compaction.c 22 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830
  1. /*
  2. * linux/mm/compaction.c
  3. *
  4. * Memory compaction for the reduction of external fragmentation. Note that
  5. * this heavily depends upon page migration to do all the real heavy
  6. * lifting
  7. *
  8. * Copyright IBM Corp. 2007-2010 Mel Gorman <mel@csn.ul.ie>
  9. */
  10. #include <linux/swap.h>
  11. #include <linux/migrate.h>
  12. #include <linux/compaction.h>
  13. #include <linux/mm_inline.h>
  14. #include <linux/backing-dev.h>
  15. #include <linux/sysctl.h>
  16. #include <linux/sysfs.h>
  17. #include "internal.h"
  18. #define CREATE_TRACE_POINTS
  19. #include <trace/events/compaction.h>
  20. /*
  21. * compact_control is used to track pages being migrated and the free pages
  22. * they are being migrated to during memory compaction. The free_pfn starts
  23. * at the end of a zone and migrate_pfn begins at the start. Movable pages
  24. * are moved to the end of a zone during a compaction run and the run
  25. * completes when free_pfn <= migrate_pfn
  26. */
  27. struct compact_control {
  28. struct list_head freepages; /* List of free pages to migrate to */
  29. struct list_head migratepages; /* List of pages being migrated */
  30. unsigned long nr_freepages; /* Number of isolated free pages */
  31. unsigned long nr_migratepages; /* Number of pages to migrate */
  32. unsigned long free_pfn; /* isolate_freepages search base */
  33. unsigned long migrate_pfn; /* isolate_migratepages search base */
  34. bool sync; /* Synchronous migration */
  35. int order; /* order a direct compactor needs */
  36. int migratetype; /* MOVABLE, RECLAIMABLE etc */
  37. struct zone *zone;
  38. };
  39. static unsigned long release_freepages(struct list_head *freelist)
  40. {
  41. struct page *page, *next;
  42. unsigned long count = 0;
  43. list_for_each_entry_safe(page, next, freelist, lru) {
  44. list_del(&page->lru);
  45. __free_page(page);
  46. count++;
  47. }
  48. return count;
  49. }
  50. /* Isolate free pages onto a private freelist. Must hold zone->lock */
  51. static unsigned long isolate_freepages_block(struct zone *zone,
  52. unsigned long blockpfn,
  53. struct list_head *freelist)
  54. {
  55. unsigned long zone_end_pfn, end_pfn;
  56. int nr_scanned = 0, total_isolated = 0;
  57. struct page *cursor;
  58. /* Get the last PFN we should scan for free pages at */
  59. zone_end_pfn = zone->zone_start_pfn + zone->spanned_pages;
  60. end_pfn = min(blockpfn + pageblock_nr_pages, zone_end_pfn);
  61. /* Find the first usable PFN in the block to initialse page cursor */
  62. for (; blockpfn < end_pfn; blockpfn++) {
  63. if (pfn_valid_within(blockpfn))
  64. break;
  65. }
  66. cursor = pfn_to_page(blockpfn);
  67. /* Isolate free pages. This assumes the block is valid */
  68. for (; blockpfn < end_pfn; blockpfn++, cursor++) {
  69. int isolated, i;
  70. struct page *page = cursor;
  71. if (!pfn_valid_within(blockpfn))
  72. continue;
  73. nr_scanned++;
  74. if (!PageBuddy(page))
  75. continue;
  76. /* Found a free page, break it into order-0 pages */
  77. isolated = split_free_page(page);
  78. total_isolated += isolated;
  79. for (i = 0; i < isolated; i++) {
  80. list_add(&page->lru, freelist);
  81. page++;
  82. }
  83. /* If a page was split, advance to the end of it */
  84. if (isolated) {
  85. blockpfn += isolated - 1;
  86. cursor += isolated - 1;
  87. }
  88. }
  89. trace_mm_compaction_isolate_freepages(nr_scanned, total_isolated);
  90. return total_isolated;
  91. }
  92. /* Returns true if the page is within a block suitable for migration to */
  93. static bool suitable_migration_target(struct page *page)
  94. {
  95. int migratetype = get_pageblock_migratetype(page);
  96. /* Don't interfere with memory hot-remove or the min_free_kbytes blocks */
  97. if (migratetype == MIGRATE_ISOLATE || migratetype == MIGRATE_RESERVE)
  98. return false;
  99. /* If the page is a large free page, then allow migration */
  100. if (PageBuddy(page) && page_order(page) >= pageblock_order)
  101. return true;
  102. /* If the block is MIGRATE_MOVABLE, allow migration */
  103. if (migratetype == MIGRATE_MOVABLE)
  104. return true;
  105. /* Otherwise skip the block */
  106. return false;
  107. }
  108. /*
  109. * Based on information in the current compact_control, find blocks
  110. * suitable for isolating free pages from and then isolate them.
