pgtable.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463
  1. #include <linux/mm.h>
  2. #include <linux/gfp.h>
  3. #include <asm/pgalloc.h>
  4. #include <asm/pgtable.h>
  5. #include <asm/tlb.h>
  6. #include <asm/fixmap.h>
  7. #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
  8. #ifdef CONFIG_HIGHPTE
  9. #define PGALLOC_USER_GFP __GFP_HIGHMEM
  10. #else
  11. #define PGALLOC_USER_GFP 0
  12. #endif
  13. gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
  14. pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
  15. {
  16. return (pte_t *)__get_free_page(PGALLOC_GFP);
  17. }
  18. pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
  19. {
  20. struct page *pte;
  21. pte = alloc_pages(__userpte_alloc_gfp, 0);
  22. if (pte)
  23. pgtable_page_ctor(pte);
  24. return pte;
  25. }
  26. static int __init setup_userpte(char *arg)
  27. {
  28. if (!arg)
  29. return -EINVAL;
  30. /*
  31. * "userpte=nohigh" disables allocation of user pagetables in
  32. * high memory.
  33. */
  34. if (strcmp(arg, "nohigh") == 0)
  35. __userpte_alloc_gfp &= ~__GFP_HIGHMEM;
  36. else
  37. return -EINVAL;
  38. return 0;
  39. }
  40. early_param("userpte", setup_userpte);
  41. void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
  42. {
  43. pgtable_page_dtor(pte);
  44. paravirt_release_pte(page_to_pfn(pte));
  45. tlb_remove_page(tlb, pte);
  46. }
  47. #if PAGETABLE_LEVELS > 2
  48. void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
  49. {
  50. paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
  51. /*
  52. * NOTE! For PAE, any changes to the top page-directory-pointer-table
  53. * entries need a full cr3 reload to flush.
  54. */
  55. #ifdef CONFIG_X86_PAE
  56. tlb->need_flush_all = 1;
  57. #endif
  58. tlb_remove_page(tlb, virt_to_page(pmd));
  59. }
  60. #if PAGETABLE_LEVELS > 3
  61. void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
  62. {
  63. paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
  64. tlb_remove_page(tlb, virt_to_page(pud));
  65. }
  66. #endif /* PAGETABLE_LEVELS > 3 */
  67. #endif /* PAGETABLE_LEVELS > 2 */
  68. static inline void pgd_list_add(pgd_t *pgd)
  69. {
  70. struct page *page = virt_to_page(pgd);
  71. list_add(&page->lru, &pgd_list);
  72. }
  73. static inline void pgd_list_del(pgd_t *pgd)
  74. {
  75. struct page *page = virt_to_page(pgd);
  76. list_del(&page->lru);
  77. }
  78. #define UNSHARED_PTRS_PER_PGD \
  79. (SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
  80. static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
  81. {
  82. BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
  83. virt_to_page(pgd)->index = (pgoff_t)mm;
  84. }
  85. struct mm_struct *pgd_page_get_mm(struct page *page)
  86. {
  87. return (struct mm_struct *)page->index;
  88. }
  89. static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
  90. {
  91. /* If the pgd points to a shared pagetable level (either the
  92. ptes in non-PAE, or shared PMD in PAE), then just copy the
  93. references from swapper_pg_dir. */
  94. if (PAGETABLE_LEVELS == 2 ||
  95. (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
  96. PAGETABLE_LEVELS == 4) {
  97. clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
  98. swapper_pg_dir + KERNEL_PGD_BOUNDARY,
  99. KERNEL_PGD_PTRS);
  100. }
  101. /* list required to sync kernel mapping updates */
  102. if (!SHARED_KERNEL_PMD) {
  103. pgd_set_mm(pgd, mm);
  104. pgd_list_add(pgd);
  105. }
  106. }
  107. static void pgd_dtor(pgd_t *pgd)
  108. {
  109. if (SHARED_KERNEL_PMD)
  110. return;
  111. spin_lock(&pgd_lock);
  112. pgd_list_del(pgd);
  113. spin_unlock(&pgd_lock);
  114. }
  115. /*
  116. * List of all pgd's needed for non-PAE so it can invalidate entries
  117. * in both cached and uncached pgd's; not needed for PAE since the
  118. * kernel pmd is shared. If PAE were not to share the pmd a similar
  119. * tactic would be needed. This is essentially codepath-based locking
  120. * against pageattr.c; it is the unique case in which a valid change
  121. * of kernel pagetables can't be lazily synchronized by vmalloc faults.
  122. * vmalloc faults work because attached pagetables are never freed.
