numa.c 20 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821
  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: node_to_cpumask() 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 u64 nd_low = PFN_PHYS(MAX_DMA_PFN);
  168. const u64 nd_high = PFN_PHYS(max_pfn_mapped);
  169. const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
  170. bool remapped = false;
  171. u64 nd_pa;
  172. void *nd;
  173. int tnid;
  174. /*
  175. * Don't confuse VM with a node that doesn't have the
  176. * minimum amount of memory:
  177. */
  178. if (end && (end - start) < NODE_MIN_SIZE)
  179. return;
  180. /* initialize remap allocator before aligning to ZONE_ALIGN */
  181. init_alloc_remap(nid, start, end);
  182. start = roundup(start, ZONE_ALIGN);
  183. printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n",
  184. nid, start, end);
  185. /*
  186. * Allocate node data. Try remap allocator first, node-local
  187. * memory and then any node. Never allocate in DMA zone.
  188. */
  189. nd = alloc_remap(nid, nd_size);
  190. if (nd) {
  191. nd_pa = __pa(nd);
  192. remapped = true;
  193. } else {
  194. nd_pa = memblock_x86_find_in_range_node(nid, nd_low, nd_high,
  195. nd_size, SMP_CACHE_BYTES);
  196. if (nd_pa == MEMBLOCK_ERROR)
  197. nd_pa = memblock_find_in_range(nd_low, nd_high,
  198. nd_size, SMP_CACHE_BYTES);
  199. if (nd_pa == MEMBLOCK_ERROR) {
  200. pr_err("Cannot find %zu bytes in node %d\n",
  201. nd_size, nid);
  202. return;
  203. }
  204. memblock_x86_reserve_range(nd_pa, nd_pa + nd_size, "NODE_DATA");
  205. nd = __va(nd_pa);
  206. }
  207. /* report and initialize */
  208. printk(KERN_INFO " NODE_DATA [%016Lx - %016Lx]%s\n",
  209. nd_pa, nd_pa + nd_size - 1, remapped ? " (remapped)" : "");
  210. tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
  211. if (!remapped && tnid != nid)
  212. printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
  213. node_data[nid] = nd;
  214. memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
  215. NODE_DATA(nid)->node_id = nid;
  216. NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
  217. NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
  218. node_set_online(nid);
  219. }
  220. /**
  221. * numa_cleanup_meminfo - Cleanup a numa_meminfo
  222. * @mi: numa_meminfo to clean up
  223. *
  224. * Sanitize @mi by merging and removing unncessary memblks. Also check for
  225. * conflicts and clear unused memblks.
  226. *
  227. * RETURNS:
  228. * 0 on success, -errno on failure.
  229. */
  230. int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
  231. {
  232. const u64 low = 0;
  233. const u64 high = PFN_PHYS(max_pfn);
  234. int i, j, k;
  235. for (i = 0; i < mi->nr_blks; i++) {
  236. struct numa_memblk *bi = &mi->blk[i];
  237. /* make sure all blocks are inside the limits */
  238. bi->start = max(bi->start, low);
  239. bi->end = min(bi->end, high);
  240. /* and there's no empty block */
  241. if (bi->start >= bi->end) {
  242. numa_remove_memblk_from(i--, mi);
  243. continue;
  244. }
  245. for (j = i + 1; j < mi->nr_blks; j++) {
  246. struct numa_memblk *bj = &mi->blk[j];
  247. u64 start, end;
  248. /*
  249. * See whether there are overlapping blocks. Whine
  250. * about but allow overlaps of the same nid. They
  251. * will be merged below.
  252. */
  253. if (bi->end > bj->start && bi->start < bj->end) {
  254. if (bi->nid != bj->nid) {
  255. pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
  256. bi->nid, bi->start, bi->end,
  257. bj->nid, bj->start, bj->end);
  258. return -EINVAL;
  259. }
  260. pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
  261. bi->nid, bi->start, bi->end,
  262. bj->start, bj->end);
  263. }
  264. /*
  265. * Join together blocks on the same node, holes
  266. * between which don't overlap with memory on other
  267. * nodes.
