page-writeback.c 35 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238
  1. /*
  2. * mm/page-writeback.c
  3. *
  4. * Copyright (C) 2002, Linus Torvalds.
  5. * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  6. *
  7. * Contains functions related to writing back dirty pages at the
  8. * address_space level.
  9. *
  10. * 10Apr2002 akpm@zip.com.au
  11. * Initial version
  12. */
  13. #include <linux/kernel.h>
  14. #include <linux/module.h>
  15. #include <linux/spinlock.h>
  16. #include <linux/fs.h>
  17. #include <linux/mm.h>
  18. #include <linux/swap.h>
  19. #include <linux/slab.h>
  20. #include <linux/pagemap.h>
  21. #include <linux/writeback.h>
  22. #include <linux/init.h>
  23. #include <linux/backing-dev.h>
  24. #include <linux/task_io_accounting_ops.h>
  25. #include <linux/blkdev.h>
  26. #include <linux/mpage.h>
  27. #include <linux/rmap.h>
  28. #include <linux/percpu.h>
  29. #include <linux/notifier.h>
  30. #include <linux/smp.h>
  31. #include <linux/sysctl.h>
  32. #include <linux/cpu.h>
  33. #include <linux/syscalls.h>
  34. #include <linux/buffer_head.h>
  35. #include <linux/pagevec.h>
  36. /*
  37. * The maximum number of pages to writeout in a single bdflush/kupdate
  38. * operation. We do this so we don't hold I_SYNC against an inode for
  39. * enormous amounts of time, which would block a userspace task which has
  40. * been forced to throttle against that inode. Also, the code reevaluates
  41. * the dirty each time it has written this many pages.
  42. */
  43. #define MAX_WRITEBACK_PAGES 1024
  44. /*
  45. * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  46. * will look to see if it needs to force writeback or throttling.
  47. */
  48. static long ratelimit_pages = 32;
  49. /*
  50. * When balance_dirty_pages decides that the caller needs to perform some
  51. * non-background writeback, this is how many pages it will attempt to write.
  52. * It should be somewhat larger than RATELIMIT_PAGES to ensure that reasonably
  53. * large amounts of I/O are submitted.
  54. */
  55. static inline long sync_writeback_pages(void)
  56. {
  57. return ratelimit_pages + ratelimit_pages / 2;
  58. }
  59. /* The following parameters are exported via /proc/sys/vm */
  60. /*
  61. * Start background writeback (via pdflush) at this percentage
  62. */
  63. int dirty_background_ratio = 5;
  64. /*
  65. * The generator of dirty data starts writeback at this percentage
  66. */
  67. int vm_dirty_ratio = 10;
  68. /*
  69. * The interval between `kupdate'-style writebacks, in jiffies
  70. */
  71. int dirty_writeback_interval = 5 * HZ;
  72. /*
  73. * The longest number of jiffies for which data is allowed to remain dirty
  74. */
  75. int dirty_expire_interval = 30 * HZ;
  76. /*
  77. * Flag that makes the machine dump writes/reads and block dirtyings.
  78. */
  79. int block_dump;
  80. /*
  81. * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
  82. * a full sync is triggered after this time elapses without any disk activity.
  83. */
  84. int laptop_mode;
  85. EXPORT_SYMBOL(laptop_mode);
  86. /* End of sysctl-exported parameters */
  87. static void background_writeout(unsigned long _min_pages);
  88. /*
  89. * Scale the writeback cache size proportional to the relative writeout speeds.
  90. *
  91. * We do this by keeping a floating proportion between BDIs, based on page
  92. * writeback completions [end_page_writeback()]. Those devices that write out
  93. * pages fastest will get the larger share, while the slower will get a smaller
  94. * share.
  95. *
  96. * We use page writeout completions because we are interested in getting rid of
  97. * dirty pages. Having them written out is the primary goal.
  98. *
  99. * We introduce a concept of time, a period over which we measure these events,
  100. * because demand can/will vary over time. The length of this period itself is
  101. * measured in page writeback completions.
  102. *
  103. */
  104. static struct prop_descriptor vm_completions;
  105. static struct prop_descriptor vm_dirties;
  106. static unsigned long determine_dirtyable_memory(void);
  107. /*
  108. * couple the period to the dirty_ratio:
  109. *
  110. * period/2 ~ roundup_pow_of_two(dirty limit)
  111. */
  112. static int calc_period_shift(void)
  113. {
  114. unsigned long dirty_total;
  115. dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
  116. return 2 + ilog2(dirty_total - 1);
  117. }
  118. /*
  119. * update the period when the dirty ratio changes.
  120. */
  121. int dirty_ratio_handler(struct ctl_table *table, int write,
  122. struct file *filp, void __user *buffer, size_t *lenp,
  123. loff_t *ppos)
  124. {
  125. int old_ratio = vm_dirty_ratio;
  126. int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
  127. if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
  128. int shift = calc_period_shift();
  129. prop_change_shift(&vm_completions, shift);
  130. prop_change_shift(&vm_dirties, shift);
  131. }
  132. return ret;
  133. }
  134. /*
  135. * Increment the BDI's writeout completion count and the global writeout
  136. * completion count. Called from test_clear_page_writeback().
  137. */
  138. static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
  139. {
  140. __prop_inc_percpu(&vm_completions, &bdi->completions);
  141. }
  142. static inline void task_dirty_inc(struct task_struct *tsk)
  143. {
  144. prop_inc_single(&vm_dirties, &tsk->dirties);
  145. }
  146. /*
  147. * Obtain an accurate fraction of the BDI's portion.
