hmm.c 17 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599
  1. // SPDX-License-Identifier: GPL-2.0-or-later
  2. /*
  3. * Copyright 2013 Red Hat Inc.
  4. *
  5. * Authors: Jérôme Glisse <jglisse@redhat.com>
  6. */
  7. /*
  8. * Refer to include/linux/hmm.h for information about heterogeneous memory
  9. * management or HMM for short.
  10. */
  11. #include <linux/pagewalk.h>
  12. #include <linux/hmm.h>
  13. #include <linux/init.h>
  14. #include <linux/rmap.h>
  15. #include <linux/swap.h>
  16. #include <linux/slab.h>
  17. #include <linux/sched.h>
  18. #include <linux/mmzone.h>
  19. #include <linux/pagemap.h>
  20. #include <linux/swapops.h>
  21. #include <linux/hugetlb.h>
  22. #include <linux/memremap.h>
  23. #include <linux/sched/mm.h>
  24. #include <linux/jump_label.h>
  25. #include <linux/dma-mapping.h>
  26. #include <linux/mmu_notifier.h>
  27. #include <linux/memory_hotplug.h>
  28. struct hmm_vma_walk {
  29. struct hmm_range *range;
  30. unsigned long last;
  31. };
  32. enum {
  33. HMM_NEED_FAULT = 1 << 0,
  34. HMM_NEED_WRITE_FAULT = 1 << 1,
  35. HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
  36. };
  37. static int hmm_pfns_fill(unsigned long addr, unsigned long end,
  38. struct hmm_range *range, unsigned long cpu_flags)
  39. {
  40. unsigned long i = (addr - range->start) >> PAGE_SHIFT;
  41. for (; addr < end; addr += PAGE_SIZE, i++)
  42. range->hmm_pfns[i] = cpu_flags;
  43. return 0;
  44. }
  45. /*
  46. * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
  47. * @addr: range virtual start address (inclusive)
  48. * @end: range virtual end address (exclusive)
  49. * @required_fault: HMM_NEED_* flags
  50. * @walk: mm_walk structure
  51. * Return: -EBUSY after page fault, or page fault error
  52. *
  53. * This function will be called whenever pmd_none() or pte_none() returns true,
  54. * or whenever there is no page directory covering the virtual address range.
  55. */
  56. static int hmm_vma_fault(unsigned long addr, unsigned long end,
  57. unsigned int required_fault, struct mm_walk *walk)
  58. {
  59. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  60. struct vm_area_struct *vma = walk->vma;
  61. unsigned int fault_flags = FAULT_FLAG_REMOTE;
  62. WARN_ON_ONCE(!required_fault);
  63. hmm_vma_walk->last = addr;
  64. if (required_fault & HMM_NEED_WRITE_FAULT) {
  65. if (!(vma->vm_flags & VM_WRITE))
  66. return -EPERM;
  67. fault_flags |= FAULT_FLAG_WRITE;
  68. }
  69. for (; addr < end; addr += PAGE_SIZE)
  70. if (handle_mm_fault(vma, addr, fault_flags, NULL) &
  71. VM_FAULT_ERROR)
  72. return -EFAULT;
  73. return -EBUSY;
  74. }
  75. static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
  76. unsigned long pfn_req_flags,
  77. unsigned long cpu_flags)
  78. {
  79. struct hmm_range *range = hmm_vma_walk->range;
  80. /*
  81. * So we not only consider the individual per page request we also
  82. * consider the default flags requested for the range. The API can
  83. * be used 2 ways. The first one where the HMM user coalesces
  84. * multiple page faults into one request and sets flags per pfn for
  85. * those faults. The second one where the HMM user wants to pre-
  86. * fault a range with specific flags. For the latter one it is a
  87. * waste to have the user pre-fill the pfn arrays with a default
  88. * flags value.
