fs-writeback.c 75 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611
  1. // SPDX-License-Identifier: GPL-2.0-only
  2. /*
  3. * fs/fs-writeback.c
  4. *
  5. * Copyright (C) 2002, Linus Torvalds.
  6. *
  7. * Contains all the functions related to writing back and waiting
  8. * upon dirty inodes against superblocks, and writing back dirty
  9. * pages against inodes. ie: data writeback. Writeout of the
  10. * inode itself is not handled here.
  11. *
  12. * 10Apr2002 Andrew Morton
  13. * Split out of fs/inode.c
  14. * Additions for address_space-based writeback
  15. */
  16. #include <linux/kernel.h>
  17. #include <linux/export.h>
  18. #include <linux/spinlock.h>
  19. #include <linux/slab.h>
  20. #include <linux/sched.h>
  21. #include <linux/fs.h>
  22. #include <linux/mm.h>
  23. #include <linux/pagemap.h>
  24. #include <linux/kthread.h>
  25. #include <linux/writeback.h>
  26. #include <linux/blkdev.h>
  27. #include <linux/backing-dev.h>
  28. #include <linux/tracepoint.h>
  29. #include <linux/device.h>
  30. #include <linux/memcontrol.h>
  31. #include "internal.h"
  32. /*
  33. * 4MB minimal write chunk size
  34. */
  35. #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10))
  36. /*
  37. * Passed into wb_writeback(), essentially a subset of writeback_control
  38. */
  39. struct wb_writeback_work {
  40. long nr_pages;
  41. struct super_block *sb;
  42. enum writeback_sync_modes sync_mode;
  43. unsigned int tagged_writepages:1;
  44. unsigned int for_kupdate:1;
  45. unsigned int range_cyclic:1;
  46. unsigned int for_background:1;
  47. unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */
  48. unsigned int auto_free:1; /* free on completion */
  49. enum wb_reason reason; /* why was writeback initiated? */
  50. struct list_head list; /* pending work list */
  51. struct wb_completion *done; /* set if the caller waits */
  52. };
  53. /*
  54. * If an inode is constantly having its pages dirtied, but then the
  55. * updates stop dirtytime_expire_interval seconds in the past, it's
  56. * possible for the worst case time between when an inode has its
  57. * timestamps updated and when they finally get written out to be two
  58. * dirtytime_expire_intervals. We set the default to 12 hours (in
  59. * seconds), which means most of the time inodes will have their
  60. * timestamps written to disk after 12 hours, but in the worst case a
  61. * few inodes might not their timestamps updated for 24 hours.
  62. */
  63. unsigned int dirtytime_expire_interval = 12 * 60 * 60;
  64. static inline struct inode *wb_inode(struct list_head *head)
  65. {
  66. return list_entry(head, struct inode, i_io_list);
  67. }
  68. /*
  69. * Include the creation of the trace points after defining the
  70. * wb_writeback_work structure and inline functions so that the definition
  71. * remains local to this file.
  72. */
  73. #define CREATE_TRACE_POINTS
  74. #include <trace/events/writeback.h>
  75. EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage);
  76. static bool wb_io_lists_populated(struct bdi_writeback *wb)
  77. {
  78. if (wb_has_dirty_io(wb)) {
  79. return false;
  80. } else {
  81. set_bit(WB_has_dirty_io, &wb->state);
  82. WARN_ON_ONCE(!wb->avg_write_bandwidth);
  83. atomic_long_add(wb->avg_write_bandwidth,
  84. &wb->bdi->tot_write_bandwidth);
  85. return true;
  86. }
  87. }
  88. static void wb_io_lists_depopulated(struct bdi_writeback *wb)
  89. {
  90. if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
  91. list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
  92. clear_bit(WB_has_dirty_io, &wb->state);
  93. WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
  94. &wb->bdi->tot_write_bandwidth) < 0);
  95. }
  96. }
  97. /**
  98. * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list
  99. * @inode: inode to be moved
  100. * @wb: target bdi_writeback
  101. * @head: one of @wb->b_{dirty|io|more_io|dirty_time}
  102. *
  103. * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io.
  104. * Returns %true if @inode is the first occupant of the !dirty_time IO
  105. * lists; otherwise, %false.
  106. */
  107. static bool inode_io_list_move_locked(struct inode *inode,
  108. struct bdi_writeback *wb,
  109. struct list_head *head)
  110. {
  111. assert_spin_locked(&wb->list_lock);
  112. list_move(&inode->i_io_list, head);
  113. /* dirty_time doesn't count as dirty_io until expiration */
  114. if (head != &wb->b_dirty_time)
  115. return wb_io_lists_populated(wb);
  116. wb_io_lists_depopulated(wb);
  117. return false;
  118. }
  119. /**
  120. * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list
  121. * @inode: inode to be removed
  122. * @wb: bdi_writeback @inode is being removed from
  123. *
  124. * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and
  125. * clear %WB_has_dirty_io if all are empty afterwards.
  126. */
  127. static void inode_io_list_del_locked(struct inode *inode,
  128. struct bdi_writeback *wb)
  129. {
  130. assert_spin_locked(&wb->list_lock);
  131. assert_spin_locked(&inode->i_lock);
  132. inode->i_state &= ~I_SYNC_QUEUED;
  133. list_del_init(&inode->i_io_list);
  134. wb_io_lists_depopulated(wb);
  135. }
  136. static void wb_wakeup(struct bdi_writeback *wb)
  137. {
  138. spin_lock_bh(&wb->work_lock);
  139. if (test_bit(WB_registered, &wb->state))
  140. mod_delayed_work(bdi_wq, &wb->dwork, 0);
  141. spin_unlock_bh(&wb->work_lock);
  142. }
  143. static void finish_writeback_work(struct bdi_writeback *wb,
  144. struct wb_writeback_work *work)
  145. {
  146. struct wb_completion *done = work->done;
  147. if (work->auto_free)
  148. kfree(work);
  149. if (done) {
  150. wait_queue_head_t *waitq = done->waitq;
  151. /* @done can't be accessed after the following dec */
  152. if (atomic_dec_and_test(&done->cnt))
  153. wake_up_all(waitq);
  154. }
  155. }
  156. static void wb_queue_work(struct bdi_writeback *wb,
  157. struct wb_writeback_work *work)
  158. {
  159. trace_writeback_queue(wb, work);
  160. if (work->done)
  161. atomic_inc(&work->done->cnt);
  162. spin_lock_bh(&wb->work_lock);
  163. if (test_bit(WB_registered, &wb->state)) {
  164. list_add_tail(&work->list, &wb->work_list);
  165. mod_delayed_work(bdi_wq, &wb->dwork, 0);
  166. } else
  167. finish_writeback_work(wb, work);
  168. spin_unlock_bh(&wb->work_lock);
  169. }
  170. /**
  171. * wb_wait_for_completion - wait for completion of bdi_writeback_works
  172. * @done: target wb_completion
  173. *
  174. * Wait for one or more work items issued to @bdi with their ->done field
  175. * set to @done, which should have been initialized with
  176. * DEFINE_WB_COMPLETION(). This function returns after all such work items
  177. * are completed. Work items which are waited upon aren't freed
  178. * automatically on completion.
  179. */
  180. void wb_wait_for_completion(struct wb_completion *done)
  181. {
  182. atomic_dec(&done->cnt); /* put down the initial count */
  183. wait_event(*done->waitq, !atomic_read(&done->cnt));
  184. }
  185. #ifdef CONFIG_CGROUP_WRITEBACK
  186. /*
  187. * Parameters for foreign inode detection, see wbc_detach_inode() to see
  188. * how they're used.
  189. *
  190. * These paramters are inherently heuristical as the detection target
  191. * itself is fuzzy. All we want to do is detaching an inode from the
  192. * current owner if it's being written to by some other cgroups too much.
  193. *
  194. * The current cgroup writeback is built on the assumption that multiple
  195. * cgroups writing to the same inode concurrently is very rare and a mode
  196. * of operation which isn't well supported. As such, the goal is not
  197. * taking too long when a different cgroup takes over an inode while
  198. * avoiding too aggressive flip-flops from occasional foreign writes.
  199. *
  200. * We record, very roughly, 2s worth of IO time history and if more than
  201. * half of that is foreign, trigger the switch. The recording is quantized
  202. * to 16 slots. To avoid tiny writes from swinging the decision too much,
  203. * writes smaller than 1/8 of avg size are ignored.
  204. */
  205. #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */
  206. #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */
  207. #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */
  208. #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */
  209. #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */
  210. #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
  211. /* each slot's duration is 2s / 16 */
  212. #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2)
  213. /* if foreign slots >= 8, switch */
  214. #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1)
  215. /* one round can affect upto 5 slots */
  216. #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */
  217. static atomic_t isw_nr_in_flight = ATOMIC_INIT(0);
  218. static struct workqueue_struct *isw_wq;
  219. void __inode_attach_wb(struct inode *inode, struct page *page)
  220. {
  221. struct backing_dev_info *bdi = inode_to_bdi(inode);
  222. struct bdi_writeback *wb = NULL;
  223. if (inode_cgwb_enabled(inode)) {
  224. struct cgroup_subsys_state *memcg_css;
  225. if (page) {
  226. memcg_css = mem_cgroup_css_from_page(page);
  227. wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
  228. } else {
  229. /* must pin memcg_css, see wb_get_create() */
  230. memcg_css = task_get_css(current, memory_cgrp_id);
  231. wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
  232. css_put(memcg_css);
  233. }
  234. }
  235. if (!wb)
  236. wb = &bdi->wb;
  237. /*
  238. * There may be multiple instances of this function racing to
  239. * update the same inode. Use cmpxchg() to tell the winner.
  240. */
  241. if (unlikely(cmpxchg(&inode->i_wb, NULL, wb)))
  242. wb_put(wb);
  243. }
  244. EXPORT_SYMBOL_GPL(__inode_attach_wb);
  245. /**
  246. * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it
  247. * @inode: inode of interest with i_lock held
  248. *
  249. * Returns @inode's wb with its list_lock held. @inode->i_lock must be
  250. * held on entry and is released on return. The returned wb is guaranteed
  251. * to stay @inode's associated wb until its list_lock is released.
  252. */
  253. static struct bdi_writeback *
  254. locked_inode_to_wb_and_lock_list(struct inode *inode)
  255. __releases(&inode->i_lock)
  256. __acquires(&wb->list_lock)
  257. {
  258. while (true) {
  259. struct bdi_writeback *wb = inode_to_wb(inode);
  260. /*
  261. * inode_to_wb() association is protected by both
  262. * @inode->i_lock and @wb->list_lock but list_lock nests
  263. * outside i_lock. Drop i_lock and verify that the
  264. * association hasn't changed after acquiring list_lock.
  265. */
  266. wb_get(wb);
  267. spin_unlock(&inode->i_lock);
  268. spin_lock(&wb->list_lock);
  269. /* i_wb may have changed inbetween, can't use inode_to_wb() */
  270. if (likely(wb == inode->i_wb)) {
  271. wb_put(wb); /* @inode already has ref */
  272. return wb;
  273. }
  274. spin_unlock(&wb->list_lock);
  275. wb_put(wb);
  276. cpu_relax();
  277. spin_lock(&inode->i_lock);
  278. }
  279. }
  280. /**
  281. * inode_to_wb_and_lock_list - determine an inode's wb and lock it
  282. * @inode: inode of interest
  283. *
  284. * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held
  285. * on entry.