  268. */
  269. if (bi->nid != bj->nid)
  270. continue;
  271. start = max(min(bi->start, bj->start), low);
  272. end = min(max(bi->end, bj->end), high);
  273. for (k = 0; k < mi->nr_blks; k++) {
  274. struct numa_memblk *bk = &mi->blk[k];
  275. if (bi->nid == bk->nid)
  276. continue;
  277. if (start < bk->end && end > bk->start)
  278. break;
  279. }
  280. if (k < mi->nr_blks)
  281. continue;
  282. printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n",
  283. bi->nid, bi->start, bi->end, bj->start, bj->end,
  284. start, end);
  285. bi->start = start;
  286. bi->end = end;
  287. numa_remove_memblk_from(j--, mi);
  288. }
  289. }
  290. for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
  291. mi->blk[i].start = mi->blk[i].end = 0;
  292. mi->blk[i].nid = NUMA_NO_NODE;
  293. }
  294. return 0;
  295. }
  296. /*
  297. * Set nodes, which have memory in @mi, in *@nodemask.
  298. */
  299. static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
  300. const struct numa_meminfo *mi)
  301. {
  302. int i;
  303. for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
  304. if (mi->blk[i].start != mi->blk[i].end &&
  305. mi->blk[i].nid != NUMA_NO_NODE)
  306. node_set(mi->blk[i].nid, *nodemask);
  307. }
  308. /**
  309. * numa_reset_distance - Reset NUMA distance table
  310. *
  311. * The current table is freed. The next numa_set_distance() call will
  312. * create a new one.
  313. */
  314. void __init numa_reset_distance(void)
  315. {
  316. size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
  317. /* numa_distance could be 1LU marking allocation failure, test cnt */
  318. if (numa_distance_cnt)
  319. memblock_x86_free_range(__pa(numa_distance),
  320. __pa(numa_distance) + size);
  321. numa_distance_cnt = 0;
  322. numa_distance = NULL; /* enable table creation */
  323. }
  324. static int __init numa_alloc_distance(void)
  325. {
  326. nodemask_t nodes_parsed;
  327. size_t size;
  328. int i, j, cnt = 0;
  329. u64 phys;
  330. /* size the new table and allocate it */
  331. nodes_parsed = numa_nodes_parsed;
  332. numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
  333. for_each_node_mask(i, nodes_parsed)
  334. cnt = i;
  335. cnt++;
  336. size = cnt * cnt * sizeof(numa_distance[0]);
  337. phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
  338. size, PAGE_SIZE);
  339. if (phys == MEMBLOCK_ERROR) {
  340. pr_warning("NUMA: Warning: can't allocate distance table!\n");
  341. /* don't retry until explicitly reset */
  342. numa_distance = (void *)1LU;
  343. return -ENOMEM;
  344. }
  345. memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
  346. numa_distance = __va(phys);
  347. numa_distance_cnt = cnt;
  348. /* fill with the default distances */
  349. for (i = 0; i < cnt; i++)
  350. for (j = 0; j < cnt; j++)
  351. numa_distance[i * cnt + j] = i == j ?
  352. LOCAL_DISTANCE : REMOTE_DISTANCE;
  353. printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
  354. return 0;
  355. }
  356. /**
  357. * numa_set_distance - Set NUMA distance from one NUMA to another
  358. * @from: the 'from' node to set distance
  359. * @to: the 'to' node to set distance
  360. * @distance: NUMA distance
  361. *
  362. * Set the distance from node @from to @to to @distance. If distance table
  363. * doesn't exist, one which is large enough to accommodate all the currently
  364. * known nodes will be created.
  365. *
  366. * If such table cannot be allocated, a warning is printed and further
  367. * calls are ignored until the distance table is reset with
  368. * numa_reset_distance().