  123. * -- nyc
  124. */
  125. #ifdef CONFIG_X86_PAE
  126. /*
  127. * In PAE mode, we need to do a cr3 reload (=tlb flush) when
  128. * updating the top-level pagetable entries to guarantee the
  129. * processor notices the update. Since this is expensive, and
  130. * all 4 top-level entries are used almost immediately in a
  131. * new process's life, we just pre-populate them here.
  132. *
  133. * Also, if we're in a paravirt environment where the kernel pmd is
  134. * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
  135. * and initialize the kernel pmds here.
  136. */
  137. #define PREALLOCATED_PMDS UNSHARED_PTRS_PER_PGD
  138. void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
  139. {
  140. paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
  141. /* Note: almost everything apart from _PAGE_PRESENT is
  142. reserved at the pmd (PDPT) level. */
  143. set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
  144. /*
  145. * According to Intel App note "TLBs, Paging-Structure Caches,
  146. * and Their Invalidation", April 2007, document 317080-001,
  147. * section 8.1: in PAE mode we explicitly have to flush the
  148. * TLB via cr3 if the top-level pgd is changed...
  149. */
  150. flush_tlb_mm(mm);
  151. }
  152. #else /* !CONFIG_X86_PAE */
  153. /* No need to prepopulate any pagetable entries in non-PAE modes. */
  154. #define PREALLOCATED_PMDS 0
  155. #endif /* CONFIG_X86_PAE */
  156. static void free_pmds(pmd_t *pmds[])
  157. {
  158. int i;
  159. for(i = 0; i < PREALLOCATED_PMDS; i++)
  160. if (pmds[i])
  161. free_page((unsigned long)pmds[i]);
  162. }
  163. static int preallocate_pmds(pmd_t *pmds[])
  164. {
  165. int i;
  166. bool failed = false;
  167. for(i = 0; i < PREALLOCATED_PMDS; i++) {
  168. pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
  169. if (pmd == NULL)
  170. failed = true;
  171. pmds[i] = pmd;
  172. }
  173. if (failed) {
  174. free_pmds(pmds);
  175. return -ENOMEM;
  176. }
  177. return 0;
  178. }
  179. /*
  180. * Mop up any pmd pages which may still be attached to the pgd.
  181. * Normally they will be freed by munmap/exit_mmap, but any pmd we
  182. * preallocate which never got a corresponding vma will need to be
  183. * freed manually.
  184. */
  185. static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
  186. {
  187. int i;
  188. for(i = 0; i < PREALLOCATED_PMDS; i++) {
  189. pgd_t pgd = pgdp[i];
  190. if (pgd_val(pgd) != 0) {
  191. pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
  192. pgdp[i] = native_make_pgd(0);
  193. paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
  194. pmd_free(mm, pmd);
  195. }
  196. }
  197. }
  198. static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
  199. {
  200. pud_t *pud;
  201. int i;
  202. if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
  203. return;
  204. pud = pud_offset(pgd, 0);
  205. for (i = 0; i < PREALLOCATED_PMDS; i++, pud++) {
  206. pmd_t *pmd = pmds[i];
  207. if (i >= KERNEL_PGD_BOUNDARY)
  208. memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
  209. sizeof(pmd_t) * PTRS_PER_PMD);
  210. pud_populate(mm, pud, pmd);
  211. }
  212. }
  213. pgd_t *pgd_alloc(struct mm_struct *mm)
  214. {
  215. pgd_t *pgd;
  216. pmd_t *pmds[PREALLOCATED_PMDS];
  217. pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
  218. if (pgd == NULL)
  219. goto out;
  220. mm->pgd = pgd;
  221. if (preallocate_pmds(pmds) != 0)
  222. goto out_free_pgd;
  223. if (paravirt_pgd_alloc(mm) != 0)
  224. goto out_free_pmds;
  225. /*
  226. * Make sure that pre-populating the pmds is atomic with
  227. * respect to anything walking the pgd_list, so that they
  228. * never see a partially populated pgd.
  229. */
  230. spin_lock(&pgd_lock);
  231. pgd_ctor(mm, pgd);
  232. pgd_prepopulate_pmd(mm, pgd, pmds);
  233. spin_unlock(&pgd_lock);
  234. return pgd;
  235. out_free_pmds:
  236. free_pmds(pmds);
  237. out_free_pgd:
  238. free_page((unsigned long)pgd);
  239. out:
  240. return NULL;
  241. }
  242. void pgd_free(struct mm_struct *mm, pgd_t *pgd)
  243. {
  244. pgd_mop_up_pmds(mm, pgd);
  245. pgd_dtor(pgd);
  246. paravirt_pgd_free(mm, pgd);
  247. free_page((unsigned long)pgd);
  248. }
  249. /*
  250. * Used to set accessed or dirty bits in the page table entries
  251. * on other architectures. On x86, the accessed and dirty bits
  252. * are tracked by hardware. However, do_wp_page calls this function
  253. * to also make the pte writeable at the same time the dirty bit is
  254. * set. In that case we do actually need to write the PTE.
