workqueue.c 106 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890
  1. /*
  2. * kernel/workqueue.c - generic async execution with shared worker pool
  3. *
  4. * Copyright (C) 2002 Ingo Molnar
  5. *
  6. * Derived from the taskqueue/keventd code by:
  7. * David Woodhouse <dwmw2@infradead.org>
  8. * Andrew Morton
  9. * Kai Petzke <wpp@marie.physik.tu-berlin.de>
  10. * Theodore Ts'o <tytso@mit.edu>
  11. *
  12. * Made to use alloc_percpu by Christoph Lameter.
  13. *
  14. * Copyright (C) 2010 SUSE Linux Products GmbH
  15. * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
  16. *
  17. * This is the generic async execution mechanism. Work items as are
  18. * executed in process context. The worker pool is shared and
  19. * automatically managed. There is one worker pool for each CPU and
  20. * one extra for works which are better served by workers which are
  21. * not bound to any specific CPU.
  22. *
  23. * Please read Documentation/workqueue.txt for details.
  24. */
  25. #include <linux/export.h>
  26. #include <linux/kernel.h>
  27. #include <linux/sched.h>
  28. #include <linux/init.h>
  29. #include <linux/signal.h>
  30. #include <linux/completion.h>
  31. #include <linux/workqueue.h>
  32. #include <linux/slab.h>
  33. #include <linux/cpu.h>
  34. #include <linux/notifier.h>
  35. #include <linux/kthread.h>
  36. #include <linux/hardirq.h>
  37. #include <linux/mempolicy.h>
  38. #include <linux/freezer.h>
  39. #include <linux/kallsyms.h>
  40. #include <linux/debug_locks.h>
  41. #include <linux/lockdep.h>
  42. #include <linux/idr.h>
  43. #include "workqueue_sched.h"
  44. enum {
  45. /*
  46. * global_cwq flags
  47. *
  48. * A bound gcwq is either associated or disassociated with its CPU.
  49. * While associated (!DISASSOCIATED), all workers are bound to the
  50. * CPU and none has %WORKER_UNBOUND set and concurrency management
  51. * is in effect.
  52. *
  53. * While DISASSOCIATED, the cpu may be offline and all workers have
  54. * %WORKER_UNBOUND set and concurrency management disabled, and may
  55. * be executing on any CPU. The gcwq behaves as an unbound one.
  56. *
  57. * Note that DISASSOCIATED can be flipped only while holding
  58. * assoc_mutex of all pools on the gcwq to avoid changing binding
  59. * state while create_worker() is in progress.
  60. */
  61. GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
  62. GCWQ_FREEZING = 1 << 1, /* freeze in progress */
  63. /* pool flags */
  64. POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
  65. POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
  66. /* worker flags */
  67. WORKER_STARTED = 1 << 0, /* started */
  68. WORKER_DIE = 1 << 1, /* die die die */
  69. WORKER_IDLE = 1 << 2, /* is idle */
  70. WORKER_PREP = 1 << 3, /* preparing to run works */
  71. WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
  72. WORKER_UNBOUND = 1 << 7, /* worker is unbound */
  73. WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
  74. WORKER_CPU_INTENSIVE,
  75. NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
  76. BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
  77. BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
  78. BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
  79. MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
  80. IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
  81. MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
  82. /* call for help after 10ms
  83. (min two ticks) */
  84. MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
  85. CREATE_COOLDOWN = HZ, /* time to breath after fail */
  86. /*
  87. * Rescue workers are used only on emergencies and shared by
  88. * all cpus. Give -20.
  89. */
  90. RESCUER_NICE_LEVEL = -20,
  91. HIGHPRI_NICE_LEVEL = -20,
  92. };
  93. /*
  94. * Structure fields follow one of the following exclusion rules.
  95. *
  96. * I: Modifiable by initialization/destruction paths and read-only for
  97. * everyone else.
  98. *
  99. * P: Preemption protected. Disabling preemption is enough and should
  100. * only be modified and accessed from the local cpu.
  101. *
  102. * L: gcwq->lock protected. Access with gcwq->lock held.
  103. *
  104. * X: During normal operation, modification requires gcwq->lock and
  105. * should be done only from local cpu. Either disabling preemption
  106. * on local cpu or grabbing gcwq->lock is enough for read access.
  107. * If GCWQ_DISASSOCIATED is set, it's identical to L.
  108. *
  109. * F: wq->flush_mutex protected.
  110. *
  111. * W: workqueue_lock protected.
  112. */
  113. struct global_cwq;
  114. struct worker_pool;
  115. /*
  116. * The poor guys doing the actual heavy lifting. All on-duty workers
  117. * are either serving the manager role, on idle list or on busy hash.
  118. */
  119. struct worker {
  120. /* on idle list while idle, on busy hash table while busy */
  121. union {
  122. struct list_head entry; /* L: while idle */
  123. struct hlist_node hentry; /* L: while busy */
  124. };
  125. struct work_struct *current_work; /* L: work being processed */
  126. struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
  127. struct list_head scheduled; /* L: scheduled works */
  128. struct task_struct *task; /* I: worker task */
  129. struct worker_pool *pool; /* I: the associated pool */
  130. /* 64 bytes boundary on 64bit, 32 on 32bit */
  131. unsigned long last_active; /* L: last active timestamp */
  132. unsigned int flags; /* X: flags */
  133. int id; /* I: worker id */
  134. /* for rebinding worker to CPU */
  135. struct work_struct rebind_work; /* L: for busy worker */
  136. };
  137. struct worker_pool {
  138. struct global_cwq *gcwq; /* I: the owning gcwq */
  139. unsigned int flags; /* X: flags */
  140. struct list_head worklist; /* L: list of pending works */
  141. int nr_workers; /* L: total number of workers */
  142. /* nr_idle includes the ones off idle_list for rebinding */
  143. int nr_idle; /* L: currently idle ones */
  144. struct list_head idle_list; /* X: list of idle workers */
  145. struct timer_list idle_timer; /* L: worker idle timeout */
  146. struct timer_list mayday_timer; /* L: SOS timer for workers */
  147. struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */
  148. struct ida worker_ida; /* L: for worker IDs */
  149. };
  150. /*
  151. * Global per-cpu workqueue. There's one and only one for each cpu
  152. * and all works are queued and processed here regardless of their
  153. * target workqueues.
  154. */
  155. struct global_cwq {
  156. spinlock_t lock; /* the gcwq lock */
  157. unsigned int cpu; /* I: the associated cpu */
  158. unsigned int flags; /* L: GCWQ_* flags */
  159. /* workers are chained either in busy_hash or pool idle_list */
  160. struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
  161. /* L: hash of busy workers */
  162. struct worker_pool pools[NR_WORKER_POOLS];
  163. /* normal and highpri pools */
  164. } ____cacheline_aligned_in_smp;
  165. /*
  166. * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
  167. * work_struct->data are used for flags and thus cwqs need to be
  168. * aligned at two's power of the number of flag bits.
  169. */
  170. struct cpu_workqueue_struct {
  171. struct worker_pool *pool; /* I: the associated pool */
  172. struct workqueue_struct *wq; /* I: the owning workqueue */
  173. int work_color; /* L: current color */
  174. int flush_color; /* L: flushing color */
  175. int nr_in_flight[WORK_NR_COLORS];
  176. /* L: nr of in_flight works */
  177. int nr_active; /* L: nr of active works */
  178. int max_active; /* L: max active works */
  179. struct list_head delayed_works; /* L: delayed works */
  180. };
  181. /*
  182. * Structure used to wait for workqueue flush.
  183. */
  184. struct wq_flusher {
  185. struct list_head list; /* F: list of flushers */
  186. int flush_color; /* F: flush color waiting for */
  187. struct completion done; /* flush completion */
  188. };
  189. /*
  190. * All cpumasks are assumed to be always set on UP and thus can't be
  191. * used to determine whether there's something to be done.
  192. */
  193. #ifdef CONFIG_SMP
  194. typedef cpumask_var_t mayday_mask_t;
  195. #define mayday_test_and_set_cpu(cpu, mask) \
  196. cpumask_test_and_set_cpu((cpu), (mask))
  197. #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
  198. #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
  199. #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
  200. #define free_mayday_mask(mask) free_cpumask_var((mask))
  201. #else
  202. typedef unsigned long mayday_mask_t;
  203. #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
  204. #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
  205. #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
  206. #define alloc_mayday_mask(maskp, gfp) true
  207. #define free_mayday_mask(mask) do { } while (0)
  208. #endif
  209. /*
  210. * The externally visible workqueue abstraction is an array of
  211. * per-CPU workqueues:
  212. */
  213. struct workqueue_struct {
  214. unsigned int flags; /* W: WQ_* flags */
  215. union {
  216. struct cpu_workqueue_struct __percpu *pcpu;
  217. struct cpu_workqueue_struct *single;
  218. unsigned long v;
  219. } cpu_wq; /* I: cwq's */
  220. struct list_head list; /* W: list of all workqueues */
  221. struct mutex flush_mutex; /* protects wq flushing */
  222. int work_color; /* F: current work color */
  223. int flush_color; /* F: current flush color */
  224. atomic_t nr_cwqs_to_flush; /* flush in progress */
  225. struct wq_flusher *first_flusher; /* F: first flusher */
  226. struct list_head flusher_queue; /* F: flush waiters */
  227. struct list_head flusher_overflow; /* F: flush overflow list */
  228. mayday_mask_t mayday_mask; /* cpus requesting rescue */
  229. struct worker *rescuer; /* I: rescue worker */
  230. int nr_drainers; /* W: drain in progress */
  231. int saved_max_active; /* W: saved cwq max_active */
  232. #ifdef CONFIG_LOCKDEP
  233. struct lockdep_map lockdep_map;
  234. #endif
  235. char name[]; /* I: workqueue name */
  236. };
  237. struct workqueue_struct *system_wq __read_mostly;
  238. EXPORT_SYMBOL_GPL(system_wq);
  239. struct workqueue_struct *system_highpri_wq __read_mostly;
  240. EXPORT_SYMBOL_GPL(system_highpri_wq);
  241. struct workqueue_struct *system_long_wq __read_mostly;
  242. EXPORT_SYMBOL_GPL(system_long_wq);
  243. struct workqueue_struct *system_unbound_wq __read_mostly;
  244. EXPORT_SYMBOL_GPL(system_unbound_wq);
  245. struct workqueue_struct *system_freezable_wq __read_mostly;
  246. EXPORT_SYMBOL_GPL(system_freezable_wq);
  247. #define CREATE_TRACE_POINTS
  248. #include <trace/events/workqueue.h>
  249. #define for_each_worker_pool(pool, gcwq) \
  250. for ((pool) = &(gcwq)->pools[0]; \
  251. (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
  252. #define for_each_busy_worker(worker, i, pos, gcwq) \
  253. for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
  254. hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
  255. static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
  256. unsigned int sw)
  257. {
  258. if (cpu < nr_cpu_ids) {
  259. if (sw & 1) {
  260. cpu = cpumask_next(cpu, mask);
  261. if (cpu < nr_cpu_ids)
  262. return cpu;
  263. }
  264. if (sw & 2)
  265. return WORK_CPU_UNBOUND;
  266. }
  267. return WORK_CPU_NONE;
  268. }
  269. static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
  270. struct workqueue_struct *wq)
  271. {
  272. return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
  273. }
  274. /*
  275. * CPU iterators
  276. *
  277. * An extra gcwq is defined for an invalid cpu number
  278. * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
  279. * specific CPU. The following iterators are similar to
  280. * for_each_*_cpu() iterators but also considers the unbound gcwq.
  281. *
  282. * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
  283. * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
  284. * for_each_cwq_cpu() : possible CPUs for bound workqueues,
  285. * WORK_CPU_UNBOUND for unbound workqueues
  286. */
  287. #define for_each_gcwq_cpu(cpu) \
  288. for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
  289. (cpu) < WORK_CPU_NONE; \
  290. (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
  291. #define for_each_online_gcwq_cpu(cpu) \
  292. for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
  293. (cpu) < WORK_CPU_NONE; \
  294. (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
  295. #define for_each_cwq_cpu(cpu, wq) \
  296. for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
  297. (cpu) < WORK_CPU_NONE; \
  298. (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
  299. #ifdef CONFIG_DEBUG_OBJECTS_WORK
  300. static struct debug_obj_descr work_debug_descr;
  301. static void *work_debug_hint(void *addr)
  302. {
  303. return ((struct work_struct *) addr)->func;
  304. }
  305. /*
  306. * fixup_init is called when:
  307. * - an active object is initialized
  308. */
  309. static int work_fixup_init(void *addr, enum debug_obj_state state)
  310. {
  311. struct work_struct *work = addr;
  312. switch (state) {
  313. case ODEBUG_STATE_ACTIVE:
  314. cancel_work_sync(work);
  315. debug_object_init(work, &work_debug_descr);
  316. return 1;
  317. default:
  318. return 0;
  319. }
  320. }
  321. /*
  322. * fixup_activate is called when:
  323. * - an active object is activated
  324. * - an unknown object is activated (might be a statically initialized object)
  325. */
  326. static int work_fixup_activate(void *addr, enum debug_obj_state state)
  327. {
  328. struct work_struct *work = addr;
  329. switch (state) {
  330. case ODEBUG_STATE_NOTAVAILABLE:
  331. /*
  332. * This is not really a fixup. The work struct was
  333. * statically initialized. We just make sure that it
  334. * is tracked in the object tracker.
  335. */
  336. if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
  337. debug_object_init(work, &work_debug_descr);
  338. debug_object_activate(work, &work_debug_descr);
  339. return 0;
  340. }
  341. WARN_ON_ONCE(1);
  342. return 0;
  343. case ODEBUG_STATE_ACTIVE:
  344. WARN_ON(1);
  345. default:
  346. return 0;
  347. }
  348. }
  349. /*
  350. * fixup_free is called when:
  351. * - an active object is freed
  352. */
  353. static int work_fixup_free(void *addr, enum debug_obj_state state)
  354. {
  355. struct work_struct *work = addr;
  356. switch (state) {
  357. case ODEBUG_STATE_ACTIVE:
  358. cancel_work_sync(work);
  359. debug_object_free(work, &work_debug_descr);
  360. return 1;
  361. default:
  362. return 0;
  363. }
  364. }
  365. static struct debug_obj_descr work_debug_descr = {
  366. .name = "work_struct",
  367. .debug_hint = work_debug_hint,
  368. .fixup_init = work_fixup_init,
  369. .fixup_activate = work_fixup_activate,
  370. .fixup_free = work_fixup_free,
  371. };
  372. static inline void debug_work_activate(struct work_struct *work)
  373. {
  374. debug_object_activate(work, &work_debug_descr);
  375. }
  376. static inline void debug_work_deactivate(struct work_struct *work)
  377. {
  378. debug_object_deactivate(work, &work_debug_descr);
  379. }
  380. void __init_work(struct work_struct *work, int onstack)
  381. {
  382. if (onstack)
  383. debug_object_init_on_stack(work, &work_debug_descr);
  384. else
  385. debug_object_init(work, &work_debug_descr);
  386. }
  387. EXPORT_SYMBOL_GPL(__init_work);
  388. void destroy_work_on_stack(struct work_struct *work)
  389. {
  390. debug_object_free(work, &work_debug_descr);
  391. }
  392. EXPORT_SYMBOL_GPL(destroy_work_on_stack);
  393. #else
  394. static inline void debug_work_activate(struct work_struct *work) { }
  395. static inline void debug_work_deactivate(struct work_struct *work) { }
  396. #endif
  397. /* Serializes the accesses to the list of workqueues. */
  398. static DEFINE_SPINLOCK(workqueue_lock);
  399. static LIST_HEAD(workqueues);
  400. static bool workqueue_freezing; /* W: have wqs started freezing? */
  401. /*
  402. * The almighty global cpu workqueues. nr_running is the only field
  403. * which is expected to be used frequently by other cpus via
  404. * try_to_wake_up(). Put it in a separate cacheline.