  148. */
  149. static void bdi_writeout_fraction(struct backing_dev_info *bdi,
  150. long *numerator, long *denominator)
  151. {
  152. if (bdi_cap_writeback_dirty(bdi)) {
  153. prop_fraction_percpu(&vm_completions, &bdi->completions,
  154. numerator, denominator);
  155. } else {
  156. *numerator = 0;
  157. *denominator = 1;
  158. }
  159. }
  160. /*
  161. * Clip the earned share of dirty pages to that which is actually available.
  162. * This avoids exceeding the total dirty_limit when the floating averages
  163. * fluctuate too quickly.
  164. */
  165. static void
  166. clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
  167. {
  168. long avail_dirty;
  169. avail_dirty = dirty -
  170. (global_page_state(NR_FILE_DIRTY) +
  171. global_page_state(NR_WRITEBACK) +
  172. global_page_state(NR_UNSTABLE_NFS));
  173. if (avail_dirty < 0)
  174. avail_dirty = 0;
  175. avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
  176. bdi_stat(bdi, BDI_WRITEBACK);
  177. *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
  178. }
  179. static inline void task_dirties_fraction(struct task_struct *tsk,
  180. long *numerator, long *denominator)
  181. {
  182. prop_fraction_single(&vm_dirties, &tsk->dirties,
  183. numerator, denominator);
  184. }
  185. /*
  186. * scale the dirty limit
  187. *
  188. * task specific dirty limit:
  189. *
  190. * dirty -= (dirty/8) * p_{t}
  191. */
  192. void task_dirty_limit(struct task_struct *tsk, long *pdirty)
  193. {
  194. long numerator, denominator;
  195. long dirty = *pdirty;
  196. u64 inv = dirty >> 3;
  197. task_dirties_fraction(tsk, &numerator, &denominator);
  198. inv *= numerator;
  199. do_div(inv, denominator);
  200. dirty -= inv;
  201. if (dirty < *pdirty/2)
  202. dirty = *pdirty/2;
  203. *pdirty = dirty;
  204. }
  205. /*
  206. * Work out the current dirty-memory clamping and background writeout
  207. * thresholds.
  208. *
  209. * The main aim here is to lower them aggressively if there is a lot of mapped
  210. * memory around. To avoid stressing page reclaim with lots of unreclaimable
  211. * pages. It is better to clamp down on writers than to start swapping, and
  212. * performing lots of scanning.
  213. *
  214. * We only allow 1/2 of the currently-unmapped memory to be dirtied.
  215. *
  216. * We don't permit the clamping level to fall below 5% - that is getting rather
  217. * excessive.
  218. *
  219. * We make sure that the background writeout level is below the adjusted
  220. * clamping level.
  221. */
  222. static unsigned long highmem_dirtyable_memory(unsigned long total)
  223. {
  224. #ifdef CONFIG_HIGHMEM
  225. int node;
  226. unsigned long x = 0;
  227. for_each_node_state(node, N_HIGH_MEMORY) {
  228. struct zone *z =
  229. &NODE_DATA(node)->node_zones[ZONE_HIGHMEM];
  230. x += zone_page_state(z, NR_FREE_PAGES)
  231. + zone_page_state(z, NR_INACTIVE)
  232. + zone_page_state(z, NR_ACTIVE);
  233. }
  234. /*
  235. * Make sure that the number of highmem pages is never larger
  236. * than the number of the total dirtyable memory. This can only
  237. * occur in very strange VM situations but we want to make sure
  238. * that this does not occur.
  239. */
  240. return min(x, total);
  241. #else
  242. return 0;
  243. #endif
  244. }
  245. static unsigned long determine_dirtyable_memory(void)
  246. {
  247. unsigned long x;
  248. x = global_page_state(NR_FREE_PAGES)
  249. + global_page_state(NR_INACTIVE)
  250. + global_page_state(NR_ACTIVE);
  251. x -= highmem_dirtyable_memory(x);
  252. return x + 1; /* Ensure that we never return 0 */
  253. }
  254. static void
  255. get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
  256. struct backing_dev_info *bdi)
  257. {
  258. int background_ratio; /* Percentages */
  259. int dirty_ratio;
  260. int unmapped_ratio;
  261. long background;
  262. long dirty;
  263. unsigned long available_memory = determine_dirtyable_memory();
  264. struct task_struct *tsk;
  265. unmapped_ratio = 100 - ((global_page_state(NR_FILE_MAPPED) +
  266. global_page_state(NR_ANON_PAGES)) * 100) /
  267. available_memory;
  268. dirty_ratio = vm_dirty_ratio;
  269. if (dirty_ratio > unmapped_ratio / 2)
  270. dirty_ratio = unmapped_ratio / 2;
  271. if (dirty_ratio < 5)
  272. dirty_ratio = 5;
  273. background_ratio = dirty_background_ratio;
  274. if (background_ratio >= dirty_ratio)
  275. background_ratio = dirty_ratio / 2;
  276. background = (background_ratio * available_memory) / 100;
  277. dirty = (dirty_ratio * available_memory) / 100;
  278. tsk = current;
  279. if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
  280. background += background / 4;
  281. dirty += dirty / 4;
  282. }
  283. *pbackground = background;
  284. *pdirty = dirty;
  285. if (bdi) {
  286. u64 bdi_dirty = dirty;
  287. long numerator, denominator;
  288. /*
  289. * Calculate this BDI's share of the dirty ratio.
