sparse.c 8.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362
  1. /*
  2. * sparse memory mappings.
  3. */
  4. #include <linux/mm.h>
  5. #include <linux/mmzone.h>
  6. #include <linux/bootmem.h>
  7. #include <linux/highmem.h>
  8. #include <linux/module.h>
  9. #include <linux/spinlock.h>
  10. #include <linux/vmalloc.h>
  11. #include <asm/dma.h>
  12. #include <asm/pgalloc.h>
  13. #include <asm/pgtable.h>
  14. /*
  15. * Permanent SPARSEMEM data:
  16. *
  17. * 1) mem_section - memory sections, mem_map's for valid memory
  18. */
  19. #ifdef CONFIG_SPARSEMEM_EXTREME
  20. struct mem_section *mem_section[NR_SECTION_ROOTS]
  21. ____cacheline_internodealigned_in_smp;
  22. #else
  23. struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  24. ____cacheline_internodealigned_in_smp;
  25. #endif
  26. EXPORT_SYMBOL(mem_section);
  27. #ifdef NODE_NOT_IN_PAGE_FLAGS
  28. /*
  29. * If we did not store the node number in the page then we have to
  30. * do a lookup in the section_to_node_table in order to find which
  31. * node the page belongs to.
  32. */
  33. #if MAX_NUMNODES <= 256
  34. static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  35. #else
  36. static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  37. #endif
  38. int page_to_nid(struct page *page)
  39. {
  40. return section_to_node_table[page_to_section(page)];
  41. }
  42. EXPORT_SYMBOL(page_to_nid);
  43. static void set_section_nid(unsigned long section_nr, int nid)
  44. {
  45. section_to_node_table[section_nr] = nid;
  46. }
  47. #else /* !NODE_NOT_IN_PAGE_FLAGS */
  48. static inline void set_section_nid(unsigned long section_nr, int nid)
  49. {
  50. }
  51. #endif
  52. #ifdef CONFIG_SPARSEMEM_EXTREME
  53. static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  54. {
  55. struct mem_section *section = NULL;
  56. unsigned long array_size = SECTIONS_PER_ROOT *
  57. sizeof(struct mem_section);
  58. if (slab_is_available())
  59. section = kmalloc_node(array_size, GFP_KERNEL, nid);
  60. else
  61. section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  62. if (section)
  63. memset(section, 0, array_size);
  64. return section;
  65. }
  66. static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  67. {
  68. static DEFINE_SPINLOCK(index_init_lock);
  69. unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  70. struct mem_section *section;
  71. int ret = 0;
  72. if (mem_section[root])
  73. return -EEXIST;
  74. section = sparse_index_alloc(nid);
  75. /*
  76. * This lock keeps two different sections from
  77. * reallocating for the same index
  78. */
  79. spin_lock(&index_init_lock);
  80. if (mem_section[root]) {
  81. ret = -EEXIST;
  82. goto out;
  83. }
  84. mem_section[root] = section;
  85. out:
  86. spin_unlock(&index_init_lock);
  87. return ret;
  88. }
  89. #else /* !SPARSEMEM_EXTREME */
  90. static inline int sparse_index_init(unsigned long section_nr, int nid)
  91. {
  92. return 0;
  93. }
  94. #endif
  95. /*
  96. * Although written for the SPARSEMEM_EXTREME case, this happens
  97. * to also work for the flat array case because
  98. * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
  99. */
  100. int __section_nr(struct mem_section* ms)
  101. {
  102. unsigned long root_nr;
  103. struct mem_section* root;
  104. for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
  105. root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
  106. if (!root)
  107. continue;
  108. if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
  109. break;
  110. }
  111. return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
  112. }
  113. /*
  114. * During early boot, before section_mem_map is used for an actual
  115. * mem_map, we use section_mem_map to store the section's NUMA
  116. * node. This keeps us from having to use another data structure. The
  117. * node information is cleared just before we store the real mem_map.
  118. */
  119. static inline unsigned long sparse_encode_early_nid(int nid)
  120. {
  121. return (nid << SECTION_NID_SHIFT);
  122. }
  123. static inline int sparse_early_nid(struct mem_section *section)
  124. {
  125. return (section->section_mem_map >> SECTION_NID_SHIFT);
  126. }
  127. /* Record a memory area against a node. */
  128. void __init memory_present(int nid, unsigned long start, unsigned long end)
  129. {
  130. unsigned long pfn;
  131. start &= PAGE_SECTION_MASK;
  132. for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
  133. unsigned long section = pfn_to_section_nr(pfn);
  134. struct mem_section *ms;
  135. sparse_index_init(section, nid);
  136. set_section_nid(section, nid);
  137. ms = __nr_to_section(section);
  138. if (!ms->section_mem_map)
  139. ms->section_mem_map = sparse_encode_early_nid(nid) |
  140. SECTION_MARKED_PRESENT;
  141. }
  142. }
  143. /*
  144. * Only used by the i386 NUMA architecures, but relatively
  145. * generic code.
