numa_64.c 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767
  1. /*
  2. * Generic VM initialization for x86-64 NUMA setups.
  3. * Copyright 2002,2003 Andi Kleen, SuSE Labs.
  4. */
  5. #include <linux/kernel.h>
  6. #include <linux/mm.h>
  7. #include <linux/string.h>
  8. #include <linux/init.h>
  9. #include <linux/bootmem.h>
  10. #include <linux/mmzone.h>
  11. #include <linux/ctype.h>
  12. #include <linux/module.h>
  13. #include <linux/nodemask.h>
  14. #include <linux/sched.h>
  15. #include <asm/e820.h>
  16. #include <asm/proto.h>
  17. #include <asm/dma.h>
  18. #include <asm/numa.h>
  19. #include <asm/acpi.h>
  20. #include <asm/k8.h>
  21. struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
  22. EXPORT_SYMBOL(node_data);
  23. struct memnode memnode;
  24. s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
  25. [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
  26. };
  27. int numa_off __initdata;
  28. static unsigned long __initdata nodemap_addr;
  29. static unsigned long __initdata nodemap_size;
  30. DEFINE_PER_CPU(int, node_number) = 0;
  31. EXPORT_PER_CPU_SYMBOL(node_number);
  32. /*
  33. * Map cpu index to node index
  34. */
  35. DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
  36. EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
  37. /*
  38. * Given a shift value, try to populate memnodemap[]
  39. * Returns :
  40. * 1 if OK
  41. * 0 if memnodmap[] too small (of shift too small)
  42. * -1 if node overlap or lost ram (shift too big)
  43. */
  44. static int __init populate_memnodemap(const struct bootnode *nodes,
  45. int numnodes, int shift, int *nodeids)
  46. {
  47. unsigned long addr, end;
  48. int i, res = -1;
  49. memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
  50. for (i = 0; i < numnodes; i++) {
  51. addr = nodes[i].start;
  52. end = nodes[i].end;
  53. if (addr >= end)
  54. continue;
  55. if ((end >> shift) >= memnodemapsize)
  56. return 0;
  57. do {
  58. if (memnodemap[addr >> shift] != NUMA_NO_NODE)
  59. return -1;
  60. if (!nodeids)
  61. memnodemap[addr >> shift] = i;
  62. else
  63. memnodemap[addr >> shift] = nodeids[i];
  64. addr += (1UL << shift);
  65. } while (addr < end);
  66. res = 1;
  67. }
  68. return res;
  69. }
  70. static int __init allocate_cachealigned_memnodemap(void)
  71. {
  72. unsigned long addr;
  73. memnodemap = memnode.embedded_map;
  74. if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
  75. return 0;
  76. addr = 0x8000;
  77. nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
  78. nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT,
  79. nodemap_size, L1_CACHE_BYTES);
  80. if (nodemap_addr == -1UL) {
  81. printk(KERN_ERR
  82. "NUMA: Unable to allocate Memory to Node hash map\n");
  83. nodemap_addr = nodemap_size = 0;
  84. return -1;
  85. }
  86. memnodemap = phys_to_virt(nodemap_addr);
  87. reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
  88. printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
  89. nodemap_addr, nodemap_addr + nodemap_size);
  90. return 0;
  91. }
  92. /*
  93. * The LSB of all start and end addresses in the node map is the value of the
  94. * maximum possible shift.