  290. */
  291. bdi_writeout_fraction(bdi, &numerator, &denominator);
  292. bdi_dirty *= numerator;
  293. do_div(bdi_dirty, denominator);
  294. *pbdi_dirty = bdi_dirty;
  295. clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
  296. task_dirty_limit(current, pbdi_dirty);
  297. }
  298. }
  299. /*
  300. * balance_dirty_pages() must be called by processes which are generating dirty
  301. * data. It looks at the number of dirty pages in the machine and will force
  302. * the caller to perform writeback if the system is over `vm_dirty_ratio'.
  303. * If we're over `background_thresh' then pdflush is woken to perform some
  304. * writeout.
  305. */
  306. static void balance_dirty_pages(struct address_space *mapping)
  307. {
  308. long bdi_nr_reclaimable;
  309. long bdi_nr_writeback;
  310. long background_thresh;
  311. long dirty_thresh;
  312. long bdi_thresh;
  313. unsigned long pages_written = 0;
  314. unsigned long write_chunk = sync_writeback_pages();
  315. struct backing_dev_info *bdi = mapping->backing_dev_info;
  316. for (;;) {
  317. struct writeback_control wbc = {
  318. .bdi = bdi,
  319. .sync_mode = WB_SYNC_NONE,
  320. .older_than_this = NULL,
  321. .nr_to_write = write_chunk,
  322. .range_cyclic = 1,
  323. };
  324. get_dirty_limits(&background_thresh, &dirty_thresh,
  325. &bdi_thresh, bdi);
  326. bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
  327. bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
  328. if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
  329. break;
  330. if (!bdi->dirty_exceeded)
  331. bdi->dirty_exceeded = 1;
  332. /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
  333. * Unstable writes are a feature of certain networked
  334. * filesystems (i.e. NFS) in which data may have been
  335. * written to the server's write cache, but has not yet
  336. * been flushed to permanent storage.
  337. */
  338. if (bdi_nr_reclaimable) {
  339. writeback_inodes(&wbc);
  340. pages_written += write_chunk - wbc.nr_to_write;
  341. get_dirty_limits(&background_thresh, &dirty_thresh,
  342. &bdi_thresh, bdi);
  343. }
  344. /*
  345. * In order to avoid the stacked BDI deadlock we need
  346. * to ensure we accurately count the 'dirty' pages when
  347. * the threshold is low.
  348. *
  349. * Otherwise it would be possible to get thresh+n pages
  350. * reported dirty, even though there are thresh-m pages
  351. * actually dirty; with m+n sitting in the percpu
  352. * deltas.
  353. */
  354. if (bdi_thresh < 2*bdi_stat_error(bdi)) {
  355. bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
  356. bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
  357. } else if (bdi_nr_reclaimable) {
  358. bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
  359. bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
  360. }
  361. if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
  362. break;
  363. if (pages_written >= write_chunk)
  364. break; /* We've done our duty */
  365. congestion_wait(WRITE, HZ/10);
  366. }
  367. if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
  368. bdi->dirty_exceeded)
  369. bdi->dirty_exceeded = 0;
  370. if (writeback_in_progress(bdi))
  371. return; /* pdflush is already working this queue */
  372. /*
  373. * In laptop mode, we wait until hitting the higher threshold before
  374. * starting background writeout, and then write out all the way down
  375. * to the lower threshold. So slow writers cause minimal disk activity.
  376. *
  377. * In normal mode, we start background writeout at the lower
  378. * background_thresh, to keep the amount of dirty memory low.
  379. */
  380. if ((laptop_mode && pages_written) ||
  381. (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
  382. + global_page_state(NR_UNSTABLE_NFS)
  383. > background_thresh)))
  384. pdflush_operation(background_writeout, 0);
  385. }
  386. void set_page_dirty_balance(struct page *page, int page_mkwrite)
  387. {
  388. if (set_page_dirty(page) || page_mkwrite) {
  389. struct address_space *mapping = page_mapping(page);
  390. if (mapping)
  391. balance_dirty_pages_ratelimited(mapping);
  392. }
  393. }
  394. /**
  395. * balance_dirty_pages_ratelimited_nr - balance dirty memory state
  396. * @mapping: address_space which was dirtied
  397. * @nr_pages_dirtied: number of pages which the caller has just dirtied
  398. *
  399. * Processes which are dirtying memory should call in here once for each page
  400. * which was newly dirtied. The function will periodically check the system's
  401. * dirty state and will initiate writeback if needed.
  402. *
  403. * On really big machines, get_writeback_state is expensive, so try to avoid
  404. * calling it too often (ratelimiting). But once we're over the dirty memory
  405. * limit we decrease the ratelimiting by a lot, to prevent individual processes
  406. * from overshooting the limit by (ratelimit_pages) each.
  407. */
  408. void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
  409. unsigned long nr_pages_dirtied)
  410. {
  411. static DEFINE_PER_CPU(unsigned long, ratelimits) = 0;
  412. unsigned long ratelimit;
  413. unsigned long *p;
  414. ratelimit = ratelimit_pages;
  415. if (mapping->backing_dev_info->dirty_exceeded)
  416. ratelimit = 8;
  417. /*
  418. * Check the rate limiting. Also, we do not want to throttle real-time
  419. * tasks in balance_dirty_pages(). Period.