  405. */
  406. static DEFINE_PER_CPU(struct global_cwq, global_cwq);
  407. static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
  408. /*
  409. * Global cpu workqueue and nr_running counter for unbound gcwq. The
  410. * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
  411. * workers have WORKER_UNBOUND set.
  412. */
  413. static struct global_cwq unbound_global_cwq;
  414. static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
  415. [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
  416. };
  417. static int worker_thread(void *__worker);
  418. static int worker_pool_pri(struct worker_pool *pool)
  419. {
  420. return pool - pool->gcwq->pools;
  421. }
  422. static struct global_cwq *get_gcwq(unsigned int cpu)
  423. {
  424. if (cpu != WORK_CPU_UNBOUND)
  425. return &per_cpu(global_cwq, cpu);
  426. else
  427. return &unbound_global_cwq;
  428. }
  429. static atomic_t *get_pool_nr_running(struct worker_pool *pool)
  430. {
  431. int cpu = pool->gcwq->cpu;
  432. int idx = worker_pool_pri(pool);
  433. if (cpu != WORK_CPU_UNBOUND)
  434. return &per_cpu(pool_nr_running, cpu)[idx];
  435. else
  436. return &unbound_pool_nr_running[idx];
  437. }
  438. static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
  439. struct workqueue_struct *wq)
  440. {
  441. if (!(wq->flags & WQ_UNBOUND)) {
  442. if (likely(cpu < nr_cpu_ids))
  443. return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
  444. } else if (likely(cpu == WORK_CPU_UNBOUND))
  445. return wq->cpu_wq.single;
  446. return NULL;
  447. }
  448. static unsigned int work_color_to_flags(int color)
  449. {
  450. return color << WORK_STRUCT_COLOR_SHIFT;
  451. }
  452. static int get_work_color(struct work_struct *work)
  453. {
  454. return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
  455. ((1 << WORK_STRUCT_COLOR_BITS) - 1);
  456. }
  457. static int work_next_color(int color)
  458. {
  459. return (color + 1) % WORK_NR_COLORS;
  460. }
  461. /*
  462. * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
  463. * contain the pointer to the queued cwq. Once execution starts, the flag
  464. * is cleared and the high bits contain OFFQ flags and CPU number.
  465. *
  466. * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
  467. * and clear_work_data() can be used to set the cwq, cpu or clear
  468. * work->data. These functions should only be called while the work is
  469. * owned - ie. while the PENDING bit is set.
  470. *
  471. * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
  472. * a work. gcwq is available once the work has been queued anywhere after
  473. * initialization until it is sync canceled. cwq is available only while
  474. * the work item is queued.
  475. *
  476. * %WORK_OFFQ_CANCELING is used to mark a work item which is being
  477. * canceled. While being canceled, a work item may have its PENDING set
  478. * but stay off timer and worklist for arbitrarily long and nobody should
  479. * try to steal the PENDING bit.
  480. */
  481. static inline void set_work_data(struct work_struct *work, unsigned long data,
  482. unsigned long flags)
  483. {
  484. BUG_ON(!work_pending(work));
  485. atomic_long_set(&work->data, data | flags | work_static(work));
  486. }
  487. static void set_work_cwq(struct work_struct *work,
  488. struct cpu_workqueue_struct *cwq,
  489. unsigned long extra_flags)
  490. {
  491. set_work_data(work, (unsigned long)cwq,
  492. WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
  493. }
  494. static void set_work_cpu_and_clear_pending(struct work_struct *work,
  495. unsigned int cpu)
  496. {
  497. /*
  498. * The following wmb is paired with the implied mb in
  499. * test_and_set_bit(PENDING) and ensures all updates to @work made
  500. * here are visible to and precede any updates by the next PENDING
  501. * owner.
  502. */
  503. smp_wmb();
  504. set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
  505. }
  506. static void clear_work_data(struct work_struct *work)
  507. {
  508. smp_wmb(); /* see set_work_cpu_and_clear_pending() */
  509. set_work_data(work, WORK_STRUCT_NO_CPU, 0);
  510. }
  511. static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
  512. {
  513. unsigned long data = atomic_long_read(&work->data);
  514. if (data & WORK_STRUCT_CWQ)
  515. return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
  516. else
  517. return NULL;
  518. }
  519. static struct global_cwq *get_work_gcwq(struct work_struct *work)
  520. {
  521. unsigned long data = atomic_long_read(&work->data);
  522. unsigned int cpu;
  523. if (data & WORK_STRUCT_CWQ)
  524. return ((struct cpu_workqueue_struct *)
  525. (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
  526. cpu = data >> WORK_OFFQ_CPU_SHIFT;
  527. if (cpu == WORK_CPU_NONE)
  528. return NULL;
  529. BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
  530. return get_gcwq(cpu);
  531. }
  532. static void mark_work_canceling(struct work_struct *work)
  533. {
  534. struct global_cwq *gcwq = get_work_gcwq(work);
  535. unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
  536. set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
  537. WORK_STRUCT_PENDING);
  538. }
  539. static bool work_is_canceling(struct work_struct *work)
  540. {
  541. unsigned long data = atomic_long_read(&work->data);
  542. return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
  543. }
  544. /*
  545. * Policy functions. These define the policies on how the global worker
  546. * pools are managed. Unless noted otherwise, these functions assume that
  547. * they're being called with gcwq->lock held.
  548. */
  549. static bool __need_more_worker(struct worker_pool *pool)
  550. {
  551. return !atomic_read(get_pool_nr_running(pool));
  552. }
  553. /*
  554. * Need to wake up a worker? Called from anything but currently
  555. * running workers.
  556. *
  557. * Note that, because unbound workers never contribute to nr_running, this
  558. * function will always return %true for unbound gcwq as long as the
  559. * worklist isn't empty.
  560. */
  561. static bool need_more_worker(struct worker_pool *pool)
  562. {
  563. return !list_empty(&pool->worklist) && __need_more_worker(pool);
  564. }
  565. /* Can I start working? Called from busy but !running workers. */
  566. static bool may_start_working(struct worker_pool *pool)
  567. {
  568. return pool->nr_idle;
  569. }
  570. /* Do I need to keep working? Called from currently running workers. */
  571. static bool keep_working(struct worker_pool *pool)
  572. {
  573. atomic_t *nr_running = get_pool_nr_running(pool);
  574. return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
  575. }
  576. /* Do we need a new worker? Called from manager. */
  577. static bool need_to_create_worker(struct worker_pool *pool)
  578. {
  579. return need_more_worker(pool) && !may_start_working(pool);
  580. }
  581. /* Do I need to be the manager? */
  582. static bool need_to_manage_workers(struct worker_pool *pool)
  583. {
  584. return need_to_create_worker(pool) ||
  585. (pool->flags & POOL_MANAGE_WORKERS);
  586. }
  587. /* Do we have too many workers and should some go away? */
  588. static bool too_many_workers(struct worker_pool *pool)
  589. {
  590. bool managing = pool->flags & POOL_MANAGING_WORKERS;
  591. int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
  592. int nr_busy = pool->nr_workers - nr_idle;
  593. /*
  594. * nr_idle and idle_list may disagree if idle rebinding is in
  595. * progress. Never return %true if idle_list is empty.
  596. */
  597. if (list_empty(&pool->idle_list))
  598. return false;
  599. return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
  600. }
  601. /*
  602. * Wake up functions.
  603. */
  604. /* Return the first worker. Safe with preemption disabled */
  605. static struct worker *first_worker(struct worker_pool *pool)
  606. {
  607. if (unlikely(list_empty(&pool->idle_list)))
  608. return NULL;
  609. return list_first_entry(&pool->idle_list, struct worker, entry);
  610. }
  611. /**
  612. * wake_up_worker - wake up an idle worker
  613. * @pool: worker pool to wake worker from
  614. *
  615. * Wake up the first idle worker of @pool.
  616. *
  617. * CONTEXT:
  618. * spin_lock_irq(gcwq->lock).
  619. */
  620. static void wake_up_worker(struct worker_pool *pool)
  621. {
  622. struct worker *worker = first_worker(pool);
  623. if (likely(worker))
  624. wake_up_process(worker->task);
  625. }
  626. /**
  627. * wq_worker_waking_up - a worker is waking up
  628. * @task: task waking up
  629. * @cpu: CPU @task is waking up to
  630. *
  631. * This function is called during try_to_wake_up() when a worker is
  632. * being awoken.
  633. *
  634. * CONTEXT:
  635. * spin_lock_irq(rq->lock)
  636. */
  637. void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
  638. {
  639. struct worker *worker = kthread_data(task);
  640. if (!(worker->flags & WORKER_NOT_RUNNING))
  641. atomic_inc(get_pool_nr_running(worker->pool));
  642. }
  643. /**
  644. * wq_worker_sleeping - a worker is going to sleep
  645. * @task: task going to sleep
  646. * @cpu: CPU in question, must be the current CPU number
  647. *
  648. * This function is called during schedule() when a busy worker is
  649. * going to sleep. Worker on the same cpu can be woken up by
  650. * returning pointer to its task.
  651. *
  652. * CONTEXT:
  653. * spin_lock_irq(rq->lock)
  654. *
  655. * RETURNS:
  656. * Worker task on @cpu to wake up, %NULL if none.
  657. */
  658. struct task_struct *wq_worker_sleeping(struct task_struct *task,
  659. unsigned int cpu)
  660. {
  661. struct worker *worker = kthread_data(task), *to_wakeup = NULL;
  662. struct worker_pool *pool = worker->pool;
  663. atomic_t *nr_running = get_pool_nr_running(pool);
  664. if (worker->flags & WORKER_NOT_RUNNING)
  665. return NULL;
  666. /* this can only happen on the local cpu */
  667. BUG_ON(cpu != raw_smp_processor_id());
  668. /*
  669. * The counterpart of the following dec_and_test, implied mb,
  670. * worklist not empty test sequence is in insert_work().
  671. * Please read comment there.
  672. *
  673. * NOT_RUNNING is clear. This means that we're bound to and
  674. * running on the local cpu w/ rq lock held and preemption
  675. * disabled, which in turn means that none else could be
  676. * manipulating idle_list, so dereferencing idle_list without gcwq
  677. * lock is safe.
  678. */
  679. if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
  680. to_wakeup = first_worker(pool);
  681. return to_wakeup ? to_wakeup->task : NULL;
  682. }
  683. /**
  684. * worker_set_flags - set worker flags and adjust nr_running accordingly
  685. * @worker: self
  686. * @flags: flags to set
  687. * @wakeup: wakeup an idle worker if necessary
  688. *
  689. * Set @flags in @worker->flags and adjust nr_running accordingly. If
  690. * nr_running becomes zero and @wakeup is %true, an idle worker is
  691. * woken up.
  692. *
  693. * CONTEXT:
  694. * spin_lock_irq(gcwq->lock)
  695. */
  696. static inline void worker_set_flags(struct worker *worker, unsigned int flags,
  697. bool wakeup)
  698. {
  699. struct worker_pool *pool = worker->pool;
  700. WARN_ON_ONCE(worker->task != current);
  701. /*
  702. * If transitioning into NOT_RUNNING, adjust nr_running and
  703. * wake up an idle worker as necessary if requested by
  704. * @wakeup.
  705. */
  706. if ((flags & WORKER_NOT_RUNNING) &&
  707. !(worker->flags & WORKER_NOT_RUNNING)) {
  708. atomic_t *nr_running = get_pool_nr_running(pool);
  709. if (wakeup) {
  710. if (atomic_dec_and_test(nr_running) &&
  711. !list_empty(&pool->worklist))
  712. wake_up_worker(pool);
  713. } else
  714. atomic_dec(nr_running);
  715. }
  716. worker->flags |= flags;
  717. }
  718. /**
  719. * worker_clr_flags - clear worker flags and adjust nr_running accordingly
  720. * @worker: self
  721. * @flags: flags to clear
  722. *
  723. * Clear @flags in @worker->flags and adjust nr_running accordingly.
  724. *
  725. * CONTEXT:
  726. * spin_lock_irq(gcwq->lock)
  727. */
  728. static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
  729. {
  730. struct worker_pool *pool = worker->pool;
  731. unsigned int oflags = worker->flags;
  732. WARN_ON_ONCE(worker->task != current);
  733. worker->flags &= ~flags;
  734. /*
  735. * If transitioning out of NOT_RUNNING, increment nr_running. Note
  736. * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
  737. * of multiple flags, not a single flag.
  738. */
  739. if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
  740. if (!(worker->flags & WORKER_NOT_RUNNING))
  741. atomic_inc(get_pool_nr_running(pool));
  742. }
  743. /**
  744. * busy_worker_head - return the busy hash head for a work
  745. * @gcwq: gcwq of interest
  746. * @work: work to be hashed
  747. *
  748. * Return hash head of @gcwq for @work.
  749. *
  750. * CONTEXT:
  751. * spin_lock_irq(gcwq->lock).
  752. *
  753. * RETURNS:
  754. * Pointer to the hash head.
  755. */
  756. static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
  757. struct work_struct *work)
  758. {
  759. const int base_shift = ilog2(sizeof(struct work_struct));
  760. unsigned long v = (unsigned long)work;
  761. /* simple shift and fold hash, do we need something better? */
  762. v >>= base_shift;
  763. v += v >> BUSY_WORKER_HASH_ORDER;
  764. v &= BUSY_WORKER_HASH_MASK;
  765. return &gcwq->busy_hash[v];
  766. }
  767. /**
  768. * __find_worker_executing_work - find worker which is executing a work
  769. * @gcwq: gcwq of interest
  770. * @bwh: hash head as returned by busy_worker_head()
  771. * @work: work to find worker for
  772. *
  773. * Find a worker which is executing @work on @gcwq. @bwh should be
  774. * the hash head obtained by calling busy_worker_head() with the same
  775. * work.
  776. *
  777. * CONTEXT:
  778. * spin_lock_irq(gcwq->lock).
  779. *
  780. * RETURNS:
  781. * Pointer to worker which is executing @work if found, NULL
  782. * otherwise.
  783. */
  784. static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
  785. struct hlist_head *bwh,
  786. struct work_struct *work)
  787. {
  788. struct worker *worker;
  789. struct hlist_node *tmp;
  790. hlist_for_each_entry(worker, tmp, bwh, hentry)
  791. if (worker->current_work == work)
  792. return worker;
  793. return NULL;
  794. }
  795. /**
  796. * find_worker_executing_work - find worker which is executing a work
  797. * @gcwq: gcwq of interest
  798. * @work: work to find worker for
  799. *
  800. * Find a worker which is executing @work on @gcwq. This function is
  801. * identical to __find_worker_executing_work() except that this
  802. * function calculates @bwh itself.
  803. *
  804. * CONTEXT:
  805. * spin_lock_irq(gcwq->lock).
  806. *
  807. * RETURNS:
  808. * Pointer to worker which is executing @work if found, NULL
  809. * otherwise.
  810. */
  811. static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
  812. struct work_struct *work)
  813. {
  814. return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
  815. work);
  816. }
  817. /**
  818. * move_linked_works - move linked works to a list
  819. * @work: start of series of works to be scheduled
  820. * @head: target list to append @work to
  821. * @nextp: out paramter for nested worklist walking
  822. *
  823. * Schedule linked works starting from @work to @head. Work series to
  824. * be scheduled starts at @work and includes any consecutive work with
  825. * WORK_STRUCT_LINKED set in its predecessor.
  826. *
  827. * If @nextp is not NULL, it's updated to point to the next work of
  828. * the last scheduled work. This allows move_linked_works() to be
  829. * nested inside outer list_for_each_entry_safe().
  830. *
  831. * CONTEXT:
  832. * spin_lock_irq(gcwq->lock).
  833. */
  834. static void move_linked_works(struct work_struct *work, struct list_head *head,
  835. struct work_struct **nextp)
  836. {
  837. struct work_struct *n;
  838. /*
  839. * Linked worklist will always end before the end of the list,
  840. * use NULL for list head.
