memremap.c 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531
  1. /* SPDX-License-Identifier: GPL-2.0 */
  2. /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
  3. #include <linux/device.h>
  4. #include <linux/io.h>
  5. #include <linux/kasan.h>
  6. #include <linux/memory_hotplug.h>
  7. #include <linux/mm.h>
  8. #include <linux/pfn_t.h>
  9. #include <linux/swap.h>
  10. #include <linux/mmzone.h>
  11. #include <linux/swapops.h>
  12. #include <linux/types.h>
  13. #include <linux/wait_bit.h>
  14. #include <linux/xarray.h>
  15. static DEFINE_XARRAY(pgmap_array);
  16. /*
  17. * The memremap() and memremap_pages() interfaces are alternately used
  18. * to map persistent memory namespaces. These interfaces place different
  19. * constraints on the alignment and size of the mapping (namespace).
  20. * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
  21. * only map subsections (2MB), and at least one architecture (PowerPC)
  22. * the minimum mapping granularity of memremap_pages() is 16MB.
  23. *
  24. * The role of memremap_compat_align() is to communicate the minimum
  25. * arch supported alignment of a namespace such that it can freely
  26. * switch modes without violating the arch constraint. Namely, do not
  27. * allow a namespace to be PAGE_SIZE aligned since that namespace may be
  28. * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
  29. */
  30. #ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
  31. unsigned long memremap_compat_align(void)
  32. {
  33. return SUBSECTION_SIZE;
  34. }
  35. EXPORT_SYMBOL_GPL(memremap_compat_align);
  36. #endif
  37. #ifdef CONFIG_DEV_PAGEMAP_OPS
  38. DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
  39. EXPORT_SYMBOL(devmap_managed_key);
  40. static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  41. {
  42. if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  43. pgmap->type == MEMORY_DEVICE_FS_DAX)
  44. static_branch_dec(&devmap_managed_key);
  45. }
  46. static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  47. {
  48. if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  49. pgmap->type == MEMORY_DEVICE_FS_DAX)
  50. static_branch_inc(&devmap_managed_key);
  51. }
  52. #else
  53. static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  54. {
  55. }
  56. static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  57. {
  58. }
  59. #endif /* CONFIG_DEV_PAGEMAP_OPS */
  60. static void pgmap_array_delete(struct range *range)
  61. {
  62. xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
  63. NULL, GFP_KERNEL);
  64. synchronize_rcu();
  65. }
  66. static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
  67. {
  68. struct range *range = &pgmap->ranges[range_id];
  69. unsigned long pfn = PHYS_PFN(range->start);
  70. if (range_id)
  71. return pfn;
  72. return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
  73. }
  74. bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn)
  75. {
  76. int i;
  77. for (i = 0; i < pgmap->nr_range; i++) {
  78. struct range *range = &pgmap->ranges[i];
  79. if (pfn >= PHYS_PFN(range->start) &&
  80. pfn <= PHYS_PFN(range->end))
  81. return pfn >= pfn_first(pgmap, i);
  82. }
  83. return false;
  84. }
  85. static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
  86. {
  87. const struct range *range = &pgmap->ranges[range_id];
  88. return (range->start + range_len(range)) >> PAGE_SHIFT;
  89. }
  90. static unsigned long pfn_next(unsigned long pfn)
  91. {
  92. if (pfn % 1024 == 0)
  93. cond_resched();
  94. return pfn + 1;
  95. }
  96. #define for_each_device_pfn(pfn, map, i) \
  97. for (pfn = pfn_first(map, i); pfn < pfn_end(map, i); pfn = pfn_next(pfn))
  98. static void dev_pagemap_kill(struct dev_pagemap *pgmap)
  99. {
  100. if (pgmap->ops && pgmap->ops->kill)
  101. pgmap->ops->kill(pgmap);
  102. else
  103. percpu_ref_kill(pgmap->ref);
  104. }
  105. static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
  106. {
  107. if (pgmap->ops && pgmap->ops->cleanup) {
  108. pgmap->ops->cleanup(pgmap);
  109. } else {
  110. wait_for_completion(&pgmap->done);
  111. percpu_ref_exit(pgmap->ref);
  112. }
  113. /*
  114. * Undo the pgmap ref assignment for the internal case as the
  115. * caller may re-enable the same pgmap.