  420. */
  421. preempt_disable();
  422. p = &__get_cpu_var(ratelimits);
  423. *p += nr_pages_dirtied;
  424. if (unlikely(*p >= ratelimit)) {
  425. *p = 0;
  426. preempt_enable();
  427. balance_dirty_pages(mapping);
  428. return;
  429. }
  430. preempt_enable();
  431. }
  432. EXPORT_SYMBOL(balance_dirty_pages_ratelimited_nr);
  433. void throttle_vm_writeout(gfp_t gfp_mask)
  434. {
  435. long background_thresh;
  436. long dirty_thresh;
  437. for ( ; ; ) {
  438. get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
  439. /*
  440. * Boost the allowable dirty threshold a bit for page
  441. * allocators so they don't get DoS'ed by heavy writers
  442. */
  443. dirty_thresh += dirty_thresh / 10; /* wheeee... */
  444. if (global_page_state(NR_UNSTABLE_NFS) +
  445. global_page_state(NR_WRITEBACK) <= dirty_thresh)
  446. break;
  447. congestion_wait(WRITE, HZ/10);
  448. /*
  449. * The caller might hold locks which can prevent IO completion
  450. * or progress in the filesystem. So we cannot just sit here
  451. * waiting for IO to complete.
  452. */
  453. if ((gfp_mask & (__GFP_FS|__GFP_IO)) != (__GFP_FS|__GFP_IO))
  454. break;
  455. }
  456. }
  457. /*
  458. * writeback at least _min_pages, and keep writing until the amount of dirty
  459. * memory is less than the background threshold, or until we're all clean.
  460. */
  461. static void background_writeout(unsigned long _min_pages)
  462. {
  463. long min_pages = _min_pages;
  464. struct writeback_control wbc = {
  465. .bdi = NULL,
  466. .sync_mode = WB_SYNC_NONE,
  467. .older_than_this = NULL,
  468. .nr_to_write = 0,
  469. .nonblocking = 1,
  470. .range_cyclic = 1,
  471. };
  472. for ( ; ; ) {
  473. long background_thresh;
  474. long dirty_thresh;
  475. get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
  476. if (global_page_state(NR_FILE_DIRTY) +
  477. global_page_state(NR_UNSTABLE_NFS) < background_thresh
  478. && min_pages <= 0)
  479. break;
  480. wbc.more_io = 0;
  481. wbc.encountered_congestion = 0;
  482. wbc.nr_to_write = MAX_WRITEBACK_PAGES;
  483. wbc.pages_skipped = 0;
  484. writeback_inodes(&wbc);
  485. min_pages -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
  486. if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
  487. /* Wrote less than expected */
  488. if (wbc.encountered_congestion || wbc.more_io)
  489. congestion_wait(WRITE, HZ/10);
  490. else
  491. break;
  492. }
  493. }
  494. }
  495. /*
  496. * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back
  497. * the whole world. Returns 0 if a pdflush thread was dispatched. Returns
  498. * -1 if all pdflush threads were busy.
  499. */
  500. int wakeup_pdflush(long nr_pages)
  501. {
  502. if (nr_pages == 0)
  503. nr_pages = global_page_state(NR_FILE_DIRTY) +
  504. global_page_state(NR_UNSTABLE_NFS);
  505. return pdflush_operation(background_writeout, nr_pages);
  506. }
  507. static void wb_timer_fn(unsigned long unused);
  508. static void laptop_timer_fn(unsigned long unused);
  509. static DEFINE_TIMER(wb_timer, wb_timer_fn, 0, 0);
  510. static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
  511. /*
  512. * Periodic writeback of "old" data.
  513. *
  514. * Define "old": the first time one of an inode's pages is dirtied, we mark the
  515. * dirtying-time in the inode's address_space. So this periodic writeback code
  516. * just walks the superblock inode list, writing back any inodes which are
  517. * older than a specific point in time.
  518. *
  519. * Try to run once per dirty_writeback_interval. But if a writeback event
  520. * takes longer than a dirty_writeback_interval interval, then leave a
  521. * one-second gap.
  522. *
  523. * older_than_this takes precedence over nr_to_write. So we'll only write back
  524. * all dirty pages if they are all attached to "old" mappings.