  369. *
  370. * If @from or @to is higher than the highest known node at the time of
  371. * table creation or @distance doesn't make sense, the call is ignored.
  372. * This is to allow simplification of specific NUMA config implementations.
  373. */
  374. void __init numa_set_distance(int from, int to, int distance)
  375. {
  376. if (!numa_distance && numa_alloc_distance() < 0)
  377. return;
  378. if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
  379. printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
  380. from, to, distance);
  381. return;
  382. }
  383. if ((u8)distance != distance ||
  384. (from == to && distance != LOCAL_DISTANCE)) {
  385. pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
  386. from, to, distance);
  387. return;
  388. }
  389. numa_distance[from * numa_distance_cnt + to] = distance;
  390. }
  391. int __node_distance(int from, int to)
  392. {
  393. if (from >= numa_distance_cnt || to >= numa_distance_cnt)
  394. return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
  395. return numa_distance[from * numa_distance_cnt + to];
  396. }
  397. EXPORT_SYMBOL(__node_distance);
  398. /*
  399. * Sanity check to catch more bad NUMA configurations (they are amazingly
  400. * common). Make sure the nodes cover all memory.
  401. */
  402. static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
  403. {
  404. u64 numaram, e820ram;
  405. int i;
  406. numaram = 0;
  407. for (i = 0; i < mi->nr_blks; i++) {
  408. u64 s = mi->blk[i].start >> PAGE_SHIFT;
  409. u64 e = mi->blk[i].end >> PAGE_SHIFT;
  410. numaram += e - s;
  411. numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
  412. if ((s64)numaram < 0)
  413. numaram = 0;
  414. }
  415. e820ram = max_pfn - (memblock_x86_hole_size(0,
  416. PFN_PHYS(max_pfn)) >> PAGE_SHIFT);
  417. /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
  418. if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
  419. printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
  420. (numaram << PAGE_SHIFT) >> 20,
  421. (e820ram << PAGE_SHIFT) >> 20);
  422. return false;
  423. }
  424. return true;
  425. }
  426. static int __init numa_register_memblks(struct numa_meminfo *mi)
  427. {
  428. int i, nid;
  429. /* Account for nodes with cpus and no memory */
  430. node_possible_map = numa_nodes_parsed;
  431. numa_nodemask_from_meminfo(&node_possible_map, mi);
  432. if (WARN_ON(nodes_empty(node_possible_map)))
  433. return -EINVAL;
  434. for (i = 0; i < mi->nr_blks; i++)
  435. memblock_x86_register_active_regions(mi->blk[i].nid,
  436. mi->blk[i].start >> PAGE_SHIFT,
  437. mi->blk[i].end >> PAGE_SHIFT);
  438. /* for out of order entries */
  439. sort_node_map();
  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. return 0;
  456. }
  457. /*
  458. * There are unfortunately some poorly designed mainboards around that
  459. * only connect memory to a single CPU. This breaks the 1:1 cpu->node
  460. * mapping. To avoid this fill in the mapping for all possible CPUs,
  461. * as the number of CPUs is not known yet. We round robin the existing
  462. * nodes.