  286. */
  287. static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
  288. __acquires(&wb->list_lock)
  289. {
  290. spin_lock(&inode->i_lock);
  291. return locked_inode_to_wb_and_lock_list(inode);
  292. }
  293. struct inode_switch_wbs_context {
  294. struct inode *inode;
  295. struct bdi_writeback *new_wb;
  296. struct rcu_head rcu_head;
  297. struct work_struct work;
  298. };
  299. static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi)
  300. {
  301. down_write(&bdi->wb_switch_rwsem);
  302. }
  303. static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi)
  304. {
  305. up_write(&bdi->wb_switch_rwsem);
  306. }
  307. static void inode_switch_wbs_work_fn(struct work_struct *work)
  308. {
  309. struct inode_switch_wbs_context *isw =
  310. container_of(work, struct inode_switch_wbs_context, work);
  311. struct inode *inode = isw->inode;
  312. struct backing_dev_info *bdi = inode_to_bdi(inode);
  313. struct address_space *mapping = inode->i_mapping;
  314. struct bdi_writeback *old_wb = inode->i_wb;
  315. struct bdi_writeback *new_wb = isw->new_wb;
  316. XA_STATE(xas, &mapping->i_pages, 0);
  317. struct page *page;
  318. bool switched = false;
  319. /*
  320. * If @inode switches cgwb membership while sync_inodes_sb() is
  321. * being issued, sync_inodes_sb() might miss it. Synchronize.
  322. */
  323. down_read(&bdi->wb_switch_rwsem);
  324. /*
  325. * By the time control reaches here, RCU grace period has passed
  326. * since I_WB_SWITCH assertion and all wb stat update transactions
  327. * between unlocked_inode_to_wb_begin/end() are guaranteed to be
  328. * synchronizing against the i_pages lock.
  329. *
  330. * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock
  331. * gives us exclusion against all wb related operations on @inode
  332. * including IO list manipulations and stat updates.
  333. */
  334. if (old_wb < new_wb) {
  335. spin_lock(&old_wb->list_lock);
  336. spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING);
  337. } else {
  338. spin_lock(&new_wb->list_lock);
  339. spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING);
  340. }
  341. spin_lock(&inode->i_lock);
  342. xa_lock_irq(&mapping->i_pages);
  343. /*
  344. * Once I_FREEING is visible under i_lock, the eviction path owns
  345. * the inode and we shouldn't modify ->i_io_list.
  346. */
  347. if (unlikely(inode->i_state & I_FREEING))
  348. goto skip_switch;
  349. trace_inode_switch_wbs(inode, old_wb, new_wb);
  350. /*
  351. * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points
  352. * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to
  353. * pages actually under writeback.
  354. */
  355. xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) {
  356. if (PageDirty(page)) {
  357. dec_wb_stat(old_wb, WB_RECLAIMABLE);
  358. inc_wb_stat(new_wb, WB_RECLAIMABLE);
  359. }
  360. }
  361. xas_set(&xas, 0);
  362. xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) {
  363. WARN_ON_ONCE(!PageWriteback(page));
  364. dec_wb_stat(old_wb, WB_WRITEBACK);
  365. inc_wb_stat(new_wb, WB_WRITEBACK);
  366. }
  367. wb_get(new_wb);
  368. /*
  369. * Transfer to @new_wb's IO list if necessary. The specific list
  370. * @inode was on is ignored and the inode is put on ->b_dirty which
  371. * is always correct including from ->b_dirty_time. The transfer
  372. * preserves @inode->dirtied_when ordering.
  373. */
  374. if (!list_empty(&inode->i_io_list)) {
  375. struct inode *pos;
  376. inode_io_list_del_locked(inode, old_wb);
  377. inode->i_wb = new_wb;
  378. list_for_each_entry(pos, &new_wb->b_dirty, i_io_list)
  379. if (time_after_eq(inode->dirtied_when,
  380. pos->dirtied_when))
  381. break;
  382. inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev);
  383. } else {
  384. inode->i_wb = new_wb;
  385. }
  386. /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */
  387. inode->i_wb_frn_winner = 0;
  388. inode->i_wb_frn_avg_time = 0;
  389. inode->i_wb_frn_history = 0;
  390. switched = true;
  391. skip_switch:
  392. /*
  393. * Paired with load_acquire in unlocked_inode_to_wb_begin() and
  394. * ensures that the new wb is visible if they see !I_WB_SWITCH.
  395. */
  396. smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH);
  397. xa_unlock_irq(&mapping->i_pages);
  398. spin_unlock(&inode->i_lock);
  399. spin_unlock(&new_wb->list_lock);
  400. spin_unlock(&old_wb->list_lock);
  401. up_read(&bdi->wb_switch_rwsem);
  402. if (switched) {
  403. wb_wakeup(new_wb);
  404. wb_put(old_wb);
  405. }
  406. wb_put(new_wb);
  407. iput(inode);
  408. kfree(isw);
  409. atomic_dec(&isw_nr_in_flight);
  410. }
  411. static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head)
  412. {
  413. struct inode_switch_wbs_context *isw = container_of(rcu_head,
  414. struct inode_switch_wbs_context, rcu_head);
  415. /* needs to grab bh-unsafe locks, bounce to work item */
  416. INIT_WORK(&isw->work, inode_switch_wbs_work_fn);
  417. queue_work(isw_wq, &isw->work);
  418. }
  419. /**
  420. * inode_switch_wbs - change the wb association of an inode
  421. * @inode: target inode
  422. * @new_wb_id: ID of the new wb
  423. *
  424. * Switch @inode's wb association to the wb identified by @new_wb_id. The
  425. * switching is performed asynchronously and may fail silently.
  426. */
  427. static void inode_switch_wbs(struct inode *inode, int new_wb_id)
  428. {
  429. struct backing_dev_info *bdi = inode_to_bdi(inode);
  430. struct cgroup_subsys_state *memcg_css;
  431. struct inode_switch_wbs_context *isw;
  432. /* noop if seems to be already in progress */
  433. if (inode->i_state & I_WB_SWITCH)
  434. return;
  435. /* avoid queueing a new switch if too many are already in flight */
  436. if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT)
  437. return;
  438. isw = kzalloc(sizeof(*isw), GFP_ATOMIC);
  439. if (!isw)
  440. return;
  441. atomic_inc(&isw_nr_in_flight);
  442. /* find and pin the new wb */
  443. rcu_read_lock();
  444. memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
  445. if (memcg_css && !css_tryget(memcg_css))
  446. memcg_css = NULL;
  447. rcu_read_unlock();
  448. if (!memcg_css)
  449. goto out_free;
  450. isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC);
  451. css_put(memcg_css);
  452. if (!isw->new_wb)
  453. goto out_free;
  454. /* while holding I_WB_SWITCH, no one else can update the association */
  455. spin_lock(&inode->i_lock);
  456. if (!(inode->i_sb->s_flags & SB_ACTIVE) ||
  457. inode->i_state & (I_WB_SWITCH | I_FREEING) ||
  458. inode_to_wb(inode) == isw->new_wb) {
  459. spin_unlock(&inode->i_lock);
  460. goto out_free;
  461. }
  462. inode->i_state |= I_WB_SWITCH;
  463. __iget(inode);
  464. spin_unlock(&inode->i_lock);
  465. isw->inode = inode;
  466. /*
  467. * In addition to synchronizing among switchers, I_WB_SWITCH tells
  468. * the RCU protected stat update paths to grab the i_page
  469. * lock so that stat transfer can synchronize against them.
  470. * Let's continue after I_WB_SWITCH is guaranteed to be visible.
  471. */
  472. call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn);
  473. return;
  474. out_free:
  475. atomic_dec(&isw_nr_in_flight);
  476. if (isw->new_wb)
  477. wb_put(isw->new_wb);
  478. kfree(isw);
  479. }
  480. /**
  481. * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it
  482. * @wbc: writeback_control of interest
  483. * @inode: target inode
  484. *
  485. * @inode is locked and about to be written back under the control of @wbc.
  486. * Record @inode's writeback context into @wbc and unlock the i_lock. On
  487. * writeback completion, wbc_detach_inode() should be called. This is used
  488. * to track the cgroup writeback context.
  489. */
  490. void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
  491. struct inode *inode)
  492. {
  493. if (!inode_cgwb_enabled(inode)) {
  494. spin_unlock(&inode->i_lock);
  495. return;
  496. }
  497. wbc->wb = inode_to_wb(inode);
  498. wbc->inode = inode;
  499. wbc->wb_id = wbc->wb->memcg_css->id;
  500. wbc->wb_lcand_id = inode->i_wb_frn_winner;
  501. wbc->wb_tcand_id = 0;
  502. wbc->wb_bytes = 0;
  503. wbc->wb_lcand_bytes = 0;
  504. wbc->wb_tcand_bytes = 0;
  505. wb_get(wbc->wb);
  506. spin_unlock(&inode->i_lock);
  507. /*
  508. * A dying wb indicates that either the blkcg associated with the
  509. * memcg changed or the associated memcg is dying. In the first
  510. * case, a replacement wb should already be available and we should
  511. * refresh the wb immediately. In the second case, trying to
  512. * refresh will keep failing.
  513. */
  514. if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css)))
  515. inode_switch_wbs(inode, wbc->wb_id);
  516. }
  517. EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode);
  518. /**
  519. * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
  520. * @wbc: writeback_control of the just finished writeback
  521. *
  522. * To be called after a writeback attempt of an inode finishes and undoes
  523. * wbc_attach_and_unlock_inode(). Can be called under any context.
  524. *
  525. * As concurrent write sharing of an inode is expected to be very rare and
  526. * memcg only tracks page ownership on first-use basis severely confining
  527. * the usefulness of such sharing, cgroup writeback tracks ownership
  528. * per-inode. While the support for concurrent write sharing of an inode
  529. * is deemed unnecessary, an inode being written to by different cgroups at
  530. * different points in time is a lot more common, and, more importantly,
  531. * charging only by first-use can too readily lead to grossly incorrect
  532. * behaviors (single foreign page can lead to gigabytes of writeback to be
  533. * incorrectly attributed).
  534. *
  535. * To resolve this issue, cgroup writeback detects the majority dirtier of
  536. * an inode and transfers the ownership to it. To avoid unnnecessary
  537. * oscillation, the detection mechanism keeps track of history and gives
  538. * out the switch verdict only if the foreign usage pattern is stable over
  539. * a certain amount of time and/or writeback attempts.
  540. *
  541. * On each writeback attempt, @wbc tries to detect the majority writer
  542. * using Boyer-Moore majority vote algorithm. In addition to the byte
  543. * count from the majority voting, it also counts the bytes written for the
  544. * current wb and the last round's winner wb (max of last round's current
  545. * wb, the winner from two rounds ago, and the last round's majority
  546. * candidate). Keeping track of the historical winner helps the algorithm
  547. * to semi-reliably detect the most active writer even when it's not the
  548. * absolute majority.
  549. *
  550. * Once the winner of the round is determined, whether the winner is
  551. * foreign or not and how much IO time the round consumed is recorded in
  552. * inode->i_wb_frn_history. If the amount of recorded foreign IO time is
  553. * over a certain threshold, the switch verdict is given.
  554. */
  555. void wbc_detach_inode(struct writeback_control *wbc)
  556. {
  557. struct bdi_writeback *wb = wbc->wb;
  558. struct inode *inode = wbc->inode;
  559. unsigned long avg_time, max_bytes, max_time;
  560. u16 history;
  561. int max_id;
  562. if (!wb)
  563. return;
  564. history = inode->i_wb_frn_history;
  565. avg_time = inode->i_wb_frn_avg_time;
  566. /* pick the winner of this round */
  567. if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
  568. wbc->wb_bytes >= wbc->wb_tcand_bytes) {
  569. max_id = wbc->wb_id;
  570. max_bytes = wbc->wb_bytes;
  571. } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
  572. max_id = wbc->wb_lcand_id;
  573. max_bytes = wbc->wb_lcand_bytes;
  574. } else {
  575. max_id = wbc->wb_tcand_id;
  576. max_bytes = wbc->wb_tcand_bytes;
  577. }
  578. /*
  579. * Calculate the amount of IO time the winner consumed and fold it
  580. * into the running average kept per inode. If the consumed IO
  581. * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
  582. * deciding whether to switch or not. This is to prevent one-off
  583. * small dirtiers from skewing the verdict.