  525. */
  526. static void wb_kupdate(unsigned long arg)
  527. {
  528. unsigned long oldest_jif;
  529. unsigned long start_jif;
  530. unsigned long next_jif;
  531. long nr_to_write;
  532. struct writeback_control wbc = {
  533. .bdi = NULL,
  534. .sync_mode = WB_SYNC_NONE,
  535. .older_than_this = &oldest_jif,
  536. .nr_to_write = 0,
  537. .nonblocking = 1,
  538. .for_kupdate = 1,
  539. .range_cyclic = 1,
  540. };
  541. sync_supers();
  542. oldest_jif = jiffies - dirty_expire_interval;
  543. start_jif = jiffies;
  544. next_jif = start_jif + dirty_writeback_interval;
  545. nr_to_write = global_page_state(NR_FILE_DIRTY) +
  546. global_page_state(NR_UNSTABLE_NFS) +
  547. (inodes_stat.nr_inodes - inodes_stat.nr_unused);
  548. while (nr_to_write > 0) {
  549. wbc.more_io = 0;
  550. wbc.encountered_congestion = 0;
  551. wbc.nr_to_write = MAX_WRITEBACK_PAGES;
  552. writeback_inodes(&wbc);
  553. if (wbc.nr_to_write > 0) {
  554. if (wbc.encountered_congestion || wbc.more_io)
  555. congestion_wait(WRITE, HZ/10);
  556. else
  557. break; /* All the old data is written */
  558. }
  559. nr_to_write -= MAX_WRITEBACK_PAGES - wbc.nr_to_write;
  560. }
  561. if (time_before(next_jif, jiffies + HZ))
  562. next_jif = jiffies + HZ;
  563. if (dirty_writeback_interval)
  564. mod_timer(&wb_timer, next_jif);
  565. }
  566. /*
  567. * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
  568. */
  569. int dirty_writeback_centisecs_handler(ctl_table *table, int write,
  570. struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
  571. {
  572. proc_dointvec_userhz_jiffies(table, write, file, buffer, length, ppos);
  573. if (dirty_writeback_interval)
  574. mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
  575. else
  576. del_timer(&wb_timer);
  577. return 0;
  578. }
  579. static void wb_timer_fn(unsigned long unused)
  580. {
  581. if (pdflush_operation(wb_kupdate, 0) < 0)
  582. mod_timer(&wb_timer, jiffies + HZ); /* delay 1 second */
  583. }
  584. static void laptop_flush(unsigned long unused)
  585. {
  586. sys_sync();
  587. }
  588. static void laptop_timer_fn(unsigned long unused)
  589. {
  590. pdflush_operation(laptop_flush, 0);
  591. }
  592. /*
  593. * We've spun up the disk and we're in laptop mode: schedule writeback
  594. * of all dirty data a few seconds from now. If the flush is already scheduled
  595. * then push it back - the user is still using the disk.
  596. */
  597. void laptop_io_completion(void)
  598. {
  599. mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
  600. }
  601. /*
  602. * We're in laptop mode and we've just synced. The sync's writes will have
  603. * caused another writeback to be scheduled by laptop_io_completion.
  604. * Nothing needs to be written back anymore, so we unschedule the writeback.
  605. */
  606. void laptop_sync_completion(void)
  607. {
  608. del_timer(&laptop_mode_wb_timer);
  609. }
  610. /*
  611. * If ratelimit_pages is too high then we can get into dirty-data overload
  612. * if a large number of processes all perform writes at the same time.
  613. * If it is too low then SMP machines will call the (expensive)
  614. * get_writeback_state too often.
  615. *
  616. * Here we set ratelimit_pages to a level which ensures that when all CPUs are
  617. * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
  618. * thresholds before writeback cuts in.
  619. *
  620. * But the limit should not be set too high. Because it also controls the
  621. * amount of memory which the balance_dirty_pages() caller has to write back.
  622. * If this is too large then the caller will block on the IO queue all the
  623. * time. So limit it to four megabytes - the balance_dirty_pages() caller
  624. * will write six megabyte chunks, max.
  625. */
  626. void writeback_set_ratelimit(void)
  627. {
  628. ratelimit_pages = vm_total_pages / (num_online_cpus() * 32);
  629. if (ratelimit_pages < 16)
  630. ratelimit_pages = 16;
  631. if (ratelimit_pages * PAGE_CACHE_SIZE > 4096 * 1024)
  632. ratelimit_pages = (4096 * 1024) / PAGE_CACHE_SIZE;
  633. }
  634. static int __cpuinit
  635. ratelimit_handler(struct notifier_block *self, unsigned long u, void *v)
  636. {
  637. writeback_set_ratelimit();
  638. return NOTIFY_DONE;
  639. }
  640. static struct notifier_block __cpuinitdata ratelimit_nb = {
  641. .notifier_call = ratelimit_handler,
  642. .next = NULL,
  643. };
  644. /*
  645. * Called early on to tune the page writeback dirty limits.
  646. *
  647. * We used to scale dirty pages according to how total memory
  648. * related to pages that could be allocated for buffers (by
  649. * comparing nr_free_buffer_pages() to vm_total_pages.
  650. *
  651. * However, that was when we used "dirty_ratio" to scale with
  652. * all memory, and we don't do that any more. "dirty_ratio"
  653. * is now applied to total non-HIGHPAGE memory (by subtracting
  654. * totalhigh_pages from vm_total_pages), and as such we can't
  655. * get into the old insane situation any more where we had
  656. * large amounts of dirty pages compared to a small amount of
  657. * non-HIGHMEM memory.
  658. *
  659. * But we might still want to scale the dirty_ratio by how
  660. * much memory the box has..
  661. */
  662. void __init page_writeback_init(void)
  663. {
  664. int shift;
  665. mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
  666. writeback_set_ratelimit();
  667. register_cpu_notifier(&ratelimit_nb);
  668. shift = calc_period_shift();
  669. prop_descriptor_init(&vm_completions, shift);
  670. prop_descriptor_init(&vm_dirties, shift);
  671. }
  672. /**
  673. * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
  674. * @mapping: address space structure to write
  675. * @wbc: subtract the number of written pages from *@wbc->nr_to_write
  676. * @writepage: function called for each page
  677. * @data: data passed to writepage function
  678. *
  679. * If a page is already under I/O, write_cache_pages() skips it, even
  680. * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
  681. * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
  682. * and msync() need to guarantee that all the data which was dirty at the time
  683. * the call was made get new I/O started against them. If wbc->sync_mode is
  684. * WB_SYNC_ALL then we were called for data integrity and we must wait for
  685. * existing IO to complete.