  463. */
  464. static void __init numa_init_array(void)
  465. {
  466. int rr, i;
  467. rr = first_node(node_online_map);
  468. for (i = 0; i < nr_cpu_ids; i++) {
  469. if (early_cpu_to_node(i) != NUMA_NO_NODE)
  470. continue;
  471. numa_set_node(i, rr);
  472. rr = next_node(rr, node_online_map);
  473. if (rr == MAX_NUMNODES)
  474. rr = first_node(node_online_map);
  475. }
  476. }
  477. static int __init numa_init(int (*init_func)(void))
  478. {
  479. int i;
  480. int ret;
  481. for (i = 0; i < MAX_LOCAL_APIC; i++)
  482. set_apicid_to_node(i, NUMA_NO_NODE);
  483. nodes_clear(numa_nodes_parsed);
  484. nodes_clear(node_possible_map);
  485. nodes_clear(node_online_map);
  486. memset(&numa_meminfo, 0, sizeof(numa_meminfo));
  487. remove_all_active_ranges();
  488. numa_reset_distance();
  489. ret = init_func();
  490. if (ret < 0)
  491. return ret;
  492. ret = numa_cleanup_meminfo(&numa_meminfo);
  493. if (ret < 0)
  494. return ret;
  495. numa_emulation(&numa_meminfo, numa_distance_cnt);
  496. ret = numa_register_memblks(&numa_meminfo);
  497. if (ret < 0)
  498. return ret;
  499. for (i = 0; i < nr_cpu_ids; i++) {
  500. int nid = early_cpu_to_node(i);
  501. if (nid == NUMA_NO_NODE)
  502. continue;
  503. if (!node_online(nid))
  504. numa_clear_node(i);
  505. }
  506. numa_init_array();
  507. return 0;
  508. }
  509. /**
  510. * dummy_numa_init - Fallback dummy NUMA init
  511. *
  512. * Used if there's no underlying NUMA architecture, NUMA initialization
  513. * fails, or NUMA is disabled on the command line.
  514. *
  515. * Must online at least one node and add memory blocks that cover all
  516. * allowed memory. This function must not fail.
  517. */
  518. static int __init dummy_numa_init(void)
  519. {
  520. printk(KERN_INFO "%s\n",
  521. numa_off ? "NUMA turned off" : "No NUMA configuration found");
  522. printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n",
  523. 0LLU, PFN_PHYS(max_pfn));
  524. node_set(0, numa_nodes_parsed);
  525. numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
  526. return 0;
  527. }
  528. /**
  529. * x86_numa_init - Initialize NUMA
  530. *
  531. * Try each configured NUMA initialization method until one succeeds. The
  532. * last fallback is dummy single node config encomapssing whole memory and
  533. * never fails.
  534. */
  535. void __init x86_numa_init(void)
  536. {
  537. if (!numa_off) {
  538. #ifdef CONFIG_X86_NUMAQ
  539. if (!numa_init(numaq_numa_init))
  540. return;
  541. #endif
  542. #ifdef CONFIG_ACPI_NUMA
  543. if (!numa_init(x86_acpi_numa_init))
  544. return;
  545. #endif
  546. #ifdef CONFIG_AMD_NUMA
  547. if (!numa_init(amd_numa_init))
  548. return;
  549. #endif
  550. }
  551. numa_init(dummy_numa_init);
  552. }
  553. static __init int find_near_online_node(int node)
  554. {
  555. int n, val;
  556. int min_val = INT_MAX;
  557. int best_node = -1;
  558. for_each_online_node(n) {
  559. val = node_distance(node, n);
  560. if (val < min_val) {
  561. min_val = val;
  562. best_node = n;
  563. }
  564. }
  565. return best_node;
  566. }
  567. /*
  568. * Setup early cpu_to_node.
  569. *
  570. * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
  571. * and apicid_to_node[] tables have valid entries for a CPU.
  572. * This means we skip cpu_to_node[] initialisation for NUMA
  573. * emulation and faking node case (when running a kernel compiled
  574. * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
  575. * is already initialized in a round robin manner at numa_init_array,
  576. * prior to this call, and this initialization is good enough
  577. * for the fake NUMA cases.
  578. *
  579. * Called before the per_cpu areas are setup.
