hugetlbpage.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508
  1. // SPDX-License-Identifier: GPL-2.0
  2. /*
  3. * SPARC64 Huge TLB page support.
  4. *
  5. * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
  6. */
  7. #include <linux/fs.h>
  8. #include <linux/mm.h>
  9. #include <linux/sched/mm.h>
  10. #include <linux/hugetlb.h>
  11. #include <linux/pagemap.h>
  12. #include <linux/sysctl.h>
  13. #include <asm/mman.h>
  14. #include <asm/pgalloc.h>
  15. #include <asm/pgtable.h>
  16. #include <asm/tlb.h>
  17. #include <asm/tlbflush.h>
  18. #include <asm/cacheflush.h>
  19. #include <asm/mmu_context.h>
  20. /* Slightly simplified from the non-hugepage variant because by
  21. * definition we don't have to worry about any page coloring stuff
  22. */
  23. static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
  24. unsigned long addr,
  25. unsigned long len,
  26. unsigned long pgoff,
  27. unsigned long flags)
  28. {
  29. struct hstate *h = hstate_file(filp);
  30. unsigned long task_size = TASK_SIZE;
  31. struct vm_unmapped_area_info info;
  32. if (test_thread_flag(TIF_32BIT))
  33. task_size = STACK_TOP32;
  34. info.flags = 0;
  35. info.length = len;
  36. info.low_limit = TASK_UNMAPPED_BASE;
  37. info.high_limit = min(task_size, VA_EXCLUDE_START);
  38. info.align_mask = PAGE_MASK & ~huge_page_mask(h);
  39. info.align_offset = 0;
  40. addr = vm_unmapped_area(&info);
  41. if ((addr & ~PAGE_MASK) && task_size > VA_EXCLUDE_END) {
  42. VM_BUG_ON(addr != -ENOMEM);
  43. info.low_limit = VA_EXCLUDE_END;
  44. info.high_limit = task_size;
  45. addr = vm_unmapped_area(&info);
  46. }
  47. return addr;
  48. }
  49. static unsigned long
  50. hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
  51. const unsigned long len,
  52. const unsigned long pgoff,
  53. const unsigned long flags)
  54. {
  55. struct hstate *h = hstate_file(filp);
  56. struct mm_struct *mm = current->mm;
  57. unsigned long addr = addr0;
  58. struct vm_unmapped_area_info info;
  59. /* This should only ever run for 32-bit processes. */
  60. BUG_ON(!test_thread_flag(TIF_32BIT));
  61. info.flags = VM_UNMAPPED_AREA_TOPDOWN;
  62. info.length = len;
  63. info.low_limit = PAGE_SIZE;
  64. info.high_limit = mm->mmap_base;
  65. info.align_mask = PAGE_MASK & ~huge_page_mask(h);
  66. info.align_offset = 0;
  67. addr = vm_unmapped_area(&info);
  68. /*
  69. * A failed mmap() very likely causes application failure,
  70. * so fall back to the bottom-up function here. This scenario
  71. * can happen with large stack limits and large mmap()
  72. * allocations.