  686. */
  687. int write_cache_pages(struct address_space *mapping,
  688. struct writeback_control *wbc, writepage_t writepage,
  689. void *data)
  690. {
  691. struct backing_dev_info *bdi = mapping->backing_dev_info;
  692. int ret = 0;
  693. int done = 0;
  694. struct pagevec pvec;
  695. int nr_pages;
  696. pgoff_t index;
  697. pgoff_t end; /* Inclusive */
  698. int scanned = 0;
  699. int range_whole = 0;
  700. if (wbc->nonblocking && bdi_write_congested(bdi)) {
  701. wbc->encountered_congestion = 1;
  702. return 0;
  703. }
  704. pagevec_init(&pvec, 0);
  705. if (wbc->range_cyclic) {
  706. index = mapping->writeback_index; /* Start from prev offset */
  707. end = -1;
  708. } else {
  709. index = wbc->range_start >> PAGE_CACHE_SHIFT;
  710. end = wbc->range_end >> PAGE_CACHE_SHIFT;
  711. if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
  712. range_whole = 1;
  713. scanned = 1;
  714. }
  715. retry:
  716. while (!done && (index <= end) &&
  717. (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
  718. PAGECACHE_TAG_DIRTY,
  719. min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
  720. unsigned i;
  721. scanned = 1;
  722. for (i = 0; i < nr_pages; i++) {
  723. struct page *page = pvec.pages[i];
  724. /*
  725. * At this point we hold neither mapping->tree_lock nor
  726. * lock on the page itself: the page may be truncated or
  727. * invalidated (changing page->mapping to NULL), or even
  728. * swizzled back from swapper_space to tmpfs file
  729. * mapping
  730. */
  731. lock_page(page);
  732. if (unlikely(page->mapping != mapping)) {
  733. unlock_page(page);
  734. continue;
  735. }
  736. if (!wbc->range_cyclic && page->index > end) {
  737. done = 1;
  738. unlock_page(page);
  739. continue;
  740. }
  741. if (wbc->sync_mode != WB_SYNC_NONE)
  742. wait_on_page_writeback(page);
  743. if (PageWriteback(page) ||
  744. !clear_page_dirty_for_io(page)) {
  745. unlock_page(page);
  746. continue;
  747. }
  748. ret = (*writepage)(page, wbc, data);
  749. if (unlikely(ret == AOP_WRITEPAGE_ACTIVATE)) {
  750. unlock_page(page);
  751. ret = 0;
  752. }
  753. if (ret || (--(wbc->nr_to_write) <= 0))
  754. done = 1;
  755. if (wbc->nonblocking && bdi_write_congested(bdi)) {
  756. wbc->encountered_congestion = 1;
  757. done = 1;
  758. }
  759. }
  760. pagevec_release(&pvec);
  761. cond_resched();
  762. }
  763. if (!scanned && !done) {
  764. /*
  765. * We hit the last page and there is more work to be done: wrap
  766. * back to the start of the file
  767. */
  768. scanned = 1;
  769. index = 0;
  770. goto retry;
  771. }
  772. if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
  773. mapping->writeback_index = index;
  774. return ret;
  775. }
  776. EXPORT_SYMBOL(write_cache_pages);
  777. /*
  778. * Function used by generic_writepages to call the real writepage
  779. * function and set the mapping flags on error
  780. */
  781. static int __writepage(struct page *page, struct writeback_control *wbc,
  782. void *data)
  783. {
  784. struct address_space *mapping = data;
  785. int ret = mapping->a_ops->writepage(page, wbc);
  786. mapping_set_error(mapping, ret);
  787. return ret;
  788. }
  789. /**
  790. * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
  791. * @mapping: address space structure to write
  792. * @wbc: subtract the number of written pages from *@wbc->nr_to_write
  793. *
  794. * This is a library function, which implements the writepages()
  795. * address_space_operation.
  796. */
  797. int generic_writepages(struct address_space *mapping,
  798. struct writeback_control *wbc)
  799. {
  800. /* deal with chardevs and other special file */
  801. if (!mapping->a_ops->writepage)
  802. return 0;
  803. return write_cache_pages(mapping, wbc, __writepage, mapping);
  804. }
  805. EXPORT_SYMBOL(generic_writepages);
  806. int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
  807. {
  808. int ret;
  809. if (wbc->nr_to_write <= 0)
  810. return 0;
  811. wbc->for_writepages = 1;
  812. if (mapping->a_ops->writepages)
  813. ret = mapping->a_ops->writepages(mapping, wbc);
  814. else
  815. ret = generic_writepages(mapping, wbc);
  816. wbc->for_writepages = 0;
  817. return ret;
  818. }
  819. /**
  820. * write_one_page - write out a single page and optionally wait on I/O
  821. * @page: the page to write
  822. * @wait: if true, wait on writeout
  823. *
  824. * The page must be locked by the caller and will be unlocked upon return.
  825. *
  826. * write_one_page() returns a negative error code if I/O failed.