  89. */
  90. pfn_req_flags &= range->pfn_flags_mask;
  91. pfn_req_flags |= range->default_flags;
  92. /* We aren't ask to do anything ... */
  93. if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
  94. return 0;
  95. /* Need to write fault ? */
  96. if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
  97. !(cpu_flags & HMM_PFN_WRITE))
  98. return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
  99. /* If CPU page table is not valid then we need to fault */
  100. if (!(cpu_flags & HMM_PFN_VALID))
  101. return HMM_NEED_FAULT;
  102. return 0;
  103. }
  104. static unsigned int
  105. hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
  106. const unsigned long hmm_pfns[], unsigned long npages,
  107. unsigned long cpu_flags)
  108. {
  109. struct hmm_range *range = hmm_vma_walk->range;
  110. unsigned int required_fault = 0;
  111. unsigned long i;
  112. /*
  113. * If the default flags do not request to fault pages, and the mask does
  114. * not allow for individual pages to be faulted, then
  115. * hmm_pte_need_fault() will always return 0.
  116. */
  117. if (!((range->default_flags | range->pfn_flags_mask) &
  118. HMM_PFN_REQ_FAULT))
  119. return 0;
  120. for (i = 0; i < npages; ++i) {
  121. required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
  122. cpu_flags);
  123. if (required_fault == HMM_NEED_ALL_BITS)
  124. return required_fault;
  125. }
  126. return required_fault;
  127. }
  128. static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
  129. __always_unused int depth, struct mm_walk *walk)
  130. {
  131. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  132. struct hmm_range *range = hmm_vma_walk->range;
  133. unsigned int required_fault;
  134. unsigned long i, npages;
  135. unsigned long *hmm_pfns;
  136. i = (addr - range->start) >> PAGE_SHIFT;
  137. npages = (end - addr) >> PAGE_SHIFT;
  138. hmm_pfns = &range->hmm_pfns[i];
  139. required_fault =
  140. hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
  141. if (!walk->vma) {
  142. if (required_fault)
  143. return -EFAULT;
  144. return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
  145. }
  146. if (required_fault)
  147. return hmm_vma_fault(addr, end, required_fault, walk);
  148. return hmm_pfns_fill(addr, end, range, 0);
  149. }
  150. static inline unsigned long hmm_pfn_flags_order(unsigned long order)
  151. {
  152. return order << HMM_PFN_ORDER_SHIFT;
  153. }
  154. static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
  155. pmd_t pmd)
  156. {
  157. if (pmd_protnone(pmd))
  158. return 0;
  159. return (pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
  160. HMM_PFN_VALID) |
  161. hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
  162. }
  163. #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  164. static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
  165. unsigned long end, unsigned long hmm_pfns[],
  166. pmd_t pmd)
  167. {
  168. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  169. struct hmm_range *range = hmm_vma_walk->range;
  170. unsigned long pfn, npages, i;
  171. unsigned int required_fault;
  172. unsigned long cpu_flags;
  173. npages = (end - addr) >> PAGE_SHIFT;
  174. cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
  175. required_fault =
  176. hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
  177. if (required_fault)
  178. return hmm_vma_fault(addr, end, required_fault, walk);
  179. pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
  180. for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
  181. hmm_pfns[i] = pfn | cpu_flags;
  182. return 0;
  183. }
  184. #else /* CONFIG_TRANSPARENT_HUGEPAGE */
  185. /* stub to allow the code below to compile */
  186. int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
  187. unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
  188. #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
  189. static inline bool hmm_is_device_private_entry(struct hmm_range *range,
  190. swp_entry_t entry)
  191. {
  192. return is_device_private_entry(entry) &&
  193. device_private_entry_to_page(entry)->pgmap->owner ==
  194. range->dev_private_owner;
  195. }
  196. static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
  197. pte_t pte)
  198. {
  199. if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
  200. return 0;
  201. return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
  202. }
  203. static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
  204. unsigned long end, pmd_t *pmdp, pte_t *ptep,
  205. unsigned long *hmm_pfn)
  206. {
  207. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  208. struct hmm_range *range = hmm_vma_walk->range;
  209. unsigned int required_fault;
  210. unsigned long cpu_flags;
  211. pte_t pte = *ptep;
  212. uint64_t pfn_req_flags = *hmm_pfn;
  213. if (pte_none(pte)) {
  214. required_fault =
  215. hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
  216. if (required_fault)
  217. goto fault;
  218. *hmm_pfn = 0;
  219. return 0;
  220. }
  221. if (!pte_present(pte)) {
  222. swp_entry_t entry = pte_to_swp_entry(pte);
  223. /*
  224. * Never fault in device private pages, but just report
  225. * the PFN even if not present.