  584. */
  585. max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
  586. wb->avg_write_bandwidth);
  587. if (avg_time)
  588. avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
  589. (avg_time >> WB_FRN_TIME_AVG_SHIFT);
  590. else
  591. avg_time = max_time; /* immediate catch up on first run */
  592. if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
  593. int slots;
  594. /*
  595. * The switch verdict is reached if foreign wb's consume
  596. * more than a certain proportion of IO time in a
  597. * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot
  598. * history mask where each bit represents one sixteenth of
  599. * the period. Determine the number of slots to shift into
  600. * history from @max_time.
  601. */
  602. slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
  603. (unsigned long)WB_FRN_HIST_MAX_SLOTS);
  604. history <<= slots;
  605. if (wbc->wb_id != max_id)
  606. history |= (1U << slots) - 1;
  607. if (history)
  608. trace_inode_foreign_history(inode, wbc, history);
  609. /*
  610. * Switch if the current wb isn't the consistent winner.
  611. * If there are multiple closely competing dirtiers, the
  612. * inode may switch across them repeatedly over time, which
  613. * is okay. The main goal is avoiding keeping an inode on
  614. * the wrong wb for an extended period of time.
  615. */
  616. if (hweight32(history) > WB_FRN_HIST_THR_SLOTS)
  617. inode_switch_wbs(inode, max_id);
  618. }
  619. /*
  620. * Multiple instances of this function may race to update the
  621. * following fields but we don't mind occassional inaccuracies.
  622. */
  623. inode->i_wb_frn_winner = max_id;
  624. inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
  625. inode->i_wb_frn_history = history;
  626. wb_put(wbc->wb);
  627. wbc->wb = NULL;
  628. }
  629. EXPORT_SYMBOL_GPL(wbc_detach_inode);
  630. /**
  631. * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership
  632. * @wbc: writeback_control of the writeback in progress
  633. * @page: page being written out
  634. * @bytes: number of bytes being written out
  635. *
  636. * @bytes from @page are about to written out during the writeback
  637. * controlled by @wbc. Keep the book for foreign inode detection. See
  638. * wbc_detach_inode().
  639. */
  640. void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page,
  641. size_t bytes)
  642. {
  643. struct cgroup_subsys_state *css;
  644. int id;
  645. /*
  646. * pageout() path doesn't attach @wbc to the inode being written
  647. * out. This is intentional as we don't want the function to block
  648. * behind a slow cgroup. Ultimately, we want pageout() to kick off
  649. * regular writeback instead of writing things out itself.
  650. */
  651. if (!wbc->wb || wbc->no_cgroup_owner)
  652. return;
  653. css = mem_cgroup_css_from_page(page);
  654. /* dead cgroups shouldn't contribute to inode ownership arbitration */
  655. if (!(css->flags & CSS_ONLINE))
  656. return;
  657. id = css->id;
  658. if (id == wbc->wb_id) {
  659. wbc->wb_bytes += bytes;
  660. return;
  661. }
  662. if (id == wbc->wb_lcand_id)
  663. wbc->wb_lcand_bytes += bytes;
  664. /* Boyer-Moore majority vote algorithm */
  665. if (!wbc->wb_tcand_bytes)
  666. wbc->wb_tcand_id = id;
  667. if (id == wbc->wb_tcand_id)
  668. wbc->wb_tcand_bytes += bytes;
  669. else
  670. wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
  671. }
  672. EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner);
  673. /**
  674. * inode_congested - test whether an inode is congested
  675. * @inode: inode to test for congestion (may be NULL)
  676. * @cong_bits: mask of WB_[a]sync_congested bits to test
  677. *
  678. * Tests whether @inode is congested. @cong_bits is the mask of congestion
  679. * bits to test and the return value is the mask of set bits.
  680. *
  681. * If cgroup writeback is enabled for @inode, the congestion state is
  682. * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg
  683. * associated with @inode is congested; otherwise, the root wb's congestion
  684. * state is used.
  685. *
  686. * @inode is allowed to be NULL as this function is often called on
  687. * mapping->host which is NULL for the swapper space.
  688. */
  689. int inode_congested(struct inode *inode, int cong_bits)
  690. {
  691. /*
  692. * Once set, ->i_wb never becomes NULL while the inode is alive.
  693. * Start transaction iff ->i_wb is visible.
  694. */
  695. if (inode && inode_to_wb_is_valid(inode)) {
  696. struct bdi_writeback *wb;
  697. struct wb_lock_cookie lock_cookie = {};
  698. bool congested;
  699. wb = unlocked_inode_to_wb_begin(inode, &lock_cookie);
  700. congested = wb_congested(wb, cong_bits);
  701. unlocked_inode_to_wb_end(inode, &lock_cookie);
  702. return congested;
  703. }
  704. return wb_congested(&inode_to_bdi(inode)->wb, cong_bits);
  705. }
  706. EXPORT_SYMBOL_GPL(inode_congested);
  707. /**
  708. * wb_split_bdi_pages - split nr_pages to write according to bandwidth
  709. * @wb: target bdi_writeback to split @nr_pages to
  710. * @nr_pages: number of pages to write for the whole bdi
  711. *
  712. * Split @wb's portion of @nr_pages according to @wb's write bandwidth in
  713. * relation to the total write bandwidth of all wb's w/ dirty inodes on
  714. * @wb->bdi.
  715. */
  716. static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
  717. {
  718. unsigned long this_bw = wb->avg_write_bandwidth;
  719. unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
  720. if (nr_pages == LONG_MAX)
  721. return LONG_MAX;
  722. /*
  723. * This may be called on clean wb's and proportional distribution
  724. * may not make sense, just use the original @nr_pages in those
  725. * cases. In general, we wanna err on the side of writing more.
  726. */
  727. if (!tot_bw || this_bw >= tot_bw)
  728. return nr_pages;
  729. else
  730. return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw);
  731. }
  732. /**
  733. * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi
  734. * @bdi: target backing_dev_info
  735. * @base_work: wb_writeback_work to issue
  736. * @skip_if_busy: skip wb's which already have writeback in progress
  737. *
  738. * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which
  739. * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's
  740. * distributed to the busy wbs according to each wb's proportion in the
  741. * total active write bandwidth of @bdi.
  742. */
  743. static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
  744. struct wb_writeback_work *base_work,
  745. bool skip_if_busy)
  746. {
  747. struct bdi_writeback *last_wb = NULL;
  748. struct bdi_writeback *wb = list_entry(&bdi->wb_list,
  749. struct bdi_writeback, bdi_node);
  750. might_sleep();
  751. restart:
  752. rcu_read_lock();
  753. list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) {
  754. DEFINE_WB_COMPLETION(fallback_work_done, bdi);
  755. struct wb_writeback_work fallback_work;
  756. struct wb_writeback_work *work;
  757. long nr_pages;
  758. if (last_wb) {
  759. wb_put(last_wb);
  760. last_wb = NULL;
  761. }
  762. /* SYNC_ALL writes out I_DIRTY_TIME too */
  763. if (!wb_has_dirty_io(wb) &&
  764. (base_work->sync_mode == WB_SYNC_NONE ||
  765. list_empty(&wb->b_dirty_time)))
  766. continue;
  767. if (skip_if_busy && writeback_in_progress(wb))
  768. continue;
  769. nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages);
  770. work = kmalloc(sizeof(*work), GFP_ATOMIC);
  771. if (work) {
  772. *work = *base_work;
  773. work->nr_pages = nr_pages;
  774. work->auto_free = 1;
  775. wb_queue_work(wb, work);
  776. continue;
  777. }
  778. /* alloc failed, execute synchronously using on-stack fallback */
  779. work = &fallback_work;
  780. *work = *base_work;
  781. work->nr_pages = nr_pages;
  782. work->auto_free = 0;
  783. work->done = &fallback_work_done;
  784. wb_queue_work(wb, work);
  785. /*
  786. * Pin @wb so that it stays on @bdi->wb_list. This allows
  787. * continuing iteration from @wb after dropping and
  788. * regrabbing rcu read lock.
  789. */
  790. wb_get(wb);
  791. last_wb = wb;
  792. rcu_read_unlock();
  793. wb_wait_for_completion(&fallback_work_done);
  794. goto restart;
  795. }
  796. rcu_read_unlock();
  797. if (last_wb)
  798. wb_put(last_wb);
  799. }
  800. /**
  801. * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs
  802. * @bdi_id: target bdi id
  803. * @memcg_id: target memcg css id
  804. * @nr: number of pages to write, 0 for best-effort dirty flushing
  805. * @reason: reason why some writeback work initiated
  806. * @done: target wb_completion
  807. *
  808. * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id
  809. * with the specified parameters.
  810. */
  811. int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, unsigned long nr,
  812. enum wb_reason reason, struct wb_completion *done)
  813. {
  814. struct backing_dev_info *bdi;
  815. struct cgroup_subsys_state *memcg_css;
  816. struct bdi_writeback *wb;
  817. struct wb_writeback_work *work;
  818. int ret;
  819. /* lookup bdi and memcg */
  820. bdi = bdi_get_by_id(bdi_id);
  821. if (!bdi)
  822. return -ENOENT;
  823. rcu_read_lock();
  824. memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys);
  825. if (memcg_css && !css_tryget(memcg_css))
  826. memcg_css = NULL;
  827. rcu_read_unlock();
  828. if (!memcg_css) {
  829. ret = -ENOENT;
  830. goto out_bdi_put;
  831. }
  832. /*
  833. * And find the associated wb. If the wb isn't there already
  834. * there's nothing to flush, don't create one.
  835. */
  836. wb = wb_get_lookup(bdi, memcg_css);
  837. if (!wb) {
  838. ret = -ENOENT;
  839. goto out_css_put;
  840. }
  841. /*
  842. * If @nr is zero, the caller is attempting to write out most of
  843. * the currently dirty pages. Let's take the current dirty page
  844. * count and inflate it by 25% which should be large enough to
  845. * flush out most dirty pages while avoiding getting livelocked by
  846. * concurrent dirtiers.
  847. */
  848. if (!nr) {
  849. unsigned long filepages, headroom, dirty, writeback;
  850. mem_cgroup_wb_stats(wb, &filepages, &headroom, &dirty,
  851. &writeback);
  852. nr = dirty * 10 / 8;
  853. }
  854. /* issue the writeback work */
  855. work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN);
  856. if (work) {
  857. work->nr_pages = nr;
  858. work->sync_mode = WB_SYNC_NONE;
  859. work->range_cyclic = 1;
  860. work->reason = reason;
  861. work->done = done;
  862. work->auto_free = 1;
  863. wb_queue_work(wb, work);
  864. ret = 0;
  865. } else {
  866. ret = -ENOMEM;
  867. }
  868. wb_put(wb);
  869. out_css_put:
  870. css_put(memcg_css);
  871. out_bdi_put:
  872. bdi_put(bdi);
  873. return ret;
  874. }
  875. /**
  876. * cgroup_writeback_umount - flush inode wb switches for umount
  877. *
  878. * This function is called when a super_block is about to be destroyed and
  879. * flushes in-flight inode wb switches. An inode wb switch goes through
  880. * RCU and then workqueue, so the two need to be flushed in order to ensure
  881. * that all previously scheduled switches are finished. As wb switches are
  882. * rare occurrences and synchronize_rcu() can take a while, perform
  883. * flushing iff wb switches are in flight.