  580. */
  581. void __init init_cpu_to_node(void)
  582. {
  583. int cpu;
  584. u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
  585. BUG_ON(cpu_to_apicid == NULL);
  586. for_each_possible_cpu(cpu) {
  587. int node = numa_cpu_node(cpu);
  588. if (node == NUMA_NO_NODE)
  589. continue;
  590. if (!node_online(node))
  591. node = find_near_online_node(node);
  592. numa_set_node(cpu, node);
  593. }
  594. }
  595. #ifndef CONFIG_DEBUG_PER_CPU_MAPS
  596. # ifndef CONFIG_NUMA_EMU
  597. void __cpuinit numa_add_cpu(int cpu)
  598. {
  599. cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  600. }
  601. void __cpuinit numa_remove_cpu(int cpu)
  602. {
  603. cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  604. }
  605. # endif /* !CONFIG_NUMA_EMU */
  606. #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
  607. int __cpu_to_node(int cpu)
  608. {
  609. if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
  610. printk(KERN_WARNING
  611. "cpu_to_node(%d): usage too early!\n", cpu);
  612. dump_stack();
  613. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  614. }
  615. return per_cpu(x86_cpu_to_node_map, cpu);
  616. }
  617. EXPORT_SYMBOL(__cpu_to_node);
  618. /*
  619. * Same function as cpu_to_node() but used if called before the
  620. * per_cpu areas are setup.
  621. */
  622. int early_cpu_to_node(int cpu)
  623. {
  624. if (early_per_cpu_ptr(x86_cpu_to_node_map))
  625. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  626. if (!cpu_possible(cpu)) {
  627. printk(KERN_WARNING
  628. "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
  629. dump_stack();
  630. return NUMA_NO_NODE;
  631. }
  632. return per_cpu(x86_cpu_to_node_map, cpu);
  633. }
  634. void debug_cpumask_set_cpu(int cpu, int node, bool enable)
  635. {
  636. struct cpumask *mask;
  637. char buf[64];
  638. if (node == NUMA_NO_NODE) {
  639. /* early_cpu_to_node() already emits a warning and trace */
  640. return;
  641. }
  642. mask = node_to_cpumask_map[node];
  643. if (!mask) {
  644. pr_err("node_to_cpumask_map[%i] NULL\n", node);
  645. dump_stack();
  646. return;
  647. }
  648. if (enable)
  649. cpumask_set_cpu(cpu, mask);
  650. else
  651. cpumask_clear_cpu(cpu, mask);
  652. cpulist_scnprintf(buf, sizeof(buf), mask);
  653. printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
  654. enable ? "numa_add_cpu" : "numa_remove_cpu",
  655. cpu, node, buf);
  656. return;
  657. }
  658. # ifndef CONFIG_NUMA_EMU
  659. static void __cpuinit numa_set_cpumask(int cpu, bool enable)
  660. {
  661. debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
  662. }
  663. void __cpuinit numa_add_cpu(int cpu)
  664. {
  665. numa_set_cpumask(cpu, true);
  666. }
  667. void __cpuinit numa_remove_cpu(int cpu)
  668. {
  669. numa_set_cpumask(cpu, false);
  670. }
  671. # endif /* !CONFIG_NUMA_EMU */
  672. /*
  673. * Returns a pointer to the bitmask of CPUs on Node 'node'.
  674. */
  675. const struct cpumask *cpumask_of_node(int node)
  676. {
  677. if (node >= nr_node_ids) {
  678. printk(KERN_WARNING
  679. "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
  680. node, nr_node_ids);
  681. dump_stack();
  682. return cpu_none_mask;
  683. }
  684. if (node_to_cpumask_map[node] == NULL) {
  685. printk(KERN_WARNING
  686. "cpumask_of_node(%d): no node_to_cpumask_map!\n",
  687. node);
  688. dump_stack();
  689. return cpu_online_mask;
  690. }
  691. return node_to_cpumask_map[node];
  692. }
  693. EXPORT_SYMBOL(cpumask_of_node);
  694. #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
  695. #ifdef CONFIG_MEMORY_HOTPLUG
  696. int memory_add_physaddr_to_nid(u64 start)
  697. {
  698. struct numa_meminfo *mi = &numa_meminfo;
  699. int nid = mi->blk[0].nid;
  700. int i;
  701. for (i = 0; i < mi->nr_blks; i++)
  702. if (mi->blk[i].start <= start && mi->blk[i].end > start)
  703. nid = mi->blk[i].nid;
  704. return nid;
  705. }
  706. EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
  707. #endif