  841. */
  842. list_for_each_entry_safe_from(work, n, NULL, entry) {
  843. list_move_tail(&work->entry, head);
  844. if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
  845. break;
  846. }
  847. /*
  848. * If we're already inside safe list traversal and have moved
  849. * multiple works to the scheduled queue, the next position
  850. * needs to be updated.
  851. */
  852. if (nextp)
  853. *nextp = n;
  854. }
  855. static void cwq_activate_delayed_work(struct work_struct *work)
  856. {
  857. struct cpu_workqueue_struct *cwq = get_work_cwq(work);
  858. trace_workqueue_activate_work(work);
  859. move_linked_works(work, &cwq->pool->worklist, NULL);
  860. __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
  861. cwq->nr_active++;
  862. }
  863. static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
  864. {
  865. struct work_struct *work = list_first_entry(&cwq->delayed_works,
  866. struct work_struct, entry);
  867. cwq_activate_delayed_work(work);
  868. }
  869. /**
  870. * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
  871. * @cwq: cwq of interest
  872. * @color: color of work which left the queue
  873. *
  874. * A work either has completed or is removed from pending queue,
  875. * decrement nr_in_flight of its cwq and handle workqueue flushing.
  876. *
  877. * CONTEXT:
  878. * spin_lock_irq(gcwq->lock).
  879. */
  880. static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
  881. {
  882. /* ignore uncolored works */
  883. if (color == WORK_NO_COLOR)
  884. return;
  885. cwq->nr_in_flight[color]--;
  886. cwq->nr_active--;
  887. if (!list_empty(&cwq->delayed_works)) {
  888. /* one down, submit a delayed one */
  889. if (cwq->nr_active < cwq->max_active)
  890. cwq_activate_first_delayed(cwq);
  891. }
  892. /* is flush in progress and are we at the flushing tip? */
  893. if (likely(cwq->flush_color != color))
  894. return;
  895. /* are there still in-flight works? */
  896. if (cwq->nr_in_flight[color])
  897. return;
  898. /* this cwq is done, clear flush_color */
  899. cwq->flush_color = -1;
  900. /*
  901. * If this was the last cwq, wake up the first flusher. It
  902. * will handle the rest.
  903. */
  904. if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
  905. complete(&cwq->wq->first_flusher->done);
  906. }
  907. /**
  908. * try_to_grab_pending - steal work item from worklist and disable irq
  909. * @work: work item to steal
  910. * @is_dwork: @work is a delayed_work
  911. * @flags: place to store irq state
  912. *
  913. * Try to grab PENDING bit of @work. This function can handle @work in any
  914. * stable state - idle, on timer or on worklist. Return values are
  915. *
  916. * 1 if @work was pending and we successfully stole PENDING
  917. * 0 if @work was idle and we claimed PENDING
  918. * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
  919. * -ENOENT if someone else is canceling @work, this state may persist
  920. * for arbitrarily long
  921. *
  922. * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
  923. * interrupted while holding PENDING and @work off queue, irq must be
  924. * disabled on entry. This, combined with delayed_work->timer being
  925. * irqsafe, ensures that we return -EAGAIN for finite short period of time.
  926. *
  927. * On successful return, >= 0, irq is disabled and the caller is
  928. * responsible for releasing it using local_irq_restore(*@flags).
  929. *
  930. * This function is safe to call from any context including IRQ handler.
  931. */
  932. static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
  933. unsigned long *flags)
  934. {
  935. struct global_cwq *gcwq;
  936. local_irq_save(*flags);
  937. /* try to steal the timer if it exists */
  938. if (is_dwork) {
  939. struct delayed_work *dwork = to_delayed_work(work);
  940. /*
  941. * dwork->timer is irqsafe. If del_timer() fails, it's
  942. * guaranteed that the timer is not queued anywhere and not
  943. * running on the local CPU.
  944. */
  945. if (likely(del_timer(&dwork->timer)))
  946. return 1;
  947. }
  948. /* try to claim PENDING the normal way */
  949. if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
  950. return 0;
  951. /*
  952. * The queueing is in progress, or it is already queued. Try to
  953. * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
  954. */
  955. gcwq = get_work_gcwq(work);
  956. if (!gcwq)
  957. goto fail;
  958. spin_lock(&gcwq->lock);
  959. if (!list_empty(&work->entry)) {
  960. /*
  961. * This work is queued, but perhaps we locked the wrong gcwq.
  962. * In that case we must see the new value after rmb(), see
  963. * insert_work()->wmb().
  964. */
  965. smp_rmb();
  966. if (gcwq == get_work_gcwq(work)) {
  967. debug_work_deactivate(work);
  968. /*
  969. * A delayed work item cannot be grabbed directly
  970. * because it might have linked NO_COLOR work items
  971. * which, if left on the delayed_list, will confuse
  972. * cwq->nr_active management later on and cause
  973. * stall. Make sure the work item is activated
  974. * before grabbing.
  975. */
  976. if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
  977. cwq_activate_delayed_work(work);
  978. list_del_init(&work->entry);
  979. cwq_dec_nr_in_flight(get_work_cwq(work),
  980. get_work_color(work));
  981. spin_unlock(&gcwq->lock);
  982. return 1;
  983. }
  984. }
  985. spin_unlock(&gcwq->lock);
  986. fail:
  987. local_irq_restore(*flags);
  988. if (work_is_canceling(work))
  989. return -ENOENT;
  990. cpu_relax();
  991. return -EAGAIN;
  992. }
  993. /**
  994. * insert_work - insert a work into gcwq
  995. * @cwq: cwq @work belongs to
  996. * @work: work to insert
  997. * @head: insertion point
  998. * @extra_flags: extra WORK_STRUCT_* flags to set
  999. *
  1000. * Insert @work which belongs to @cwq into @gcwq after @head.
  1001. * @extra_flags is or'd to work_struct flags.
  1002. *
  1003. * CONTEXT:
  1004. * spin_lock_irq(gcwq->lock).
  1005. */
  1006. static void insert_work(struct cpu_workqueue_struct *cwq,
  1007. struct work_struct *work, struct list_head *head,
  1008. unsigned int extra_flags)
  1009. {
  1010. struct worker_pool *pool = cwq->pool;
  1011. /* we own @work, set data and link */
  1012. set_work_cwq(work, cwq, extra_flags);
  1013. /*
  1014. * Ensure that we get the right work->data if we see the
  1015. * result of list_add() below, see try_to_grab_pending().
  1016. */
  1017. smp_wmb();
  1018. list_add_tail(&work->entry, head);
  1019. /*
  1020. * Ensure either worker_sched_deactivated() sees the above
  1021. * list_add_tail() or we see zero nr_running to avoid workers
  1022. * lying around lazily while there are works to be processed.
  1023. */
  1024. smp_mb();
  1025. if (__need_more_worker(pool))
  1026. wake_up_worker(pool);
  1027. }
  1028. /*
  1029. * Test whether @work is being queued from another work executing on the
  1030. * same workqueue. This is rather expensive and should only be used from
  1031. * cold paths.
  1032. */
  1033. static bool is_chained_work(struct workqueue_struct *wq)
  1034. {
  1035. unsigned long flags;
  1036. unsigned int cpu;
  1037. for_each_gcwq_cpu(cpu) {
  1038. struct global_cwq *gcwq = get_gcwq(cpu);
  1039. struct worker *worker;
  1040. struct hlist_node *pos;
  1041. int i;
  1042. spin_lock_irqsave(&gcwq->lock, flags);
  1043. for_each_busy_worker(worker, i, pos, gcwq) {
  1044. if (worker->task != current)
  1045. continue;
  1046. spin_unlock_irqrestore(&gcwq->lock, flags);
  1047. /*
  1048. * I'm @worker, no locking necessary. See if @work
  1049. * is headed to the same workqueue.
  1050. */
  1051. return worker->current_cwq->wq == wq;
  1052. }
  1053. spin_unlock_irqrestore(&gcwq->lock, flags);
  1054. }
  1055. return false;
  1056. }
  1057. static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
  1058. struct work_struct *work)
  1059. {
  1060. struct global_cwq *gcwq;
  1061. struct cpu_workqueue_struct *cwq;
  1062. struct list_head *worklist;
  1063. unsigned int work_flags;
  1064. unsigned int req_cpu = cpu;
  1065. /*
  1066. * While a work item is PENDING && off queue, a task trying to
  1067. * steal the PENDING will busy-loop waiting for it to either get
  1068. * queued or lose PENDING. Grabbing PENDING and queueing should
  1069. * happen with IRQ disabled.
  1070. */
  1071. WARN_ON_ONCE(!irqs_disabled());
  1072. debug_work_activate(work);
  1073. /* if dying, only works from the same workqueue are allowed */
  1074. if (unlikely(wq->flags & WQ_DRAINING) &&
  1075. WARN_ON_ONCE(!is_chained_work(wq)))
  1076. return;
  1077. /* determine gcwq to use */
  1078. if (!(wq->flags & WQ_UNBOUND)) {
  1079. struct global_cwq *last_gcwq;
  1080. if (cpu == WORK_CPU_UNBOUND)
  1081. cpu = raw_smp_processor_id();
  1082. /*
  1083. * It's multi cpu. If @work was previously on a different
  1084. * cpu, it might still be running there, in which case the
  1085. * work needs to be queued on that cpu to guarantee
  1086. * non-reentrancy.
  1087. */
  1088. gcwq = get_gcwq(cpu);
  1089. last_gcwq = get_work_gcwq(work);
  1090. if (last_gcwq && last_gcwq != gcwq) {
  1091. struct worker *worker;
  1092. spin_lock(&last_gcwq->lock);
  1093. worker = find_worker_executing_work(last_gcwq, work);
  1094. if (worker && worker->current_cwq->wq == wq)
  1095. gcwq = last_gcwq;
  1096. else {
  1097. /* meh... not running there, queue here */
  1098. spin_unlock(&last_gcwq->lock);
  1099. spin_lock(&gcwq->lock);
  1100. }
  1101. } else {
  1102. spin_lock(&gcwq->lock);
  1103. }
  1104. } else {
  1105. gcwq = get_gcwq(WORK_CPU_UNBOUND);
  1106. spin_lock(&gcwq->lock);
  1107. }
  1108. /* gcwq determined, get cwq and queue */
  1109. cwq = get_cwq(gcwq->cpu, wq);
  1110. trace_workqueue_queue_work(req_cpu, cwq, work);
  1111. if (WARN_ON(!list_empty(&work->entry))) {
  1112. spin_unlock(&gcwq->lock);
  1113. return;
  1114. }
  1115. cwq->nr_in_flight[cwq->work_color]++;
  1116. work_flags = work_color_to_flags(cwq->work_color);
  1117. if (likely(cwq->nr_active < cwq->max_active)) {
  1118. trace_workqueue_activate_work(work);
  1119. cwq->nr_active++;
  1120. worklist = &cwq->pool->worklist;
  1121. } else {
  1122. work_flags |= WORK_STRUCT_DELAYED;
  1123. worklist = &cwq->delayed_works;
  1124. }
  1125. insert_work(cwq, work, worklist, work_flags);
  1126. spin_unlock(&gcwq->lock);
  1127. }
  1128. /**
  1129. * queue_work_on - queue work on specific cpu
  1130. * @cpu: CPU number to execute work on
  1131. * @wq: workqueue to use
  1132. * @work: work to queue
  1133. *
  1134. * Returns %false if @work was already on a queue, %true otherwise.
  1135. *
  1136. * We queue the work to a specific CPU, the caller must ensure it
  1137. * can't go away.
  1138. */
  1139. bool queue_work_on(int cpu, struct workqueue_struct *wq,
  1140. struct work_struct *work)
  1141. {
  1142. bool ret = false;
  1143. unsigned long flags;
  1144. local_irq_save(flags);
  1145. if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
  1146. __queue_work(cpu, wq, work);
  1147. ret = true;
  1148. }
  1149. local_irq_restore(flags);
  1150. return ret;
  1151. }
  1152. EXPORT_SYMBOL_GPL(queue_work_on);
  1153. /**
  1154. * queue_work - queue work on a workqueue
  1155. * @wq: workqueue to use
  1156. * @work: work to queue
  1157. *
  1158. * Returns %false if @work was already on a queue, %true otherwise.
  1159. *
  1160. * We queue the work to the CPU on which it was submitted, but if the CPU dies
  1161. * it can be processed by another CPU.
  1162. */
  1163. bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
  1164. {
  1165. return queue_work_on(WORK_CPU_UNBOUND, wq, work);
  1166. }
  1167. EXPORT_SYMBOL_GPL(queue_work);
  1168. void delayed_work_timer_fn(unsigned long __data)
  1169. {
  1170. struct delayed_work *dwork = (struct delayed_work *)__data;
  1171. struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
  1172. /* should have been called from irqsafe timer with irq already off */
  1173. __queue_work(dwork->cpu, cwq->wq, &dwork->work);
  1174. }
  1175. EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
  1176. static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
  1177. struct delayed_work *dwork, unsigned long delay)
  1178. {
  1179. struct timer_list *timer = &dwork->timer;
  1180. struct work_struct *work = &dwork->work;
  1181. unsigned int lcpu;
  1182. WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
  1183. timer->data != (unsigned long)dwork);
  1184. BUG_ON(timer_pending(timer));
  1185. BUG_ON(!list_empty(&work->entry));
  1186. timer_stats_timer_set_start_info(&dwork->timer);
  1187. /*
  1188. * This stores cwq for the moment, for the timer_fn. Note that the
  1189. * work's gcwq is preserved to allow reentrance detection for
  1190. * delayed works.
  1191. */
  1192. if (!(wq->flags & WQ_UNBOUND)) {
  1193. struct global_cwq *gcwq = get_work_gcwq(work);
  1194. /*
  1195. * If we cannot get the last gcwq from @work directly,
  1196. * select the last CPU such that it avoids unnecessarily
  1197. * triggering non-reentrancy check in __queue_work().
  1198. */
  1199. lcpu = cpu;
  1200. if (gcwq)
  1201. lcpu = gcwq->cpu;
  1202. if (lcpu == WORK_CPU_UNBOUND)
  1203. lcpu = raw_smp_processor_id();
  1204. } else {
  1205. lcpu = WORK_CPU_UNBOUND;
  1206. }
  1207. set_work_cwq(work, get_cwq(lcpu, wq), 0);
  1208. dwork->cpu = cpu;
  1209. timer->expires = jiffies + delay;
  1210. if (unlikely(cpu != WORK_CPU_UNBOUND))
  1211. add_timer_on(timer, cpu);
  1212. else
  1213. add_timer(timer);
  1214. }
  1215. /**
  1216. * queue_delayed_work_on - queue work on specific CPU after delay
  1217. * @cpu: CPU number to execute work on
  1218. * @wq: workqueue to use
  1219. * @dwork: work to queue
  1220. * @delay: number of jiffies to wait before queueing
  1221. *
  1222. * Returns %false if @work was already on a queue, %true otherwise. If
  1223. * @delay is zero and @dwork is idle, it will be scheduled for immediate
  1224. * execution.
  1225. */
  1226. bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
  1227. struct delayed_work *dwork, unsigned long delay)
  1228. {
  1229. struct work_struct *work = &dwork->work;
  1230. bool ret = false;
  1231. unsigned long flags;
  1232. if (!delay)
  1233. return queue_work_on(cpu, wq, &dwork->work);
  1234. /* read the comment in __queue_work() */
  1235. local_irq_save(flags);
  1236. if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
  1237. __queue_delayed_work(cpu, wq, dwork, delay);
  1238. ret = true;
  1239. }
  1240. local_irq_restore(flags);
  1241. return ret;
  1242. }
  1243. EXPORT_SYMBOL_GPL(queue_delayed_work_on);
  1244. /**
  1245. * queue_delayed_work - queue work on a workqueue after delay
  1246. * @wq: workqueue to use
  1247. * @dwork: delayable work to queue
  1248. * @delay: number of jiffies to wait before queueing
  1249. *
  1250. * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
  1251. */
  1252. bool queue_delayed_work(struct workqueue_struct *wq,
  1253. struct delayed_work *dwork, unsigned long delay)
  1254. {
  1255. return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
  1256. }
  1257. EXPORT_SYMBOL_GPL(queue_delayed_work);
  1258. /**
  1259. * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
  1260. * @cpu: CPU number to execute work on
  1261. * @wq: workqueue to use
  1262. * @dwork: work to queue
  1263. * @delay: number of jiffies to wait before queueing
  1264. *
  1265. * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
  1266. * modify @dwork's timer so that it expires after @delay. If @delay is
  1267. * zero, @work is guaranteed to be scheduled immediately regardless of its
  1268. * current state.
