bi = &mi->blk[i];
  232. for (j = i + 1; j < mi->nr_blks; j++) {
  233. struct numa_memblk *bj = &mi->blk[j];
  234. u64 start, end;
  235. /*
  236. * See whether there are overlapping blocks. Whine
  237. * about but allow overlaps of the same nid. They
  238. * will be merged below.
  239. */
  240. if (bi->end > bj->start && bi->start < bj->end) {
  241. if (bi->nid != bj->nid) {
  242. pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
  243. bi->nid, bi->start, bi->end,
  244. bj->nid, bj->start, bj->end);
  245. return -EINVAL;
  246. }
  247. pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
  248. bi->nid, bi->start, bi->end,
  249. bj->start, bj->end);
  250. }
  251. /*
  252. * Join together blocks on the same node, holes
  253. * between which don't overlap with memory on other
  254. * nodes.
  255. */
  256. if (bi->nid != bj->nid)
  257. continue;
  258. start = min(bi->start, bj->start);
  259. end = max(bi->end, bj->end);
  260. for (k = 0; k < mi->nr_blks; k++) {
  261. struct numa_memblk *bk = &mi->blk[k];
  262. if (bi->nid == bk->nid)
  263. continue;
  264. if (start < bk->end && end > bk->start)
  265. break;
  266. }
  267. if (k < mi->nr_blks)
  268. continue;
  269. printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n",
  270. bi->nid, bi->start, bi->end, bj->start, bj->end,
  271. start, end);
  272. bi->start = start;
  273. bi->end = end;
  274. numa_remove_memblk_from(j--, mi);
  275. }
  276. }
  277. /* clear unused ones */
  278. for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
  279. mi->blk[i].start = mi->blk[i].end = 0;
  280. mi->blk[i].nid = NUMA_NO_NODE;
  281. }
  282. return 0;
  283. }
  284. /*
  285. * Set nodes, which have memory in @mi, in *@nodemask.
  286. */
  287. static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
  288. const struct numa_meminfo *mi)
  289. {
  290. int i;
  291. for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
  292. if (mi->blk[i].start != mi->blk[i].end &&
  293. mi->blk[i].nid != NUMA_NO_NODE)
  294. node_set(mi->blk[i].nid, *nodemask);
  295. }
  296. /**
  297. * numa_reset_distance - Reset NUMA distance table
  298. *
  299. * The current table is freed. The next numa_set_distance() call will
  300. * create a new one.
  301. */
  302. void __init numa_reset_distance(void)
  303. {
  304. size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
  305. /* numa_distance could be 1LU marking allocation failure, test cnt */
  306. if (numa_distance_cnt)
  307. memblock_free(__pa(numa_distance), size);
  308. numa_distance_cnt = 0;
  309. numa_distance = NULL; /* enable table creation */
  310. }
  311. static int __init numa_alloc_distance(void)
  312. {
  313. nodemask_t nodes_parsed;
  314. size_t size;
  315. int i, j, cnt = 0;
  316. u64 phys;
  317. /* size the new table and allocate it */
  318. nodes_parsed = numa_nodes_parsed;
  319. numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
  320. for_each_node_mask(i, nodes_parsed)
  321. cnt = i;
  322. cnt++;
  323. size = cnt * cnt * sizeof(numa_distance[0]);
  324. phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
  325. size, PAGE_SIZE);
  326. if (!phys) {
  327. pr_warning("NUMA: Warning: can't allocate distance table!\n");
  328. /* don't retry until explicitly reset */
  329. numa_distance = (void *)1LU;
  330. return -ENOMEM;
  331. }
  332. memblock_reserve(phys, size);
  333. numa_distance = __va(phys);
  334. numa_distance_cnt = cnt;
  335. /* fill with the default distances */
  336. for (i = 0; i < cnt; i++)
  337. for (j = 0; j < cnt; j++)
  338. numa_distance[i * cnt + j] = i == j ?
  339. LOCAL_DISTANCE : REMOTE_DISTANCE;
  340. printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
  341. return 0;
  342. }
  343. /**
  344. * numa_set_distance - Set NUMA distance from one NUMA to another
  345. * @from: the 'from' node to set distance
  346. * @to: the 'to' node to set distance
  347. * @distance: NUMA distance
  348. *
  349. * Set the distance from node @from to @to to @distance. If distance table
  350. * doesn't exist, one which is large enough to accommodate all the currently
  351. * known nodes will be created.