  95. */
  96. static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
  97. int numnodes)
  98. {
  99. int i, nodes_used = 0;
  100. unsigned long start, end;
  101. unsigned long bitfield = 0, memtop = 0;
  102. for (i = 0; i < numnodes; i++) {
  103. start = nodes[i].start;
  104. end = nodes[i].end;
  105. if (start >= end)
  106. continue;
  107. bitfield |= start;
  108. nodes_used++;
  109. if (end > memtop)
  110. memtop = end;
  111. }
  112. if (nodes_used <= 1)
  113. i = 63;
  114. else
  115. i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
  116. memnodemapsize = (memtop >> i)+1;
  117. return i;
  118. }
  119. int __init compute_hash_shift(struct bootnode *nodes, int numnodes,
  120. int *nodeids)
  121. {
  122. int shift;
  123. shift = extract_lsb_from_nodes(nodes, numnodes);
  124. if (allocate_cachealigned_memnodemap())
  125. return -1;
  126. printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
  127. shift);
  128. if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) {
  129. printk(KERN_INFO "Your memory is not aligned you need to "
  130. "rebuild your kernel with a bigger NODEMAPSIZE "
  131. "shift=%d\n", shift);
  132. return -1;
  133. }
  134. return shift;
  135. }
  136. int __meminit __early_pfn_to_nid(unsigned long pfn)
  137. {
  138. return phys_to_nid(pfn << PAGE_SHIFT);
  139. }
  140. static void * __init early_node_mem(int nodeid, unsigned long start,
  141. unsigned long end, unsigned long size,
  142. unsigned long align)
  143. {
  144. unsigned long mem = find_e820_area(start, end, size, align);
  145. void *ptr;
  146. if (mem != -1L)
  147. return __va(mem);
  148. ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
  149. if (ptr == NULL) {
  150. printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
  151. size, nodeid);
  152. return NULL;
  153. }
  154. return ptr;
  155. }
  156. /* Initialize bootmem allocator for a node */
  157. void __init setup_node_bootmem(int nodeid, unsigned long start,
  158. unsigned long end)
  159. {
  160. unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size;
  161. unsigned long bootmap_start, nodedata_phys;
  162. void *bootmap;
  163. const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
  164. int nid;
  165. if (!end)
  166. return;
  167. start = roundup(start, ZONE_ALIGN);
  168. printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
  169. start, end);
  170. start_pfn = start >> PAGE_SHIFT;
  171. last_pfn = end >> PAGE_SHIFT;
  172. node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
  173. SMP_CACHE_BYTES);
  174. if (node_data[nodeid] == NULL)
  175. return;
  176. nodedata_phys = __pa(node_data[nodeid]);
  177. printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
  178. nodedata_phys + pgdat_size - 1);
  179. memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
  180. NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid];
  181. NODE_DATA(nodeid)->node_start_pfn = start_pfn;
  182. NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
  183. /*
  184. * Find a place for the bootmem map
  185. * nodedata_phys could be on other nodes by alloc_bootmem,
  186. * so need to sure bootmap_start not to be small, otherwise
  187. * early_node_mem will get that with find_e820_area instead
  188. * of alloc_bootmem, that could clash with reserved range
  189. */
  190. bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn);
  191. nid = phys_to_nid(nodedata_phys);
  192. if (nid == nodeid)
  193. bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE);
  194. else
  195. bootmap_start = roundup(start, PAGE_SIZE);
  196. /*
  197. * SMP_CACHE_BYTES could be enough, but init_bootmem_node like
  198. * to use that to align to PAGE_SIZE
  199. */
  200. bootmap = early_node_mem(nodeid, bootmap_start, end,
  201. bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
  202. if (bootmap == NULL) {
  203. if (nodedata_phys < start || nodedata_phys >= end)
  204. free_bootmem(nodedata_phys, pgdat_size);
  205. node_data[nodeid] = NULL;
  206. return;
  207. }
  208. bootmap_start = __pa(bootmap);
  209. bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
  210. bootmap_start >> PAGE_SHIFT,
  211. start_pfn, last_pfn);
  212. printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n",
  213. bootmap_start, bootmap_start + bootmap_size - 1,
  214. bootmap_pages);
  215. free_bootmem_with_active_regions(nodeid, end);
  216. /*
  217. * convert early reserve to bootmem reserve earlier
  218. * otherwise early_node_mem could use early reserved mem
  219. * on previous node
  220. */
  221. early_res_to_bootmem(start, end);
  222. /*
  223. * in some case early_node_mem could use alloc_bootmem
  224. * to get range on other node, don't reserve that again
  225. */
  226. if (nid != nodeid)
  227. printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
  228. else
  229. reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys,
  230. pgdat_size, BOOTMEM_DEFAULT);
  231. nid = phys_to_nid(bootmap_start);
  232. if (nid != nodeid)
  233. printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid);
  234. else
  235. reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
  236. bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
  237. #ifdef CONFIG_ACPI_NUMA
  238. srat_reserve_add_area(nodeid);
  239. #endif
  240. node_set_online(nodeid);
  241. }
  242. /*
  243. * There are unfortunately some poorly designed mainboards around that
  244. * only connect memory to a single CPU. This breaks the 1:1 cpu->node
  245. * mapping. To avoid this fill in the mapping for all possible CPUs,
  246. * as the number of CPUs is not known yet. We round robin the existing
  247. * nodes.