  111. */
  112. static void isolate_freepages(struct zone *zone,
  113. struct compact_control *cc)
  114. {
  115. struct page *page;
  116. unsigned long high_pfn, low_pfn, pfn;
  117. unsigned long flags;
  118. int nr_freepages = cc->nr_freepages;
  119. struct list_head *freelist = &cc->freepages;
  120. /*
  121. * Initialise the free scanner. The starting point is where we last
  122. * scanned from (or the end of the zone if starting). The low point
  123. * is the end of the pageblock the migration scanner is using.
  124. */
  125. pfn = cc->free_pfn;
  126. low_pfn = cc->migrate_pfn + pageblock_nr_pages;
  127. /*
  128. * Take care that if the migration scanner is at the end of the zone
  129. * that the free scanner does not accidentally move to the next zone
  130. * in the next isolation cycle.
  131. */
  132. high_pfn = min(low_pfn, pfn);
  133. /*
  134. * Isolate free pages until enough are available to migrate the
  135. * pages on cc->migratepages. We stop searching if the migrate
  136. * and free page scanners meet or enough free pages are isolated.
  137. */
  138. for (; pfn > low_pfn && cc->nr_migratepages > nr_freepages;
  139. pfn -= pageblock_nr_pages) {
  140. unsigned long isolated;
  141. if (!pfn_valid(pfn))
  142. continue;
  143. /*
  144. * Check for overlapping nodes/zones. It's possible on some
  145. * configurations to have a setup like
  146. * node0 node1 node0
  147. * i.e. it's possible that all pages within a zones range of
  148. * pages do not belong to a single zone.
  149. */
  150. page = pfn_to_page(pfn);
  151. if (page_zone(page) != zone)
  152. continue;
  153. /* Check the block is suitable for migration */
  154. if (!suitable_migration_target(page))
  155. continue;
  156. /*
  157. * Found a block suitable for isolating free pages from. Now
  158. * we disabled interrupts, double check things are ok and
  159. * isolate the pages. This is to minimise the time IRQs
  160. * are disabled
  161. */
  162. isolated = 0;
  163. spin_lock_irqsave(&zone->lock, flags);
  164. if (suitable_migration_target(page)) {
  165. isolated = isolate_freepages_block(zone, pfn, freelist);
  166. nr_freepages += isolated;
  167. }
  168. spin_unlock_irqrestore(&zone->lock, flags);
  169. /*
  170. * Record the highest PFN we isolated pages from. When next
  171. * looking for free pages, the search will restart here as
  172. * page migration may have returned some pages to the allocator
  173. */
  174. if (isolated)
  175. high_pfn = max(high_pfn, pfn);
  176. }
  177. /* split_free_page does not map the pages */
  178. list_for_each_entry(page, freelist, lru) {
  179. arch_alloc_page(page, 0);
  180. kernel_map_pages(page, 1, 1);
  181. }
  182. cc->free_pfn = high_pfn;
  183. cc->nr_freepages = nr_freepages;
  184. }
  185. /* Update the number of anon and file isolated pages in the zone */
  186. static void acct_isolated(struct zone *zone, struct compact_control *cc)
  187. {
  188. struct page *page;
  189. unsigned int count[2] = { 0, };
  190. list_for_each_entry(page, &cc->migratepages, lru)
  191. count[!!page_is_file_cache(page)]++;
  192. __mod_zone_page_state(zone, NR_ISOLATED_ANON, count[0]);
  193. __mod_zone_page_state(zone, NR_ISOLATED_FILE, count[1]);
  194. }
  195. /* Similar to reclaim, but different enough that they don't share logic */
  196. static bool too_many_isolated(struct zone *zone)
  197. {
  198. unsigned long active, inactive, isolated;
  199. inactive = zone_page_state(zone, NR_INACTIVE_FILE) +
  200. zone_page_state(zone, NR_INACTIVE_ANON);
  201. active = zone_page_state(zone, NR_ACTIVE_FILE) +
  202. zone_page_state(zone, NR_ACTIVE_ANON);
  203. isolated = zone_page_state(zone, NR_ISOLATED_FILE) +
  204. zone_page_state(zone, NR_ISOLATED_ANON);
  205. return isolated > (inactive + active) / 2;
  206. }
  207. /* possible outcome of isolate_migratepages */
  208. typedef enum {
  209. ISOLATE_ABORT, /* Abort compaction now */
  210. ISOLATE_NONE, /* No pages isolated, continue scanning */
  211. ISOLATE_SUCCESS, /* Pages isolated, migrate */
  212. } isolate_migrate_t;
  213. /**
  214. * isolate_migratepages_range() - isolate all migrate-able pages in range.