  226. */
  227. if (hmm_is_device_private_entry(range, entry)) {
  228. cpu_flags = HMM_PFN_VALID;
  229. if (is_write_device_private_entry(entry))
  230. cpu_flags |= HMM_PFN_WRITE;
  231. *hmm_pfn = device_private_entry_to_pfn(entry) |
  232. cpu_flags;
  233. return 0;
  234. }
  235. required_fault =
  236. hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
  237. if (!required_fault) {
  238. *hmm_pfn = 0;
  239. return 0;
  240. }
  241. if (!non_swap_entry(entry))
  242. goto fault;
  243. if (is_migration_entry(entry)) {
  244. pte_unmap(ptep);
  245. hmm_vma_walk->last = addr;
  246. migration_entry_wait(walk->mm, pmdp, addr);
  247. return -EBUSY;
  248. }
  249. /* Report error for everything else */
  250. pte_unmap(ptep);
  251. return -EFAULT;
  252. }
  253. cpu_flags = pte_to_hmm_pfn_flags(range, pte);
  254. required_fault =
  255. hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
  256. if (required_fault)
  257. goto fault;
  258. /*
  259. * Bypass devmap pte such as DAX page when all pfn requested
  260. * flags(pfn_req_flags) are fulfilled.
  261. * Since each architecture defines a struct page for the zero page, just
  262. * fall through and treat it like a normal page.
  263. */
  264. if (!vm_normal_page(walk->vma, addr, pte) &&
  265. !pte_devmap(pte) &&
  266. !is_zero_pfn(pte_pfn(pte))) {
  267. if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
  268. pte_unmap(ptep);
  269. return -EFAULT;
  270. }
  271. *hmm_pfn = HMM_PFN_ERROR;
  272. return 0;
  273. }
  274. *hmm_pfn = pte_pfn(pte) | cpu_flags;
  275. return 0;
  276. fault:
  277. pte_unmap(ptep);
  278. /* Fault any virtual address we were asked to fault */
  279. return hmm_vma_fault(addr, end, required_fault, walk);
  280. }
  281. static int hmm_vma_walk_pmd(pmd_t *pmdp,
  282. unsigned long start,
  283. unsigned long end,
  284. struct mm_walk *walk)
  285. {
  286. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  287. struct hmm_range *range = hmm_vma_walk->range;
  288. unsigned long *hmm_pfns =
  289. &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
  290. unsigned long npages = (end - start) >> PAGE_SHIFT;
  291. unsigned long addr = start;
  292. pte_t *ptep;
  293. pmd_t pmd;
  294. again:
  295. pmd = READ_ONCE(*pmdp);
  296. if (pmd_none(pmd))
  297. return hmm_vma_walk_hole(start, end, -1, walk);
  298. if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
  299. if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
  300. hmm_vma_walk->last = addr;
  301. pmd_migration_entry_wait(walk->mm, pmdp);
  302. return -EBUSY;
  303. }
  304. return hmm_pfns_fill(start, end, range, 0);
  305. }
  306. if (!pmd_present(pmd)) {
  307. if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
  308. return -EFAULT;
  309. return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
  310. }
  311. if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
  312. /*
  313. * No need to take pmd_lock here, even if some other thread
  314. * is splitting the huge pmd we will get that event through
  315. * mmu_notifier callback.