  248. */
  249. void __init numa_init_array(void)
  250. {
  251. int rr, i;
  252. rr = first_node(node_online_map);
  253. for (i = 0; i < nr_cpu_ids; i++) {
  254. if (early_cpu_to_node(i) != NUMA_NO_NODE)
  255. continue;
  256. numa_set_node(i, rr);
  257. rr = next_node(rr, node_online_map);
  258. if (rr == MAX_NUMNODES)
  259. rr = first_node(node_online_map);
  260. }
  261. }
  262. #ifdef CONFIG_NUMA_EMU
  263. /* Numa emulation */
  264. static char *cmdline __initdata;
  265. /*
  266. * Setups up nid to range from addr to addr + size. If the end
  267. * boundary is greater than max_addr, then max_addr is used instead.
  268. * The return value is 0 if there is additional memory left for
  269. * allocation past addr and -1 otherwise. addr is adjusted to be at
  270. * the end of the node.
  271. */
  272. static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
  273. u64 size, u64 max_addr)
  274. {
  275. int ret = 0;
  276. nodes[nid].start = *addr;
  277. *addr += size;
  278. if (*addr >= max_addr) {
  279. *addr = max_addr;
  280. ret = -1;
  281. }
  282. nodes[nid].end = *addr;
  283. node_set(nid, node_possible_map);
  284. printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
  285. nodes[nid].start, nodes[nid].end,
  286. (nodes[nid].end - nodes[nid].start) >> 20);
  287. return ret;
  288. }
  289. /*
  290. * Splits num_nodes nodes up equally starting at node_start. The return value
  291. * is the number of nodes split up and addr is adjusted to be at the end of the
  292. * last node allocated.
  293. */
  294. static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
  295. u64 max_addr, int node_start,
  296. int num_nodes)
  297. {
  298. unsigned int big;
  299. u64 size;
  300. int i;
  301. if (num_nodes <= 0)
  302. return -1;
  303. if (num_nodes > MAX_NUMNODES)
  304. num_nodes = MAX_NUMNODES;
  305. size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
  306. num_nodes;
  307. /*
  308. * Calculate the number of big nodes that can be allocated as a result
  309. * of consolidating the leftovers.
  310. */
  311. big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
  312. FAKE_NODE_MIN_SIZE;
  313. /* Round down to nearest FAKE_NODE_MIN_SIZE. */
  314. size &= FAKE_NODE_MIN_HASH_MASK;
  315. if (!size) {
  316. printk(KERN_ERR "Not enough memory for each node. "
  317. "NUMA emulation disabled.\n");
  318. return -1;
  319. }
  320. for (i = node_start; i < num_nodes + node_start; i++) {
  321. u64 end = *addr + size;
  322. if (i < big)
  323. end += FAKE_NODE_MIN_SIZE;
  324. /*
  325. * The final node can have the remaining system RAM. Other
  326. * nodes receive roughly the same amount of available pages.