  884. */
  885. void cgroup_writeback_umount(void)
  886. {
  887. if (atomic_read(&isw_nr_in_flight)) {
  888. /*
  889. * Use rcu_barrier() to wait for all pending callbacks to
  890. * ensure that all in-flight wb switches are in the workqueue.
  891. */
  892. rcu_barrier();
  893. flush_workqueue(isw_wq);
  894. }
  895. }
  896. static int __init cgroup_writeback_init(void)
  897. {
  898. isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0);
  899. if (!isw_wq)
  900. return -ENOMEM;
  901. return 0;
  902. }
  903. fs_initcall(cgroup_writeback_init);
  904. #else /* CONFIG_CGROUP_WRITEBACK */
  905. static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
  906. static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { }
  907. static struct bdi_writeback *
  908. locked_inode_to_wb_and_lock_list(struct inode *inode)
  909. __releases(&inode->i_lock)
  910. __acquires(&wb->list_lock)
  911. {
  912. struct bdi_writeback *wb = inode_to_wb(inode);
  913. spin_unlock(&inode->i_lock);
  914. spin_lock(&wb->list_lock);
  915. return wb;
  916. }
  917. static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode)
  918. __acquires(&wb->list_lock)
  919. {
  920. struct bdi_writeback *wb = inode_to_wb(inode);
  921. spin_lock(&wb->list_lock);
  922. return wb;
  923. }
  924. static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages)
  925. {
  926. return nr_pages;
  927. }
  928. static void bdi_split_work_to_wbs(struct backing_dev_info *bdi,
  929. struct wb_writeback_work *base_work,
  930. bool skip_if_busy)
  931. {
  932. might_sleep();
  933. if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) {
  934. base_work->auto_free = 0;
  935. wb_queue_work(&bdi->wb, base_work);
  936. }
  937. }
  938. #endif /* CONFIG_CGROUP_WRITEBACK */
  939. /*
  940. * Add in the number of potentially dirty inodes, because each inode
  941. * write can dirty pagecache in the underlying blockdev.
  942. */
  943. static unsigned long get_nr_dirty_pages(void)
  944. {
  945. return global_node_page_state(NR_FILE_DIRTY) +
  946. get_nr_dirty_inodes();
  947. }
  948. static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason)
  949. {
  950. if (!wb_has_dirty_io(wb))
  951. return;
  952. /*
  953. * All callers of this function want to start writeback of all
  954. * dirty pages. Places like vmscan can call this at a very
  955. * high frequency, causing pointless allocations of tons of
  956. * work items and keeping the flusher threads busy retrieving
  957. * that work. Ensure that we only allow one of them pending and
  958. * inflight at the time.
  959. */
  960. if (test_bit(WB_start_all, &wb->state) ||
  961. test_and_set_bit(WB_start_all, &wb->state))
  962. return;
  963. wb->start_all_reason = reason;
  964. wb_wakeup(wb);
  965. }
  966. /**
  967. * wb_start_background_writeback - start background writeback
  968. * @wb: bdi_writback to write from
  969. *
  970. * Description:
  971. * This makes sure WB_SYNC_NONE background writeback happens. When
  972. * this function returns, it is only guaranteed that for given wb
  973. * some IO is happening if we are over background dirty threshold.
  974. * Caller need not hold sb s_umount semaphore.
  975. */
  976. void wb_start_background_writeback(struct bdi_writeback *wb)
  977. {
  978. /*
  979. * We just wake up the flusher thread. It will perform background
  980. * writeback as soon as there is no other work to do.
  981. */
  982. trace_writeback_wake_background(wb);
  983. wb_wakeup(wb);
  984. }
  985. /*
  986. * Remove the inode from the writeback list it is on.
  987. */
  988. void inode_io_list_del(struct inode *inode)
  989. {
  990. struct bdi_writeback *wb;
  991. wb = inode_to_wb_and_lock_list(inode);
  992. spin_lock(&inode->i_lock);
  993. inode_io_list_del_locked(inode, wb);
  994. spin_unlock(&inode->i_lock);
  995. spin_unlock(&wb->list_lock);
  996. }
  997. EXPORT_SYMBOL(inode_io_list_del);
  998. /*
  999. * mark an inode as under writeback on the sb
  1000. */
  1001. void sb_mark_inode_writeback(struct inode *inode)
  1002. {
  1003. struct super_block *sb = inode->i_sb;
  1004. unsigned long flags;
  1005. if (list_empty(&inode->i_wb_list)) {
  1006. spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
  1007. if (list_empty(&inode->i_wb_list)) {
  1008. list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb);
  1009. trace_sb_mark_inode_writeback(inode);
  1010. }
  1011. spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
  1012. }
  1013. }
  1014. /*
  1015. * clear an inode as under writeback on the sb
  1016. */
  1017. void sb_clear_inode_writeback(struct inode *inode)
  1018. {
  1019. struct super_block *sb = inode->i_sb;
  1020. unsigned long flags;
  1021. if (!list_empty(&inode->i_wb_list)) {
  1022. spin_lock_irqsave(&sb->s_inode_wblist_lock, flags);
  1023. if (!list_empty(&inode->i_wb_list)) {
  1024. list_del_init(&inode->i_wb_list);
  1025. trace_sb_clear_inode_writeback(inode);
  1026. }
  1027. spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags);
  1028. }
  1029. }
  1030. /*
  1031. * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
  1032. * furthest end of its superblock's dirty-inode list.
  1033. *
  1034. * Before stamping the inode's ->dirtied_when, we check to see whether it is
  1035. * already the most-recently-dirtied inode on the b_dirty list. If that is
  1036. * the case then the inode must have been redirtied while it was being written
  1037. * out and we don't reset its dirtied_when.
  1038. */
  1039. static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb)
  1040. {
  1041. assert_spin_locked(&inode->i_lock);
  1042. if (!list_empty(&wb->b_dirty)) {
  1043. struct inode *tail;
  1044. tail = wb_inode(wb->b_dirty.next);
  1045. if (time_before(inode->dirtied_when, tail->dirtied_when))
  1046. inode->dirtied_when = jiffies;
  1047. }
  1048. inode_io_list_move_locked(inode, wb, &wb->b_dirty);
  1049. inode->i_state &= ~I_SYNC_QUEUED;
  1050. }
  1051. static void redirty_tail(struct inode *inode, struct bdi_writeback *wb)
  1052. {
  1053. spin_lock(&inode->i_lock);
  1054. redirty_tail_locked(inode, wb);
  1055. spin_unlock(&inode->i_lock);
  1056. }
  1057. /*
  1058. * requeue inode for re-scanning after bdi->b_io list is exhausted.
  1059. */
  1060. static void requeue_io(struct inode *inode, struct bdi_writeback *wb)
  1061. {
  1062. inode_io_list_move_locked(inode, wb, &wb->b_more_io);
  1063. }
  1064. static void inode_sync_complete(struct inode *inode)
  1065. {
  1066. inode->i_state &= ~I_SYNC;
  1067. /* If inode is clean an unused, put it into LRU now... */
  1068. inode_add_lru(inode);
  1069. /* Waiters must see I_SYNC cleared before being woken up */
  1070. smp_mb();
  1071. wake_up_bit(&inode->i_state, __I_SYNC);
  1072. }
  1073. static bool inode_dirtied_after(struct inode *inode, unsigned long t)
  1074. {
  1075. bool ret = time_after(inode->dirtied_when, t);
  1076. #ifndef CONFIG_64BIT
  1077. /*
  1078. * For inodes being constantly redirtied, dirtied_when can get stuck.
  1079. * It _appears_ to be in the future, but is actually in distant past.
  1080. * This test is necessary to prevent such wrapped-around relative times
  1081. * from permanently stopping the whole bdi writeback.
  1082. */
  1083. ret = ret && time_before_eq(inode->dirtied_when, jiffies);
  1084. #endif
  1085. return ret;
  1086. }
  1087. #define EXPIRE_DIRTY_ATIME 0x0001
  1088. /*
  1089. * Move expired (dirtied before dirtied_before) dirty inodes from
  1090. * @delaying_queue to @dispatch_queue.
  1091. */
  1092. static int move_expired_inodes(struct list_head *delaying_queue,
  1093. struct list_head *dispatch_queue,
  1094. unsigned long dirtied_before)
  1095. {
  1096. LIST_HEAD(tmp);
  1097. struct list_head *pos, *node;
  1098. struct super_block *sb = NULL;
  1099. struct inode *inode;
  1100. int do_sb_sort = 0;
  1101. int moved = 0;
  1102. while (!list_empty(delaying_queue)) {
  1103. inode = wb_inode(delaying_queue->prev);
  1104. if (inode_dirtied_after(inode, dirtied_before))
  1105. break;
  1106. list_move(&inode->i_io_list, &tmp);
  1107. moved++;
  1108. spin_lock(&inode->i_lock);
  1109. inode->i_state |= I_SYNC_QUEUED;
  1110. spin_unlock(&inode->i_lock);
  1111. if (sb_is_blkdev_sb(inode->i_sb))
  1112. continue;
  1113. if (sb && sb != inode->i_sb)
  1114. do_sb_sort = 1;
  1115. sb = inode->i_sb;
  1116. }
  1117. /* just one sb in list, splice to dispatch_queue and we're done */
  1118. if (!do_sb_sort) {
  1119. list_splice(&tmp, dispatch_queue);
  1120. goto out;
  1121. }
  1122. /* Move inodes from one superblock together */
  1123. while (!list_empty(&tmp)) {
  1124. sb = wb_inode(tmp.prev)->i_sb;
  1125. list_for_each_prev_safe(pos, node, &tmp) {
  1126. inode = wb_inode(pos);
  1127. if (inode->i_sb == sb)
  1128. list_move(&inode->i_io_list, dispatch_queue);
  1129. }
  1130. }
  1131. out:
  1132. return moved;
  1133. }
  1134. /*
  1135. * Queue all expired dirty inodes for io, eldest first.
  1136. * Before
  1137. * newly dirtied b_dirty b_io b_more_io
  1138. * =============> gf edc BA
  1139. * After
  1140. * newly dirtied b_dirty b_io b_more_io
  1141. * =============> g fBAedc
  1142. * |
  1143. * +--> dequeue for IO
  1144. */
  1145. static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work,
  1146. unsigned long dirtied_before)
  1147. {
  1148. int moved;
  1149. unsigned long time_expire_jif = dirtied_before;
  1150. assert_spin_locked(&wb->list_lock);
  1151. list_splice_init(&wb->b_more_io, &wb->b_io);
  1152. moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before);
  1153. if (!work->for_sync)
  1154. time_expire_jif = jiffies - dirtytime_expire_interval * HZ;
  1155. moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io,
  1156. time_expire_jif);
  1157. if (moved)
  1158. wb_io_lists_populated(wb);
  1159. trace_writeback_queue_io(wb, work, dirtied_before, moved);
  1160. }
  1161. static int write_inode(struct inode *inode, struct writeback_control *wbc)
  1162. {
  1163. int ret;
  1164. if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) {
  1165. trace_writeback_write_inode_start(inode, wbc);
  1166. ret = inode->i_sb->s_op->write_inode(inode, wbc);
  1167. trace_writeback_write_inode(inode, wbc);
  1168. return ret;
  1169. }
  1170. return 0;
  1171. }
  1172. /*
  1173. * Wait for writeback on an inode to complete. Called with i_lock held.
  1174. * Caller must make sure inode cannot go away when we drop i_lock.