  116. */
  117. if (pgmap->ref == &pgmap->internal_ref)
  118. pgmap->ref = NULL;
  119. }
  120. static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
  121. {
  122. struct range *range = &pgmap->ranges[range_id];
  123. struct page *first_page;
  124. int nid;
  125. /* make sure to access a memmap that was actually initialized */
  126. first_page = pfn_to_page(pfn_first(pgmap, range_id));
  127. /* pages are dead and unused, undo the arch mapping */
  128. nid = page_to_nid(first_page);
  129. mem_hotplug_begin();
  130. remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
  131. PHYS_PFN(range_len(range)));
  132. if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
  133. __remove_pages(PHYS_PFN(range->start),
  134. PHYS_PFN(range_len(range)), NULL);
  135. } else {
  136. arch_remove_memory(nid, range->start, range_len(range),
  137. pgmap_altmap(pgmap));
  138. kasan_remove_zero_shadow(__va(range->start), range_len(range));
  139. }
  140. mem_hotplug_done();
  141. untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
  142. pgmap_array_delete(range);
  143. }
  144. void memunmap_pages(struct dev_pagemap *pgmap)
  145. {
  146. unsigned long pfn;
  147. int i;
  148. dev_pagemap_kill(pgmap);
  149. for (i = 0; i < pgmap->nr_range; i++)
  150. for_each_device_pfn(pfn, pgmap, i)
  151. put_page(pfn_to_page(pfn));
  152. dev_pagemap_cleanup(pgmap);
  153. for (i = 0; i < pgmap->nr_range; i++)
  154. pageunmap_range(pgmap, i);
  155. WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
  156. devmap_managed_enable_put(pgmap);
  157. }
  158. EXPORT_SYMBOL_GPL(memunmap_pages);
  159. static void devm_memremap_pages_release(void *data)
  160. {
  161. memunmap_pages(data);
  162. }
  163. static void dev_pagemap_percpu_release(struct percpu_ref *ref)
  164. {
  165. struct dev_pagemap *pgmap =
  166. container_of(ref, struct dev_pagemap, internal_ref);
  167. complete(&pgmap->done);
  168. }
  169. static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
  170. int range_id, int nid)
  171. {
  172. struct range *range = &pgmap->ranges[range_id];
  173. struct dev_pagemap *conflict_pgmap;
  174. int error, is_ram;
  175. if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
  176. "altmap not supported for multiple ranges\n"))
  177. return -EINVAL;
  178. conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start), NULL);
  179. if (conflict_pgmap) {
  180. WARN(1, "Conflicting mapping in same section\n");
  181. put_dev_pagemap(conflict_pgmap);
  182. return -ENOMEM;
  183. }
  184. conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end), NULL);
  185. if (conflict_pgmap) {
  186. WARN(1, "Conflicting mapping in same section\n");
  187. put_dev_pagemap(conflict_pgmap);
  188. return -ENOMEM;
  189. }
  190. is_ram = region_intersects(range->start, range_len(range),
  191. IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
  192. if (is_ram != REGION_DISJOINT) {
  193. WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
  194. is_ram == REGION_MIXED ? "mixed" : "ram",
  195. range->start, range->end);
  196. return -ENXIO;
  197. }
  198. error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
  199. PHYS_PFN(range->end), pgmap, GFP_KERNEL));
  200. if (error)
  201. return error;
  202. if (nid < 0)
  203. nid = numa_mem_id();
  204. error = track_pfn_remap(NULL, &params->pgprot, PHYS_PFN(range->start), 0,
  205. range_len(range));
  206. if (error)
  207. goto err_pfn_remap;
  208. mem_hotplug_begin();
  209. /*
  210. * For device private memory we call add_pages() as we only need to
  211. * allocate and initialize struct page for the device memory. More-
  212. * over the device memory is un-accessible thus we do not want to
  213. * create a linear mapping for the memory like arch_add_memory()
  214. * would do.
  215. *
  216. * For all other device memory types, which are accessible by
  217. * the CPU, we do want the linear mapping and thus use
  218. * arch_add_memory().