  73. */
  74. if (addr & ~PAGE_MASK) {
  75. VM_BUG_ON(addr != -ENOMEM);
  76. info.flags = 0;
  77. info.low_limit = TASK_UNMAPPED_BASE;
  78. info.high_limit = STACK_TOP32;
  79. addr = vm_unmapped_area(&info);
  80. }
  81. return addr;
  82. }
  83. unsigned long
  84. hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
  85. unsigned long len, unsigned long pgoff, unsigned long flags)
  86. {
  87. struct hstate *h = hstate_file(file);
  88. struct mm_struct *mm = current->mm;
  89. struct vm_area_struct *vma;
  90. unsigned long task_size = TASK_SIZE;
  91. if (test_thread_flag(TIF_32BIT))
  92. task_size = STACK_TOP32;
  93. if (len & ~huge_page_mask(h))
  94. return -EINVAL;
  95. if (len > task_size)
  96. return -ENOMEM;
  97. if (flags & MAP_FIXED) {
  98. if (prepare_hugepage_range(file, addr, len))
  99. return -EINVAL;
  100. return addr;
  101. }
  102. if (addr) {
  103. addr = ALIGN(addr, huge_page_size(h));
  104. vma = find_vma(mm, addr);
  105. if (task_size - len >= addr &&
  106. (!vma || addr + len <= vm_start_gap(vma)))
  107. return addr;
  108. }
  109. if (mm->get_unmapped_area == arch_get_unmapped_area)
  110. return hugetlb_get_unmapped_area_bottomup(file, addr, len,
  111. pgoff, flags);
  112. else
  113. return hugetlb_get_unmapped_area_topdown(file, addr, len,
  114. pgoff, flags);
  115. }
  116. static pte_t sun4u_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
  117. {
  118. return entry;
  119. }
  120. static pte_t sun4v_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
  121. {
  122. unsigned long hugepage_size = _PAGE_SZ4MB_4V;
  123. pte_val(entry) = pte_val(entry) & ~_PAGE_SZALL_4V;
  124. switch (shift) {
  125. case HPAGE_16GB_SHIFT:
  126. hugepage_size = _PAGE_SZ16GB_4V;
  127. pte_val(entry) |= _PAGE_PUD_HUGE;
  128. break;
  129. case HPAGE_2GB_SHIFT:
  130. hugepage_size = _PAGE_SZ2GB_4V;
  131. pte_val(entry) |= _PAGE_PMD_HUGE;
  132. break;
  133. case HPAGE_256MB_SHIFT:
  134. hugepage_size = _PAGE_SZ256MB_4V;
  135. pte_val(entry) |= _PAGE_PMD_HUGE;
  136. break;
  137. case HPAGE_SHIFT:
  138. pte_val(entry) |= _PAGE_PMD_HUGE;
  139. break;
  140. case HPAGE_64K_SHIFT:
  141. hugepage_size = _PAGE_SZ64K_4V;
  142. break;
  143. default:
  144. WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift);
  145. }
  146. pte_val(entry) = pte_val(entry) | hugepage_size;
  147. return entry;
  148. }
  149. static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift)
  150. {
  151. if (tlb_type == hypervisor)
  152. return sun4v_hugepage_shift_to_tte(entry, shift);
  153. else
  154. return sun4u_hugepage_shift_to_tte(entry, shift);
  155. }
  156. pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
  157. struct page *page, int writeable)
  158. {
  159. unsigned int shift = huge_page_shift(hstate_vma(vma));
  160. return hugepage_shift_to_tte(entry, shift);
  161. }
  162. static unsigned int sun4v_huge_tte_to_shift(pte_t entry)
  163. {
  164. unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V;
  165. unsigned int shift;
  166. switch (tte_szbits) {
  167. case _PAGE_SZ16GB_4V:
  168. shift = HPAGE_16GB_SHIFT;
  169. break;
  170. case _PAGE_SZ2GB_4V:
  171. shift = HPAGE_2GB_SHIFT;
  172. break;
  173. case _PAGE_SZ256MB_4V:
  174. shift = HPAGE_256MB_SHIFT;
  175. break;
  176. case _PAGE_SZ4MB_4V:
  177. shift = REAL_HPAGE_SHIFT;
  178. break;
  179. case _PAGE_SZ64K_4V:
  180. shift = HPAGE_64K_SHIFT;
  181. break;
  182. default:
  183. shift = PAGE_SHIFT;
  184. break;
  185. }
  186. return shift;
  187. }
  188. static unsigned int sun4u_huge_tte_to_shift(pte_t entry)
  189. {