  327. */
  328. if (i == num_nodes + node_start - 1)
  329. end = max_addr;
  330. else
  331. while (end - *addr - e820_hole_size(*addr, end) <
  332. size) {
  333. end += FAKE_NODE_MIN_SIZE;
  334. if (end > max_addr) {
  335. end = max_addr;
  336. break;
  337. }
  338. }
  339. if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
  340. break;
  341. }
  342. return i - node_start + 1;
  343. }
  344. /*
  345. * Splits the remaining system RAM into chunks of size. The remaining memory is
  346. * always assigned to a final node and can be asymmetric. Returns the number of
  347. * nodes split.
  348. */
  349. static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
  350. u64 max_addr, int node_start, u64 size)
  351. {
  352. int i = node_start;
  353. size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
  354. while (!setup_node_range(i++, nodes, addr, size, max_addr))
  355. ;
  356. return i - node_start;
  357. }
  358. /*
  359. * Sets up the system RAM area from start_pfn to last_pfn according to the
  360. * numa=fake command-line option.
  361. */
  362. static struct bootnode nodes[MAX_NUMNODES] __initdata;
  363. static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn)
  364. {
  365. u64 size, addr = start_pfn << PAGE_SHIFT;
  366. u64 max_addr = last_pfn << PAGE_SHIFT;
  367. int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
  368. memset(&nodes, 0, sizeof(nodes));
  369. /*
  370. * If the numa=fake command-line is just a single number N, split the
  371. * system RAM into N fake nodes.
  372. */
  373. if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
  374. long n = simple_strtol(cmdline, NULL, 0);
  375. num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
  376. if (num_nodes < 0)
  377. return num_nodes;
  378. goto out;
  379. }
  380. /* Parse the command line. */
  381. for (coeff_flag = 0; ; cmdline++) {
  382. if (*cmdline && isdigit(*cmdline)) {
  383. num = num * 10 + *cmdline - '0';
  384. continue;
  385. }
  386. if (*cmdline == '*') {
  387. if (num > 0)
  388. coeff = num;
  389. coeff_flag = 1;
  390. }
  391. if (!*cmdline || *cmdline == ',') {
  392. if (!coeff_flag)
  393. coeff = 1;
  394. /*
  395. * Round down to the nearest FAKE_NODE_MIN_SIZE.
  396. * Command-line coefficients are in megabytes.
  397. */
  398. size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
  399. if (size)
  400. for (i = 0; i < coeff; i++, num_nodes++)
  401. if (setup_node_range(num_nodes, nodes,
  402. &addr, size, max_addr) < 0)
  403. goto done;
  404. if (!*cmdline)
  405. break;
  406. coeff_flag = 0;
  407. coeff = -1;
  408. }
  409. num = 0;
  410. }
  411. done:
  412. if (!num_nodes)
  413. return -1;
  414. /* Fill remainder of system RAM, if appropriate. */
  415. if (addr < max_addr) {
  416. if (coeff_flag && coeff < 0) {
  417. /* Split remaining nodes into num-sized chunks */
  418. num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
  419. num_nodes, num);
  420. goto out;
  421. }
  422. switch (*(cmdline - 1)) {
  423. case '*':
  424. /* Split remaining nodes into coeff chunks */
  425. if (coeff <= 0)
  426. break;
  427. num_nodes += split_nodes_equally(nodes, &addr, max_addr,
  428. num_nodes, coeff);
  429. break;
  430. case ',':
  431. /* Do not allocate remaining system RAM */
  432. break;
  433. default:
  434. /* Give one final node */
  435. setup_node_range(num_nodes, nodes, &addr,
  436. max_addr - addr, max_addr);
  437. num_nodes++;
  438. }
  439. }
  440. out:
  441. memnode_shift = compute_hash_shift(nodes, num_nodes, NULL);
  442. if (memnode_shift < 0) {
  443. memnode_shift = 0;
  444. printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
  445. "disabled.\n");
  446. return -1;
  447. }
  448. /*
  449. * We need to vacate all active ranges that may have been registered by
  450. * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
  451. * true. NUMA emulation has succeeded so we will not scan ACPI nodes.