  215. * @zone: Zone pages are in.
  216. * @cc: Compaction control structure.
  217. * @low_pfn: The first PFN of the range.
  218. * @end_pfn: The one-past-the-last PFN of the range.
  219. *
  220. * Isolate all pages that can be migrated from the range specified by
  221. * [low_pfn, end_pfn). Returns zero if there is a fatal signal
  222. * pending), otherwise PFN of the first page that was not scanned
  223. * (which may be both less, equal to or more then end_pfn).
  224. *
  225. * Assumes that cc->migratepages is empty and cc->nr_migratepages is
  226. * zero.
  227. *
  228. * Apart from cc->migratepages and cc->nr_migratetypes this function
  229. * does not modify any cc's fields, in particular it does not modify
  230. * (or read for that matter) cc->migrate_pfn.
  231. */
  232. static unsigned long
  233. isolate_migratepages_range(struct zone *zone, struct compact_control *cc,
  234. unsigned long low_pfn, unsigned long end_pfn)
  235. {
  236. unsigned long last_pageblock_nr = 0, pageblock_nr;
  237. unsigned long nr_scanned = 0, nr_isolated = 0;
  238. struct list_head *migratelist = &cc->migratepages;
  239. isolate_mode_t mode = ISOLATE_ACTIVE|ISOLATE_INACTIVE;
  240. /*
  241. * Ensure that there are not too many pages isolated from the LRU
  242. * list by either parallel reclaimers or compaction. If there are,
  243. * delay for some time until fewer pages are isolated
  244. */
  245. while (unlikely(too_many_isolated(zone))) {
  246. /* async migration should just abort */
  247. if (!cc->sync)
  248. return 0;
  249. congestion_wait(BLK_RW_ASYNC, HZ/10);
  250. if (fatal_signal_pending(current))
  251. return 0;
  252. }
  253. /* Time to isolate some pages for migration */
  254. cond_resched();
  255. spin_lock_irq(&zone->lru_lock);
  256. for (; low_pfn < end_pfn; low_pfn++) {
  257. struct page *page;
  258. bool locked = true;
  259. /* give a chance to irqs before checking need_resched() */
  260. if (!((low_pfn+1) % SWAP_CLUSTER_MAX)) {
  261. spin_unlock_irq(&zone->lru_lock);
  262. locked = false;
  263. }
  264. if (need_resched() || spin_is_contended(&zone->lru_lock)) {
  265. if (locked)
  266. spin_unlock_irq(&zone->lru_lock);
  267. cond_resched();
  268. spin_lock_irq(&zone->lru_lock);
  269. if (fatal_signal_pending(current))
  270. break;
  271. } else if (!locked)
  272. spin_lock_irq(&zone->lru_lock);
  273. /*
  274. * migrate_pfn does not necessarily start aligned to a
  275. * pageblock. Ensure that pfn_valid is called when moving
  276. * into a new MAX_ORDER_NR_PAGES range in case of large
  277. * memory holes within the zone
  278. */
  279. if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
  280. if (!pfn_valid(low_pfn)) {
  281. low_pfn += MAX_ORDER_NR_PAGES - 1;
  282. continue;
  283. }
  284. }
  285. if (!pfn_valid_within(low_pfn))
  286. continue;
  287. nr_scanned++;
  288. /*
  289. * Get the page and ensure the page is within the same zone.