  219. */
  220. if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
  221. error = add_pages(nid, PHYS_PFN(range->start),
  222. PHYS_PFN(range_len(range)), params);
  223. } else {
  224. error = kasan_add_zero_shadow(__va(range->start), range_len(range));
  225. if (error) {
  226. mem_hotplug_done();
  227. goto err_kasan;
  228. }
  229. error = arch_add_memory(nid, range->start, range_len(range),
  230. params);
  231. }
  232. if (!error) {
  233. struct zone *zone;
  234. zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
  235. move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
  236. PHYS_PFN(range_len(range)), params->altmap,
  237. MIGRATE_MOVABLE);
  238. }
  239. mem_hotplug_done();
  240. if (error)
  241. goto err_add_memory;
  242. /*
  243. * Initialization of the pages has been deferred until now in order
  244. * to allow us to do the work while not holding the hotplug lock.
  245. */
  246. memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
  247. PHYS_PFN(range->start),
  248. PHYS_PFN(range_len(range)), pgmap);
  249. percpu_ref_get_many(pgmap->ref, pfn_end(pgmap, range_id)
  250. - pfn_first(pgmap, range_id));
  251. return 0;
  252. err_add_memory:
  253. kasan_remove_zero_shadow(__va(range->start), range_len(range));
  254. err_kasan:
  255. untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
  256. err_pfn_remap:
  257. pgmap_array_delete(range);
  258. return error;
  259. }
  260. /*
  261. * Not device managed version of dev_memremap_pages, undone by
  262. * memunmap_pages(). Please use dev_memremap_pages if you have a struct
  263. * device available.
  264. */
  265. void *memremap_pages(struct dev_pagemap *pgmap, int nid)
  266. {
  267. struct mhp_params params = {
  268. .altmap = pgmap_altmap(pgmap),
  269. .pgprot = PAGE_KERNEL,
  270. };
  271. const int nr_range = pgmap->nr_range;
  272. int error, i;
  273. if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
  274. return ERR_PTR(-EINVAL);
  275. switch (pgmap->type) {
  276. case MEMORY_DEVICE_PRIVATE:
  277. if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
  278. WARN(1, "Device private memory not supported\n");
  279. return ERR_PTR(-EINVAL);
  280. }
  281. if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
  282. WARN(1, "Missing migrate_to_ram method\n");
  283. return ERR_PTR(-EINVAL);
  284. }
  285. if (!pgmap->ops->page_free) {
  286. WARN(1, "Missing page_free method\n");
  287. return ERR_PTR(-EINVAL);
  288. }
  289. if (!pgmap->owner) {
  290. WARN(1, "Missing owner\n");
  291. return ERR_PTR(-EINVAL);
  292. }
  293. break;
  294. case MEMORY_DEVICE_FS_DAX:
  295. if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
  296. IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
  297. WARN(1, "File system DAX not supported\n");
  298. return ERR_PTR(-EINVAL);
  299. }
  300. break;
  301. case MEMORY_DEVICE_GENERIC:
  302. break;
  303. case MEMORY_DEVICE_PCI_P2PDMA:
  304. params.pgprot = pgprot_noncached(params.pgprot);
  305. break;
  306. default:
  307. WARN(1, "Invalid pgmap type %d\n", pgmap->type);
  308. break;
  309. }
  310. if (!pgmap->ref) {
  311. if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
  312. return ERR_PTR(-EINVAL);
  313. init_completion(&pgmap->done);
  314. error = percpu_ref_init(&pgmap->internal_ref,
  315. dev_pagemap_percpu_release, 0, GFP_KERNEL);
  316. if (error)
  317. return ERR_PTR(error);
  318. pgmap->ref = &pgmap->internal_ref;
  319. } else {
  320. if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
  321. WARN(1, "Missing reference count teardown definition\n");
  322. return ERR_PTR(-EINVAL);
  323. }
  324. }
  325. devmap_managed_enable_get(pgmap);
  326. /*
  327. * Clear the pgmap nr_range as it will be incremented for each
  328. * successfully processed range. This communicates how many
  329. * regions to unwind in the abort case.