  452. */
  453. remove_all_active_ranges();
  454. #ifdef CONFIG_ACPI_NUMA
  455. acpi_numa = -1;
  456. #endif
  457. for_each_node_mask(i, node_possible_map) {
  458. e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
  459. nodes[i].end >> PAGE_SHIFT);
  460. setup_node_bootmem(i, nodes[i].start, nodes[i].end);
  461. }
  462. acpi_fake_nodes(nodes, num_nodes);
  463. numa_init_array();
  464. return 0;
  465. }
  466. #endif /* CONFIG_NUMA_EMU */
  467. void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn)
  468. {
  469. int i;
  470. nodes_clear(node_possible_map);
  471. nodes_clear(node_online_map);
  472. #ifdef CONFIG_NUMA_EMU
  473. if (cmdline && !numa_emulation(start_pfn, last_pfn))
  474. return;
  475. nodes_clear(node_possible_map);
  476. nodes_clear(node_online_map);
  477. #endif
  478. #ifdef CONFIG_ACPI_NUMA
  479. if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
  480. last_pfn << PAGE_SHIFT))
  481. return;
  482. nodes_clear(node_possible_map);
  483. nodes_clear(node_online_map);
  484. #endif
  485. #ifdef CONFIG_K8_NUMA
  486. if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
  487. last_pfn<<PAGE_SHIFT))
  488. return;
  489. nodes_clear(node_possible_map);
  490. nodes_clear(node_online_map);
  491. #endif
  492. printk(KERN_INFO "%s\n",
  493. numa_off ? "NUMA turned off" : "No NUMA configuration found");
  494. printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
  495. start_pfn << PAGE_SHIFT,
  496. last_pfn << PAGE_SHIFT);
  497. /* setup dummy node covering all memory */
  498. memnode_shift = 63;
  499. memnodemap = memnode.embedded_map;
  500. memnodemap[0] = 0;
  501. node_set_online(0);
  502. node_set(0, node_possible_map);
  503. for (i = 0; i < nr_cpu_ids; i++)
  504. numa_set_node(i, 0);
  505. e820_register_active_regions(0, start_pfn, last_pfn);
  506. setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT);
  507. }
  508. unsigned long __init numa_free_all_bootmem(void)
  509. {
  510. unsigned long pages = 0;
  511. int i;
  512. for_each_online_node(i)
  513. pages += free_all_bootmem_node(NODE_DATA(i));
  514. return pages;
  515. }
  516. void __init paging_init(void)
  517. {
  518. unsigned long max_zone_pfns[MAX_NR_ZONES];
  519. memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
  520. max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
  521. max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
  522. max_zone_pfns[ZONE_NORMAL] = max_pfn;
  523. sparse_memory_present_with_active_regions(MAX_NUMNODES);
  524. sparse_init();
  525. free_area_init_nodes(max_zone_pfns);
  526. }
  527. static __init int numa_setup(char *opt)
  528. {
  529. if (!opt)
  530. return -EINVAL;
  531. if (!strncmp(opt, "off", 3))
  532. numa_off = 1;
  533. #ifdef CONFIG_NUMA_EMU
  534. if (!strncmp(opt, "fake=", 5))
  535. cmdline = opt + 5;
  536. #endif
  537. #ifdef CONFIG_ACPI_NUMA
  538. if (!strncmp(opt, "noacpi", 6))
  539. acpi_numa = -1;
  540. if (!strncmp(opt, "hotadd=", 7))
  541. hotadd_percent = simple_strtoul(opt+7, NULL, 10);
  542. #endif
  543. return 0;
  544. }
  545. early_param("numa", numa_setup);
  546. #ifdef CONFIG_NUMA
  547. /*
  548. * Setup early cpu_to_node.
  549. *
  550. * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
  551. * and apicid_to_node[] tables have valid entries for a CPU.
  552. * This means we skip cpu_to_node[] initialisation for NUMA
  553. * emulation and faking node case (when running a kernel compiled
  554. * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
  555. * is already initialized in a round robin manner at numa_init_array,
  556. * prior to this call, and this initialization is good enough
  557. * for the fake NUMA cases.