  316. *
  317. * So just read pmd value and check again it's a transparent
  318. * huge or device mapping one and compute corresponding pfn
  319. * values.
  320. */
  321. pmd = pmd_read_atomic(pmdp);
  322. barrier();
  323. if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
  324. goto again;
  325. return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
  326. }
  327. /*
  328. * We have handled all the valid cases above ie either none, migration,
  329. * huge or transparent huge. At this point either it is a valid pmd
  330. * entry pointing to pte directory or it is a bad pmd that will not
  331. * recover.
  332. */
  333. if (pmd_bad(pmd)) {
  334. if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
  335. return -EFAULT;
  336. return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
  337. }
  338. ptep = pte_offset_map(pmdp, addr);
  339. for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
  340. int r;
  341. r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
  342. if (r) {
  343. /* hmm_vma_handle_pte() did pte_unmap() */
  344. return r;
  345. }
  346. }
  347. pte_unmap(ptep - 1);
  348. return 0;
  349. }
  350. #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
  351. defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
  352. static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
  353. pud_t pud)
  354. {
  355. if (!pud_present(pud))
  356. return 0;
  357. return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
  358. HMM_PFN_VALID) |
  359. hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
  360. }
  361. static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
  362. struct mm_walk *walk)
  363. {
  364. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  365. struct hmm_range *range = hmm_vma_walk->range;
  366. unsigned long addr = start;
  367. pud_t pud;
  368. int ret = 0;
  369. spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
  370. if (!ptl)
  371. return 0;
  372. /* Normally we don't want to split the huge page */
  373. walk->action = ACTION_CONTINUE;
  374. pud = READ_ONCE(*pudp);
  375. if (pud_none(pud)) {
  376. spin_unlock(ptl);
  377. return hmm_vma_walk_hole(start, end, -1, walk);
  378. }
  379. if (pud_huge(pud) && pud_devmap(pud)) {
  380. unsigned long i, npages, pfn;
  381. unsigned int required_fault;
  382. unsigned long *hmm_pfns;
  383. unsigned long cpu_flags;
  384. if (!pud_present(pud)) {
  385. spin_unlock(ptl);
  386. return hmm_vma_walk_hole(start, end, -1, walk);
  387. }
  388. i = (addr - range->start) >> PAGE_SHIFT;
  389. npages = (end - addr) >> PAGE_SHIFT;
  390. hmm_pfns = &range->hmm_pfns[i];
  391. cpu_flags = pud_to_hmm_pfn_flags(range, pud);
  392. required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
  393. npages, cpu_flags);
  394. if (required_fault) {
  395. spin_unlock(ptl);
  396. return hmm_vma_fault(addr, end, required_fault, walk);
  397. }
  398. pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
  399. for (i = 0; i < npages; ++i, ++pfn)
  400. hmm_pfns[i] = pfn | cpu_flags;
  401. goto out_unlock;
  402. }
  403. /* Ask for the PUD to be split */
  404. walk->action = ACTION_SUBTREE;
  405. out_unlock:
  406. spin_unlock(ptl);
  407. return ret;
  408. }
  409. #else
  410. #define hmm_vma_walk_pud NULL
  411. #endif
  412. #ifdef CONFIG_HUGETLB_PAGE
  413. static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
  414. unsigned long start, unsigned long end,
  415. struct mm_walk *walk)
  416. {
  417. unsigned long addr = start, i, pfn;
  418. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  419. struct hmm_range *range = hmm_vma_walk->range;
  420. struct vm_area_struct *vma = walk->vma;
  421. unsigned int required_fault;
  422. unsigned long pfn_req_flags;
  423. unsigned long cpu_flags;
  424. spinlock_t *ptl;
  425. pte_t entry;
  426. ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
  427. entry = huge_ptep_get(pte);