  1175. */
  1176. static void __inode_wait_for_writeback(struct inode *inode)
  1177. __releases(inode->i_lock)
  1178. __acquires(inode->i_lock)
  1179. {
  1180. DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC);
  1181. wait_queue_head_t *wqh;
  1182. wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
  1183. while (inode->i_state & I_SYNC) {
  1184. spin_unlock(&inode->i_lock);
  1185. __wait_on_bit(wqh, &wq, bit_wait,
  1186. TASK_UNINTERRUPTIBLE);
  1187. spin_lock(&inode->i_lock);
  1188. }
  1189. }
  1190. /*
  1191. * Wait for writeback on an inode to complete. Caller must have inode pinned.
  1192. */
  1193. void inode_wait_for_writeback(struct inode *inode)
  1194. {
  1195. spin_lock(&inode->i_lock);
  1196. __inode_wait_for_writeback(inode);
  1197. spin_unlock(&inode->i_lock);
  1198. }
  1199. /*
  1200. * Sleep until I_SYNC is cleared. This function must be called with i_lock
  1201. * held and drops it. It is aimed for callers not holding any inode reference
  1202. * so once i_lock is dropped, inode can go away.
  1203. */
  1204. static void inode_sleep_on_writeback(struct inode *inode)
  1205. __releases(inode->i_lock)
  1206. {
  1207. DEFINE_WAIT(wait);
  1208. wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC);
  1209. int sleep;
  1210. prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
  1211. sleep = inode->i_state & I_SYNC;
  1212. spin_unlock(&inode->i_lock);
  1213. if (sleep)
  1214. schedule();
  1215. finish_wait(wqh, &wait);
  1216. }
  1217. /*
  1218. * Find proper writeback list for the inode depending on its current state and
  1219. * possibly also change of its state while we were doing writeback. Here we
  1220. * handle things such as livelock prevention or fairness of writeback among
  1221. * inodes. This function can be called only by flusher thread - noone else
  1222. * processes all inodes in writeback lists and requeueing inodes behind flusher
  1223. * thread's back can have unexpected consequences.
  1224. */
  1225. static void requeue_inode(struct inode *inode, struct bdi_writeback *wb,
  1226. struct writeback_control *wbc)
  1227. {
  1228. if (inode->i_state & I_FREEING)
  1229. return;
  1230. /*
  1231. * Sync livelock prevention. Each inode is tagged and synced in one
  1232. * shot. If still dirty, it will be redirty_tail()'ed below. Update
  1233. * the dirty time to prevent enqueue and sync it again.
  1234. */
  1235. if ((inode->i_state & I_DIRTY) &&
  1236. (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages))
  1237. inode->dirtied_when = jiffies;
  1238. if (wbc->pages_skipped) {
  1239. /*
  1240. * writeback is not making progress due to locked
  1241. * buffers. Skip this inode for now.
  1242. */
  1243. redirty_tail_locked(inode, wb);
  1244. return;
  1245. }
  1246. if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
  1247. /*
  1248. * We didn't write back all the pages. nfs_writepages()
  1249. * sometimes bales out without doing anything.
  1250. */
  1251. if (wbc->nr_to_write <= 0) {
  1252. /* Slice used up. Queue for next turn. */
  1253. requeue_io(inode, wb);
  1254. } else {
  1255. /*
  1256. * Writeback blocked by something other than
  1257. * congestion. Delay the inode for some time to
  1258. * avoid spinning on the CPU (100% iowait)
  1259. * retrying writeback of the dirty page/inode
  1260. * that cannot be performed immediately.
  1261. */
  1262. redirty_tail_locked(inode, wb);
  1263. }
  1264. } else if (inode->i_state & I_DIRTY) {
  1265. /*
  1266. * Filesystems can dirty the inode during writeback operations,
  1267. * such as delayed allocation during submission or metadata
  1268. * updates after data IO completion.
  1269. */
  1270. redirty_tail_locked(inode, wb);
  1271. } else if (inode->i_state & I_DIRTY_TIME) {
  1272. inode->dirtied_when = jiffies;
  1273. inode_io_list_move_locked(inode, wb, &wb->b_dirty_time);
  1274. inode->i_state &= ~I_SYNC_QUEUED;
  1275. } else {
  1276. /* The inode is clean. Remove from writeback lists. */
  1277. inode_io_list_del_locked(inode, wb);
  1278. }
  1279. }
  1280. /*
  1281. * Write out an inode and its dirty pages. Do not update the writeback list
  1282. * linkage. That is left to the caller. The caller is also responsible for
  1283. * setting I_SYNC flag and calling inode_sync_complete() to clear it.
  1284. */
  1285. static int
  1286. __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
  1287. {
  1288. struct address_space *mapping = inode->i_mapping;
  1289. long nr_to_write = wbc->nr_to_write;
  1290. unsigned dirty;
  1291. int ret;
  1292. WARN_ON(!(inode->i_state & I_SYNC));
  1293. trace_writeback_single_inode_start(inode, wbc, nr_to_write);
  1294. ret = do_writepages(mapping, wbc);
  1295. /*
  1296. * Make sure to wait on the data before writing out the metadata.
  1297. * This is important for filesystems that modify metadata on data
  1298. * I/O completion. We don't do it for sync(2) writeback because it has a
  1299. * separate, external IO completion path and ->sync_fs for guaranteeing
  1300. * inode metadata is written back correctly.
  1301. */
  1302. if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) {
  1303. int err = filemap_fdatawait(mapping);
  1304. if (ret == 0)
  1305. ret = err;
  1306. }
  1307. /*
  1308. * If the inode has dirty timestamps and we need to write them, call
  1309. * mark_inode_dirty_sync() to notify the filesystem about it and to
  1310. * change I_DIRTY_TIME into I_DIRTY_SYNC.
  1311. */
  1312. if ((inode->i_state & I_DIRTY_TIME) &&
  1313. (wbc->sync_mode == WB_SYNC_ALL || wbc->for_sync ||
  1314. time_after(jiffies, inode->dirtied_time_when +
  1315. dirtytime_expire_interval * HZ))) {
  1316. trace_writeback_lazytime(inode);
  1317. mark_inode_dirty_sync(inode);
  1318. }
  1319. /*
  1320. * Some filesystems may redirty the inode during the writeback
  1321. * due to delalloc, clear dirty metadata flags right before
  1322. * write_inode()
  1323. */
  1324. spin_lock(&inode->i_lock);
  1325. dirty = inode->i_state & I_DIRTY;
  1326. inode->i_state &= ~dirty;
  1327. /*
  1328. * Paired with smp_mb() in __mark_inode_dirty(). This allows
  1329. * __mark_inode_dirty() to test i_state without grabbing i_lock -
  1330. * either they see the I_DIRTY bits cleared or we see the dirtied
  1331. * inode.
  1332. *
  1333. * I_DIRTY_PAGES is always cleared together above even if @mapping
  1334. * still has dirty pages. The flag is reinstated after smp_mb() if
  1335. * necessary. This guarantees that either __mark_inode_dirty()
  1336. * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY.
  1337. */
  1338. smp_mb();
  1339. if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
  1340. inode->i_state |= I_DIRTY_PAGES;
  1341. spin_unlock(&inode->i_lock);
  1342. /* Don't write the inode if only I_DIRTY_PAGES was set */
  1343. if (dirty & ~I_DIRTY_PAGES) {
  1344. int err = write_inode(inode, wbc);
  1345. if (ret == 0)
  1346. ret = err;
  1347. }
  1348. trace_writeback_single_inode(inode, wbc, nr_to_write);
  1349. return ret;
  1350. }
  1351. /*
  1352. * Write out an inode's dirty pages. Either the caller has an active reference
  1353. * on the inode or the inode has I_WILL_FREE set.
  1354. *
  1355. * This function is designed to be called for writing back one inode which
  1356. * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode()
  1357. * and does more profound writeback list handling in writeback_sb_inodes().
  1358. */
  1359. static int writeback_single_inode(struct inode *inode,
  1360. struct writeback_control *wbc)
  1361. {
  1362. struct bdi_writeback *wb;
  1363. int ret = 0;
  1364. spin_lock(&inode->i_lock);
  1365. if (!atomic_read(&inode->i_count))
  1366. WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING)));
  1367. else
  1368. WARN_ON(inode->i_state & I_WILL_FREE);
  1369. if (inode->i_state & I_SYNC) {
  1370. if (wbc->sync_mode != WB_SYNC_ALL)
  1371. goto out;
  1372. /*
  1373. * It's a data-integrity sync. We must wait. Since callers hold
  1374. * inode reference or inode has I_WILL_FREE set, it cannot go
  1375. * away under us.
  1376. */
  1377. __inode_wait_for_writeback(inode);
  1378. }
  1379. WARN_ON(inode->i_state & I_SYNC);
  1380. /*
  1381. * Skip inode if it is clean and we have no outstanding writeback in
  1382. * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this
  1383. * function since flusher thread may be doing for example sync in
  1384. * parallel and if we move the inode, it could get skipped. So here we
  1385. * make sure inode is on some writeback list and leave it there unless
  1386. * we have completely cleaned the inode.
  1387. */
  1388. if (!(inode->i_state & I_DIRTY_ALL) &&
  1389. (wbc->sync_mode != WB_SYNC_ALL ||
  1390. !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK)))
  1391. goto out;
  1392. inode->i_state |= I_SYNC;
  1393. wbc_attach_and_unlock_inode(wbc, inode);
  1394. ret = __writeback_single_inode(inode, wbc);
  1395. wbc_detach_inode(wbc);
  1396. wb = inode_to_wb_and_lock_list(inode);
  1397. spin_lock(&inode->i_lock);
  1398. /*
  1399. * If inode is clean, remove it from writeback lists. Otherwise don't
  1400. * touch it. See comment above for explanation.
  1401. */
  1402. if (!(inode->i_state & I_DIRTY_ALL))
  1403. inode_io_list_del_locked(inode, wb);
  1404. spin_unlock(&wb->list_lock);
  1405. inode_sync_complete(inode);
  1406. out:
  1407. spin_unlock(&inode->i_lock);
  1408. return ret;
  1409. }
  1410. static long writeback_chunk_size(struct bdi_writeback *wb,
  1411. struct wb_writeback_work *work)
  1412. {
  1413. long pages;
  1414. /*
  1415. * WB_SYNC_ALL mode does livelock avoidance by syncing dirty
  1416. * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX
  1417. * here avoids calling into writeback_inodes_wb() more than once.
  1418. *
  1419. * The intended call sequence for WB_SYNC_ALL writeback is:
  1420. *
  1421. * wb_writeback()
  1422. * writeback_sb_inodes() <== called only once
  1423. * write_cache_pages() <== called once for each inode
  1424. * (quickly) tag currently dirty pages
  1425. * (maybe slowly) sync all tagged pages
  1426. */
  1427. if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages)
  1428. pages = LONG_MAX;
  1429. else {
  1430. pages = min(wb->avg_write_bandwidth / 2,
  1431. global_wb_domain.dirty_limit / DIRTY_SCOPE);
  1432. pages = min(pages, work->nr_pages);
  1433. pages = round_down(pages + MIN_WRITEBACK_PAGES,
  1434. MIN_WRITEBACK_PAGES);
  1435. }
  1436. return pages;
  1437. }
  1438. /*
  1439. * Write a portion of b_io inodes which belong to @sb.
  1440. *
  1441. * Return the number of pages and/or inodes written.
  1442. *
  1443. * NOTE! This is called with wb->list_lock held, and will
  1444. * unlock and relock that for each inode it ends up doing
  1445. * IO for.