  1269. *
  1270. * Returns %false if @dwork was idle and queued, %true if @dwork was
  1271. * pending and its timer was modified.
  1272. *
  1273. * This function is safe to call from any context including IRQ handler.
  1274. * See try_to_grab_pending() for details.
  1275. */
  1276. bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
  1277. struct delayed_work *dwork, unsigned long delay)
  1278. {
  1279. unsigned long flags;
  1280. int ret;
  1281. do {
  1282. ret = try_to_grab_pending(&dwork->work, true, &flags);
  1283. } while (unlikely(ret == -EAGAIN));
  1284. if (likely(ret >= 0)) {
  1285. __queue_delayed_work(cpu, wq, dwork, delay);
  1286. local_irq_restore(flags);
  1287. }
  1288. /* -ENOENT from try_to_grab_pending() becomes %true */
  1289. return ret;
  1290. }
  1291. EXPORT_SYMBOL_GPL(mod_delayed_work_on);
  1292. /**
  1293. * mod_delayed_work - modify delay of or queue a delayed work
  1294. * @wq: workqueue to use
  1295. * @dwork: work to queue
  1296. * @delay: number of jiffies to wait before queueing
  1297. *
  1298. * mod_delayed_work_on() on local CPU.
  1299. */
  1300. bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
  1301. unsigned long delay)
  1302. {
  1303. return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
  1304. }
  1305. EXPORT_SYMBOL_GPL(mod_delayed_work);
  1306. /**
  1307. * worker_enter_idle - enter idle state
  1308. * @worker: worker which is entering idle state
  1309. *
  1310. * @worker is entering idle state. Update stats and idle timer if
  1311. * necessary.
  1312. *
  1313. * LOCKING:
  1314. * spin_lock_irq(gcwq->lock).
  1315. */
  1316. static void worker_enter_idle(struct worker *worker)
  1317. {
  1318. struct worker_pool *pool = worker->pool;
  1319. struct global_cwq *gcwq = pool->gcwq;
  1320. BUG_ON(worker->flags & WORKER_IDLE);
  1321. BUG_ON(!list_empty(&worker->entry) &&
  1322. (worker->hentry.next || worker->hentry.pprev));
  1323. /* can't use worker_set_flags(), also called from start_worker() */
  1324. worker->flags |= WORKER_IDLE;
  1325. pool->nr_idle++;
  1326. worker->last_active = jiffies;
  1327. /* idle_list is LIFO */
  1328. list_add(&worker->entry, &pool->idle_list);
  1329. if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
  1330. mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
  1331. /*
  1332. * Sanity check nr_running. Because gcwq_unbind_fn() releases
  1333. * gcwq->lock between setting %WORKER_UNBOUND and zapping
  1334. * nr_running, the warning may trigger spuriously. Check iff
  1335. * unbind is not in progress.
  1336. */
  1337. WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
  1338. pool->nr_workers == pool->nr_idle &&
  1339. atomic_read(get_pool_nr_running(pool)));
  1340. }
  1341. /**
  1342. * worker_leave_idle - leave idle state
  1343. * @worker: worker which is leaving idle state
  1344. *
  1345. * @worker is leaving idle state. Update stats.
  1346. *
  1347. * LOCKING:
  1348. * spin_lock_irq(gcwq->lock).
  1349. */
  1350. static void worker_leave_idle(struct worker *worker)
  1351. {
  1352. struct worker_pool *pool = worker->pool;
  1353. BUG_ON(!(worker->flags & WORKER_IDLE));
  1354. worker_clr_flags(worker, WORKER_IDLE);
  1355. pool->nr_idle--;
  1356. list_del_init(&worker->entry);
  1357. }
  1358. /**
  1359. * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
  1360. * @worker: self
  1361. *
  1362. * Works which are scheduled while the cpu is online must at least be
  1363. * scheduled to a worker which is bound to the cpu so that if they are
  1364. * flushed from cpu callbacks while cpu is going down, they are
  1365. * guaranteed to execute on the cpu.
  1366. *
  1367. * This function is to be used by rogue workers and rescuers to bind
  1368. * themselves to the target cpu and may race with cpu going down or
  1369. * coming online. kthread_bind() can't be used because it may put the
  1370. * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
  1371. * verbatim as it's best effort and blocking and gcwq may be
  1372. * [dis]associated in the meantime.
  1373. *
  1374. * This function tries set_cpus_allowed() and locks gcwq and verifies the
  1375. * binding against %GCWQ_DISASSOCIATED which is set during
  1376. * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
  1377. * enters idle state or fetches works without dropping lock, it can
  1378. * guarantee the scheduling requirement described in the first paragraph.
  1379. *
  1380. * CONTEXT:
  1381. * Might sleep. Called without any lock but returns with gcwq->lock
  1382. * held.
  1383. *
  1384. * RETURNS:
  1385. * %true if the associated gcwq is online (@worker is successfully
  1386. * bound), %false if offline.
  1387. */
  1388. static bool worker_maybe_bind_and_lock(struct worker *worker)
  1389. __acquires(&gcwq->lock)
  1390. {
  1391. struct global_cwq *gcwq = worker->pool->gcwq;
  1392. struct task_struct *task = worker->task;
  1393. while (true) {
  1394. /*
  1395. * The following call may fail, succeed or succeed
  1396. * without actually migrating the task to the cpu if
  1397. * it races with cpu hotunplug operation. Verify
  1398. * against GCWQ_DISASSOCIATED.
  1399. */
  1400. if (!(gcwq->flags & GCWQ_DISASSOCIATED))
  1401. set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
  1402. spin_lock_irq(&gcwq->lock);
  1403. if (gcwq->flags & GCWQ_DISASSOCIATED)
  1404. return false;
  1405. if (task_cpu(task) == gcwq->cpu &&
  1406. cpumask_equal(&current->cpus_allowed,
  1407. get_cpu_mask(gcwq->cpu)))
  1408. return true;
  1409. spin_unlock_irq(&gcwq->lock);
  1410. /*
  1411. * We've raced with CPU hot[un]plug. Give it a breather
  1412. * and retry migration. cond_resched() is required here;
  1413. * otherwise, we might deadlock against cpu_stop trying to
  1414. * bring down the CPU on non-preemptive kernel.
  1415. */
  1416. cpu_relax();
  1417. cond_resched();
  1418. }
  1419. }
  1420. /*
  1421. * Rebind an idle @worker to its CPU. worker_thread() will test
  1422. * list_empty(@worker->entry) before leaving idle and call this function.
  1423. */
  1424. static void idle_worker_rebind(struct worker *worker)
  1425. {
  1426. struct global_cwq *gcwq = worker->pool->gcwq;
  1427. /* CPU may go down again inbetween, clear UNBOUND only on success */
  1428. if (worker_maybe_bind_and_lock(worker))
  1429. worker_clr_flags(worker, WORKER_UNBOUND);
  1430. /* rebind complete, become available again */
  1431. list_add(&worker->entry, &worker->pool->idle_list);
  1432. spin_unlock_irq(&gcwq->lock);
  1433. }
  1434. /*
  1435. * Function for @worker->rebind.work used to rebind unbound busy workers to
  1436. * the associated cpu which is coming back online. This is scheduled by
  1437. * cpu up but can race with other cpu hotplug operations and may be
  1438. * executed twice without intervening cpu down.
  1439. */
  1440. static void busy_worker_rebind_fn(struct work_struct *work)
  1441. {
  1442. struct worker *worker = container_of(work, struct worker, rebind_work);
  1443. struct global_cwq *gcwq = worker->pool->gcwq;
  1444. if (worker_maybe_bind_and_lock(worker))
  1445. worker_clr_flags(worker, WORKER_UNBOUND);
  1446. spin_unlock_irq(&gcwq->lock);
  1447. }
  1448. /**
  1449. * rebind_workers - rebind all workers of a gcwq to the associated CPU
  1450. * @gcwq: gcwq of interest
  1451. *
  1452. * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
  1453. * is different for idle and busy ones.
  1454. *
  1455. * Idle ones will be removed from the idle_list and woken up. They will
  1456. * add themselves back after completing rebind. This ensures that the
  1457. * idle_list doesn't contain any unbound workers when re-bound busy workers
  1458. * try to perform local wake-ups for concurrency management.
  1459. *
  1460. * Busy workers can rebind after they finish their current work items.
  1461. * Queueing the rebind work item at the head of the scheduled list is
  1462. * enough. Note that nr_running will be properly bumped as busy workers
  1463. * rebind.
  1464. *
  1465. * On return, all non-manager workers are scheduled for rebind - see
  1466. * manage_workers() for the manager special case. Any idle worker
  1467. * including the manager will not appear on @idle_list until rebind is
  1468. * complete, making local wake-ups safe.
  1469. */
  1470. static void rebind_workers(struct global_cwq *gcwq)
  1471. {
  1472. struct worker_pool *pool;
  1473. struct worker *worker, *n;
  1474. struct hlist_node *pos;
  1475. int i;
  1476. lockdep_assert_held(&gcwq->lock);
  1477. for_each_worker_pool(pool, gcwq)
  1478. lockdep_assert_held(&pool->assoc_mutex);
  1479. /* dequeue and kick idle ones */
  1480. for_each_worker_pool(pool, gcwq) {
  1481. list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
  1482. /*
  1483. * idle workers should be off @pool->idle_list
  1484. * until rebind is complete to avoid receiving
  1485. * premature local wake-ups.
  1486. */
  1487. list_del_init(&worker->entry);
  1488. /*
  1489. * worker_thread() will see the above dequeuing
  1490. * and call idle_worker_rebind().
  1491. */
  1492. wake_up_process(worker->task);
  1493. }
  1494. }
  1495. /* rebind busy workers */
  1496. for_each_busy_worker(worker, i, pos, gcwq) {
  1497. struct work_struct *rebind_work = &worker->rebind_work;
  1498. struct workqueue_struct *wq;
  1499. if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
  1500. work_data_bits(rebind_work)))
  1501. continue;
  1502. debug_work_activate(rebind_work);
  1503. /*
  1504. * wq doesn't really matter but let's keep @worker->pool
  1505. * and @cwq->pool consistent for sanity.
  1506. */
  1507. if (worker_pool_pri(worker->pool))
  1508. wq = system_highpri_wq;
  1509. else
  1510. wq = system_wq;
  1511. insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
  1512. worker->scheduled.next,
  1513. work_color_to_flags(WORK_NO_COLOR));
  1514. }
  1515. }
  1516. static struct worker *alloc_worker(void)
  1517. {
  1518. struct worker *worker;
  1519. worker = kzalloc(sizeof(*worker), GFP_KERNEL);
  1520. if (worker) {
  1521. INIT_LIST_HEAD(&worker->entry);
  1522. INIT_LIST_HEAD(&worker->scheduled);
  1523. INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
  1524. /* on creation a worker is in !idle && prep state */
  1525. worker->flags = WORKER_PREP;
  1526. }
  1527. return worker;
  1528. }
  1529. /**
  1530. * create_worker - create a new workqueue worker
  1531. * @pool: pool the new worker will belong to
  1532. *
  1533. * Create a new worker which is bound to @pool. The returned worker
  1534. * can be started by calling start_worker() or destroyed using
  1535. * destroy_worker().
  1536. *
  1537. * CONTEXT:
  1538. * Might sleep. Does GFP_KERNEL allocations.
  1539. *
  1540. * RETURNS:
  1541. * Pointer to the newly created worker.
  1542. */
  1543. static struct worker *create_worker(struct worker_pool *pool)
  1544. {
  1545. struct global_cwq *gcwq = pool->gcwq;
  1546. const char *pri = worker_pool_pri(pool) ? "H" : "";
  1547. struct worker *worker = NULL;
  1548. int id = -1;
  1549. spin_lock_irq(&gcwq->lock);
  1550. while (ida_get_new(&pool->worker_ida, &id)) {
  1551. spin_unlock_irq(&gcwq->lock);
  1552. if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
  1553. goto fail;
  1554. spin_lock_irq(&gcwq->lock);
  1555. }
  1556. spin_unlock_irq(&gcwq->lock);
  1557. worker = alloc_worker();
  1558. if (!worker)
  1559. goto fail;
  1560. worker->pool = pool;
  1561. worker->id = id;
  1562. if (gcwq->cpu != WORK_CPU_UNBOUND)
  1563. worker->task = kthread_create_on_node(worker_thread,
  1564. worker, cpu_to_node(gcwq->cpu),
  1565. "kworker/%u:%d%s", gcwq->cpu, id, pri);
  1566. else
  1567. worker->task = kthread_create(worker_thread, worker,
  1568. "kworker/u:%d%s", id, pri);
  1569. if (IS_ERR(worker->task))
  1570. goto fail;
  1571. if (worker_pool_pri(pool))
  1572. set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
  1573. /*
  1574. * Determine CPU binding of the new worker depending on
  1575. * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
  1576. * flag remains stable across this function. See the comments
  1577. * above the flag definition for details.
  1578. *
  1579. * As an unbound worker may later become a regular one if CPU comes
  1580. * online, make sure every worker has %PF_THREAD_BOUND set.
  1581. */
  1582. if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
  1583. kthread_bind(worker->task, gcwq->cpu);
  1584. } else {
  1585. worker->task->flags |= PF_THREAD_BOUND;
  1586. worker->flags |= WORKER_UNBOUND;
  1587. }
  1588. return worker;
  1589. fail:
  1590. if (id >= 0) {
  1591. spin_lock_irq(&gcwq->lock);
  1592. ida_remove(&pool->worker_ida, id);
  1593. spin_unlock_irq(&gcwq->lock);
  1594. }
  1595. kfree(worker);
  1596. return NULL;
  1597. }
  1598. /**
  1599. * start_worker - start a newly created worker
  1600. * @worker: worker to start
  1601. *
  1602. * Make the gcwq aware of @worker and start it.
  1603. *
  1604. * CONTEXT:
  1605. * spin_lock_irq(gcwq->lock).
  1606. */
  1607. static void start_worker(struct worker *worker)
  1608. {
  1609. worker->flags |= WORKER_STARTED;
  1610. worker->pool->nr_workers++;
  1611. worker_enter_idle(worker);
  1612. wake_up_process(worker->task);
  1613. }
  1614. /**
  1615. * destroy_worker - destroy a workqueue worker
  1616. * @worker: worker to be destroyed
  1617. *
  1618. * Destroy @worker and adjust @gcwq stats accordingly.
  1619. *
  1620. * CONTEXT:
  1621. * spin_lock_irq(gcwq->lock) which is released and regrabbed.
