page_ext.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427
  1. // SPDX-License-Identifier: GPL-2.0
  2. #include <linux/mm.h>
  3. #include <linux/mmzone.h>
  4. #include <linux/memblock.h>
  5. #include <linux/page_ext.h>
  6. #include <linux/memory.h>
  7. #include <linux/vmalloc.h>
  8. #include <linux/kmemleak.h>
  9. #include <linux/page_owner.h>
  10. #include <linux/page_idle.h>
  11. /*
  12. * struct page extension
  13. *
  14. * This is the feature to manage memory for extended data per page.
  15. *
  16. * Until now, we must modify struct page itself to store extra data per page.
  17. * This requires rebuilding the kernel and it is really time consuming process.
  18. * And, sometimes, rebuild is impossible due to third party module dependency.
  19. * At last, enlarging struct page could cause un-wanted system behaviour change.
  20. *
  21. * This feature is intended to overcome above mentioned problems. This feature
  22. * allocates memory for extended data per page in certain place rather than
  23. * the struct page itself. This memory can be accessed by the accessor
  24. * functions provided by this code. During the boot process, it checks whether
  25. * allocation of huge chunk of memory is needed or not. If not, it avoids
  26. * allocating memory at all. With this advantage, we can include this feature
  27. * into the kernel in default and can avoid rebuild and solve related problems.
  28. *
  29. * To help these things to work well, there are two callbacks for clients. One
  30. * is the need callback which is mandatory if user wants to avoid useless
  31. * memory allocation at boot-time. The other is optional, init callback, which
  32. * is used to do proper initialization after memory is allocated.
  33. *
  34. * The need callback is used to decide whether extended memory allocation is
  35. * needed or not. Sometimes users want to deactivate some features in this
  36. * boot and extra memory would be unneccessary. In this case, to avoid
  37. * allocating huge chunk of memory, each clients represent their need of
  38. * extra memory through the need callback. If one of the need callbacks
  39. * returns true, it means that someone needs extra memory so that
  40. * page extension core should allocates memory for page extension. If
  41. * none of need callbacks return true, memory isn't needed at all in this boot
  42. * and page extension core can skip to allocate memory. As result,
  43. * none of memory is wasted.
  44. *
  45. * When need callback returns true, page_ext checks if there is a request for
  46. * extra memory through size in struct page_ext_operations. If it is non-zero,
  47. * extra space is allocated for each page_ext entry and offset is returned to
  48. * user through offset in struct page_ext_operations.
  49. *
  50. * The init callback is used to do proper initialization after page extension
  51. * is completely initialized. In sparse memory system, extra memory is
  52. * allocated some time later than memmap is allocated. In other words, lifetime
  53. * of memory for page extension isn't same with memmap for struct page.
  54. * Therefore, clients can't store extra data until page extension is
  55. * initialized, even if pages are allocated and used freely. This could
  56. * cause inadequate state of extra data per page, so, to prevent it, client
  57. * can utilize this callback to initialize the state of it correctly.
  58. */
  59. #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
  60. static bool need_page_idle(void)
  61. {
  62. return true;
  63. }
  64. struct page_ext_operations page_idle_ops = {
  65. .need = need_page_idle,
  66. };
  67. #endif
  68. static struct page_ext_operations *page_ext_ops[] = {
  69. #ifdef CONFIG_PAGE_OWNER
  70. &page_owner_ops,
  71. #endif
  72. #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
  73. &page_idle_ops,
  74. #endif
  75. #ifdef CONFIG_PAGE_PINNER
  76. &page_pinner_ops,
  77. #endif
  78. };
  79. unsigned long page_ext_size = sizeof(struct page_ext);
  80. static unsigned long total_usage;
  81. static bool __init invoke_need_callbacks(void)
  82. {
  83. int i;
  84. int entries = ARRAY_SIZE(page_ext_ops);
  85. bool need = false;
  86. for (i = 0; i < entries; i++) {
  87. if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
  88. page_ext_ops[i]->offset = page_ext_size;
  89. page_ext_size += page_ext_ops[i]->size;
  90. need = true;
  91. }
  92. }
  93. return need;
  94. }
  95. static void __init invoke_init_callbacks(void)
  96. {
  97. int i;
  98. int entries = ARRAY_SIZE(page_ext_ops);
  99. for (i = 0; i < entries; i++) {
  100. if (page_ext_ops[i]->init)
  101. page_ext_ops[i]->init();
  102. }
  103. }
  104. static inline struct page_ext *get_entry(void *base, unsigned long index)
  105. {
  106. return base + page_ext_size * index;
  107. }
  108. #if !defined(CONFIG_SPARSEMEM)
  109. void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
  110. {
  111. pgdat->node_page_ext = NULL;
  112. }
  113. struct page_ext *lookup_page_ext(const struct page *page)
  114. {
  115. unsigned long pfn = page_to_pfn(page);
  116. unsigned long index;
  117. struct page_ext *base;
  118. base = NODE_DATA(page_to_nid(page))->node_page_ext;
  119. /*
  120. * The sanity checks the page allocator does upon freeing a
  121. * page can reach here before the page_ext arrays are
  122. * allocated when feeding a range of pages to the allocator
  123. * for the first time during bootup or memory hotplug.