  255. */
  256. int ptep_set_access_flags(struct vm_area_struct *vma,
  257. unsigned long address, pte_t *ptep,
  258. pte_t entry, int dirty)
  259. {
  260. int changed = !pte_same(*ptep, entry);
  261. if (changed && dirty) {
  262. *ptep = entry;
  263. pte_update_defer(vma->vm_mm, address, ptep);
  264. }
  265. return changed;
  266. }
  267. #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  268. int pmdp_set_access_flags(struct vm_area_struct *vma,
  269. unsigned long address, pmd_t *pmdp,
  270. pmd_t entry, int dirty)
  271. {
  272. int changed = !pmd_same(*pmdp, entry);
  273. VM_BUG_ON(address & ~HPAGE_PMD_MASK);
  274. if (changed && dirty) {
  275. *pmdp = entry;
  276. pmd_update_defer(vma->vm_mm, address, pmdp);
  277. /*
  278. * We had a write-protection fault here and changed the pmd
  279. * to to more permissive. No need to flush the TLB for that,
  280. * #PF is architecturally guaranteed to do that and in the
  281. * worst-case we'll generate a spurious fault.
  282. */
  283. }
  284. return changed;
  285. }
  286. #endif
  287. int ptep_test_and_clear_young(struct vm_area_struct *vma,
  288. unsigned long addr, pte_t *ptep)
  289. {
  290. int ret = 0;
  291. if (pte_young(*ptep))
  292. ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
  293. (unsigned long *) &ptep->pte);
  294. if (ret)
  295. pte_update(vma->vm_mm, addr, ptep);
  296. return ret;
  297. }
  298. #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  299. int pmdp_test_and_clear_young(struct vm_area_struct *vma,
  300. unsigned long addr, pmd_t *pmdp)
  301. {
  302. int ret = 0;
  303. if (pmd_young(*pmdp))
  304. ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
  305. (unsigned long *)pmdp);
  306. if (ret)
  307. pmd_update(vma->vm_mm, addr, pmdp);
  308. return ret;
  309. }
  310. #endif
  311. int ptep_clear_flush_young(struct vm_area_struct *vma,
  312. unsigned long address, pte_t *ptep)
  313. {
  314. int young;
  315. young = ptep_test_and_clear_young(vma, address, ptep);
  316. if (young)
  317. flush_tlb_page(vma, address);
  318. return young;
  319. }
  320. #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  321. int pmdp_clear_flush_young(struct vm_area_struct *vma,
  322. unsigned long address, pmd_t *pmdp)
  323. {
  324. int young;
  325. VM_BUG_ON(address & ~HPAGE_PMD_MASK);
  326. young = pmdp_test_and_clear_young(vma, address, pmdp);
  327. if (young)
  328. flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
  329. return young;
  330. }
  331. void pmdp_splitting_flush(struct vm_area_struct *vma,
  332. unsigned long address, pmd_t *pmdp)
  333. {
  334. int set;
  335. VM_BUG_ON(address & ~HPAGE_PMD_MASK);
  336. set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
  337. (unsigned long *)pmdp);
  338. if (set) {
  339. pmd_update(vma->vm_mm, address, pmdp);
  340. /* need tlb flush only to serialize against gup-fast */
  341. flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
  342. }
  343. }
  344. #endif
  345. /**
  346. * reserve_top_address - reserves a hole in the top of kernel address space
  347. * @reserve - size of hole to reserve
  348. *
  349. * Can be used to relocate the fixmap area and poke a hole in the top
  350. * of kernel address space to make room for a hypervisor.
  351. */
  352. void __init reserve_top_address(unsigned long reserve)
  353. {
  354. #ifdef CONFIG_X86_32
  355. BUG_ON(fixmaps_set > 0);
  356. printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
  357. (int)-reserve);
  358. __FIXADDR_TOP = -reserve - PAGE_SIZE;
  359. #endif
  360. }
  361. int fixmaps_set;
  362. void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
  363. {
  364. unsigned long address = __fix_to_virt(idx);
  365. if (idx >= __end_of_fixed_addresses) {
  366. BUG();
  367. return;
  368. }
  369. set_pte_vaddr(address, pte);
  370. fixmaps_set++;
  371. }
  372. void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
  373. pgprot_t flags)
  374. {
  375. __native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
  376. }