  190. unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U;
  191. unsigned int shift;
  192. switch (tte_szbits) {
  193. case _PAGE_SZ256MB_4U:
  194. shift = HPAGE_256MB_SHIFT;
  195. break;
  196. case _PAGE_SZ4MB_4U:
  197. shift = REAL_HPAGE_SHIFT;
  198. break;
  199. case _PAGE_SZ64K_4U:
  200. shift = HPAGE_64K_SHIFT;
  201. break;
  202. default:
  203. shift = PAGE_SHIFT;
  204. break;
  205. }
  206. return shift;
  207. }
  208. static unsigned int huge_tte_to_shift(pte_t entry)
  209. {
  210. unsigned long shift;
  211. if (tlb_type == hypervisor)
  212. shift = sun4v_huge_tte_to_shift(entry);
  213. else
  214. shift = sun4u_huge_tte_to_shift(entry);
  215. if (shift == PAGE_SHIFT)
  216. WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n",
  217. pte_val(entry));
  218. return shift;
  219. }
  220. static unsigned long huge_tte_to_size(pte_t pte)
  221. {
  222. unsigned long size = 1UL << huge_tte_to_shift(pte);
  223. if (size == REAL_HPAGE_SIZE)
  224. size = HPAGE_SIZE;
  225. return size;
  226. }
  227. pte_t *huge_pte_alloc(struct mm_struct *mm,
  228. unsigned long addr, unsigned long sz)
  229. {
  230. pgd_t *pgd;
  231. pud_t *pud;
  232. pmd_t *pmd;
  233. pgd = pgd_offset(mm, addr);
  234. pud = pud_alloc(mm, pgd, addr);
  235. if (!pud)
  236. return NULL;
  237. if (sz >= PUD_SIZE)
  238. return (pte_t *)pud;
  239. pmd = pmd_alloc(mm, pud, addr);
  240. if (!pmd)
  241. return NULL;
  242. if (sz >= PMD_SIZE)
  243. return (pte_t *)pmd;
  244. return pte_alloc_map(mm, pmd, addr);
  245. }
  246. pte_t *huge_pte_offset(struct mm_struct *mm,
  247. unsigned long addr, unsigned long sz)
  248. {
  249. pgd_t *pgd;
  250. pud_t *pud;
  251. pmd_t *pmd;
  252. pgd = pgd_offset(mm, addr);
  253. if (pgd_none(*pgd))
  254. return NULL;
  255. pud = pud_offset(pgd, addr);
  256. if (pud_none(*pud))
  257. return NULL;
  258. if (is_hugetlb_pud(*pud))
  259. return (pte_t *)pud;
  260. pmd = pmd_offset(pud, addr);
  261. if (pmd_none(*pmd))
  262. return NULL;
  263. if (is_hugetlb_pmd(*pmd))
  264. return (pte_t *)pmd;
  265. return pte_offset_map(pmd, addr);
  266. }
  267. void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
  268. pte_t *ptep, pte_t entry)
  269. {
  270. unsigned int nptes, orig_shift, shift;
  271. unsigned long i, size;
  272. pte_t orig;
  273. size = huge_tte_to_size(entry);
  274. shift = PAGE_SHIFT;
  275. if (size >= PUD_SIZE)
  276. shift = PUD_SHIFT;
  277. else if (size >= PMD_SIZE)
  278. shift = PMD_SHIFT;
  279. else
  280. shift = PAGE_SHIFT;
  281. nptes = size >> shift;
  282. if (!pte_present(*ptep) && pte_present(entry))
  283. mm->context.hugetlb_pte_count += nptes;
  284. addr &= ~(size - 1);
  285. orig = *ptep;
  286. orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig);
  287. for (i = 0; i < nptes; i++)
  288. ptep[i] = __pte(pte_val(entry) + (i << shift));
  289. maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift);
  290. /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
  291. if (size == HPAGE_SIZE)
  292. maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0,
  293. orig_shift);
  294. }
  295. pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
  296. pte_t *ptep)
  297. {
  298. unsigned int i, nptes, orig_shift, shift;
  299. unsigned long size;
  300. pte_t entry;
  301. entry = *ptep;
  302. size = huge_tte_to_size(entry);
  303. shift = PAGE_SHIFT;
  304. if (size >= PUD_SIZE)
  305. shift = PUD_SHIFT;
  306. else if (size >= PMD_SIZE)
  307. shift = PMD_SHIFT;
  308. else
  309. shift = PAGE_SHIFT;
  310. nptes = size >> shift;
  311. orig_shift = pte_none(entry) ? PAGE_SHIFT : huge_tte_to_shift(entry);