  1446. */
  1447. static long writeback_sb_inodes(struct super_block *sb,
  1448. struct bdi_writeback *wb,
  1449. struct wb_writeback_work *work)
  1450. {
  1451. struct writeback_control wbc = {
  1452. .sync_mode = work->sync_mode,
  1453. .tagged_writepages = work->tagged_writepages,
  1454. .for_kupdate = work->for_kupdate,
  1455. .for_background = work->for_background,
  1456. .for_sync = work->for_sync,
  1457. .range_cyclic = work->range_cyclic,
  1458. .range_start = 0,
  1459. .range_end = LLONG_MAX,
  1460. };
  1461. unsigned long start_time = jiffies;
  1462. long write_chunk;
  1463. long wrote = 0; /* count both pages and inodes */
  1464. while (!list_empty(&wb->b_io)) {
  1465. struct inode *inode = wb_inode(wb->b_io.prev);
  1466. struct bdi_writeback *tmp_wb;
  1467. if (inode->i_sb != sb) {
  1468. if (work->sb) {
  1469. /*
  1470. * We only want to write back data for this
  1471. * superblock, move all inodes not belonging
  1472. * to it back onto the dirty list.
  1473. */
  1474. redirty_tail(inode, wb);
  1475. continue;
  1476. }
  1477. /*
  1478. * The inode belongs to a different superblock.
  1479. * Bounce back to the caller to unpin this and
  1480. * pin the next superblock.
  1481. */
  1482. break;
  1483. }
  1484. /*
  1485. * Don't bother with new inodes or inodes being freed, first
  1486. * kind does not need periodic writeout yet, and for the latter
  1487. * kind writeout is handled by the freer.
  1488. */
  1489. spin_lock(&inode->i_lock);
  1490. if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) {
  1491. redirty_tail_locked(inode, wb);
  1492. spin_unlock(&inode->i_lock);
  1493. continue;
  1494. }
  1495. if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) {
  1496. /*
  1497. * If this inode is locked for writeback and we are not
  1498. * doing writeback-for-data-integrity, move it to
  1499. * b_more_io so that writeback can proceed with the
  1500. * other inodes on s_io.
  1501. *
  1502. * We'll have another go at writing back this inode
  1503. * when we completed a full scan of b_io.
  1504. */
  1505. spin_unlock(&inode->i_lock);
  1506. requeue_io(inode, wb);
  1507. trace_writeback_sb_inodes_requeue(inode);
  1508. continue;
  1509. }
  1510. spin_unlock(&wb->list_lock);
  1511. /*
  1512. * We already requeued the inode if it had I_SYNC set and we
  1513. * are doing WB_SYNC_NONE writeback. So this catches only the
  1514. * WB_SYNC_ALL case.
  1515. */
  1516. if (inode->i_state & I_SYNC) {
  1517. /* Wait for I_SYNC. This function drops i_lock... */
  1518. inode_sleep_on_writeback(inode);
  1519. /* Inode may be gone, start again */
  1520. spin_lock(&wb->list_lock);
  1521. continue;
  1522. }
  1523. inode->i_state |= I_SYNC;
  1524. wbc_attach_and_unlock_inode(&wbc, inode);
  1525. write_chunk = writeback_chunk_size(wb, work);
  1526. wbc.nr_to_write = write_chunk;
  1527. wbc.pages_skipped = 0;
  1528. /*
  1529. * We use I_SYNC to pin the inode in memory. While it is set
  1530. * evict_inode() will wait so the inode cannot be freed.
  1531. */
  1532. __writeback_single_inode(inode, &wbc);
  1533. wbc_detach_inode(&wbc);
  1534. work->nr_pages -= write_chunk - wbc.nr_to_write;
  1535. wrote += write_chunk - wbc.nr_to_write;
  1536. if (need_resched()) {
  1537. /*
  1538. * We're trying to balance between building up a nice
  1539. * long list of IOs to improve our merge rate, and
  1540. * getting those IOs out quickly for anyone throttling
  1541. * in balance_dirty_pages(). cond_resched() doesn't
  1542. * unplug, so get our IOs out the door before we
  1543. * give up the CPU.
  1544. */
  1545. blk_flush_plug(current);
  1546. cond_resched();
  1547. }
  1548. /*
  1549. * Requeue @inode if still dirty. Be careful as @inode may
  1550. * have been switched to another wb in the meantime.
  1551. */
  1552. tmp_wb = inode_to_wb_and_lock_list(inode);
  1553. spin_lock(&inode->i_lock);
  1554. if (!(inode->i_state & I_DIRTY_ALL))
  1555. wrote++;
  1556. requeue_inode(inode, tmp_wb, &wbc);
  1557. inode_sync_complete(inode);
  1558. spin_unlock(&inode->i_lock);
  1559. if (unlikely(tmp_wb != wb)) {
  1560. spin_unlock(&tmp_wb->list_lock);
  1561. spin_lock(&wb->list_lock);
  1562. }
  1563. /*
  1564. * bail out to wb_writeback() often enough to check
  1565. * background threshold and other termination conditions.
  1566. */
  1567. if (wrote) {
  1568. if (time_is_before_jiffies(start_time + HZ / 10UL))
  1569. break;
  1570. if (work->nr_pages <= 0)
  1571. break;
  1572. }
  1573. }
  1574. return wrote;
  1575. }
  1576. static long __writeback_inodes_wb(struct bdi_writeback *wb,
  1577. struct wb_writeback_work *work)
  1578. {
  1579. unsigned long start_time = jiffies;
  1580. long wrote = 0;
  1581. while (!list_empty(&wb->b_io)) {
  1582. struct inode *inode = wb_inode(wb->b_io.prev);
  1583. struct super_block *sb = inode->i_sb;
  1584. if (!trylock_super(sb)) {
  1585. /*
  1586. * trylock_super() may fail consistently due to
  1587. * s_umount being grabbed by someone else. Don't use
  1588. * requeue_io() to avoid busy retrying the inode/sb.
  1589. */
  1590. redirty_tail(inode, wb);
  1591. continue;
  1592. }
  1593. wrote += writeback_sb_inodes(sb, wb, work);
  1594. up_read(&sb->s_umount);
  1595. /* refer to the same tests at the end of writeback_sb_inodes */
  1596. if (wrote) {
  1597. if (time_is_before_jiffies(start_time + HZ / 10UL))
  1598. break;
  1599. if (work->nr_pages <= 0)
  1600. break;
  1601. }
  1602. }
  1603. /* Leave any unwritten inodes on b_io */
  1604. return wrote;
  1605. }
  1606. static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
  1607. enum wb_reason reason)
  1608. {
  1609. struct wb_writeback_work work = {
  1610. .nr_pages = nr_pages,
  1611. .sync_mode = WB_SYNC_NONE,
  1612. .range_cyclic = 1,
  1613. .reason = reason,
  1614. };
  1615. struct blk_plug plug;
  1616. blk_start_plug(&plug);
  1617. spin_lock(&wb->list_lock);
  1618. if (list_empty(&wb->b_io))
  1619. queue_io(wb, &work, jiffies);
  1620. __writeback_inodes_wb(wb, &work);
  1621. spin_unlock(&wb->list_lock);
  1622. blk_finish_plug(&plug);
  1623. return nr_pages - work.nr_pages;
  1624. }
  1625. /*
  1626. * Explicit flushing or periodic writeback of "old" data.
  1627. *
  1628. * Define "old": the first time one of an inode's pages is dirtied, we mark the
  1629. * dirtying-time in the inode's address_space. So this periodic writeback code
  1630. * just walks the superblock inode list, writing back any inodes which are
  1631. * older than a specific point in time.
  1632. *
  1633. * Try to run once per dirty_writeback_interval. But if a writeback event
  1634. * takes longer than a dirty_writeback_interval interval, then leave a
  1635. * one-second gap.
  1636. *
  1637. * dirtied_before takes precedence over nr_to_write. So we'll only write back
  1638. * all dirty pages if they are all attached to "old" mappings.
  1639. */
  1640. static long wb_writeback(struct bdi_writeback *wb,
  1641. struct wb_writeback_work *work)
  1642. {
  1643. unsigned long wb_start = jiffies;
  1644. long nr_pages = work->nr_pages;
  1645. unsigned long dirtied_before = jiffies;
  1646. struct inode *inode;
  1647. long progress;
  1648. struct blk_plug plug;
  1649. blk_start_plug(&plug);
  1650. spin_lock(&wb->list_lock);
  1651. for (;;) {
  1652. /*
  1653. * Stop writeback when nr_pages has been consumed
  1654. */
  1655. if (work->nr_pages <= 0)
  1656. break;
  1657. /*
  1658. * Background writeout and kupdate-style writeback may
  1659. * run forever. Stop them if there is other work to do
  1660. * so that e.g. sync can proceed. They'll be restarted
  1661. * after the other works are all done.
  1662. */
  1663. if ((work->for_background || work->for_kupdate) &&
  1664. !list_empty(&wb->work_list))
  1665. break;
  1666. /*
  1667. * For background writeout, stop when we are below the
  1668. * background dirty threshold
  1669. */
  1670. if (work->for_background && !wb_over_bg_thresh(wb))
  1671. break;
  1672. /*
  1673. * Kupdate and background works are special and we want to
  1674. * include all inodes that need writing. Livelock avoidance is
  1675. * handled by these works yielding to any other work so we are
  1676. * safe.
  1677. */
  1678. if (work->for_kupdate) {
  1679. dirtied_before = jiffies -
  1680. msecs_to_jiffies(dirty_expire_interval * 10);
  1681. } else if (work->for_background)
  1682. dirtied_before = jiffies;
  1683. trace_writeback_start(wb, work);
  1684. if (list_empty(&wb->b_io))
  1685. queue_io(wb, work, dirtied_before);
  1686. if (work->sb)
  1687. progress = writeback_sb_inodes(work->sb, wb, work);
  1688. else
  1689. progress = __writeback_inodes_wb(wb, work);
  1690. trace_writeback_written(wb, work);
  1691. wb_update_bandwidth(wb, wb_start);
  1692. /*
  1693. * Did we write something? Try for more
  1694. *
  1695. * Dirty inodes are moved to b_io for writeback in batches.
  1696. * The completion of the current batch does not necessarily
  1697. * mean the overall work is done. So we keep looping as long
  1698. * as made some progress on cleaning pages or inodes.
  1699. */
  1700. if (progress)
  1701. continue;
  1702. /*
  1703. * No more inodes for IO, bail
  1704. */
  1705. if (list_empty(&wb->b_more_io))
  1706. break;
  1707. /*
  1708. * Nothing written. Wait for some inode to
  1709. * become available for writeback. Otherwise
  1710. * we'll just busyloop.
  1711. */
  1712. trace_writeback_wait(wb, work);
  1713. inode = wb_inode(wb->b_more_io.prev);
  1714. spin_lock(&inode->i_lock);
  1715. spin_unlock(&wb->list_lock);
  1716. /* This function drops i_lock... */
  1717. inode_sleep_on_writeback(inode);
  1718. spin_lock(&wb->list_lock);
  1719. }
  1720. spin_unlock(&wb->list_lock);
  1721. blk_finish_plug(&plug);
  1722. return nr_pages - work->nr_pages;
  1723. }
  1724. /*
  1725. * Return the next wb_writeback_work struct that hasn't been processed yet.