  352. *
  353. * If such table cannot be allocated, a warning is printed and further
  354. * calls are ignored until the distance table is reset with
  355. * numa_reset_distance().
  356. *
  357. * If @from or @to is higher than the highest known node or lower than zero
  358. * at the time of table creation or @distance doesn't make sense, the call
  359. * is ignored.
  360. * This is to allow simplification of specific NUMA config implementations.
  361. */
  362. void __init numa_set_distance(int from, int to, int distance)
  363. {
  364. if (!numa_distance && numa_alloc_distance() < 0)
  365. return;
  366. if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
  367. from < 0 || to < 0) {
  368. pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
  369. from, to, distance);
  370. return;
  371. }
  372. if ((u8)distance != distance ||
  373. (from == to && distance != LOCAL_DISTANCE)) {
  374. pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
  375. from, to, distance);
  376. return;
  377. }
  378. numa_distance[from * numa_distance_cnt + to] = distance;
  379. }
  380. int __node_distance(int from, int to)
  381. {
  382. if (from >= numa_distance_cnt || to >= numa_distance_cnt)
  383. return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
  384. return numa_distance[from * numa_distance_cnt + to];
  385. }
  386. EXPORT_SYMBOL(__node_distance);
  387. /*
  388. * Sanity check to catch more bad NUMA configurations (they are amazingly
  389. * common). Make sure the nodes cover all memory.
  390. */
  391. static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
  392. {
  393. u64 numaram, e820ram;
  394. int i;
  395. numaram = 0;
  396. for (i = 0; i < mi->nr_blks; i++) {
  397. u64 s = mi->blk[i].start >> PAGE_SHIFT;
  398. u64 e = mi->blk[i].end >> PAGE_SHIFT;
  399. numaram += e - s;
  400. numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
  401. if ((s64)numaram < 0)
  402. numaram = 0;
  403. }
  404. e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
  405. /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
  406. if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
  407. printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
  408. (numaram << PAGE_SHIFT) >> 20,
  409. (e820ram << PAGE_SHIFT) >> 20);
  410. return false;
  411. }
  412. return true;
  413. }
  414. static int __init numa_register_memblks(struct numa_meminfo *mi)
  415. {
  416. unsigned long uninitialized_var(pfn_align);
  417. int i, nid;
  418. /* Account for nodes with cpus and no memory */
  419. node_possible_map = numa_nodes_parsed;
  420. numa_nodemask_from_meminfo(&node_possible_map, mi);
  421. if (WARN_ON(nodes_empty(node_possible_map)))
  422. return -EINVAL;
  423. for (i = 0; i < mi->nr_blks; i++) {
  424. struct numa_memblk *mb = &mi->blk[i];
  425. memblock_set_node(mb->start, mb->end - mb->start, mb->nid);
  426. }
  427. /*
  428. * If sections array is gonna be used for pfn -> nid mapping, check
  429. * whether its granularity is fine enough.
  430. */
  431. #ifdef NODE_NOT_IN_PAGE_FLAGS
  432. pfn_align = node_map_pfn_alignment();
  433. if (pfn_align && pfn_align < PAGES_PER_SECTION) {
  434. printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
  435. PFN_PHYS(pfn_align) >> 20,
  436. PFN_PHYS(PAGES_PER_SECTION) >> 20);
  437. return -EINVAL;
  438. }
  439. #endif
  440. if (!numa_meminfo_cover_memory(mi))
  441. return -EINVAL;
  442. /* Finally register nodes. */
  443. for_each_node_mask(nid, node_possible_map) {
  444. u64 start = PFN_PHYS(max_pfn);
  445. u64 end = 0;
  446. for (i = 0; i < mi->nr_blks; i++) {
  447. if (nid != mi->blk[i].nid)
  448. continue;
  449. start = min(mi->blk[i].start, start);
  450. end = max(mi->blk[i].end, end);
  451. }
  452. if (start < end)
  453. setup_node_data(nid, start, end);
  454. }
  455. /* Dump memblock with node info and return. */
  456. memblock_dump_all();
  457. return 0;
  458. }
  459. /*
  460. * There are unfortunately some poorly designed mainboards around that
  461. * only connect memory to a single CPU. This breaks the 1:1 cpu->node
  462. * mapping. To avoid this fill in the mapping for all possible CPUs,
  463. * as the number of CPUs is not known yet. We round robin the existing
  464. * nodes.