  146. */
  147. unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
  148. unsigned long end_pfn)
  149. {
  150. unsigned long pfn;
  151. unsigned long nr_pages = 0;
  152. for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
  153. if (nid != early_pfn_to_nid(pfn))
  154. continue;
  155. if (pfn_present(pfn))
  156. nr_pages += PAGES_PER_SECTION;
  157. }
  158. return nr_pages * sizeof(struct page);
  159. }
  160. /*
  161. * Subtle, we encode the real pfn into the mem_map such that
  162. * the identity pfn - section_mem_map will return the actual
  163. * physical page frame number.
  164. */
  165. static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
  166. {
  167. return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
  168. }
  169. /*
  170. * We need this if we ever free the mem_maps. While not implemented yet,
  171. * this function is included for parity with its sibling.
  172. */
  173. static __attribute((unused))
  174. struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
  175. {
  176. return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
  177. }
  178. static int __meminit sparse_init_one_section(struct mem_section *ms,
  179. unsigned long pnum, struct page *mem_map)
  180. {
  181. if (!present_section(ms))
  182. return -EINVAL;
  183. ms->section_mem_map &= ~SECTION_MAP_MASK;
  184. ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
  185. SECTION_HAS_MEM_MAP;
  186. return 1;
  187. }
  188. __attribute__((weak)) __init
  189. void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size)
  190. {
  191. return NULL;
  192. }
  193. #ifndef CONFIG_SPARSEMEM_VMEMMAP
  194. struct page __init *sparse_early_mem_map_populate(unsigned long pnum, int nid)
  195. {
  196. struct page *map;
  197. map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
  198. if (map)
  199. return map;
  200. map = alloc_bootmem_high_node(NODE_DATA(nid),
  201. sizeof(struct page) * PAGES_PER_SECTION);
  202. if (map)
  203. return map;
  204. map = alloc_bootmem_node(NODE_DATA(nid),
  205. sizeof(struct page) * PAGES_PER_SECTION);
  206. return map;
  207. }
  208. #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
  209. struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
  210. {
  211. struct page *map;
  212. struct mem_section *ms = __nr_to_section(pnum);
  213. int nid = sparse_early_nid(ms);
  214. map = sparse_early_mem_map_populate(pnum, nid);
  215. if (map)
  216. return map;
  217. printk(KERN_ERR "%s: sparsemem memory map backing failed "
  218. "some memory will not be available.\n", __FUNCTION__);
  219. ms->section_mem_map = 0;
  220. return NULL;
  221. }
  222. /*
  223. * Allocate the accumulated non-linear sections, allocate a mem_map
  224. * for each and record the physical to section mapping.
  225. */
  226. void __init sparse_init(void)
  227. {
  228. unsigned long pnum;
  229. struct page *map;
  230. for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
  231. if (!present_section_nr(pnum))
  232. continue;
  233. map = sparse_early_mem_map_alloc(pnum);
  234. if (!map)
  235. continue;
  236. sparse_init_one_section(__nr_to_section(pnum), pnum, map);
  237. }
  238. }
  239. #ifdef CONFIG_MEMORY_HOTPLUG
  240. static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
  241. {
  242. struct page *page, *ret;
  243. unsigned long memmap_size = sizeof(struct page) * nr_pages;
  244. page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
  245. if (page)
  246. goto got_map_page;
  247. ret = vmalloc(memmap_size);
  248. if (ret)
  249. goto got_map_ptr;
  250. return NULL;
  251. got_map_page:
  252. ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
  253. got_map_ptr:
  254. memset(ret, 0, memmap_size);
  255. return ret;
  256. }
  257. static int vaddr_in_vmalloc_area(void *addr)
  258. {
  259. if (addr >= (void *)VMALLOC_START &&
  260. addr < (void *)VMALLOC_END)
  261. return 1;
  262. return 0;
  263. }
  264. static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
  265. {
  266. if (vaddr_in_vmalloc_area(memmap))
  267. vfree(memmap);
  268. else
  269. free_pages((unsigned long)memmap,
  270. get_order(sizeof(struct page) * nr_pages));
  271. }
  272. /*
  273. * returns the number of sections whose mem_maps were properly
  274. * set. If this is <=0, then that means that the passed-in
  275. * map was not consumed and must be freed.
  276. */
  277. int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
  278. int nr_pages)
  279. {
  280. unsigned long section_nr = pfn_to_section_nr(start_pfn);
  281. struct pglist_data *pgdat = zone->zone_pgdat;
  282. struct mem_section *ms;
  283. struct page *memmap;
  284. unsigned long flags;
  285. int ret;
  286. /*
  287. * no locking for this, because it does its own
  288. * plus, it does a kmalloc
  289. */
  290. sparse_index_init(section_nr, pgdat->node_id);
  291. memmap = __kmalloc_section_memmap(nr_pages);
  292. pgdat_resize_lock(pgdat, &flags);
  293. ms = __pfn_to_section(start_pfn);
  294. if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
  295. ret = -EEXIST;
  296. goto out;
  297. }
  298. ms->section_mem_map |= SECTION_MARKED_PRESENT;
  299. ret = sparse_init_one_section(ms, section_nr, memmap);
  300. out:
  301. pgdat_resize_unlock(pgdat, &flags);
  302. if (ret <= 0)
  303. __kfree_section_memmap(memmap, nr_pages);
  304. return ret;
  305. }
  306. #endif