  290. * See the comment in isolate_freepages about overlapping
  291. * nodes. It is deliberate that the new zone lock is not taken
  292. * as memory compaction should not move pages between nodes.
  293. */
  294. page = pfn_to_page(low_pfn);
  295. if (page_zone(page) != zone)
  296. continue;
  297. /* Skip if free */
  298. if (PageBuddy(page))
  299. continue;
  300. /*
  301. * For async migration, also only scan in MOVABLE blocks. Async
  302. * migration is optimistic to see if the minimum amount of work
  303. * satisfies the allocation
  304. */
  305. pageblock_nr = low_pfn >> pageblock_order;
  306. if (!cc->sync && last_pageblock_nr != pageblock_nr &&
  307. get_pageblock_migratetype(page) != MIGRATE_MOVABLE) {
  308. low_pfn += pageblock_nr_pages;
  309. low_pfn = ALIGN(low_pfn, pageblock_nr_pages) - 1;
  310. last_pageblock_nr = pageblock_nr;
  311. continue;
  312. }
  313. if (!PageLRU(page))
  314. continue;
  315. /*
  316. * PageLRU is set, and lru_lock excludes isolation,
  317. * splitting and collapsing (collapsing has already
  318. * happened if PageLRU is set).
  319. */
  320. if (PageTransHuge(page)) {
  321. low_pfn += (1 << compound_order(page)) - 1;
  322. continue;
  323. }
  324. if (!cc->sync)
  325. mode |= ISOLATE_ASYNC_MIGRATE;
  326. /* Try isolate the page */
  327. if (__isolate_lru_page(page, mode, 0) != 0)
  328. continue;
  329. VM_BUG_ON(PageTransCompound(page));
  330. /* Successfully isolated */
  331. del_page_from_lru_list(zone, page, page_lru(page));
  332. list_add(&page->lru, migratelist);
  333. cc->nr_migratepages++;
  334. nr_isolated++;
  335. /* Avoid isolating too much */
  336. if (cc->nr_migratepages == COMPACT_CLUSTER_MAX) {
  337. ++low_pfn;
  338. break;
  339. }
  340. }
  341. acct_isolated(zone, cc);
  342. spin_unlock_irq(&zone->lru_lock);
  343. trace_mm_compaction_isolate_migratepages(nr_scanned, nr_isolated);
  344. return low_pfn;
  345. }
  346. /*
  347. * Isolate all pages that can be migrated from the block pointed to by
  348. * the migrate scanner within compact_control.
  349. */
  350. static isolate_migrate_t isolate_migratepages(struct zone *zone,
  351. struct compact_control *cc)
  352. {
  353. unsigned long low_pfn, end_pfn;
  354. /* Do not scan outside zone boundaries */
  355. low_pfn = max(cc->migrate_pfn, zone->zone_start_pfn);
  356. /* Only scan within a pageblock boundary */
  357. end_pfn = ALIGN(low_pfn + pageblock_nr_pages, pageblock_nr_pages);
  358. /* Do not cross the free scanner or scan within a memory hole */
  359. if (end_pfn > cc->free_pfn || !pfn_valid(low_pfn)) {
  360. cc->migrate_pfn = end_pfn;
  361. return ISOLATE_NONE;
  362. }
  363. /* Perform the isolation */
  364. low_pfn = isolate_migratepages_range(zone, cc, low_pfn, end_pfn);
  365. if (!low_pfn)
  366. return ISOLATE_ABORT;
  367. cc->migrate_pfn = low_pfn;
  368. return ISOLATE_SUCCESS;
  369. }
  370. /*
  371. * This is a migrate-callback that "allocates" freepages by taking pages
  372. * from the isolated freelists in the block we are migrating to.
  373. */
  374. static struct page *compaction_alloc(struct page *migratepage,
  375. unsigned long data,
  376. int **result)
  377. {
  378. struct compact_control *cc = (struct compact_control *)data;
  379. struct page *freepage;
  380. /* Isolate free pages if necessary */
  381. if (list_empty(&cc->freepages)) {
  382. isolate_freepages(cc->zone, cc);
  383. if (list_empty(&cc->freepages))
  384. return NULL;
  385. }
  386. freepage = list_entry(cc->freepages.next, struct page, lru);
  387. list_del(&freepage->lru);
  388. cc->nr_freepages--;
  389. return freepage;
  390. }
  391. /*
  392. * We cannot control nr_migratepages and nr_freepages fully when migration is
  393. * running as migrate_pages() has no knowledge of compact_control. When
  394. * migration is complete, we count the number of pages on the lists by hand.