  1622. */
  1623. static void destroy_worker(struct worker *worker)
  1624. {
  1625. struct worker_pool *pool = worker->pool;
  1626. struct global_cwq *gcwq = pool->gcwq;
  1627. int id = worker->id;
  1628. /* sanity check frenzy */
  1629. BUG_ON(worker->current_work);
  1630. BUG_ON(!list_empty(&worker->scheduled));
  1631. if (worker->flags & WORKER_STARTED)
  1632. pool->nr_workers--;
  1633. if (worker->flags & WORKER_IDLE)
  1634. pool->nr_idle--;
  1635. list_del_init(&worker->entry);
  1636. worker->flags |= WORKER_DIE;
  1637. spin_unlock_irq(&gcwq->lock);
  1638. kthread_stop(worker->task);
  1639. kfree(worker);
  1640. spin_lock_irq(&gcwq->lock);
  1641. ida_remove(&pool->worker_ida, id);
  1642. }
  1643. static void idle_worker_timeout(unsigned long __pool)
  1644. {
  1645. struct worker_pool *pool = (void *)__pool;
  1646. struct global_cwq *gcwq = pool->gcwq;
  1647. spin_lock_irq(&gcwq->lock);
  1648. if (too_many_workers(pool)) {
  1649. struct worker *worker;
  1650. unsigned long expires;
  1651. /* idle_list is kept in LIFO order, check the last one */
  1652. worker = list_entry(pool->idle_list.prev, struct worker, entry);
  1653. expires = worker->last_active + IDLE_WORKER_TIMEOUT;
  1654. if (time_before(jiffies, expires))
  1655. mod_timer(&pool->idle_timer, expires);
  1656. else {
  1657. /* it's been idle for too long, wake up manager */
  1658. pool->flags |= POOL_MANAGE_WORKERS;
  1659. wake_up_worker(pool);
  1660. }
  1661. }
  1662. spin_unlock_irq(&gcwq->lock);
  1663. }
  1664. static bool send_mayday(struct work_struct *work)
  1665. {
  1666. struct cpu_workqueue_struct *cwq = get_work_cwq(work);
  1667. struct workqueue_struct *wq = cwq->wq;
  1668. unsigned int cpu;
  1669. if (!(wq->flags & WQ_RESCUER))
  1670. return false;
  1671. /* mayday mayday mayday */
  1672. cpu = cwq->pool->gcwq->cpu;
  1673. /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
  1674. if (cpu == WORK_CPU_UNBOUND)
  1675. cpu = 0;
  1676. if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
  1677. wake_up_process(wq->rescuer->task);
  1678. return true;
  1679. }
  1680. static void gcwq_mayday_timeout(unsigned long __pool)
  1681. {
  1682. struct worker_pool *pool = (void *)__pool;
  1683. struct global_cwq *gcwq = pool->gcwq;
  1684. struct work_struct *work;
  1685. spin_lock_irq(&gcwq->lock);
  1686. if (need_to_create_worker(pool)) {
  1687. /*
  1688. * We've been trying to create a new worker but
  1689. * haven't been successful. We might be hitting an
  1690. * allocation deadlock. Send distress signals to
  1691. * rescuers.
  1692. */
  1693. list_for_each_entry(work, &pool->worklist, entry)
  1694. send_mayday(work);
  1695. }
  1696. spin_unlock_irq(&gcwq->lock);
  1697. mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
  1698. }
  1699. /**
  1700. * maybe_create_worker - create a new worker if necessary
  1701. * @pool: pool to create a new worker for
  1702. *
  1703. * Create a new worker for @pool if necessary. @pool is guaranteed to
  1704. * have at least one idle worker on return from this function. If
  1705. * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
  1706. * sent to all rescuers with works scheduled on @pool to resolve
  1707. * possible allocation deadlock.
  1708. *
  1709. * On return, need_to_create_worker() is guaranteed to be false and
  1710. * may_start_working() true.
  1711. *
  1712. * LOCKING:
  1713. * spin_lock_irq(gcwq->lock) which may be released and regrabbed
  1714. * multiple times. Does GFP_KERNEL allocations. Called only from
  1715. * manager.
  1716. *
  1717. * RETURNS:
  1718. * false if no action was taken and gcwq->lock stayed locked, true
  1719. * otherwise.
  1720. */
  1721. static bool maybe_create_worker(struct worker_pool *pool)
  1722. __releases(&gcwq->lock)
  1723. __acquires(&gcwq->lock)
  1724. {
  1725. struct global_cwq *gcwq = pool->gcwq;
  1726. if (!need_to_create_worker(pool))
  1727. return false;
  1728. restart:
  1729. spin_unlock_irq(&gcwq->lock);
  1730. /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
  1731. mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
  1732. while (true) {
  1733. struct worker *worker;
  1734. worker = create_worker(pool);
  1735. if (worker) {
  1736. del_timer_sync(&pool->mayday_timer);
  1737. spin_lock_irq(&gcwq->lock);
  1738. start_worker(worker);
  1739. BUG_ON(need_to_create_worker(pool));
  1740. return true;
  1741. }
  1742. if (!need_to_create_worker(pool))
  1743. break;
  1744. __set_current_state(TASK_INTERRUPTIBLE);
  1745. schedule_timeout(CREATE_COOLDOWN);
  1746. if (!need_to_create_worker(pool))
  1747. break;
  1748. }
  1749. del_timer_sync(&pool->mayday_timer);
  1750. spin_lock_irq(&gcwq->lock);
  1751. if (need_to_create_worker(pool))
  1752. goto restart;
  1753. return true;
  1754. }
  1755. /**
  1756. * maybe_destroy_worker - destroy workers which have been idle for a while
  1757. * @pool: pool to destroy workers for
  1758. *
  1759. * Destroy @pool workers which have been idle for longer than
  1760. * IDLE_WORKER_TIMEOUT.
  1761. *
  1762. * LOCKING:
  1763. * spin_lock_irq(gcwq->lock) which may be released and regrabbed
  1764. * multiple times. Called only from manager.
  1765. *
  1766. * RETURNS:
  1767. * false if no action was taken and gcwq->lock stayed locked, true
  1768. * otherwise.
  1769. */
  1770. static bool maybe_destroy_workers(struct worker_pool *pool)
  1771. {
  1772. bool ret = false;
  1773. while (too_many_workers(pool)) {
  1774. struct worker *worker;
  1775. unsigned long expires;
  1776. worker = list_entry(pool->idle_list.prev, struct worker, entry);
  1777. expires = worker->last_active + IDLE_WORKER_TIMEOUT;
  1778. if (time_before(jiffies, expires)) {
  1779. mod_timer(&pool->idle_timer, expires);
  1780. break;
  1781. }
  1782. destroy_worker(worker);
  1783. ret = true;
  1784. }
  1785. return ret;
  1786. }
  1787. /**
  1788. * manage_workers - manage worker pool
  1789. * @worker: self
  1790. *
  1791. * Assume the manager role and manage gcwq worker pool @worker belongs
  1792. * to. At any given time, there can be only zero or one manager per
  1793. * gcwq. The exclusion is handled automatically by this function.
  1794. *
  1795. * The caller can safely start processing works on false return. On
  1796. * true return, it's guaranteed that need_to_create_worker() is false
  1797. * and may_start_working() is true.
  1798. *
  1799. * CONTEXT:
  1800. * spin_lock_irq(gcwq->lock) which may be released and regrabbed
  1801. * multiple times. Does GFP_KERNEL allocations.
  1802. *
  1803. * RETURNS:
  1804. * false if no action was taken and gcwq->lock stayed locked, true if
  1805. * some action was taken.
  1806. */
  1807. static bool manage_workers(struct worker *worker)
  1808. {
  1809. struct worker_pool *pool = worker->pool;
  1810. bool ret = false;
  1811. if (pool->flags & POOL_MANAGING_WORKERS)
  1812. return ret;
  1813. pool->flags |= POOL_MANAGING_WORKERS;
  1814. /*
  1815. * To simplify both worker management and CPU hotplug, hold off
  1816. * management while hotplug is in progress. CPU hotplug path can't
  1817. * grab %POOL_MANAGING_WORKERS to achieve this because that can
  1818. * lead to idle worker depletion (all become busy thinking someone
  1819. * else is managing) which in turn can result in deadlock under
  1820. * extreme circumstances. Use @pool->assoc_mutex to synchronize
  1821. * manager against CPU hotplug.
  1822. *
  1823. * assoc_mutex would always be free unless CPU hotplug is in
  1824. * progress. trylock first without dropping @gcwq->lock.
  1825. */
  1826. if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
  1827. spin_unlock_irq(&pool->gcwq->lock);
  1828. mutex_lock(&pool->assoc_mutex);
  1829. /*
  1830. * CPU hotplug could have happened while we were waiting
  1831. * for assoc_mutex. Hotplug itself can't handle us
  1832. * because manager isn't either on idle or busy list, and
  1833. * @gcwq's state and ours could have deviated.
  1834. *
  1835. * As hotplug is now excluded via assoc_mutex, we can
  1836. * simply try to bind. It will succeed or fail depending
  1837. * on @gcwq's current state. Try it and adjust
  1838. * %WORKER_UNBOUND accordingly.
  1839. */
  1840. if (worker_maybe_bind_and_lock(worker))
  1841. worker->flags &= ~WORKER_UNBOUND;
  1842. else
  1843. worker->flags |= WORKER_UNBOUND;
  1844. ret = true;
  1845. }
  1846. pool->flags &= ~POOL_MANAGE_WORKERS;
  1847. /*
  1848. * Destroy and then create so that may_start_working() is true
  1849. * on return.
  1850. */
  1851. ret |= maybe_destroy_workers(pool);
  1852. ret |= maybe_create_worker(pool);
  1853. pool->flags &= ~POOL_MANAGING_WORKERS;
  1854. mutex_unlock(&pool->assoc_mutex);
  1855. return ret;
  1856. }
  1857. /**
  1858. * process_one_work - process single work
  1859. * @worker: self
  1860. * @work: work to process
  1861. *
  1862. * Process @work. This function contains all the logics necessary to
  1863. * process a single work including synchronization against and
  1864. * interaction with other workers on the same cpu, queueing and
  1865. * flushing. As long as context requirement is met, any worker can
  1866. * call this function to process a work.
  1867. *
  1868. * CONTEXT:
  1869. * spin_lock_irq(gcwq->lock) which is released and regrabbed.
  1870. */
  1871. static void process_one_work(struct worker *worker, struct work_struct *work)
  1872. __releases(&gcwq->lock)
  1873. __acquires(&gcwq->lock)
  1874. {
  1875. struct cpu_workqueue_struct *cwq = get_work_cwq(work);
  1876. struct worker_pool *pool = worker->pool;
  1877. struct global_cwq *gcwq = pool->gcwq;
  1878. struct hlist_head *bwh = busy_worker_head(gcwq, work);
  1879. bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
  1880. work_func_t f = work->func;
  1881. int work_color;
  1882. struct worker *collision;
  1883. #ifdef CONFIG_LOCKDEP
  1884. /*
  1885. * It is permissible to free the struct work_struct from
  1886. * inside the function that is called from it, this we need to
  1887. * take into account for lockdep too. To avoid bogus "held
  1888. * lock freed" warnings as well as problems when looking into
  1889. * work->lockdep_map, make a copy and use that here.
  1890. */
  1891. struct lockdep_map lockdep_map;
  1892. lockdep_copy_map(&lockdep_map, &work->lockdep_map);
  1893. #endif
  1894. /*
  1895. * Ensure we're on the correct CPU. DISASSOCIATED test is
  1896. * necessary to avoid spurious warnings from rescuers servicing the
  1897. * unbound or a disassociated gcwq.
  1898. */
  1899. WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
  1900. !(gcwq->flags & GCWQ_DISASSOCIATED) &&
  1901. raw_smp_processor_id() != gcwq->cpu);
  1902. /*
  1903. * A single work shouldn't be executed concurrently by
  1904. * multiple workers on a single cpu. Check whether anyone is
  1905. * already processing the work. If so, defer the work to the
  1906. * currently executing one.
  1907. */
  1908. collision = __find_worker_executing_work(gcwq, bwh, work);
  1909. if (unlikely(collision)) {
  1910. move_linked_works(work, &collision->scheduled, NULL);
  1911. return;
  1912. }
  1913. /* claim and dequeue */
  1914. debug_work_deactivate(work);
  1915. hlist_add_head(&worker->hentry, bwh);
  1916. worker->current_work = work;
  1917. worker->current_cwq = cwq;
  1918. work_color = get_work_color(work);
  1919. list_del_init(&work->entry);
  1920. /*
  1921. * CPU intensive works don't participate in concurrency
  1922. * management. They're the scheduler's responsibility.
  1923. */
  1924. if (unlikely(cpu_intensive))
  1925. worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
  1926. /*
  1927. * Unbound gcwq isn't concurrency managed and work items should be
  1928. * executed ASAP. Wake up another worker if necessary.
  1929. */
  1930. if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
  1931. wake_up_worker(pool);
  1932. /*
  1933. * Record the last CPU and clear PENDING which should be the last
  1934. * update to @work. Also, do this inside @gcwq->lock so that
  1935. * PENDING and queued state changes happen together while IRQ is
  1936. * disabled.
  1937. */
  1938. set_work_cpu_and_clear_pending(work, gcwq->cpu);
  1939. spin_unlock_irq(&gcwq->lock);
  1940. lock_map_acquire_read(&cwq->wq->lockdep_map);
  1941. lock_map_acquire(&lockdep_map);
  1942. trace_workqueue_execute_start(work);
  1943. f(work);
  1944. /*
  1945. * While we must be careful to not use "work" after this, the trace
  1946. * point will only record its address.
  1947. */
  1948. trace_workqueue_execute_end(work);
  1949. lock_map_release(&lockdep_map);
  1950. lock_map_release(&cwq->wq->lockdep_map);
  1951. if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
  1952. pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
  1953. " last function: %pf\n",
  1954. current->comm, preempt_count(), task_pid_nr(current), f);
  1955. debug_show_held_locks(current);
  1956. dump_stack();
  1957. }
  1958. spin_lock_irq(&gcwq->lock);
  1959. /* clear cpu intensive status */
  1960. if (unlikely(cpu_intensive))
  1961. worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
  1962. /* we're done with it, release */
  1963. hlist_del_init(&worker->hentry);
  1964. worker->current_work = NULL;
  1965. worker->current_cwq = NULL;
  1966. cwq_dec_nr_in_flight(cwq, work_color);
  1967. }
  1968. /**
  1969. * process_scheduled_works - process scheduled works
  1970. * @worker: self
  1971. *
  1972. * Process all scheduled works. Please note that the scheduled list
  1973. * may change while processing a work, so this function repeatedly
  1974. * fetches a work from the top and executes it.
  1975. *
  1976. * CONTEXT:
  1977. * spin_lock_irq(gcwq->lock) which may be released and regrabbed
  1978. * multiple times.
  1979. */
  1980. static void process_scheduled_works(struct worker *worker)
  1981. {
  1982. while (!list_empty(&worker->scheduled)) {
  1983. struct work_struct *work = list_first_entry(&worker->scheduled,
  1984. struct work_struct, entry);
  1985. process_one_work(worker, work);
  1986. }
  1987. }
  1988. /**
  1989. * worker_thread - the worker thread function
  1990. * @__worker: self
  1991. *
  1992. * The gcwq worker thread function. There's a single dynamic pool of
  1993. * these per each cpu. These workers process all works regardless of
  1994. * their specific target workqueue. The only exception is works which
  1995. * belong to workqueues with a rescuer which will be explained in
  1996. * rescuer_thread().
  1997. */
  1998. static int worker_thread(void *__worker)
  1999. {
  2000. struct worker *worker = __worker;
  2001. struct worker_pool *pool = worker->pool;
  2002. struct global_cwq *gcwq = pool->gcwq;
  2003. /* tell the scheduler that this is a workqueue worker */
  2004. worker->task->flags |= PF_WQ_WORKER;
  2005. woke_up:
  2006. spin_lock_irq(&gcwq->lock);
  2007. /* we are off idle list if destruction or rebind is requested */
  2008. if (unlikely(list_empty(&worker->entry))) {
  2009. spin_unlock_irq(&gcwq->lock);
  2010. /* if DIE is set, destruction is requested */
  2011. if (worker->flags & WORKER_DIE) {
  2012. worker->task->flags &= ~PF_WQ_WORKER;
  2013. return 0;
  2014. }
  2015. /* otherwise, rebind */
  2016. idle_worker_rebind(worker);
  2017. goto woke_up;
  2018. }
  2019. worker_leave_idle(worker);
  2020. recheck:
  2021. /* no more worker necessary? */
  2022. if (!need_more_worker(pool))
  2023. goto sleep;
  2024. /* do we need to manage? */
  2025. if (unlikely(!may_start_working(pool)) && manage_workers(worker))
  2026. goto recheck;
  2027. /*
  2028. * ->scheduled list can only be filled while a worker is
  2029. * preparing to process a work or actually processing it.