  124. */
  125. if (unlikely(!base))
  126. return NULL;
  127. index = pfn - round_down(node_start_pfn(page_to_nid(page)),
  128. MAX_ORDER_NR_PAGES);
  129. return get_entry(base, index);
  130. }
  131. EXPORT_SYMBOL_GPL(lookup_page_ext);
  132. static int __init alloc_node_page_ext(int nid)
  133. {
  134. struct page_ext *base;
  135. unsigned long table_size;
  136. unsigned long nr_pages;
  137. nr_pages = NODE_DATA(nid)->node_spanned_pages;
  138. if (!nr_pages)
  139. return 0;
  140. /*
  141. * Need extra space if node range is not aligned with
  142. * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
  143. * checks buddy's status, range could be out of exact node range.
  144. */
  145. if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
  146. !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
  147. nr_pages += MAX_ORDER_NR_PAGES;
  148. table_size = page_ext_size * nr_pages;
  149. base = memblock_alloc_try_nid(
  150. table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
  151. MEMBLOCK_ALLOC_ACCESSIBLE, nid);
  152. if (!base)
  153. return -ENOMEM;
  154. NODE_DATA(nid)->node_page_ext = base;
  155. total_usage += table_size;
  156. return 0;
  157. }
  158. void __init page_ext_init_flatmem(void)
  159. {
  160. int nid, fail;
  161. if (!invoke_need_callbacks())
  162. return;
  163. for_each_online_node(nid) {
  164. fail = alloc_node_page_ext(nid);
  165. if (fail)
  166. goto fail;
  167. }
  168. pr_info("allocated %ld bytes of page_ext\n", total_usage);
  169. invoke_init_callbacks();
  170. return;
  171. fail:
  172. pr_crit("allocation of page_ext failed.\n");
  173. panic("Out of memory");
  174. }
  175. #else /* CONFIG_FLAT_NODE_MEM_MAP */
  176. struct page_ext *lookup_page_ext(const struct page *page)
  177. {
  178. unsigned long pfn = page_to_pfn(page);
  179. struct mem_section *section = __pfn_to_section(pfn);
  180. /*
  181. * The sanity checks the page allocator does upon freeing a
  182. * page can reach here before the page_ext arrays are
  183. * allocated when feeding a range of pages to the allocator
  184. * for the first time during bootup or memory hotplug.
  185. */
  186. if (!section->page_ext)
  187. return NULL;
  188. return get_entry(section->page_ext, pfn);
  189. }
  190. EXPORT_SYMBOL_GPL(lookup_page_ext);
  191. static void *__meminit alloc_page_ext(size_t size, int nid)
  192. {
  193. gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
  194. void *addr = NULL;
  195. addr = alloc_pages_exact_nid(nid, size, flags);
  196. if (addr) {
  197. kmemleak_alloc(addr, size, 1, flags);
  198. return addr;
  199. }
  200. addr = vzalloc_node(size, nid);
  201. return addr;
  202. }
  203. static int __meminit init_section_page_ext(unsigned long pfn, int nid)
  204. {
  205. struct mem_section *section;
  206. struct page_ext *base;
  207. unsigned long table_size;
  208. section = __pfn_to_section(pfn);
  209. if (section->page_ext)
  210. return 0;
  211. table_size = page_ext_size * PAGES_PER_SECTION;
  212. base = alloc_page_ext(table_size, nid);
  213. /*
  214. * The value stored in section->page_ext is (base - pfn)
  215. * and it does not point to the memory block allocated above,
  216. * causing kmemleak false positives.
  217. */
  218. kmemleak_not_leak(base);
  219. if (!base) {
  220. pr_err("page ext allocation failure\n");
  221. return -ENOMEM;
  222. }
  223. /*
  224. * The passed "pfn" may not be aligned to SECTION. For the calculation
  225. * we need to apply a mask.