  330. */
  331. pgmap->nr_range = 0;
  332. error = 0;
  333. for (i = 0; i < nr_range; i++) {
  334. error = pagemap_range(pgmap, &params, i, nid);
  335. if (error)
  336. break;
  337. pgmap->nr_range++;
  338. }
  339. if (i < nr_range) {
  340. memunmap_pages(pgmap);
  341. pgmap->nr_range = nr_range;
  342. return ERR_PTR(error);
  343. }
  344. return __va(pgmap->ranges[0].start);
  345. }
  346. EXPORT_SYMBOL_GPL(memremap_pages);
  347. /**
  348. * devm_memremap_pages - remap and provide memmap backing for the given resource
  349. * @dev: hosting device for @res
  350. * @pgmap: pointer to a struct dev_pagemap
  351. *
  352. * Notes:
  353. * 1/ At a minimum the res and type members of @pgmap must be initialized
  354. * by the caller before passing it to this function
  355. *
  356. * 2/ The altmap field may optionally be initialized, in which case
  357. * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
  358. *
  359. * 3/ The ref field may optionally be provided, in which pgmap->ref must be
  360. * 'live' on entry and will be killed and reaped at
  361. * devm_memremap_pages_release() time, or if this routine fails.
  362. *
  363. * 4/ range is expected to be a host memory range that could feasibly be
  364. * treated as a "System RAM" range, i.e. not a device mmio range, but
  365. * this is not enforced.
  366. */
  367. void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
  368. {
  369. int error;
  370. void *ret;
  371. ret = memremap_pages(pgmap, dev_to_node(dev));
  372. if (IS_ERR(ret))
  373. return ret;
  374. error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
  375. pgmap);
  376. if (error)
  377. return ERR_PTR(error);
  378. return ret;
  379. }
  380. EXPORT_SYMBOL_GPL(devm_memremap_pages);
  381. void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
  382. {
  383. devm_release_action(dev, devm_memremap_pages_release, pgmap);
  384. }
  385. EXPORT_SYMBOL_GPL(devm_memunmap_pages);
  386. unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
  387. {
  388. /* number of pfns from base where pfn_to_page() is valid */
  389. if (altmap)
  390. return altmap->reserve + altmap->free;
  391. return 0;
  392. }
  393. void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
  394. {
  395. altmap->alloc -= nr_pfns;
  396. }
  397. /**
  398. * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
  399. * @pfn: page frame number to lookup page_map
  400. * @pgmap: optional known pgmap that already has a reference
  401. *
  402. * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
  403. * is non-NULL but does not cover @pfn the reference to it will be released.
  404. */
  405. struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
  406. struct dev_pagemap *pgmap)
  407. {
  408. resource_size_t phys = PFN_PHYS(pfn);
  409. /*
  410. * In the cached case we're already holding a live reference.
  411. */
  412. if (pgmap) {
  413. if (phys >= pgmap->range.start && phys <= pgmap->range.end)
  414. return pgmap;
  415. put_dev_pagemap(pgmap);
  416. }
  417. /* fall back to slow path lookup */
  418. rcu_read_lock();
  419. pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
  420. if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
  421. pgmap = NULL;
  422. rcu_read_unlock();
  423. return pgmap;
  424. }
  425. EXPORT_SYMBOL_GPL(get_dev_pagemap);
  426. #ifdef CONFIG_DEV_PAGEMAP_OPS
  427. void free_devmap_managed_page(struct page *page)
  428. {
  429. /* notify page idle for dax */
  430. if (!is_device_private_page(page)) {
  431. wake_up_var(&page->_refcount);
  432. return;
  433. }
  434. __ClearPageWaiters(page);
  435. mem_cgroup_uncharge(page);
  436. /*
  437. * When a device_private page is freed, the page->mapping field
  438. * may still contain a (stale) mapping value. For example, the
  439. * lower bits of page->mapping may still identify the page as an
  440. * anonymous page. Ultimately, this entire field is just stale
  441. * and wrong, and it will cause errors if not cleared. One
  442. * example is:
  443. *
  444. * migrate_vma_pages()
  445. * migrate_vma_insert_page()
  446. * page_add_new_anon_rmap()
  447. * __page_set_anon_rmap()
  448. * ...checks page->mapping, via PageAnon(page) call,
  449. * and incorrectly concludes that the page is an
  450. * anonymous page. Therefore, it incorrectly,
  451. * silently fails to set up the new anon rmap.
  452. *
  453. * For other types of ZONE_DEVICE pages, migration is either
  454. * handled differently or not done at all, so there is no need
  455. * to clear page->mapping.
  456. */
  457. page->mapping = NULL;
  458. page->pgmap->ops->page_free(page);
  459. }
  460. #endif /* CONFIG_DEV_PAGEMAP_OPS */