  395. */
  396. static void update_nr_listpages(struct compact_control *cc)
  397. {
  398. int nr_migratepages = 0;
  399. int nr_freepages = 0;
  400. struct page *page;
  401. list_for_each_entry(page, &cc->migratepages, lru)
  402. nr_migratepages++;
  403. list_for_each_entry(page, &cc->freepages, lru)
  404. nr_freepages++;
  405. cc->nr_migratepages = nr_migratepages;
  406. cc->nr_freepages = nr_freepages;
  407. }
  408. static int compact_finished(struct zone *zone,
  409. struct compact_control *cc)
  410. {
  411. unsigned int order;
  412. unsigned long watermark;
  413. if (fatal_signal_pending(current))
  414. return COMPACT_PARTIAL;
  415. /* Compaction run completes if the migrate and free scanner meet */
  416. if (cc->free_pfn <= cc->migrate_pfn)
  417. return COMPACT_COMPLETE;
  418. /*
  419. * order == -1 is expected when compacting via
  420. * /proc/sys/vm/compact_memory
  421. */
  422. if (cc->order == -1)
  423. return COMPACT_CONTINUE;
  424. /* Compaction run is not finished if the watermark is not met */
  425. watermark = low_wmark_pages(zone);
  426. watermark += (1 << cc->order);
  427. if (!zone_watermark_ok(zone, cc->order, watermark, 0, 0))
  428. return COMPACT_CONTINUE;
  429. /* Direct compactor: Is a suitable page free? */
  430. for (order = cc->order; order < MAX_ORDER; order++) {
  431. /* Job done if page is free of the right migratetype */
  432. if (!list_empty(&zone->free_area[order].free_list[cc->migratetype]))
  433. return COMPACT_PARTIAL;
  434. /* Job done if allocation would set block type */
  435. if (order >= pageblock_order && zone->free_area[order].nr_free)
  436. return COMPACT_PARTIAL;
  437. }
  438. return COMPACT_CONTINUE;
  439. }
  440. /*
  441. * compaction_suitable: Is this suitable to run compaction on this zone now?
  442. * Returns
  443. * COMPACT_SKIPPED - If there are too few free pages for compaction
  444. * COMPACT_PARTIAL - If the allocation would succeed without compaction
  445. * COMPACT_CONTINUE - If compaction should run now
  446. */
  447. unsigned long compaction_suitable(struct zone *zone, int order)
  448. {
  449. int fragindex;
  450. unsigned long watermark;
  451. /*
  452. * order == -1 is expected when compacting via
  453. * /proc/sys/vm/compact_memory
  454. */
  455. if (order == -1)
  456. return COMPACT_CONTINUE;
  457. /*
  458. * Watermarks for order-0 must be met for compaction. Note the 2UL.
  459. * This is because during migration, copies of pages need to be
  460. * allocated and for a short time, the footprint is higher
  461. */
  462. watermark = low_wmark_pages(zone) + (2UL << order);
  463. if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
  464. return COMPACT_SKIPPED;
  465. /*
  466. * fragmentation index determines if allocation failures are due to
  467. * low memory or external fragmentation
  468. *
  469. * index of -1000 implies allocations might succeed depending on
  470. * watermarks
  471. * index towards 0 implies failure is due to lack of memory
  472. * index towards 1000 implies failure is due to fragmentation
  473. *
  474. * Only compact if a failure would be due to fragmentation.