  312. if (pte_present(entry))
  313. mm->context.hugetlb_pte_count -= nptes;
  314. addr &= ~(size - 1);
  315. for (i = 0; i < nptes; i++)
  316. ptep[i] = __pte(0UL);
  317. maybe_tlb_batch_add(mm, addr, ptep, entry, 0, orig_shift);
  318. /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
  319. if (size == HPAGE_SIZE)
  320. maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0,
  321. orig_shift);
  322. return entry;
  323. }
  324. int pmd_huge(pmd_t pmd)
  325. {
  326. return !pmd_none(pmd) &&
  327. (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID;
  328. }
  329. int pud_huge(pud_t pud)
  330. {
  331. return !pud_none(pud) &&
  332. (pud_val(pud) & (_PAGE_VALID|_PAGE_PUD_HUGE)) != _PAGE_VALID;
  333. }
  334. static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
  335. unsigned long addr)
  336. {
  337. pgtable_t token = pmd_pgtable(*pmd);
  338. pmd_clear(pmd);
  339. pte_free_tlb(tlb, token, addr);
  340. mm_dec_nr_ptes(tlb->mm);
  341. }
  342. static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
  343. unsigned long addr, unsigned long end,
  344. unsigned long floor, unsigned long ceiling)
  345. {
  346. pmd_t *pmd;
  347. unsigned long next;
  348. unsigned long start;
  349. start = addr;
  350. pmd = pmd_offset(pud, addr);
  351. do {
  352. next = pmd_addr_end(addr, end);
  353. if (pmd_none(*pmd))
  354. continue;
  355. if (is_hugetlb_pmd(*pmd))
  356. pmd_clear(pmd);
  357. else
  358. hugetlb_free_pte_range(tlb, pmd, addr);
  359. } while (pmd++, addr = next, addr != end);
  360. start &= PUD_MASK;
  361. if (start < floor)
  362. return;
  363. if (ceiling) {
  364. ceiling &= PUD_MASK;
  365. if (!ceiling)
  366. return;
  367. }
  368. if (end - 1 > ceiling - 1)
  369. return;
  370. pmd = pmd_offset(pud, start);
  371. pud_clear(pud);
  372. pmd_free_tlb(tlb, pmd, start);
  373. mm_dec_nr_pmds(tlb->mm);
  374. }
  375. static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
  376. unsigned long addr, unsigned long end,
  377. unsigned long floor, unsigned long ceiling)
  378. {
  379. pud_t *pud;
  380. unsigned long next;
  381. unsigned long start;
  382. start = addr;
  383. pud = pud_offset(pgd, addr);
  384. do {
  385. next = pud_addr_end(addr, end);
  386. if (pud_none_or_clear_bad(pud))
  387. continue;
  388. if (is_hugetlb_pud(*pud))
  389. pud_clear(pud);
  390. else
  391. hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
  392. ceiling);
  393. } while (pud++, addr = next, addr != end);
  394. start &= PGDIR_MASK;
  395. if (start < floor)
  396. return;
  397. if (ceiling) {
  398. ceiling &= PGDIR_MASK;
  399. if (!ceiling)
  400. return;
  401. }
  402. if (end - 1 > ceiling - 1)
  403. return;
  404. pud = pud_offset(pgd, start);
  405. pgd_clear(pgd);
  406. pud_free_tlb(tlb, pud, start);
  407. mm_dec_nr_puds(tlb->mm);
  408. }
  409. void hugetlb_free_pgd_range(struct mmu_gather *tlb,
  410. unsigned long addr, unsigned long end,
  411. unsigned long floor, unsigned long ceiling)
  412. {
  413. pgd_t *pgd;
  414. unsigned long next;
  415. addr &= PMD_MASK;
  416. if (addr < floor) {
  417. addr += PMD_SIZE;
  418. if (!addr)
  419. return;
  420. }
  421. if (ceiling) {
  422. ceiling &= PMD_MASK;
  423. if (!ceiling)
  424. return;
  425. }
  426. if (end - 1 > ceiling - 1)
  427. end -= PMD_SIZE;
  428. if (addr > end - 1)
  429. return;
  430. pgd = pgd_offset(tlb->mm, addr);
  431. do {
  432. next = pgd_addr_end(addr, end);
  433. if (pgd_none_or_clear_bad(pgd))
  434. continue;
  435. hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
  436. } while (pgd++, addr = next, addr != end);
  437. }