  827. */
  828. int write_one_page(struct page *page, int wait)
  829. {
  830. struct address_space *mapping = page->mapping;
  831. int ret = 0;
  832. struct writeback_control wbc = {
  833. .sync_mode = WB_SYNC_ALL,
  834. .nr_to_write = 1,
  835. };
  836. BUG_ON(!PageLocked(page));
  837. if (wait)
  838. wait_on_page_writeback(page);
  839. if (clear_page_dirty_for_io(page)) {
  840. page_cache_get(page);
  841. ret = mapping->a_ops->writepage(page, &wbc);
  842. if (ret == 0 && wait) {
  843. wait_on_page_writeback(page);
  844. if (PageError(page))
  845. ret = -EIO;
  846. }
  847. page_cache_release(page);
  848. } else {
  849. unlock_page(page);
  850. }
  851. return ret;
  852. }
  853. EXPORT_SYMBOL(write_one_page);
  854. /*
  855. * For address_spaces which do not use buffers nor write back.
  856. */
  857. int __set_page_dirty_no_writeback(struct page *page)
  858. {
  859. if (!PageDirty(page))
  860. SetPageDirty(page);
  861. return 0;
  862. }
  863. /*
  864. * For address_spaces which do not use buffers. Just tag the page as dirty in
  865. * its radix tree.
  866. *
  867. * This is also used when a single buffer is being dirtied: we want to set the
  868. * page dirty in that case, but not all the buffers. This is a "bottom-up"
  869. * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
  870. *
  871. * Most callers have locked the page, which pins the address_space in memory.
  872. * But zap_pte_range() does not lock the page, however in that case the
  873. * mapping is pinned by the vma's ->vm_file reference.
  874. *
  875. * We take care to handle the case where the page was truncated from the
  876. * mapping by re-checking page_mapping() inside tree_lock.
  877. */
  878. int __set_page_dirty_nobuffers(struct page *page)
  879. {
  880. if (!TestSetPageDirty(page)) {
  881. struct address_space *mapping = page_mapping(page);
  882. struct address_space *mapping2;
  883. if (!mapping)
  884. return 1;
  885. write_lock_irq(&mapping->tree_lock);
  886. mapping2 = page_mapping(page);
  887. if (mapping2) { /* Race with truncate? */
  888. BUG_ON(mapping2 != mapping);
  889. WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
  890. if (mapping_cap_account_dirty(mapping)) {
  891. __inc_zone_page_state(page, NR_FILE_DIRTY);
  892. __inc_bdi_stat(mapping->backing_dev_info,
  893. BDI_RECLAIMABLE);
  894. task_io_account_write(PAGE_CACHE_SIZE);
  895. }
  896. radix_tree_tag_set(&mapping->page_tree,
  897. page_index(page), PAGECACHE_TAG_DIRTY);
  898. }
  899. write_unlock_irq(&mapping->tree_lock);
  900. if (mapping->host) {
  901. /* !PageAnon && !swapper_space */
  902. __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
  903. }
  904. return 1;
  905. }
  906. return 0;
  907. }
  908. EXPORT_SYMBOL(__set_page_dirty_nobuffers);
  909. /*
  910. * When a writepage implementation decides that it doesn't want to write this
  911. * page for some reason, it should redirty the locked page via
  912. * redirty_page_for_writepage() and it should then unlock the page and return 0
  913. */
  914. int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
  915. {
  916. wbc->pages_skipped++;
  917. return __set_page_dirty_nobuffers(page);
  918. }
  919. EXPORT_SYMBOL(redirty_page_for_writepage);
  920. /*
  921. * If the mapping doesn't provide a set_page_dirty a_op, then
  922. * just fall through and assume that it wants buffer_heads.
  923. */
  924. static int __set_page_dirty(struct page *page)
  925. {
  926. struct address_space *mapping = page_mapping(page);
  927. if (likely(mapping)) {
  928. int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
  929. #ifdef CONFIG_BLOCK
  930. if (!spd)
  931. spd = __set_page_dirty_buffers;
  932. #endif
  933. return (*spd)(page);
  934. }
  935. if (!PageDirty(page)) {
  936. if (!TestSetPageDirty(page))
  937. return 1;
  938. }
  939. return 0;
  940. }
  941. int fastcall set_page_dirty(struct page *page)
  942. {
  943. int ret = __set_page_dirty(page);
  944. if (ret)
  945. task_dirty_inc(current);
  946. return ret;
  947. }
  948. EXPORT_SYMBOL(set_page_dirty);
  949. /*
  950. * set_page_dirty() is racy if the caller has no reference against
  951. * page->mapping->host, and if the page is unlocked. This is because another
  952. * CPU could truncate the page off the mapping and then free the mapping.
  953. *
  954. * Usually, the page _is_ locked, or the caller is a user-space process which
  955. * holds a reference on the inode by having an open file.
  956. *
  957. * In other cases, the page should be locked before running set_page_dirty().
  958. */
  959. int set_page_dirty_lock(struct page *page)
  960. {
  961. int ret;
  962. lock_page_nosync(page);
  963. ret = set_page_dirty(page);
  964. unlock_page(page);
  965. return ret;
  966. }
  967. EXPORT_SYMBOL(set_page_dirty_lock);
  968. /*
  969. * Clear a page's dirty flag, while caring for dirty memory accounting.
  970. * Returns true if the page was previously dirty.
  971. *
  972. * This is for preparing to put the page under writeout. We leave the page
  973. * tagged as dirty in the radix tree so that a concurrent write-for-sync
  974. * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
  975. * implementation will run either set_page_writeback() or set_page_dirty(),
  976. * at which stage we bring the page's dirty flag and radix-tree dirty tag
  977. * back into sync.