  2030. * Make sure nobody diddled with it while I was sleeping.
  2031. */
  2032. BUG_ON(!list_empty(&worker->scheduled));
  2033. /*
  2034. * When control reaches this point, we're guaranteed to have
  2035. * at least one idle worker or that someone else has already
  2036. * assumed the manager role.
  2037. */
  2038. worker_clr_flags(worker, WORKER_PREP);
  2039. do {
  2040. struct work_struct *work =
  2041. list_first_entry(&pool->worklist,
  2042. struct work_struct, entry);
  2043. if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
  2044. /* optimization path, not strictly necessary */
  2045. process_one_work(worker, work);
  2046. if (unlikely(!list_empty(&worker->scheduled)))
  2047. process_scheduled_works(worker);
  2048. } else {
  2049. move_linked_works(work, &worker->scheduled, NULL);
  2050. process_scheduled_works(worker);
  2051. }
  2052. } while (keep_working(pool));
  2053. worker_set_flags(worker, WORKER_PREP, false);
  2054. sleep:
  2055. if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
  2056. goto recheck;
  2057. /*
  2058. * gcwq->lock is held and there's no work to process and no
  2059. * need to manage, sleep. Workers are woken up only while
  2060. * holding gcwq->lock or from local cpu, so setting the
  2061. * current state before releasing gcwq->lock is enough to
  2062. * prevent losing any event.
  2063. */
  2064. worker_enter_idle(worker);
  2065. __set_current_state(TASK_INTERRUPTIBLE);
  2066. spin_unlock_irq(&gcwq->lock);
  2067. schedule();
  2068. goto woke_up;
  2069. }
  2070. /**
  2071. * rescuer_thread - the rescuer thread function
  2072. * @__wq: the associated workqueue
  2073. *
  2074. * Workqueue rescuer thread function. There's one rescuer for each
  2075. * workqueue which has WQ_RESCUER set.
  2076. *
  2077. * Regular work processing on a gcwq may block trying to create a new
  2078. * worker which uses GFP_KERNEL allocation which has slight chance of
  2079. * developing into deadlock if some works currently on the same queue
  2080. * need to be processed to satisfy the GFP_KERNEL allocation. This is
  2081. * the problem rescuer solves.
  2082. *
  2083. * When such condition is possible, the gcwq summons rescuers of all
  2084. * workqueues which have works queued on the gcwq and let them process
  2085. * those works so that forward progress can be guaranteed.
  2086. *
  2087. * This should happen rarely.
  2088. */
  2089. static int rescuer_thread(void *__wq)
  2090. {
  2091. struct workqueue_struct *wq = __wq;
  2092. struct worker *rescuer = wq->rescuer;
  2093. struct list_head *scheduled = &rescuer->scheduled;
  2094. bool is_unbound = wq->flags & WQ_UNBOUND;
  2095. unsigned int cpu;
  2096. set_user_nice(current, RESCUER_NICE_LEVEL);
  2097. repeat:
  2098. set_current_state(TASK_INTERRUPTIBLE);
  2099. if (kthread_should_stop())
  2100. return 0;
  2101. /*
  2102. * See whether any cpu is asking for help. Unbounded
  2103. * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
  2104. */
  2105. for_each_mayday_cpu(cpu, wq->mayday_mask) {
  2106. unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
  2107. struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
  2108. struct worker_pool *pool = cwq->pool;
  2109. struct global_cwq *gcwq = pool->gcwq;
  2110. struct work_struct *work, *n;
  2111. __set_current_state(TASK_RUNNING);
  2112. mayday_clear_cpu(cpu, wq->mayday_mask);
  2113. /* migrate to the target cpu if possible */
  2114. rescuer->pool = pool;
  2115. worker_maybe_bind_and_lock(rescuer);
  2116. /*
  2117. * Slurp in all works issued via this workqueue and
  2118. * process'em.
  2119. */
  2120. BUG_ON(!list_empty(&rescuer->scheduled));
  2121. list_for_each_entry_safe(work, n, &pool->worklist, entry)
  2122. if (get_work_cwq(work) == cwq)
  2123. move_linked_works(work, scheduled, &n);
  2124. process_scheduled_works(rescuer);
  2125. /*
  2126. * Leave this gcwq. If keep_working() is %true, notify a
  2127. * regular worker; otherwise, we end up with 0 concurrency
  2128. * and stalling the execution.
  2129. */
  2130. if (keep_working(pool))
  2131. wake_up_worker(pool);
  2132. spin_unlock_irq(&gcwq->lock);
  2133. }
  2134. schedule();
  2135. goto repeat;
  2136. }
  2137. struct wq_barrier {
  2138. struct work_struct work;
  2139. struct completion done;
  2140. };
  2141. static void wq_barrier_func(struct work_struct *work)
  2142. {
  2143. struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
  2144. complete(&barr->done);
  2145. }
  2146. /**
  2147. * insert_wq_barrier - insert a barrier work
  2148. * @cwq: cwq to insert barrier into
  2149. * @barr: wq_barrier to insert
  2150. * @target: target work to attach @barr to
  2151. * @worker: worker currently executing @target, NULL if @target is not executing
  2152. *
  2153. * @barr is linked to @target such that @barr is completed only after
  2154. * @target finishes execution. Please note that the ordering
  2155. * guarantee is observed only with respect to @target and on the local
  2156. * cpu.
  2157. *
  2158. * Currently, a queued barrier can't be canceled. This is because
  2159. * try_to_grab_pending() can't determine whether the work to be
  2160. * grabbed is at the head of the queue and thus can't clear LINKED
  2161. * flag of the previous work while there must be a valid next work
  2162. * after a work with LINKED flag set.
  2163. *
  2164. * Note that when @worker is non-NULL, @target may be modified
  2165. * underneath us, so we can't reliably determine cwq from @target.
  2166. *
  2167. * CONTEXT:
  2168. * spin_lock_irq(gcwq->lock).
  2169. */
  2170. static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
  2171. struct wq_barrier *barr,
  2172. struct work_struct *target, struct worker *worker)
  2173. {
  2174. struct list_head *head;
  2175. unsigned int linked = 0;
  2176. /*
  2177. * debugobject calls are safe here even with gcwq->lock locked
  2178. * as we know for sure that this will not trigger any of the
  2179. * checks and call back into the fixup functions where we
  2180. * might deadlock.
  2181. */
  2182. INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
  2183. __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
  2184. init_completion(&barr->done);
  2185. /*
  2186. * If @target is currently being executed, schedule the
  2187. * barrier to the worker; otherwise, put it after @target.
  2188. */
  2189. if (worker)
  2190. head = worker->scheduled.next;
  2191. else {
  2192. unsigned long *bits = work_data_bits(target);
  2193. head = target->entry.next;
  2194. /* there can already be other linked works, inherit and set */
  2195. linked = *bits & WORK_STRUCT_LINKED;
  2196. __set_bit(WORK_STRUCT_LINKED_BIT, bits);
  2197. }
  2198. debug_work_activate(&barr->work);
  2199. insert_work(cwq, &barr->work, head,
  2200. work_color_to_flags(WORK_NO_COLOR) | linked);
  2201. }
  2202. /**
  2203. * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
  2204. * @wq: workqueue being flushed
  2205. * @flush_color: new flush color, < 0 for no-op
  2206. * @work_color: new work color, < 0 for no-op
  2207. *
  2208. * Prepare cwqs for workqueue flushing.
  2209. *
  2210. * If @flush_color is non-negative, flush_color on all cwqs should be
  2211. * -1. If no cwq has in-flight commands at the specified color, all
  2212. * cwq->flush_color's stay at -1 and %false is returned. If any cwq
  2213. * has in flight commands, its cwq->flush_color is set to
  2214. * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
  2215. * wakeup logic is armed and %true is returned.
  2216. *
  2217. * The caller should have initialized @wq->first_flusher prior to
  2218. * calling this function with non-negative @flush_color. If
  2219. * @flush_color is negative, no flush color update is done and %false
  2220. * is returned.
  2221. *
  2222. * If @work_color is non-negative, all cwqs should have the same
  2223. * work_color which is previous to @work_color and all will be
  2224. * advanced to @work_color.
  2225. *
  2226. * CONTEXT:
  2227. * mutex_lock(wq->flush_mutex).
  2228. *
  2229. * RETURNS:
  2230. * %true if @flush_color >= 0 and there's something to flush. %false
  2231. * otherwise.
  2232. */
  2233. static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
  2234. int flush_color, int work_color)
  2235. {
  2236. bool wait = false;
  2237. unsigned int cpu;
  2238. if (flush_color >= 0) {
  2239. BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
  2240. atomic_set(&wq->nr_cwqs_to_flush, 1);
  2241. }
  2242. for_each_cwq_cpu(cpu, wq) {
  2243. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  2244. struct global_cwq *gcwq = cwq->pool->gcwq;
  2245. spin_lock_irq(&gcwq->lock);
  2246. if (flush_color >= 0) {
  2247. BUG_ON(cwq->flush_color != -1);
  2248. if (cwq->nr_in_flight[flush_color]) {
  2249. cwq->flush_color = flush_color;
  2250. atomic_inc(&wq->nr_cwqs_to_flush);
  2251. wait = true;
  2252. }
  2253. }
  2254. if (work_color >= 0) {
  2255. BUG_ON(work_color != work_next_color(cwq->work_color));
  2256. cwq->work_color = work_color;
  2257. }
  2258. spin_unlock_irq(&gcwq->lock);
  2259. }
  2260. if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
  2261. complete(&wq->first_flusher->done);
  2262. return wait;
  2263. }
  2264. /**
  2265. * flush_workqueue - ensure that any scheduled work has run to completion.
  2266. * @wq: workqueue to flush
  2267. *
  2268. * Forces execution of the workqueue and blocks until its completion.
  2269. * This is typically used in driver shutdown handlers.
  2270. *
  2271. * We sleep until all works which were queued on entry have been handled,
  2272. * but we are not livelocked by new incoming ones.
  2273. */
  2274. void flush_workqueue(struct workqueue_struct *wq)
  2275. {
  2276. struct wq_flusher this_flusher = {
  2277. .list = LIST_HEAD_INIT(this_flusher.list),
  2278. .flush_color = -1,
  2279. .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
  2280. };
  2281. int next_color;
  2282. lock_map_acquire(&wq->lockdep_map);
  2283. lock_map_release(&wq->lockdep_map);
  2284. mutex_lock(&wq->flush_mutex);
  2285. /*
  2286. * Start-to-wait phase
  2287. */
  2288. next_color = work_next_color(wq->work_color);
  2289. if (next_color != wq->flush_color) {
  2290. /*
  2291. * Color space is not full. The current work_color
  2292. * becomes our flush_color and work_color is advanced
  2293. * by one.
  2294. */
  2295. BUG_ON(!list_empty(&wq->flusher_overflow));
  2296. this_flusher.flush_color = wq->work_color;
  2297. wq->work_color = next_color;
  2298. if (!wq->first_flusher) {
  2299. /* no flush in progress, become the first flusher */
  2300. BUG_ON(wq->flush_color != this_flusher.flush_color);
  2301. wq->first_flusher = &this_flusher;
  2302. if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
  2303. wq->work_color)) {
  2304. /* nothing to flush, done */
  2305. wq->flush_color = next_color;
  2306. wq->first_flusher = NULL;
  2307. goto out_unlock;
  2308. }
  2309. } else {
  2310. /* wait in queue */
  2311. BUG_ON(wq->flush_color == this_flusher.flush_color);
  2312. list_add_tail(&this_flusher.list, &wq->flusher_queue);
  2313. flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
  2314. }
  2315. } else {
  2316. /*
  2317. * Oops, color space is full, wait on overflow queue.
  2318. * The next flush completion will assign us
  2319. * flush_color and transfer to flusher_queue.
  2320. */
  2321. list_add_tail(&this_flusher.list, &wq->flusher_overflow);
  2322. }
  2323. mutex_unlock(&wq->flush_mutex);
  2324. wait_for_completion(&this_flusher.done);
  2325. /*
  2326. * Wake-up-and-cascade phase
  2327. *
  2328. * First flushers are responsible for cascading flushes and
  2329. * handling overflow. Non-first flushers can simply return.
  2330. */
  2331. if (wq->first_flusher != &this_flusher)
  2332. return;
  2333. mutex_lock(&wq->flush_mutex);
  2334. /* we might have raced, check again with mutex held */
  2335. if (wq->first_flusher != &this_flusher)
  2336. goto out_unlock;
  2337. wq->first_flusher = NULL;
  2338. BUG_ON(!list_empty(&this_flusher.list));
  2339. BUG_ON(wq->flush_color != this_flusher.flush_color);
  2340. while (true) {
  2341. struct wq_flusher *next, *tmp;
  2342. /* complete all the flushers sharing the current flush color */
  2343. list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
  2344. if (next->flush_color != wq->flush_color)
  2345. break;
  2346. list_del_init(&next->list);
  2347. complete(&next->done);
  2348. }
  2349. BUG_ON(!list_empty(&wq->flusher_overflow) &&
  2350. wq->flush_color != work_next_color(wq->work_color));
  2351. /* this flush_color is finished, advance by one */
  2352. wq->flush_color = work_next_color(wq->flush_color);
  2353. /* one color has been freed, handle overflow queue */
  2354. if (!list_empty(&wq->flusher_overflow)) {
  2355. /*
  2356. * Assign the same color to all overflowed
  2357. * flushers, advance work_color and append to
  2358. * flusher_queue. This is the start-to-wait
  2359. * phase for these overflowed flushers.
  2360. */
  2361. list_for_each_entry(tmp, &wq->flusher_overflow, list)
  2362. tmp->flush_color = wq->work_color;
  2363. wq->work_color = work_next_color(wq->work_color);
  2364. list_splice_tail_init(&wq->flusher_overflow,
  2365. &wq->flusher_queue);
  2366. flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
  2367. }
  2368. if (list_empty(&wq->flusher_queue)) {
  2369. BUG_ON(wq->flush_color != wq->work_color);
  2370. break;
  2371. }
  2372. /*
  2373. * Need to flush more colors. Make the next flusher
  2374. * the new first flusher and arm cwqs.
  2375. */
  2376. BUG_ON(wq->flush_color == wq->work_color);
  2377. BUG_ON(wq->flush_color != next->flush_color);
  2378. list_del_init(&next->list);
  2379. wq->first_flusher = next;
  2380. if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
  2381. break;
  2382. /*
  2383. * Meh... this color is already done, clear first
  2384. * flusher and repeat cascading.
  2385. */
  2386. wq->first_flusher = NULL;
  2387. }
  2388. out_unlock:
  2389. mutex_unlock(&wq->flush_mutex);
  2390. }
  2391. EXPORT_SYMBOL_GPL(flush_workqueue);
  2392. /**
  2393. * drain_workqueue - drain a workqueue
  2394. * @wq: workqueue to drain
  2395. *
  2396. * Wait until the workqueue becomes empty. While draining is in progress,
  2397. * only chain queueing is allowed. IOW, only currently pending or running
  2398. * work items on @wq can queue further work items on it. @wq is flushed
  2399. * repeatedly until it becomes empty. The number of flushing is detemined
  2400. * by the depth of chaining and should be relatively short. Whine if it
  2401. * takes too long.
  2402. */
  2403. void drain_workqueue(struct workqueue_struct *wq)
  2404. {
  2405. unsigned int flush_cnt = 0;
  2406. unsigned int cpu;
  2407. /*
  2408. * __queue_work() needs to test whether there are drainers, is much
  2409. * hotter than drain_workqueue() and already looks at @wq->flags.