  1726. */
  1727. static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb)
  1728. {
  1729. struct wb_writeback_work *work = NULL;
  1730. spin_lock_bh(&wb->work_lock);
  1731. if (!list_empty(&wb->work_list)) {
  1732. work = list_entry(wb->work_list.next,
  1733. struct wb_writeback_work, list);
  1734. list_del_init(&work->list);
  1735. }
  1736. spin_unlock_bh(&wb->work_lock);
  1737. return work;
  1738. }
  1739. static long wb_check_background_flush(struct bdi_writeback *wb)
  1740. {
  1741. if (wb_over_bg_thresh(wb)) {
  1742. struct wb_writeback_work work = {
  1743. .nr_pages = LONG_MAX,
  1744. .sync_mode = WB_SYNC_NONE,
  1745. .for_background = 1,
  1746. .range_cyclic = 1,
  1747. .reason = WB_REASON_BACKGROUND,
  1748. };
  1749. return wb_writeback(wb, &work);
  1750. }
  1751. return 0;
  1752. }
  1753. static long wb_check_old_data_flush(struct bdi_writeback *wb)
  1754. {
  1755. unsigned long expired;
  1756. long nr_pages;
  1757. /*
  1758. * When set to zero, disable periodic writeback
  1759. */
  1760. if (!dirty_writeback_interval)
  1761. return 0;
  1762. expired = wb->last_old_flush +
  1763. msecs_to_jiffies(dirty_writeback_interval * 10);
  1764. if (time_before(jiffies, expired))
  1765. return 0;
  1766. wb->last_old_flush = jiffies;
  1767. nr_pages = get_nr_dirty_pages();
  1768. if (nr_pages) {
  1769. struct wb_writeback_work work = {
  1770. .nr_pages = nr_pages,
  1771. .sync_mode = WB_SYNC_NONE,
  1772. .for_kupdate = 1,
  1773. .range_cyclic = 1,
  1774. .reason = WB_REASON_PERIODIC,
  1775. };
  1776. return wb_writeback(wb, &work);
  1777. }
  1778. return 0;
  1779. }
  1780. static long wb_check_start_all(struct bdi_writeback *wb)
  1781. {
  1782. long nr_pages;
  1783. if (!test_bit(WB_start_all, &wb->state))
  1784. return 0;
  1785. nr_pages = get_nr_dirty_pages();
  1786. if (nr_pages) {
  1787. struct wb_writeback_work work = {
  1788. .nr_pages = wb_split_bdi_pages(wb, nr_pages),
  1789. .sync_mode = WB_SYNC_NONE,
  1790. .range_cyclic = 1,
  1791. .reason = wb->start_all_reason,
  1792. };
  1793. nr_pages = wb_writeback(wb, &work);
  1794. }
  1795. clear_bit(WB_start_all, &wb->state);
  1796. return nr_pages;
  1797. }
  1798. /*
  1799. * Retrieve work items and do the writeback they describe
  1800. */
  1801. static long wb_do_writeback(struct bdi_writeback *wb)
  1802. {
  1803. struct wb_writeback_work *work;
  1804. long wrote = 0;
  1805. set_bit(WB_writeback_running, &wb->state);
  1806. while ((work = get_next_work_item(wb)) != NULL) {
  1807. trace_writeback_exec(wb, work);
  1808. wrote += wb_writeback(wb, work);
  1809. finish_writeback_work(wb, work);
  1810. }
  1811. /*
  1812. * Check for a flush-everything request
  1813. */
  1814. wrote += wb_check_start_all(wb);
  1815. /*
  1816. * Check for periodic writeback, kupdated() style
  1817. */
  1818. wrote += wb_check_old_data_flush(wb);
  1819. wrote += wb_check_background_flush(wb);
  1820. clear_bit(WB_writeback_running, &wb->state);
  1821. return wrote;
  1822. }
  1823. /*
  1824. * Handle writeback of dirty data for the device backed by this bdi. Also
  1825. * reschedules periodically and does kupdated style flushing.
  1826. */
  1827. void wb_workfn(struct work_struct *work)
  1828. {
  1829. struct bdi_writeback *wb = container_of(to_delayed_work(work),
  1830. struct bdi_writeback, dwork);
  1831. long pages_written;
  1832. set_worker_desc("flush-%s", bdi_dev_name(wb->bdi));
  1833. current->flags |= PF_SWAPWRITE;
  1834. if (likely(!current_is_workqueue_rescuer() ||
  1835. !test_bit(WB_registered, &wb->state))) {
  1836. /*
  1837. * The normal path. Keep writing back @wb until its
  1838. * work_list is empty. Note that this path is also taken
  1839. * if @wb is shutting down even when we're running off the
  1840. * rescuer as work_list needs to be drained.
  1841. */
  1842. do {
  1843. pages_written = wb_do_writeback(wb);
  1844. trace_writeback_pages_written(pages_written);
  1845. } while (!list_empty(&wb->work_list));
  1846. } else {
  1847. /*
  1848. * bdi_wq can't get enough workers and we're running off
  1849. * the emergency worker. Don't hog it. Hopefully, 1024 is
  1850. * enough for efficient IO.
  1851. */
  1852. pages_written = writeback_inodes_wb(wb, 1024,
  1853. WB_REASON_FORKER_THREAD);
  1854. trace_writeback_pages_written(pages_written);
  1855. }
  1856. if (!list_empty(&wb->work_list))
  1857. wb_wakeup(wb);
  1858. else if (wb_has_dirty_io(wb) && dirty_writeback_interval)
  1859. wb_wakeup_delayed(wb);
  1860. current->flags &= ~PF_SWAPWRITE;
  1861. }
  1862. /*
  1863. * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero,
  1864. * write back the whole world.
  1865. */
  1866. static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
  1867. enum wb_reason reason)
  1868. {
  1869. struct bdi_writeback *wb;
  1870. if (!bdi_has_dirty_io(bdi))
  1871. return;
  1872. list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
  1873. wb_start_writeback(wb, reason);
  1874. }
  1875. void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi,
  1876. enum wb_reason reason)
  1877. {
  1878. rcu_read_lock();
  1879. __wakeup_flusher_threads_bdi(bdi, reason);
  1880. rcu_read_unlock();
  1881. }
  1882. /*
  1883. * Wakeup the flusher threads to start writeback of all currently dirty pages
  1884. */
  1885. void wakeup_flusher_threads(enum wb_reason reason)
  1886. {
  1887. struct backing_dev_info *bdi;
  1888. /*
  1889. * If we are expecting writeback progress we must submit plugged IO.
  1890. */
  1891. if (blk_needs_flush_plug(current))
  1892. blk_schedule_flush_plug(current);
  1893. rcu_read_lock();
  1894. list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
  1895. __wakeup_flusher_threads_bdi(bdi, reason);
  1896. rcu_read_unlock();
  1897. }
  1898. /*
  1899. * Wake up bdi's periodically to make sure dirtytime inodes gets
  1900. * written back periodically. We deliberately do *not* check the
  1901. * b_dirtytime list in wb_has_dirty_io(), since this would cause the
  1902. * kernel to be constantly waking up once there are any dirtytime
  1903. * inodes on the system. So instead we define a separate delayed work
  1904. * function which gets called much more rarely. (By default, only
  1905. * once every 12 hours.)
  1906. *
  1907. * If there is any other write activity going on in the file system,
  1908. * this function won't be necessary. But if the only thing that has
  1909. * happened on the file system is a dirtytime inode caused by an atime
  1910. * update, we need this infrastructure below to make sure that inode
  1911. * eventually gets pushed out to disk.
  1912. */
  1913. static void wakeup_dirtytime_writeback(struct work_struct *w);
  1914. static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback);
  1915. static void wakeup_dirtytime_writeback(struct work_struct *w)
  1916. {
  1917. struct backing_dev_info *bdi;
  1918. rcu_read_lock();
  1919. list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) {
  1920. struct bdi_writeback *wb;
  1921. list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node)
  1922. if (!list_empty(&wb->b_dirty_time))
  1923. wb_wakeup(wb);
  1924. }
  1925. rcu_read_unlock();
  1926. schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
  1927. }
  1928. static int __init start_dirtytime_writeback(void)
  1929. {
  1930. schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ);
  1931. return 0;
  1932. }
  1933. __initcall(start_dirtytime_writeback);
  1934. int dirtytime_interval_handler(struct ctl_table *table, int write,
  1935. void *buffer, size_t *lenp, loff_t *ppos)
  1936. {
  1937. int ret;
  1938. ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
  1939. if (ret == 0 && write)
  1940. mod_delayed_work(system_wq, &dirtytime_work, 0);
  1941. return ret;
  1942. }
  1943. /**
  1944. * __mark_inode_dirty - internal function
  1945. *
  1946. * @inode: inode to mark
  1947. * @flags: what kind of dirty (i.e. I_DIRTY_SYNC)
  1948. *
  1949. * Mark an inode as dirty. Callers should use mark_inode_dirty or
  1950. * mark_inode_dirty_sync.
  1951. *
  1952. * Put the inode on the super block's dirty list.
  1953. *
  1954. * CAREFUL! We mark it dirty unconditionally, but move it onto the
  1955. * dirty list only if it is hashed or if it refers to a blockdev.
  1956. * If it was not hashed, it will never be added to the dirty list
  1957. * even if it is later hashed, as it will have been marked dirty already.
  1958. *
  1959. * In short, make sure you hash any inodes _before_ you start marking
  1960. * them dirty.
  1961. *
  1962. * Note that for blockdevs, inode->dirtied_when represents the dirtying time of
  1963. * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of
  1964. * the kernel-internal blockdev inode represents the dirtying time of the
  1965. * blockdev's pages. This is why for I_DIRTY_PAGES we always use
  1966. * page->mapping->host, so the page-dirtying time is recorded in the internal
  1967. * blockdev inode.
  1968. */
  1969. void __mark_inode_dirty(struct inode *inode, int flags)
  1970. {
  1971. struct super_block *sb = inode->i_sb;
  1972. int dirtytime;
  1973. trace_writeback_mark_inode_dirty(inode, flags);
  1974. /*
  1975. * Don't do this for I_DIRTY_PAGES - that doesn't actually
  1976. * dirty the inode itself
  1977. */
  1978. if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) {
  1979. trace_writeback_dirty_inode_start(inode, flags);
  1980. if (sb->s_op->dirty_inode)
  1981. sb->s_op->dirty_inode(inode, flags);
  1982. trace_writeback_dirty_inode(inode, flags);
  1983. }
  1984. if (flags & I_DIRTY_INODE)
  1985. flags &= ~I_DIRTY_TIME;
  1986. dirtytime = flags & I_DIRTY_TIME;
  1987. /*
  1988. * Paired with smp_mb() in __writeback_single_inode() for the
  1989. * following lockless i_state test. See there for details.
  1990. */
  1991. smp_mb();
  1992. if (((inode->i_state & flags) == flags) ||
  1993. (dirtytime && (inode->i_state & I_DIRTY_INODE)))
  1994. return;
  1995. spin_lock(&inode->i_lock);
  1996. if (dirtytime && (inode->i_state & I_DIRTY_INODE))
  1997. goto out_unlock_inode;
  1998. if ((inode->i_state & flags) != flags) {
  1999. const int was_dirty = inode->i_state & I_DIRTY;
  2000. inode_attach_wb(inode, NULL);
  2001. if (flags & I_DIRTY_INODE)
  2002. inode->i_state &= ~I_DIRTY_TIME;
  2003. inode->i_state |= flags;
  2004. /*
  2005. * If the inode is queued for writeback by flush worker, just
  2006. * update its dirty state. Once the flush worker is done with
  2007. * the inode it will place it on the appropriate superblock
  2008. * list, based upon its state.
  2009. */
  2010. if (inode->i_state & I_SYNC_QUEUED)
  2011. goto out_unlock_inode;
  2012. /*
  2013. * Only add valid (hashed) inodes to the superblock's
  2014. * dirty list. Add blockdev inodes as well.
  2015. */
  2016. if (!S_ISBLK(inode->i_mode)) {
  2017. if (inode_unhashed(inode))
  2018. goto out_unlock_inode;
  2019. }
  2020. if (inode->i_state & I_FREEING)
  2021. goto out_unlock_inode;
  2022. /*
  2023. * If the inode was already on b_dirty/b_io/b_more_io, don't
  2024. * reposition it (that would break b_dirty time-ordering).