  475. */
  476. fragindex = fragmentation_index(zone, order);
  477. if (fragindex >= 0 && fragindex <= sysctl_extfrag_threshold)
  478. return COMPACT_SKIPPED;
  479. if (fragindex == -1000 && zone_watermark_ok(zone, order, watermark,
  480. 0, 0))
  481. return COMPACT_PARTIAL;
  482. return COMPACT_CONTINUE;
  483. }
  484. static int compact_zone(struct zone *zone, struct compact_control *cc)
  485. {
  486. int ret;
  487. ret = compaction_suitable(zone, cc->order);
  488. switch (ret) {
  489. case COMPACT_PARTIAL:
  490. case COMPACT_SKIPPED:
  491. /* Compaction is likely to fail */
  492. return ret;
  493. case COMPACT_CONTINUE:
  494. /* Fall through to compaction */
  495. ;
  496. }
  497. /* Setup to move all movable pages to the end of the zone */
  498. cc->migrate_pfn = zone->zone_start_pfn;
  499. cc->free_pfn = cc->migrate_pfn + zone->spanned_pages;
  500. cc->free_pfn &= ~(pageblock_nr_pages-1);
  501. migrate_prep_local();
  502. while ((ret = compact_finished(zone, cc)) == COMPACT_CONTINUE) {
  503. unsigned long nr_migrate, nr_remaining;
  504. int err;
  505. switch (isolate_migratepages(zone, cc)) {
  506. case ISOLATE_ABORT:
  507. ret = COMPACT_PARTIAL;
  508. goto out;
  509. case ISOLATE_NONE:
  510. continue;
  511. case ISOLATE_SUCCESS:
  512. ;
  513. }
  514. nr_migrate = cc->nr_migratepages;
  515. err = migrate_pages(&cc->migratepages, compaction_alloc,
  516. (unsigned long)cc, false,
  517. cc->sync ? MIGRATE_SYNC_LIGHT : MIGRATE_ASYNC);
  518. update_nr_listpages(cc);
  519. nr_remaining = cc->nr_migratepages;
  520. count_vm_event(COMPACTBLOCKS);
  521. count_vm_events(COMPACTPAGES, nr_migrate - nr_remaining);
  522. if (nr_remaining)
  523. count_vm_events(COMPACTPAGEFAILED, nr_remaining);
  524. trace_mm_compaction_migratepages(nr_migrate - nr_remaining,
  525. nr_remaining);
  526. /* Release LRU pages not migrated */
  527. if (err) {
  528. putback_lru_pages(&cc->migratepages);
  529. cc->nr_migratepages = 0;
  530. }
  531. }
  532. out:
  533. /* Release free pages and check accounting */
  534. cc->nr_freepages -= release_freepages(&cc->freepages);
  535. VM_BUG_ON(cc->nr_freepages != 0);
  536. return ret;
  537. }
  538. static unsigned long compact_zone_order(struct zone *zone,
  539. int order, gfp_t gfp_mask,
  540. bool sync)
  541. {
  542. struct compact_control cc = {
  543. .nr_freepages = 0,
  544. .nr_migratepages = 0,
  545. .order = order,
  546. .migratetype = allocflags_to_migratetype(gfp_mask),
  547. .zone = zone,
  548. .sync = sync,
  549. };
  550. INIT_LIST_HEAD(&cc.freepages);
  551. INIT_LIST_HEAD(&cc.migratepages);
  552. return compact_zone(zone, &cc);
  553. }
  554. int sysctl_extfrag_threshold = 500;
  555. /**
  556. * try_to_compact_pages - Direct compact to satisfy a high-order allocation
  557. * @zonelist: The zonelist used for the current allocation
  558. * @order: The order of the current allocation
  559. * @gfp_mask: The GFP mask of the current allocation
  560. * @nodemask: The allowed nodes to allocate from
  561. * @sync: Whether migration is synchronous or not
  562. *
  563. * This is the main entry point for direct page compaction.