  2410. * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
  2411. */
  2412. spin_lock(&workqueue_lock);
  2413. if (!wq->nr_drainers++)
  2414. wq->flags |= WQ_DRAINING;
  2415. spin_unlock(&workqueue_lock);
  2416. reflush:
  2417. flush_workqueue(wq);
  2418. for_each_cwq_cpu(cpu, wq) {
  2419. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  2420. bool drained;
  2421. spin_lock_irq(&cwq->pool->gcwq->lock);
  2422. drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
  2423. spin_unlock_irq(&cwq->pool->gcwq->lock);
  2424. if (drained)
  2425. continue;
  2426. if (++flush_cnt == 10 ||
  2427. (flush_cnt % 100 == 0 && flush_cnt <= 1000))
  2428. pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
  2429. wq->name, flush_cnt);
  2430. goto reflush;
  2431. }
  2432. spin_lock(&workqueue_lock);
  2433. if (!--wq->nr_drainers)
  2434. wq->flags &= ~WQ_DRAINING;
  2435. spin_unlock(&workqueue_lock);
  2436. }
  2437. EXPORT_SYMBOL_GPL(drain_workqueue);
  2438. static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
  2439. {
  2440. struct worker *worker = NULL;
  2441. struct global_cwq *gcwq;
  2442. struct cpu_workqueue_struct *cwq;
  2443. might_sleep();
  2444. gcwq = get_work_gcwq(work);
  2445. if (!gcwq)
  2446. return false;
  2447. spin_lock_irq(&gcwq->lock);
  2448. if (!list_empty(&work->entry)) {
  2449. /*
  2450. * See the comment near try_to_grab_pending()->smp_rmb().
  2451. * If it was re-queued to a different gcwq under us, we
  2452. * are not going to wait.
  2453. */
  2454. smp_rmb();
  2455. cwq = get_work_cwq(work);
  2456. if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
  2457. goto already_gone;
  2458. } else {
  2459. worker = find_worker_executing_work(gcwq, work);
  2460. if (!worker)
  2461. goto already_gone;
  2462. cwq = worker->current_cwq;
  2463. }
  2464. insert_wq_barrier(cwq, barr, work, worker);
  2465. spin_unlock_irq(&gcwq->lock);
  2466. /*
  2467. * If @max_active is 1 or rescuer is in use, flushing another work
  2468. * item on the same workqueue may lead to deadlock. Make sure the
  2469. * flusher is not running on the same workqueue by verifying write
  2470. * access.
  2471. */
  2472. if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
  2473. lock_map_acquire(&cwq->wq->lockdep_map);
  2474. else
  2475. lock_map_acquire_read(&cwq->wq->lockdep_map);
  2476. lock_map_release(&cwq->wq->lockdep_map);
  2477. return true;
  2478. already_gone:
  2479. spin_unlock_irq(&gcwq->lock);
  2480. return false;
  2481. }
  2482. /**
  2483. * flush_work - wait for a work to finish executing the last queueing instance
  2484. * @work: the work to flush
  2485. *
  2486. * Wait until @work has finished execution. @work is guaranteed to be idle
  2487. * on return if it hasn't been requeued since flush started.
  2488. *
  2489. * RETURNS:
  2490. * %true if flush_work() waited for the work to finish execution,
  2491. * %false if it was already idle.
  2492. */
  2493. bool flush_work(struct work_struct *work)
  2494. {
  2495. struct wq_barrier barr;
  2496. lock_map_acquire(&work->lockdep_map);
  2497. lock_map_release(&work->lockdep_map);
  2498. if (start_flush_work(work, &barr)) {
  2499. wait_for_completion(&barr.done);
  2500. destroy_work_on_stack(&barr.work);
  2501. return true;
  2502. } else {
  2503. return false;
  2504. }
  2505. }
  2506. EXPORT_SYMBOL_GPL(flush_work);
  2507. static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
  2508. {
  2509. unsigned long flags;
  2510. int ret;
  2511. do {
  2512. ret = try_to_grab_pending(work, is_dwork, &flags);
  2513. /*
  2514. * If someone else is canceling, wait for the same event it
  2515. * would be waiting for before retrying.
  2516. */
  2517. if (unlikely(ret == -ENOENT))
  2518. flush_work(work);
  2519. } while (unlikely(ret < 0));
  2520. /* tell other tasks trying to grab @work to back off */
  2521. mark_work_canceling(work);
  2522. local_irq_restore(flags);
  2523. flush_work(work);
  2524. clear_work_data(work);
  2525. return ret;
  2526. }
  2527. /**
  2528. * cancel_work_sync - cancel a work and wait for it to finish
  2529. * @work: the work to cancel
  2530. *
  2531. * Cancel @work and wait for its execution to finish. This function
  2532. * can be used even if the work re-queues itself or migrates to
  2533. * another workqueue. On return from this function, @work is
  2534. * guaranteed to be not pending or executing on any CPU.
  2535. *
  2536. * cancel_work_sync(&delayed_work->work) must not be used for
  2537. * delayed_work's. Use cancel_delayed_work_sync() instead.
  2538. *
  2539. * The caller must ensure that the workqueue on which @work was last
  2540. * queued can't be destroyed before this function returns.
  2541. *
  2542. * RETURNS:
  2543. * %true if @work was pending, %false otherwise.
  2544. */
  2545. bool cancel_work_sync(struct work_struct *work)
  2546. {
  2547. return __cancel_work_timer(work, false);
  2548. }
  2549. EXPORT_SYMBOL_GPL(cancel_work_sync);
  2550. /**
  2551. * flush_delayed_work - wait for a dwork to finish executing the last queueing
  2552. * @dwork: the delayed work to flush
  2553. *
  2554. * Delayed timer is cancelled and the pending work is queued for
  2555. * immediate execution. Like flush_work(), this function only
  2556. * considers the last queueing instance of @dwork.
  2557. *
  2558. * RETURNS:
  2559. * %true if flush_work() waited for the work to finish execution,
  2560. * %false if it was already idle.
  2561. */
  2562. bool flush_delayed_work(struct delayed_work *dwork)
  2563. {
  2564. local_irq_disable();
  2565. if (del_timer_sync(&dwork->timer))
  2566. __queue_work(dwork->cpu,
  2567. get_work_cwq(&dwork->work)->wq, &dwork->work);
  2568. local_irq_enable();
  2569. return flush_work(&dwork->work);
  2570. }
  2571. EXPORT_SYMBOL(flush_delayed_work);
  2572. /**
  2573. * cancel_delayed_work - cancel a delayed work
  2574. * @dwork: delayed_work to cancel
  2575. *
  2576. * Kill off a pending delayed_work. Returns %true if @dwork was pending
  2577. * and canceled; %false if wasn't pending. Note that the work callback
  2578. * function may still be running on return, unless it returns %true and the
  2579. * work doesn't re-arm itself. Explicitly flush or use
  2580. * cancel_delayed_work_sync() to wait on it.
  2581. *
  2582. * This function is safe to call from any context including IRQ handler.
  2583. */
  2584. bool cancel_delayed_work(struct delayed_work *dwork)
  2585. {
  2586. unsigned long flags;
  2587. int ret;
  2588. do {
  2589. ret = try_to_grab_pending(&dwork->work, true, &flags);
  2590. } while (unlikely(ret == -EAGAIN));
  2591. if (unlikely(ret < 0))
  2592. return false;
  2593. set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
  2594. local_irq_restore(flags);
  2595. return true;
  2596. }
  2597. EXPORT_SYMBOL(cancel_delayed_work);
  2598. /**
  2599. * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
  2600. * @dwork: the delayed work cancel
  2601. *
  2602. * This is cancel_work_sync() for delayed works.
  2603. *
  2604. * RETURNS:
  2605. * %true if @dwork was pending, %false otherwise.
  2606. */
  2607. bool cancel_delayed_work_sync(struct delayed_work *dwork)
  2608. {
  2609. return __cancel_work_timer(&dwork->work, true);
  2610. }
  2611. EXPORT_SYMBOL(cancel_delayed_work_sync);
  2612. /**
  2613. * schedule_work_on - put work task on a specific cpu
  2614. * @cpu: cpu to put the work task on
  2615. * @work: job to be done
  2616. *
  2617. * This puts a job on a specific cpu
  2618. */
  2619. bool schedule_work_on(int cpu, struct work_struct *work)
  2620. {
  2621. return queue_work_on(cpu, system_wq, work);
  2622. }
  2623. EXPORT_SYMBOL(schedule_work_on);
  2624. /**
  2625. * schedule_work - put work task in global workqueue
  2626. * @work: job to be done
  2627. *
  2628. * Returns %false if @work was already on the kernel-global workqueue and
  2629. * %true otherwise.
  2630. *
  2631. * This puts a job in the kernel-global workqueue if it was not already
  2632. * queued and leaves it in the same position on the kernel-global
  2633. * workqueue otherwise.
  2634. */
  2635. bool schedule_work(struct work_struct *work)
  2636. {
  2637. return queue_work(system_wq, work);
  2638. }
  2639. EXPORT_SYMBOL(schedule_work);
  2640. /**
  2641. * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
  2642. * @cpu: cpu to use
  2643. * @dwork: job to be done
  2644. * @delay: number of jiffies to wait
  2645. *
  2646. * After waiting for a given time this puts a job in the kernel-global
  2647. * workqueue on the specified CPU.
  2648. */
  2649. bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
  2650. unsigned long delay)
  2651. {
  2652. return queue_delayed_work_on(cpu, system_wq, dwork, delay);
  2653. }
  2654. EXPORT_SYMBOL(schedule_delayed_work_on);
  2655. /**
  2656. * schedule_delayed_work - put work task in global workqueue after delay
  2657. * @dwork: job to be done
  2658. * @delay: number of jiffies to wait or 0 for immediate execution
  2659. *
  2660. * After waiting for a given time this puts a job in the kernel-global
  2661. * workqueue.
  2662. */
  2663. bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
  2664. {
  2665. return queue_delayed_work(system_wq, dwork, delay);
  2666. }
  2667. EXPORT_SYMBOL(schedule_delayed_work);
  2668. /**
  2669. * schedule_on_each_cpu - execute a function synchronously on each online CPU
  2670. * @func: the function to call
  2671. *
  2672. * schedule_on_each_cpu() executes @func on each online CPU using the
  2673. * system workqueue and blocks until all CPUs have completed.
  2674. * schedule_on_each_cpu() is very slow.
  2675. *
  2676. * RETURNS:
  2677. * 0 on success, -errno on failure.
  2678. */
  2679. int schedule_on_each_cpu(work_func_t func)
  2680. {
  2681. int cpu;
  2682. struct work_struct __percpu *works;
  2683. works = alloc_percpu(struct work_struct);
  2684. if (!works)
  2685. return -ENOMEM;
  2686. get_online_cpus();
  2687. for_each_online_cpu(cpu) {
  2688. struct work_struct *work = per_cpu_ptr(works, cpu);
  2689. INIT_WORK(work, func);
  2690. schedule_work_on(cpu, work);
  2691. }
  2692. for_each_online_cpu(cpu)
  2693. flush_work(per_cpu_ptr(works, cpu));
  2694. put_online_cpus();
  2695. free_percpu(works);
  2696. return 0;
  2697. }
  2698. /**
  2699. * flush_scheduled_work - ensure that any scheduled work has run to completion.
  2700. *
  2701. * Forces execution of the kernel-global workqueue and blocks until its
  2702. * completion.
  2703. *
  2704. * Think twice before calling this function! It's very easy to get into
  2705. * trouble if you don't take great care. Either of the following situations
  2706. * will lead to deadlock:
  2707. *
  2708. * One of the work items currently on the workqueue needs to acquire
  2709. * a lock held by your code or its caller.
  2710. *
  2711. * Your code is running in the context of a work routine.
  2712. *
  2713. * They will be detected by lockdep when they occur, but the first might not
  2714. * occur very often. It depends on what work items are on the workqueue and
  2715. * what locks they need, which you have no control over.
  2716. *
  2717. * In most situations flushing the entire workqueue is overkill; you merely
  2718. * need to know that a particular work item isn't queued and isn't running.
  2719. * In such cases you should use cancel_delayed_work_sync() or
  2720. * cancel_work_sync() instead.
  2721. */
  2722. void flush_scheduled_work(void)
  2723. {
  2724. flush_workqueue(system_wq);
  2725. }
  2726. EXPORT_SYMBOL(flush_scheduled_work);
  2727. /**
  2728. * execute_in_process_context - reliably execute the routine with user context
  2729. * @fn: the function to execute
  2730. * @ew: guaranteed storage for the execute work structure (must
  2731. * be available when the work executes)
  2732. *
  2733. * Executes the function immediately if process context is available,
  2734. * otherwise schedules the function for delayed execution.
  2735. *
  2736. * Returns: 0 - function was executed
  2737. * 1 - function was scheduled for execution
  2738. */
  2739. int execute_in_process_context(work_func_t fn, struct execute_work *ew)
  2740. {
  2741. if (!in_interrupt()) {
  2742. fn(&ew->work);
  2743. return 0;
  2744. }
  2745. INIT_WORK(&ew->work, fn);
  2746. schedule_work(&ew->work);
  2747. return 1;
  2748. }
  2749. EXPORT_SYMBOL_GPL(execute_in_process_context);
  2750. int keventd_up(void)
  2751. {
  2752. return system_wq != NULL;
  2753. }
  2754. static int alloc_cwqs(struct workqueue_struct *wq)
  2755. {
  2756. /*
  2757. * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
  2758. * Make sure that the alignment isn't lower than that of
  2759. * unsigned long long.
  2760. */
  2761. const size_t size = sizeof(struct cpu_workqueue_struct);
  2762. const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
  2763. __alignof__(unsigned long long));
  2764. if (!(wq->flags & WQ_UNBOUND))
  2765. wq->cpu_wq.pcpu = __alloc_percpu(size, align);
  2766. else {
  2767. void *ptr;
  2768. /*
  2769. * Allocate enough room to align cwq and put an extra
  2770. * pointer at the end pointing back to the originally
  2771. * allocated pointer which will be used for free.
  2772. */
  2773. ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
  2774. if (ptr) {
  2775. wq->cpu_wq.single = PTR_ALIGN(ptr, align);
  2776. *(void **)(wq->cpu_wq.single + 1) = ptr;
  2777. }
  2778. }
  2779. /* just in case, make sure it's actually aligned */
  2780. BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
  2781. return wq->cpu_wq.v ? 0 : -ENOMEM;
  2782. }
  2783. static void free_cwqs(struct workqueue_struct *wq)
  2784. {
  2785. if (!(wq->flags & WQ_UNBOUND))
  2786. free_percpu(wq->cpu_wq.pcpu);
  2787. else if (wq->cpu_wq.single) {
  2788. /* the pointer to free is stored right after the cwq */
  2789. kfree(*(void **)(wq->cpu_wq.single + 1));
  2790. }
  2791. }
  2792. static int wq_clamp_max_active(int max_active, unsigned int flags,
  2793. const char *name)
  2794. {
  2795. int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
  2796. if (max_active < 1 || max_active > lim)
  2797. pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
  2798. max_active, name, 1, lim);
  2799. return clamp_val(max_active, 1, lim);
  2800. }
  2801. struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
  2802. unsigned int flags,
  2803. int max_active,
  2804. struct lock_class_key *key,
  2805. const char *lock_name, ...)
  2806. {
  2807. va_list args, args1;
  2808. struct workqueue_struct *wq;
  2809. unsigned int cpu;
  2810. size_t namelen;
  2811. /* determine namelen, allocate wq and format name */
  2812. va_start(args, lock_name);
  2813. va_copy(args1, args);
  2814. namelen = vsnprintf(NULL, 0, fmt, args) + 1;
  2815. wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
  2816. if (!wq)
  2817. goto err;
  2818. vsnprintf(wq->name, namelen, fmt, args1);
  2819. va_end(args);
  2820. va_end(args1);
  2821. /*
  2822. * Workqueues which may be used during memory reclaim should
  2823. * have a rescuer to guarantee forward progress.