  558. *
  559. * Called before the per_cpu areas are setup.
  560. */
  561. void __init init_cpu_to_node(void)
  562. {
  563. int cpu;
  564. u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
  565. BUG_ON(cpu_to_apicid == NULL);
  566. for_each_possible_cpu(cpu) {
  567. int node;
  568. u16 apicid = cpu_to_apicid[cpu];
  569. if (apicid == BAD_APICID)
  570. continue;
  571. node = apicid_to_node[apicid];
  572. if (node == NUMA_NO_NODE)
  573. continue;
  574. if (!node_online(node))
  575. continue;
  576. numa_set_node(cpu, node);
  577. }
  578. }
  579. #endif
  580. void __cpuinit numa_set_node(int cpu, int node)
  581. {
  582. int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
  583. /* early setting, no percpu area yet */
  584. if (cpu_to_node_map) {
  585. cpu_to_node_map[cpu] = node;
  586. return;
  587. }
  588. #ifdef CONFIG_DEBUG_PER_CPU_MAPS
  589. if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
  590. printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
  591. dump_stack();
  592. return;
  593. }
  594. #endif
  595. per_cpu(x86_cpu_to_node_map, cpu) = node;
  596. if (node != NUMA_NO_NODE)
  597. per_cpu(node_number, cpu) = node;
  598. }
  599. void __cpuinit numa_clear_node(int cpu)
  600. {
  601. numa_set_node(cpu, NUMA_NO_NODE);
  602. }
  603. #ifndef CONFIG_DEBUG_PER_CPU_MAPS
  604. void __cpuinit numa_add_cpu(int cpu)
  605. {
  606. cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  607. }
  608. void __cpuinit numa_remove_cpu(int cpu)
  609. {
  610. cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
  611. }
  612. #else /* CONFIG_DEBUG_PER_CPU_MAPS */
  613. /*
  614. * --------- debug versions of the numa functions ---------
  615. */
  616. static void __cpuinit numa_set_cpumask(int cpu, int enable)
  617. {
  618. int node = early_cpu_to_node(cpu);
  619. struct cpumask *mask;
  620. char buf[64];
  621. mask = node_to_cpumask_map[node];
  622. if (mask == NULL) {
  623. printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node);
  624. dump_stack();
  625. return;
  626. }
  627. if (enable)
  628. cpumask_set_cpu(cpu, mask);
  629. else
  630. cpumask_clear_cpu(cpu, mask);
  631. cpulist_scnprintf(buf, sizeof(buf), mask);
  632. printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
  633. enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf);
  634. }
  635. void __cpuinit numa_add_cpu(int cpu)
  636. {
  637. numa_set_cpumask(cpu, 1);
  638. }
  639. void __cpuinit numa_remove_cpu(int cpu)
  640. {
  641. numa_set_cpumask(cpu, 0);
  642. }
  643. int cpu_to_node(int cpu)
  644. {
  645. if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
  646. printk(KERN_WARNING
  647. "cpu_to_node(%d): usage too early!\n", cpu);
  648. dump_stack();
  649. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  650. }
  651. return per_cpu(x86_cpu_to_node_map, cpu);
  652. }
  653. EXPORT_SYMBOL(cpu_to_node);
  654. /*
  655. * Same function as cpu_to_node() but used if called before the
  656. * per_cpu areas are setup.
  657. */
  658. int early_cpu_to_node(int cpu)
  659. {
  660. if (early_per_cpu_ptr(x86_cpu_to_node_map))
  661. return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
  662. if (!cpu_possible(cpu)) {
  663. printk(KERN_WARNING
  664. "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
  665. dump_stack();
  666. return NUMA_NO_NODE;
  667. }
  668. return per_cpu(x86_cpu_to_node_map, cpu);
  669. }
  670. /*
  671. * --------- end of debug versions of the numa functions ---------
  672. */
  673. #endif /* CONFIG_DEBUG_PER_CPU_MAPS */