  2025. */
  2026. if (!was_dirty) {
  2027. struct bdi_writeback *wb;
  2028. struct list_head *dirty_list;
  2029. bool wakeup_bdi = false;
  2030. wb = locked_inode_to_wb_and_lock_list(inode);
  2031. WARN((wb->bdi->capabilities & BDI_CAP_WRITEBACK) &&
  2032. !test_bit(WB_registered, &wb->state),
  2033. "bdi-%s not registered\n", bdi_dev_name(wb->bdi));
  2034. inode->dirtied_when = jiffies;
  2035. if (dirtytime)
  2036. inode->dirtied_time_when = jiffies;
  2037. if (inode->i_state & I_DIRTY)
  2038. dirty_list = &wb->b_dirty;
  2039. else
  2040. dirty_list = &wb->b_dirty_time;
  2041. wakeup_bdi = inode_io_list_move_locked(inode, wb,
  2042. dirty_list);
  2043. spin_unlock(&wb->list_lock);
  2044. trace_writeback_dirty_inode_enqueue(inode);
  2045. /*
  2046. * If this is the first dirty inode for this bdi,
  2047. * we have to wake-up the corresponding bdi thread
  2048. * to make sure background write-back happens
  2049. * later.
  2050. */
  2051. if (wakeup_bdi &&
  2052. (wb->bdi->capabilities & BDI_CAP_WRITEBACK))
  2053. wb_wakeup_delayed(wb);
  2054. return;
  2055. }
  2056. }
  2057. out_unlock_inode:
  2058. spin_unlock(&inode->i_lock);
  2059. }
  2060. EXPORT_SYMBOL_NS(__mark_inode_dirty, ANDROID_GKI_VFS_EXPORT_ONLY);
  2061. /*
  2062. * The @s_sync_lock is used to serialise concurrent sync operations
  2063. * to avoid lock contention problems with concurrent wait_sb_inodes() calls.
  2064. * Concurrent callers will block on the s_sync_lock rather than doing contending
  2065. * walks. The queueing maintains sync(2) required behaviour as all the IO that
  2066. * has been issued up to the time this function is enter is guaranteed to be
  2067. * completed by the time we have gained the lock and waited for all IO that is
  2068. * in progress regardless of the order callers are granted the lock.
  2069. */
  2070. static void wait_sb_inodes(struct super_block *sb)
  2071. {
  2072. LIST_HEAD(sync_list);
  2073. /*
  2074. * We need to be protected against the filesystem going from
  2075. * r/o to r/w or vice versa.
  2076. */
  2077. WARN_ON(!rwsem_is_locked(&sb->s_umount));
  2078. mutex_lock(&sb->s_sync_lock);
  2079. /*
  2080. * Splice the writeback list onto a temporary list to avoid waiting on
  2081. * inodes that have started writeback after this point.
  2082. *
  2083. * Use rcu_read_lock() to keep the inodes around until we have a
  2084. * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as
  2085. * the local list because inodes can be dropped from either by writeback
  2086. * completion.
  2087. */
  2088. rcu_read_lock();
  2089. spin_lock_irq(&sb->s_inode_wblist_lock);
  2090. list_splice_init(&sb->s_inodes_wb, &sync_list);
  2091. /*
  2092. * Data integrity sync. Must wait for all pages under writeback, because
  2093. * there may have been pages dirtied before our sync call, but which had
  2094. * writeout started before we write it out. In which case, the inode
  2095. * may not be on the dirty list, but we still have to wait for that
  2096. * writeout.
  2097. */
  2098. while (!list_empty(&sync_list)) {
  2099. struct inode *inode = list_first_entry(&sync_list, struct inode,
  2100. i_wb_list);
  2101. struct address_space *mapping = inode->i_mapping;
  2102. /*
  2103. * Move each inode back to the wb list before we drop the lock
  2104. * to preserve consistency between i_wb_list and the mapping
  2105. * writeback tag. Writeback completion is responsible to remove
  2106. * the inode from either list once the writeback tag is cleared.
  2107. */
  2108. list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb);
  2109. /*
  2110. * The mapping can appear untagged while still on-list since we
  2111. * do not have the mapping lock. Skip it here, wb completion
  2112. * will remove it.
  2113. */
  2114. if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK))
  2115. continue;
  2116. spin_unlock_irq(&sb->s_inode_wblist_lock);
  2117. spin_lock(&inode->i_lock);
  2118. if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
  2119. spin_unlock(&inode->i_lock);
  2120. spin_lock_irq(&sb->s_inode_wblist_lock);
  2121. continue;
  2122. }
  2123. __iget(inode);
  2124. spin_unlock(&inode->i_lock);
  2125. rcu_read_unlock();
  2126. /*
  2127. * We keep the error status of individual mapping so that
  2128. * applications can catch the writeback error using fsync(2).
  2129. * See filemap_fdatawait_keep_errors() for details.
  2130. */
  2131. filemap_fdatawait_keep_errors(mapping);
  2132. cond_resched();
  2133. iput(inode);
  2134. rcu_read_lock();
  2135. spin_lock_irq(&sb->s_inode_wblist_lock);
  2136. }
  2137. spin_unlock_irq(&sb->s_inode_wblist_lock);
  2138. rcu_read_unlock();
  2139. mutex_unlock(&sb->s_sync_lock);
  2140. }
  2141. static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr,
  2142. enum wb_reason reason, bool skip_if_busy)
  2143. {
  2144. struct backing_dev_info *bdi = sb->s_bdi;
  2145. DEFINE_WB_COMPLETION(done, bdi);
  2146. struct wb_writeback_work work = {
  2147. .sb = sb,
  2148. .sync_mode = WB_SYNC_NONE,
  2149. .tagged_writepages = 1,
  2150. .done = &done,
  2151. .nr_pages = nr,
  2152. .reason = reason,
  2153. };
  2154. if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info)
  2155. return;
  2156. WARN_ON(!rwsem_is_locked(&sb->s_umount));
  2157. bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy);
  2158. wb_wait_for_completion(&done);
  2159. }
  2160. /**
  2161. * writeback_inodes_sb_nr - writeback dirty inodes from given super_block
  2162. * @sb: the superblock
  2163. * @nr: the number of pages to write
  2164. * @reason: reason why some writeback work initiated
  2165. *
  2166. * Start writeback on some inodes on this super_block. No guarantees are made
  2167. * on how many (if any) will be written, and this function does not wait
  2168. * for IO completion of submitted IO.
  2169. */
  2170. void writeback_inodes_sb_nr(struct super_block *sb,
  2171. unsigned long nr,
  2172. enum wb_reason reason)
  2173. {
  2174. __writeback_inodes_sb_nr(sb, nr, reason, false);
  2175. }
  2176. EXPORT_SYMBOL(writeback_inodes_sb_nr);
  2177. /**
  2178. * writeback_inodes_sb - writeback dirty inodes from given super_block
  2179. * @sb: the superblock
  2180. * @reason: reason why some writeback work was initiated
  2181. *
  2182. * Start writeback on some inodes on this super_block. No guarantees are made
  2183. * on how many (if any) will be written, and this function does not wait
  2184. * for IO completion of submitted IO.
  2185. */
  2186. void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
  2187. {
  2188. return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason);
  2189. }
  2190. EXPORT_SYMBOL(writeback_inodes_sb);
  2191. /**
  2192. * try_to_writeback_inodes_sb - try to start writeback if none underway
  2193. * @sb: the superblock
  2194. * @reason: reason why some writeback work was initiated
  2195. *
  2196. * Invoke __writeback_inodes_sb_nr if no writeback is currently underway.
  2197. */
  2198. void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason)
  2199. {
  2200. if (!down_read_trylock(&sb->s_umount))
  2201. return;
  2202. __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true);
  2203. up_read(&sb->s_umount);
  2204. }
  2205. EXPORT_SYMBOL_NS(try_to_writeback_inodes_sb, ANDROID_GKI_VFS_EXPORT_ONLY);
  2206. /**
  2207. * sync_inodes_sb - sync sb inode pages
  2208. * @sb: the superblock
  2209. *
  2210. * This function writes and waits on any dirty inode belonging to this
  2211. * super_block.
  2212. */
  2213. void sync_inodes_sb(struct super_block *sb)
  2214. {
  2215. struct backing_dev_info *bdi = sb->s_bdi;
  2216. DEFINE_WB_COMPLETION(done, bdi);
  2217. struct wb_writeback_work work = {
  2218. .sb = sb,
  2219. .sync_mode = WB_SYNC_ALL,
  2220. .nr_pages = LONG_MAX,
  2221. .range_cyclic = 0,
  2222. .done = &done,
  2223. .reason = WB_REASON_SYNC,
  2224. .for_sync = 1,
  2225. };
  2226. /*
  2227. * Can't skip on !bdi_has_dirty() because we should wait for !dirty
  2228. * inodes under writeback and I_DIRTY_TIME inodes ignored by
  2229. * bdi_has_dirty() need to be written out too.
  2230. */
  2231. if (bdi == &noop_backing_dev_info)
  2232. return;
  2233. WARN_ON(!rwsem_is_locked(&sb->s_umount));
  2234. /* protect against inode wb switch, see inode_switch_wbs_work_fn() */
  2235. bdi_down_write_wb_switch_rwsem(bdi);
  2236. bdi_split_work_to_wbs(bdi, &work, false);
  2237. wb_wait_for_completion(&done);
  2238. bdi_up_write_wb_switch_rwsem(bdi);
  2239. wait_sb_inodes(sb);
  2240. }
  2241. EXPORT_SYMBOL(sync_inodes_sb);
  2242. /**
  2243. * write_inode_now - write an inode to disk
  2244. * @inode: inode to write to disk
  2245. * @sync: whether the write should be synchronous or not
  2246. *
  2247. * This function commits an inode to disk immediately if it is dirty. This is
  2248. * primarily needed by knfsd.
  2249. *
  2250. * The caller must either have a ref on the inode or must have set I_WILL_FREE.
  2251. */
  2252. int write_inode_now(struct inode *inode, int sync)
  2253. {
  2254. struct writeback_control wbc = {
  2255. .nr_to_write = LONG_MAX,
  2256. .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE,
  2257. .range_start = 0,
  2258. .range_end = LLONG_MAX,
  2259. };
  2260. if (!mapping_can_writeback(inode->i_mapping))
  2261. wbc.nr_to_write = 0;
  2262. might_sleep();
  2263. return writeback_single_inode(inode, &wbc);
  2264. }
  2265. EXPORT_SYMBOL_NS(write_inode_now, ANDROID_GKI_VFS_EXPORT_ONLY);
  2266. /**
  2267. * sync_inode - write an inode and its pages to disk.
  2268. * @inode: the inode to sync
  2269. * @wbc: controls the writeback mode
  2270. *
  2271. * sync_inode() will write an inode and its pages to disk. It will also
  2272. * correctly update the inode on its superblock's dirty inode lists and will
  2273. * update inode->i_state.
  2274. *
  2275. * The caller must have a ref on the inode.
  2276. */
  2277. int sync_inode(struct inode *inode, struct writeback_control *wbc)
  2278. {
  2279. return writeback_single_inode(inode, wbc);
  2280. }
  2281. EXPORT_SYMBOL(sync_inode);
  2282. /**
  2283. * sync_inode_metadata - write an inode to disk
  2284. * @inode: the inode to sync
  2285. * @wait: wait for I/O to complete.
  2286. *
  2287. * Write an inode to disk and adjust its dirty state after completion.
  2288. *
  2289. * Note: only writes the actual inode, no associated data or other metadata.
  2290. */
  2291. int sync_inode_metadata(struct inode *inode, int wait)
  2292. {
  2293. struct writeback_control wbc = {
  2294. .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE,
  2295. .nr_to_write = 0, /* metadata-only */
  2296. };
  2297. return sync_inode(inode, &wbc);
  2298. }
  2299. EXPORT_SYMBOL_NS(sync_inode_metadata, ANDROID_GKI_VFS_EXPORT_ONLY);