  465. */
  466. static void __init numa_init_array(void)
  467. {
  468. int rr, i;
  469. rr = first_node(node_online_map);
  470. for (i = 0; i < nr_cpu_ids; i++) {
  471. if (early_cpu_to_node(i) != NUMA_NO_NODE)
  472. continue;
  473. numa_set_node(i, rr);
  474. rr = next_node(rr, node_online_map);
  475. if (rr == MAX_NUMNODES)
  476. rr = first_node(node_online_map);
  477. }
  478. }
  479. static int __init numa_init(int (*init_func)(void))
  480. {
  481. int i;
  482. int ret;
  483. for (i = 0; i < MAX_LOCAL_APIC; i++)
  484. set_apicid_to_node(i, NUMA_NO_NODE);
  485. nodes_clear(numa_nodes_parsed);
  486. nodes_clear(node_possible_map);
  487. nodes_clear(node_online_map);
  488. memset(&numa_meminfo, 0, sizeof(numa_meminfo));
  489. WARN_ON(memblock_set_node(0, ULLONG_MAX, MAX_NUMNODES));
  490. numa_reset_distance();
  491. ret = init_func();
  492. if (ret < 0)
  493. return ret;
  494. ret = numa_cleanup_meminfo(&numa_meminfo);
  495. if (ret < 0)
  496. return ret;
  497. numa_emulation(&numa_meminfo, numa_distance_cnt);
  498. ret = numa_register_memblks(&numa_meminfo);
  499. if (ret < 0)
  500. return ret;
  501. for (i = 0; i < nr_cpu_ids; i++) {
  502. int nid = early_cpu_to_node(i);
  503. if (nid == NUMA_NO_NODE)
  504. continue;
  505. if (!node_online(nid))
  506. numa_clear_node(i);
  507. }
  508. numa_init_array();
  509. return 0;
  510. }
  511. /**
  512. * dummy_numa_init - Fallback dummy NUMA init
  513. *
  514. * Used if there's no underlying NUMA architecture, NUMA initialization
  515. * fails, or NUMA is disabled on the command line.
  516. *
  517. * Must online at least one node and add memory blocks that cover all
  518. * allowed memory. This function must not fail.
  519. */
  520. static int __init dummy_numa_init(void)
  521. {
  522. printk(KERN_INFO "%s\n",
  523. numa_off ? "NUMA turned off" : "No NUMA configuration found");
  524. printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n",
  525. 0LLU, PFN_PHYS(max_pfn));
  526. node_set(0, numa_nodes_parsed);
  527. numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
  528. return 0;
  529. }
  530. /**
  531. * x86_numa_init - Initialize NUMA
  532. *
  533. * Try each configured NUMA initialization method until one succeeds. The
  534. * last fallback is dummy single node config encomapssing whole memory and
  535. * never fails.
  536. */
  537. void __init x86_numa_init(void)
  538. {
  539. if (!numa_off) {
  540. #ifdef CONFIG_X86_NUMAQ
  541. if (!numa_init(numaq_numa_init))
  542. return;
  543. #endif
  544. #ifdef CONFIG_ACPI_NUMA
  545. if (!numa_init(x86_acpi_numa_init))
  546. return;
  547. #endif
  548. #ifdef CONFIG_AMD_NUMA
  549. if (!numa_init(amd_numa_init))
  550. return;
  551. #endif
  552. }
  553. numa_init(dummy_numa_init);
  554. }
  555. static __init int find_near_online_node(int node)
  556. {
  557. int n, val;
  558. int min_val = INT_MAX;
  559. int best_node = -1;
  560. for_each_online_node(n) {
  561. val = node_distance(node, n);
  562. if (val < min_val) {
  563. min_val = val;
  564. best_node = n;
  565. }
  566. }
  567. return best_node;
  568. }
  569. /*
  570. * Setup early cpu_to_node.
  571. *
  572. * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
  573. * and apicid_to_node[] tables have valid entries for a CPU.
  574. * This means we skip cpu_to_node[] initialisation for NUMA
  575. * emulation and faking node case (when running a kernel compiled
  576. * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
  577. * is already initialized in a round robin manner at numa_init_array,
  578. * prior to this call, and this initialization is good enough
  579. * for the fake NUMA cases.
  580. *
  581. * Called before the per_cpu areas are setup.