  428. i = (start - range->start) >> PAGE_SHIFT;
  429. pfn_req_flags = range->hmm_pfns[i];
  430. cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
  431. hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
  432. required_fault =
  433. hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
  434. if (required_fault) {
  435. spin_unlock(ptl);
  436. return hmm_vma_fault(addr, end, required_fault, walk);
  437. }
  438. pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
  439. for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
  440. range->hmm_pfns[i] = pfn | cpu_flags;
  441. spin_unlock(ptl);
  442. return 0;
  443. }
  444. #else
  445. #define hmm_vma_walk_hugetlb_entry NULL
  446. #endif /* CONFIG_HUGETLB_PAGE */
  447. static int hmm_vma_walk_test(unsigned long start, unsigned long end,
  448. struct mm_walk *walk)
  449. {
  450. struct hmm_vma_walk *hmm_vma_walk = walk->private;
  451. struct hmm_range *range = hmm_vma_walk->range;
  452. struct vm_area_struct *vma = walk->vma;
  453. if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)) &&
  454. vma->vm_flags & VM_READ)
  455. return 0;
  456. /*
  457. * vma ranges that don't have struct page backing them or map I/O
  458. * devices directly cannot be handled by hmm_range_fault().
  459. *
  460. * If the vma does not allow read access, then assume that it does not
  461. * allow write access either. HMM does not support architectures that
  462. * allow write without read.
  463. *
  464. * If a fault is requested for an unsupported range then it is a hard
  465. * failure.
  466. */
  467. if (hmm_range_need_fault(hmm_vma_walk,
  468. range->hmm_pfns +
  469. ((start - range->start) >> PAGE_SHIFT),
  470. (end - start) >> PAGE_SHIFT, 0))
  471. return -EFAULT;
  472. hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
  473. /* Skip this vma and continue processing the next vma. */
  474. return 1;
  475. }
  476. static const struct mm_walk_ops hmm_walk_ops = {
  477. .pud_entry = hmm_vma_walk_pud,
  478. .pmd_entry = hmm_vma_walk_pmd,
  479. .pte_hole = hmm_vma_walk_hole,
  480. .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
  481. .test_walk = hmm_vma_walk_test,
  482. };
  483. /**
  484. * hmm_range_fault - try to fault some address in a virtual address range
  485. * @range: argument structure
  486. *
  487. * Returns 0 on success or one of the following error codes:
  488. *
  489. * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
  490. * (e.g., device file vma).
  491. * -ENOMEM: Out of memory.
  492. * -EPERM: Invalid permission (e.g., asking for write and range is read
  493. * only).
  494. * -EBUSY: The range has been invalidated and the caller needs to wait for
  495. * the invalidation to finish.
  496. * -EFAULT: A page was requested to be valid and could not be made valid
  497. * ie it has no backing VMA or it is illegal to access
  498. *
  499. * This is similar to get_user_pages(), except that it can read the page tables
  500. * without mutating them (ie causing faults).
  501. */
  502. int hmm_range_fault(struct hmm_range *range)
  503. {
  504. struct hmm_vma_walk hmm_vma_walk = {
  505. .range = range,
  506. .last = range->start,
  507. };
  508. struct mm_struct *mm = range->notifier->mm;
  509. int ret;
  510. mmap_assert_locked(mm);
  511. do {
  512. /* If range is no longer valid force retry. */
  513. if (mmu_interval_check_retry(range->notifier,
  514. range->notifier_seq))
  515. return -EBUSY;
  516. ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
  517. &hmm_walk_ops, &hmm_vma_walk);
  518. /*
  519. * When -EBUSY is returned the loop restarts with
  520. * hmm_vma_walk.last set to an address that has not been stored
  521. * in pfns. All entries < last in the pfn array are set to their
  522. * output, and all >= are still at their input values.
  523. */
  524. } while (ret == -EBUSY);
  525. return ret;
  526. }
  527. EXPORT_SYMBOL(hmm_range_fault);