  978. *
  979. * This incoherency between the page's dirty flag and radix-tree tag is
  980. * unfortunate, but it only exists while the page is locked.
  981. */
  982. int clear_page_dirty_for_io(struct page *page)
  983. {
  984. struct address_space *mapping = page_mapping(page);
  985. BUG_ON(!PageLocked(page));
  986. ClearPageReclaim(page);
  987. if (mapping && mapping_cap_account_dirty(mapping)) {
  988. /*
  989. * Yes, Virginia, this is indeed insane.
  990. *
  991. * We use this sequence to make sure that
  992. * (a) we account for dirty stats properly
  993. * (b) we tell the low-level filesystem to
  994. * mark the whole page dirty if it was
  995. * dirty in a pagetable. Only to then
  996. * (c) clean the page again and return 1 to
  997. * cause the writeback.
  998. *
  999. * This way we avoid all nasty races with the
  1000. * dirty bit in multiple places and clearing
  1001. * them concurrently from different threads.
  1002. *
  1003. * Note! Normally the "set_page_dirty(page)"
  1004. * has no effect on the actual dirty bit - since
  1005. * that will already usually be set. But we
  1006. * need the side effects, and it can help us
  1007. * avoid races.
  1008. *
  1009. * We basically use the page "master dirty bit"
  1010. * as a serialization point for all the different
  1011. * threads doing their things.
  1012. */
  1013. if (page_mkclean(page))
  1014. set_page_dirty(page);
  1015. /*
  1016. * We carefully synchronise fault handlers against
  1017. * installing a dirty pte and marking the page dirty
  1018. * at this point. We do this by having them hold the
  1019. * page lock at some point after installing their
  1020. * pte, but before marking the page dirty.
  1021. * Pages are always locked coming in here, so we get
  1022. * the desired exclusion. See mm/memory.c:do_wp_page()
  1023. * for more comments.
  1024. */
  1025. if (TestClearPageDirty(page)) {
  1026. dec_zone_page_state(page, NR_FILE_DIRTY);
  1027. dec_bdi_stat(mapping->backing_dev_info,
  1028. BDI_RECLAIMABLE);
  1029. return 1;
  1030. }
  1031. return 0;
  1032. }
  1033. return TestClearPageDirty(page);
  1034. }
  1035. EXPORT_SYMBOL(clear_page_dirty_for_io);
  1036. int test_clear_page_writeback(struct page *page)
  1037. {
  1038. struct address_space *mapping = page_mapping(page);
  1039. int ret;
  1040. if (mapping) {
  1041. struct backing_dev_info *bdi = mapping->backing_dev_info;
  1042. unsigned long flags;
  1043. write_lock_irqsave(&mapping->tree_lock, flags);
  1044. ret = TestClearPageWriteback(page);
  1045. if (ret) {
  1046. radix_tree_tag_clear(&mapping->page_tree,
  1047. page_index(page),
  1048. PAGECACHE_TAG_WRITEBACK);
  1049. if (bdi_cap_writeback_dirty(bdi)) {
  1050. __dec_bdi_stat(bdi, BDI_WRITEBACK);
  1051. __bdi_writeout_inc(bdi);
  1052. }
  1053. }
  1054. write_unlock_irqrestore(&mapping->tree_lock, flags);
  1055. } else {
  1056. ret = TestClearPageWriteback(page);
  1057. }
  1058. if (ret)
  1059. dec_zone_page_state(page, NR_WRITEBACK);
  1060. return ret;
  1061. }
  1062. int test_set_page_writeback(struct page *page)
  1063. {
  1064. struct address_space *mapping = page_mapping(page);
  1065. int ret;
  1066. if (mapping) {
  1067. struct backing_dev_info *bdi = mapping->backing_dev_info;
  1068. unsigned long flags;
  1069. write_lock_irqsave(&mapping->tree_lock, flags);
  1070. ret = TestSetPageWriteback(page);
  1071. if (!ret) {
  1072. radix_tree_tag_set(&mapping->page_tree,
  1073. page_index(page),
  1074. PAGECACHE_TAG_WRITEBACK);
  1075. if (bdi_cap_writeback_dirty(bdi))
  1076. __inc_bdi_stat(bdi, BDI_WRITEBACK);
  1077. }
  1078. if (!PageDirty(page))
  1079. radix_tree_tag_clear(&mapping->page_tree,
  1080. page_index(page),
  1081. PAGECACHE_TAG_DIRTY);
  1082. write_unlock_irqrestore(&mapping->tree_lock, flags);
  1083. } else {
  1084. ret = TestSetPageWriteback(page);
  1085. }
  1086. if (!ret)
  1087. inc_zone_page_state(page, NR_WRITEBACK);
  1088. return ret;
  1089. }
  1090. EXPORT_SYMBOL(test_set_page_writeback);
  1091. /*
  1092. * Return true if any of the pages in the mapping are marked with the
  1093. * passed tag.
  1094. */
  1095. int mapping_tagged(struct address_space *mapping, int tag)
  1096. {
  1097. int ret;
  1098. rcu_read_lock();
  1099. ret = radix_tree_tagged(&mapping->page_tree, tag);
  1100. rcu_read_unlock();
  1101. return ret;
  1102. }
  1103. EXPORT_SYMBOL(mapping_tagged);