  582. */
  583. void __init init_cpu_to_node(void)
  584. {
  585. int cpu;
  586. u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
  587. BUG_ON(cpu_to_apicid == NULL);
  588. for_each_possible_cpu(cpu) {
  589. int node = numa_cpu_node(cpu);
  590. if (node == NUMA_NO_NODE)
  591. continue;
  592. if (!node_online(node))
  593. node = find_near_online_node(node);
  594. numa_set_node(cpu, node);
  595. }
  596. }
  597. #ifndef CONFIG_DEBUG_PER_CPU_MAPS
  598. # ifndef CONFIG_NUMA_EMU
  599. void __cpuinit numa_add_cpu(int cpu)
  600. {
  601. cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  602. }
  603. void __cpuinit numa_remove_cpu(int cpu)
  604. {
  605. cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  606. }
  607. # endif /* !CONFIG_NUMA_EMU */
  608. #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
  609. int __cpu_to_node(int cpu)
  610. {
  611. if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
  612. printk(KERN_WARNING
  613. "cpu_to_node(%d): usage too early!\n", cpu);
  614. dump_stack();
  615. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  616. }
  617. return per_cpu(x86_cpu_to_node_map, cpu);
  618. }
  619. EXPORT_SYMBOL(__cpu_to_node);
  620. /*
  621. * Same function as cpu_to_node() but used if called before the
  622. * per_cpu areas are setup.
  623. */
  624. int early_cpu_to_node(int cpu)
  625. {
  626. if (early_per_cpu_ptr(x86_cpu_to_node_map))
  627. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  628. if (!cpu_possible(cpu)) {
  629. printk(KERN_WARNING
  630. "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
  631. dump_stack();
  632. return NUMA_NO_NODE;
  633. }
  634. return per_cpu(x86_cpu_to_node_map, cpu);
  635. }
  636. void debug_cpumask_set_cpu(int cpu, int node, bool enable)
  637. {
  638. struct cpumask *mask;
  639. char buf[64];
  640. if (node == NUMA_NO_NODE) {
  641. /* early_cpu_to_node() already emits a warning and trace */
  642. return;
  643. }
  644. mask = node_to_cpumask_map[node];
  645. if (!mask) {
  646. pr_err("node_to_cpumask_map[%i] NULL\n", node);
  647. dump_stack();
  648. return;
  649. }
  650. if (enable)
  651. cpumask_set_cpu(cpu, mask);
  652. else
  653. cpumask_clear_cpu(cpu, mask);
  654. cpulist_scnprintf(buf, sizeof(buf), mask);
  655. printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
  656. enable ? "numa_add_cpu" : "numa_remove_cpu",
  657. cpu, node, buf);
  658. return;
  659. }
  660. # ifndef CONFIG_NUMA_EMU
  661. static void __cpuinit numa_set_cpumask(int cpu, bool enable)
  662. {
  663. debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
  664. }
  665. void __cpuinit numa_add_cpu(int cpu)
  666. {
  667. numa_set_cpumask(cpu, true);
  668. }
  669. void __cpuinit numa_remove_cpu(int cpu)
  670. {
  671. numa_set_cpumask(cpu, false);
  672. }
  673. # endif /* !CONFIG_NUMA_EMU */
  674. /*
  675. * Returns a pointer to the bitmask of CPUs on Node 'node'.
  676. */
  677. const struct cpumask *cpumask_of_node(int node)
  678. {
  679. if (node >= nr_node_ids) {
  680. printk(KERN_WARNING
  681. "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
  682. node, nr_node_ids);
  683. dump_stack();
  684. return cpu_none_mask;
  685. }
  686. if (node_to_cpumask_map[node] == NULL) {
  687. printk(KERN_WARNING
  688. "cpumask_of_node(%d): no node_to_cpumask_map!\n",
  689. node);
  690. dump_stack();
  691. return cpu_online_mask;
  692. }
  693. return node_to_cpumask_map[node];
  694. }
  695. EXPORT_SYMBOL(cpumask_of_node);
  696. #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
  697. #ifdef CONFIG_MEMORY_HOTPLUG
  698. int memory_add_physaddr_to_nid(u64 start)
  699. {
  700. struct numa_meminfo *mi = &numa_meminfo;
  701. int nid = mi->blk[0].nid;
  702. int i;
  703. for (i = 0; i < mi->nr_blks; i++)
  704. if (mi->blk[i].start <= start && mi->blk[i].end > start)
  705. nid = mi->blk[i].nid;
  706. return nid;
  707. }
  708. EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
  709. #endif