  226. */
  227. pfn &= PAGE_SECTION_MASK;
  228. section->page_ext = (void *)base - page_ext_size * pfn;
  229. total_usage += table_size;
  230. return 0;
  231. }
  232. #ifdef CONFIG_MEMORY_HOTPLUG
  233. static void free_page_ext(void *addr)
  234. {
  235. if (is_vmalloc_addr(addr)) {
  236. vfree(addr);
  237. } else {
  238. struct page *page = virt_to_page(addr);
  239. size_t table_size;
  240. table_size = page_ext_size * PAGES_PER_SECTION;
  241. BUG_ON(PageReserved(page));
  242. kmemleak_free(addr);
  243. free_pages_exact(addr, table_size);
  244. }
  245. }
  246. static void __free_page_ext(unsigned long pfn)
  247. {
  248. struct mem_section *ms;
  249. struct page_ext *base;
  250. ms = __pfn_to_section(pfn);
  251. if (!ms || !ms->page_ext)
  252. return;
  253. base = get_entry(ms->page_ext, pfn);
  254. free_page_ext(base);
  255. ms->page_ext = NULL;
  256. }
  257. static int __meminit online_page_ext(unsigned long start_pfn,
  258. unsigned long nr_pages,
  259. int nid)
  260. {
  261. unsigned long start, end, pfn;
  262. int fail = 0;
  263. start = SECTION_ALIGN_DOWN(start_pfn);
  264. end = SECTION_ALIGN_UP(start_pfn + nr_pages);
  265. if (nid == NUMA_NO_NODE) {
  266. /*
  267. * In this case, "nid" already exists and contains valid memory.
  268. * "start_pfn" passed to us is a pfn which is an arg for
  269. * online__pages(), and start_pfn should exist.
  270. */
  271. nid = pfn_to_nid(start_pfn);
  272. VM_BUG_ON(!node_state(nid, N_ONLINE));
  273. }
  274. for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION)
  275. fail = init_section_page_ext(pfn, nid);
  276. if (!fail)
  277. return 0;
  278. /* rollback */
  279. for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
  280. __free_page_ext(pfn);
  281. return -ENOMEM;
  282. }
  283. static int __meminit offline_page_ext(unsigned long start_pfn,
  284. unsigned long nr_pages, int nid)
  285. {
  286. unsigned long start, end, pfn;
  287. start = SECTION_ALIGN_DOWN(start_pfn);
  288. end = SECTION_ALIGN_UP(start_pfn + nr_pages);
  289. for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
  290. __free_page_ext(pfn);
  291. return 0;
  292. }
  293. static int __meminit page_ext_callback(struct notifier_block *self,
  294. unsigned long action, void *arg)
  295. {
  296. struct memory_notify *mn = arg;
  297. int ret = 0;
  298. switch (action) {
  299. case MEM_GOING_ONLINE:
  300. ret = online_page_ext(mn->start_pfn,
  301. mn->nr_pages, mn->status_change_nid);
  302. break;
  303. case MEM_OFFLINE:
  304. offline_page_ext(mn->start_pfn,
  305. mn->nr_pages, mn->status_change_nid);
  306. break;
  307. case MEM_CANCEL_ONLINE:
  308. offline_page_ext(mn->start_pfn,
  309. mn->nr_pages, mn->status_change_nid);
  310. break;
  311. case MEM_GOING_OFFLINE:
  312. break;
  313. case MEM_ONLINE:
  314. case MEM_CANCEL_OFFLINE:
  315. break;
  316. }
  317. return notifier_from_errno(ret);
  318. }
  319. #endif
  320. void __init page_ext_init(void)
  321. {
  322. unsigned long pfn;
  323. int nid;
  324. if (!invoke_need_callbacks())
  325. return;
  326. for_each_node_state(nid, N_MEMORY) {
  327. unsigned long start_pfn, end_pfn;
  328. start_pfn = node_start_pfn(nid);
  329. end_pfn = node_end_pfn(nid);
  330. /*
  331. * start_pfn and end_pfn may not be aligned to SECTION and the
  332. * page->flags of out of node pages are not initialized. So we
  333. * scan [start_pfn, the biggest section's pfn < end_pfn) here.
  334. */
  335. for (pfn = start_pfn; pfn < end_pfn;
  336. pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
  337. if (!pfn_valid(pfn))
  338. continue;
  339. /*
  340. * Nodes's pfns can be overlapping.
  341. * We know some arch can have a nodes layout such as
  342. * -------------pfn-------------->
  343. * N0 | N1 | N2 | N0 | N1 | N2|....
  344. */
  345. if (pfn_to_nid(pfn) != nid)
  346. continue;
  347. if (init_section_page_ext(pfn, nid))
  348. goto oom;
  349. cond_resched();
  350. }
  351. }
  352. hotplug_memory_notifier(page_ext_callback, 0);
  353. pr_info("allocated %ld bytes of page_ext\n", total_usage);
  354. invoke_init_callbacks();
  355. return;
  356. oom:
  357. panic("Out of memory");
  358. }
  359. void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
  360. {
  361. }
  362. #endif