  564. */
  565. unsigned long try_to_compact_pages(struct zonelist *zonelist,
  566. int order, gfp_t gfp_mask, nodemask_t *nodemask,
  567. bool sync)
  568. {
  569. enum zone_type high_zoneidx = gfp_zone(gfp_mask);
  570. int may_enter_fs = gfp_mask & __GFP_FS;
  571. int may_perform_io = gfp_mask & __GFP_IO;
  572. struct zoneref *z;
  573. struct zone *zone;
  574. int rc = COMPACT_SKIPPED;
  575. /*
  576. * Check whether it is worth even starting compaction. The order check is
  577. * made because an assumption is made that the page allocator can satisfy
  578. * the "cheaper" orders without taking special steps
  579. */
  580. if (!order || !may_enter_fs || !may_perform_io)
  581. return rc;
  582. count_vm_event(COMPACTSTALL);
  583. /* Compact each zone in the list */
  584. for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
  585. nodemask) {
  586. int status;
  587. status = compact_zone_order(zone, order, gfp_mask, sync);
  588. rc = max(status, rc);
  589. /* If a normal allocation would succeed, stop compacting */
  590. if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
  591. break;
  592. }
  593. return rc;
  594. }
  595. /* Compact all zones within a node */
  596. static int __compact_pgdat(pg_data_t *pgdat, struct compact_control *cc)
  597. {
  598. int zoneid;
  599. struct zone *zone;
  600. for (zoneid = 0; zoneid < MAX_NR_ZONES; zoneid++) {
  601. zone = &pgdat->node_zones[zoneid];
  602. if (!populated_zone(zone))
  603. continue;
  604. cc->nr_freepages = 0;
  605. cc->nr_migratepages = 0;
  606. cc->zone = zone;
  607. INIT_LIST_HEAD(&cc->freepages);
  608. INIT_LIST_HEAD(&cc->migratepages);
  609. if (cc->order == -1 || !compaction_deferred(zone, cc->order))
  610. compact_zone(zone, cc);
  611. if (cc->order > 0) {
  612. int ok = zone_watermark_ok(zone, cc->order,
  613. low_wmark_pages(zone), 0, 0);
  614. if (ok && cc->order > zone->compact_order_failed)
  615. zone->compact_order_failed = cc->order + 1;
  616. /* Currently async compaction is never deferred. */
  617. else if (!ok && cc->sync)
  618. defer_compaction(zone, cc->order);
  619. }
  620. VM_BUG_ON(!list_empty(&cc->freepages));
  621. VM_BUG_ON(!list_empty(&cc->migratepages));
  622. }
  623. return 0;
  624. }
  625. int compact_pgdat(pg_data_t *pgdat, int order)
  626. {
  627. struct compact_control cc = {
  628. .order = order,
  629. .sync = false,
  630. };
  631. return __compact_pgdat(pgdat, &cc);
  632. }
  633. static int compact_node(int nid)
  634. {
  635. struct compact_control cc = {
  636. .order = -1,
  637. .sync = true,
  638. };
  639. return __compact_pgdat(NODE_DATA(nid), &cc);
  640. }
  641. /* Compact all nodes in the system */
  642. static int compact_nodes(void)
  643. {
  644. int nid;
  645. /* Flush pending updates to the LRU lists */
  646. lru_add_drain_all();
  647. for_each_online_node(nid)
  648. compact_node(nid);
  649. return COMPACT_COMPLETE;
  650. }
  651. /* The written value is actually unused, all memory is compacted */
  652. int sysctl_compact_memory;
  653. /* This is the entry point for compacting all nodes via /proc/sys/vm */
  654. int sysctl_compaction_handler(struct ctl_table *table, int write,
  655. void __user *buffer, size_t *length, loff_t *ppos)
  656. {
  657. if (write)
  658. return compact_nodes();
  659. return 0;
  660. }
  661. int sysctl_extfrag_handler(struct ctl_table *table, int write,
  662. void __user *buffer, size_t *length, loff_t *ppos)
  663. {
  664. proc_dointvec_minmax(table, write, buffer, length, ppos);
  665. return 0;
  666. }
  667. #if defined(CONFIG_SYSFS) && defined(CONFIG_NUMA)
  668. ssize_t sysfs_compact_node(struct device *dev,
  669. struct device_attribute *attr,
  670. const char *buf, size_t count)
  671. {
  672. int nid = dev->id;
  673. if (nid >= 0 && nid < nr_node_ids && node_online(nid)) {
  674. /* Flush pending updates to the LRU lists */
  675. lru_add_drain_all();
  676. compact_node(nid);
  677. }
  678. return count;
  679. }
  680. static DEVICE_ATTR(compact, S_IWUSR, NULL, sysfs_compact_node);
  681. int compaction_register_node(struct node *node)
  682. {
  683. return device_create_file(&node->dev, &dev_attr_compact);
  684. }
  685. void compaction_unregister_node(struct node *node)
  686. {
  687. return device_remove_file(&node->dev, &dev_attr_compact);
  688. }
  689. #endif /* CONFIG_SYSFS && CONFIG_NUMA */