  2824. */
  2825. if (flags & WQ_MEM_RECLAIM)
  2826. flags |= WQ_RESCUER;
  2827. max_active = max_active ?: WQ_DFL_ACTIVE;
  2828. max_active = wq_clamp_max_active(max_active, flags, wq->name);
  2829. /* init wq */
  2830. wq->flags = flags;
  2831. wq->saved_max_active = max_active;
  2832. mutex_init(&wq->flush_mutex);
  2833. atomic_set(&wq->nr_cwqs_to_flush, 0);
  2834. INIT_LIST_HEAD(&wq->flusher_queue);
  2835. INIT_LIST_HEAD(&wq->flusher_overflow);
  2836. lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
  2837. INIT_LIST_HEAD(&wq->list);
  2838. if (alloc_cwqs(wq) < 0)
  2839. goto err;
  2840. for_each_cwq_cpu(cpu, wq) {
  2841. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  2842. struct global_cwq *gcwq = get_gcwq(cpu);
  2843. int pool_idx = (bool)(flags & WQ_HIGHPRI);
  2844. BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
  2845. cwq->pool = &gcwq->pools[pool_idx];
  2846. cwq->wq = wq;
  2847. cwq->flush_color = -1;
  2848. cwq->max_active = max_active;
  2849. INIT_LIST_HEAD(&cwq->delayed_works);
  2850. }
  2851. if (flags & WQ_RESCUER) {
  2852. struct worker *rescuer;
  2853. if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
  2854. goto err;
  2855. wq->rescuer = rescuer = alloc_worker();
  2856. if (!rescuer)
  2857. goto err;
  2858. rescuer->task = kthread_create(rescuer_thread, wq, "%s",
  2859. wq->name);
  2860. if (IS_ERR(rescuer->task))
  2861. goto err;
  2862. rescuer->task->flags |= PF_THREAD_BOUND;
  2863. wake_up_process(rescuer->task);
  2864. }
  2865. /*
  2866. * workqueue_lock protects global freeze state and workqueues
  2867. * list. Grab it, set max_active accordingly and add the new
  2868. * workqueue to workqueues list.
  2869. */
  2870. spin_lock(&workqueue_lock);
  2871. if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
  2872. for_each_cwq_cpu(cpu, wq)
  2873. get_cwq(cpu, wq)->max_active = 0;
  2874. list_add(&wq->list, &workqueues);
  2875. spin_unlock(&workqueue_lock);
  2876. return wq;
  2877. err:
  2878. if (wq) {
  2879. free_cwqs(wq);
  2880. free_mayday_mask(wq->mayday_mask);
  2881. kfree(wq->rescuer);
  2882. kfree(wq);
  2883. }
  2884. return NULL;
  2885. }
  2886. EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
  2887. /**
  2888. * destroy_workqueue - safely terminate a workqueue
  2889. * @wq: target workqueue
  2890. *
  2891. * Safely destroy a workqueue. All work currently pending will be done first.
  2892. */
  2893. void destroy_workqueue(struct workqueue_struct *wq)
  2894. {
  2895. unsigned int cpu;
  2896. /* drain it before proceeding with destruction */
  2897. drain_workqueue(wq);
  2898. /*
  2899. * wq list is used to freeze wq, remove from list after
  2900. * flushing is complete in case freeze races us.
  2901. */
  2902. spin_lock(&workqueue_lock);
  2903. list_del(&wq->list);
  2904. spin_unlock(&workqueue_lock);
  2905. /* sanity check */
  2906. for_each_cwq_cpu(cpu, wq) {
  2907. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  2908. int i;
  2909. for (i = 0; i < WORK_NR_COLORS; i++)
  2910. BUG_ON(cwq->nr_in_flight[i]);
  2911. BUG_ON(cwq->nr_active);
  2912. BUG_ON(!list_empty(&cwq->delayed_works));
  2913. }
  2914. if (wq->flags & WQ_RESCUER) {
  2915. kthread_stop(wq->rescuer->task);
  2916. free_mayday_mask(wq->mayday_mask);
  2917. kfree(wq->rescuer);
  2918. }
  2919. free_cwqs(wq);
  2920. kfree(wq);
  2921. }
  2922. EXPORT_SYMBOL_GPL(destroy_workqueue);
  2923. /**
  2924. * cwq_set_max_active - adjust max_active of a cwq
  2925. * @cwq: target cpu_workqueue_struct
  2926. * @max_active: new max_active value.
  2927. *
  2928. * Set @cwq->max_active to @max_active and activate delayed works if
  2929. * increased.
  2930. *
  2931. * CONTEXT:
  2932. * spin_lock_irq(gcwq->lock).
  2933. */
  2934. static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
  2935. {
  2936. cwq->max_active = max_active;
  2937. while (!list_empty(&cwq->delayed_works) &&
  2938. cwq->nr_active < cwq->max_active)
  2939. cwq_activate_first_delayed(cwq);
  2940. }
  2941. /**
  2942. * workqueue_set_max_active - adjust max_active of a workqueue
  2943. * @wq: target workqueue
  2944. * @max_active: new max_active value.
  2945. *
  2946. * Set max_active of @wq to @max_active.
  2947. *
  2948. * CONTEXT:
  2949. * Don't call from IRQ context.
  2950. */
  2951. void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
  2952. {
  2953. unsigned int cpu;
  2954. max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
  2955. spin_lock(&workqueue_lock);
  2956. wq->saved_max_active = max_active;
  2957. for_each_cwq_cpu(cpu, wq) {
  2958. struct global_cwq *gcwq = get_gcwq(cpu);
  2959. spin_lock_irq(&gcwq->lock);
  2960. if (!(wq->flags & WQ_FREEZABLE) ||
  2961. !(gcwq->flags & GCWQ_FREEZING))
  2962. cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
  2963. spin_unlock_irq(&gcwq->lock);
  2964. }
  2965. spin_unlock(&workqueue_lock);
  2966. }
  2967. EXPORT_SYMBOL_GPL(workqueue_set_max_active);
  2968. /**
  2969. * workqueue_congested - test whether a workqueue is congested
  2970. * @cpu: CPU in question
  2971. * @wq: target workqueue
  2972. *
  2973. * Test whether @wq's cpu workqueue for @cpu is congested. There is
  2974. * no synchronization around this function and the test result is
  2975. * unreliable and only useful as advisory hints or for debugging.
  2976. *
  2977. * RETURNS:
  2978. * %true if congested, %false otherwise.
  2979. */
  2980. bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
  2981. {
  2982. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  2983. return !list_empty(&cwq->delayed_works);
  2984. }
  2985. EXPORT_SYMBOL_GPL(workqueue_congested);
  2986. /**
  2987. * work_cpu - return the last known associated cpu for @work
  2988. * @work: the work of interest
  2989. *
  2990. * RETURNS:
  2991. * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
  2992. */
  2993. unsigned int work_cpu(struct work_struct *work)
  2994. {
  2995. struct global_cwq *gcwq = get_work_gcwq(work);
  2996. return gcwq ? gcwq->cpu : WORK_CPU_NONE;
  2997. }
  2998. EXPORT_SYMBOL_GPL(work_cpu);
  2999. /**
  3000. * work_busy - test whether a work is currently pending or running
  3001. * @work: the work to be tested
  3002. *
  3003. * Test whether @work is currently pending or running. There is no
  3004. * synchronization around this function and the test result is
  3005. * unreliable and only useful as advisory hints or for debugging.
  3006. * Especially for reentrant wqs, the pending state might hide the
  3007. * running state.
  3008. *
  3009. * RETURNS:
  3010. * OR'd bitmask of WORK_BUSY_* bits.
  3011. */
  3012. unsigned int work_busy(struct work_struct *work)
  3013. {
  3014. struct global_cwq *gcwq = get_work_gcwq(work);
  3015. unsigned long flags;
  3016. unsigned int ret = 0;
  3017. if (!gcwq)
  3018. return false;
  3019. spin_lock_irqsave(&gcwq->lock, flags);
  3020. if (work_pending(work))
  3021. ret |= WORK_BUSY_PENDING;
  3022. if (find_worker_executing_work(gcwq, work))
  3023. ret |= WORK_BUSY_RUNNING;
  3024. spin_unlock_irqrestore(&gcwq->lock, flags);
  3025. return ret;
  3026. }
  3027. EXPORT_SYMBOL_GPL(work_busy);
  3028. /*
  3029. * CPU hotplug.
  3030. *
  3031. * There are two challenges in supporting CPU hotplug. Firstly, there
  3032. * are a lot of assumptions on strong associations among work, cwq and
  3033. * gcwq which make migrating pending and scheduled works very
  3034. * difficult to implement without impacting hot paths. Secondly,
  3035. * gcwqs serve mix of short, long and very long running works making
  3036. * blocked draining impractical.
  3037. *
  3038. * This is solved by allowing a gcwq to be disassociated from the CPU
  3039. * running as an unbound one and allowing it to be reattached later if the
  3040. * cpu comes back online.
  3041. */
  3042. /* claim manager positions of all pools */
  3043. static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
  3044. {
  3045. struct worker_pool *pool;
  3046. for_each_worker_pool(pool, gcwq)
  3047. mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
  3048. spin_lock_irq(&gcwq->lock);
  3049. }
  3050. /* release manager positions */
  3051. static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
  3052. {
  3053. struct worker_pool *pool;
  3054. spin_unlock_irq(&gcwq->lock);
  3055. for_each_worker_pool(pool, gcwq)
  3056. mutex_unlock(&pool->assoc_mutex);
  3057. }
  3058. static void gcwq_unbind_fn(struct work_struct *work)
  3059. {
  3060. struct global_cwq *gcwq = get_gcwq(smp_processor_id());
  3061. struct worker_pool *pool;
  3062. struct worker *worker;
  3063. struct hlist_node *pos;
  3064. int i;
  3065. BUG_ON(gcwq->cpu != smp_processor_id());
  3066. gcwq_claim_assoc_and_lock(gcwq);
  3067. /*
  3068. * We've claimed all manager positions. Make all workers unbound
  3069. * and set DISASSOCIATED. Before this, all workers except for the
  3070. * ones which are still executing works from before the last CPU
  3071. * down must be on the cpu. After this, they may become diasporas.
  3072. */
  3073. for_each_worker_pool(pool, gcwq)
  3074. list_for_each_entry(worker, &pool->idle_list, entry)
  3075. worker->flags |= WORKER_UNBOUND;
  3076. for_each_busy_worker(worker, i, pos, gcwq)
  3077. worker->flags |= WORKER_UNBOUND;
  3078. gcwq->flags |= GCWQ_DISASSOCIATED;
  3079. gcwq_release_assoc_and_unlock(gcwq);
  3080. /*
  3081. * Call schedule() so that we cross rq->lock and thus can guarantee
  3082. * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
  3083. * as scheduler callbacks may be invoked from other cpus.
  3084. */
  3085. schedule();
  3086. /*
  3087. * Sched callbacks are disabled now. Zap nr_running. After this,
  3088. * nr_running stays zero and need_more_worker() and keep_working()
  3089. * are always true as long as the worklist is not empty. @gcwq now
  3090. * behaves as unbound (in terms of concurrency management) gcwq
  3091. * which is served by workers tied to the CPU.
  3092. *
  3093. * On return from this function, the current worker would trigger
  3094. * unbound chain execution of pending work items if other workers
  3095. * didn't already.
  3096. */
  3097. for_each_worker_pool(pool, gcwq)
  3098. atomic_set(get_pool_nr_running(pool), 0);
  3099. }
  3100. /*
  3101. * Workqueues should be brought up before normal priority CPU notifiers.
  3102. * This will be registered high priority CPU notifier.
  3103. */
  3104. static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
  3105. unsigned long action,
  3106. void *hcpu)
  3107. {
  3108. unsigned int cpu = (unsigned long)hcpu;
  3109. struct global_cwq *gcwq = get_gcwq(cpu);
  3110. struct worker_pool *pool;
  3111. switch (action & ~CPU_TASKS_FROZEN) {
  3112. case CPU_UP_PREPARE:
  3113. for_each_worker_pool(pool, gcwq) {
  3114. struct worker *worker;
  3115. if (pool->nr_workers)
  3116. continue;
  3117. worker = create_worker(pool);
  3118. if (!worker)
  3119. return NOTIFY_BAD;
  3120. spin_lock_irq(&gcwq->lock);
  3121. start_worker(worker);
  3122. spin_unlock_irq(&gcwq->lock);
  3123. }
  3124. break;
  3125. case CPU_DOWN_FAILED:
  3126. case CPU_ONLINE:
  3127. gcwq_claim_assoc_and_lock(gcwq);
  3128. gcwq->flags &= ~GCWQ_DISASSOCIATED;
  3129. rebind_workers(gcwq);
  3130. gcwq_release_assoc_and_unlock(gcwq);
  3131. break;
  3132. }
  3133. return NOTIFY_OK;
  3134. }
  3135. /*
  3136. * Workqueues should be brought down after normal priority CPU notifiers.
  3137. * This will be registered as low priority CPU notifier.
  3138. */
  3139. static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
  3140. unsigned long action,
  3141. void *hcpu)
  3142. {
  3143. unsigned int cpu = (unsigned long)hcpu;
  3144. struct work_struct unbind_work;
  3145. switch (action & ~CPU_TASKS_FROZEN) {
  3146. case CPU_DOWN_PREPARE:
  3147. /* unbinding should happen on the local CPU */
  3148. INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
  3149. queue_work_on(cpu, system_highpri_wq, &unbind_work);
  3150. flush_work(&unbind_work);
  3151. break;
  3152. }
  3153. return NOTIFY_OK;
  3154. }
  3155. #ifdef CONFIG_SMP
  3156. struct work_for_cpu {
  3157. struct work_struct work;
  3158. long (*fn)(void *);
  3159. void *arg;
  3160. long ret;
  3161. };
  3162. static void work_for_cpu_fn(struct work_struct *work)
  3163. {
  3164. struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
  3165. wfc->ret = wfc->fn(wfc->arg);
  3166. }
  3167. /**
  3168. * work_on_cpu - run a function in user context on a particular cpu
  3169. * @cpu: the cpu to run on
  3170. * @fn: the function to run
  3171. * @arg: the function arg
  3172. *
  3173. * This will return the value @fn returns.
  3174. * It is up to the caller to ensure that the cpu doesn't go offline.
  3175. * The caller must not hold any locks which would prevent @fn from completing.
  3176. */
  3177. long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
  3178. {
  3179. struct work_for_cpu wfc = { .fn = fn, .arg = arg };
  3180. INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
  3181. schedule_work_on(cpu, &wfc.work);
  3182. flush_work(&wfc.work);
  3183. return wfc.ret;
  3184. }
  3185. EXPORT_SYMBOL_GPL(work_on_cpu);
  3186. #endif /* CONFIG_SMP */
  3187. #ifdef CONFIG_FREEZER
  3188. /**
  3189. * freeze_workqueues_begin - begin freezing workqueues
  3190. *
  3191. * Start freezing workqueues. After this function returns, all freezable
  3192. * workqueues will queue new works to their frozen_works list instead of
  3193. * gcwq->worklist.
  3194. *
  3195. * CONTEXT:
  3196. * Grabs and releases workqueue_lock and gcwq->lock's.
  3197. */
  3198. void freeze_workqueues_begin(void)
  3199. {
  3200. unsigned int cpu;
  3201. spin_lock(&workqueue_lock);
  3202. BUG_ON(workqueue_freezing);
  3203. workqueue_freezing = true;
  3204. for_each_gcwq_cpu(cpu) {
  3205. struct global_cwq *gcwq = get_gcwq(cpu);
  3206. struct workqueue_struct *wq;
  3207. spin_lock_irq(&gcwq->lock);
  3208. BUG_ON(gcwq->flags & GCWQ_FREEZING);
  3209. gcwq->flags |= GCWQ_FREEZING;
  3210. list_for_each_entry(wq, &workqueues, list) {
  3211. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  3212. if (cwq && wq->flags & WQ_FREEZABLE)
  3213. cwq->max_active = 0;
  3214. }
  3215. spin_unlock_irq(&gcwq->lock);
  3216. }
  3217. spin_unlock(&workqueue_lock);
  3218. }
  3219. /**
  3220. * freeze_workqueues_busy - are freezable workqueues still busy?
  3221. *
  3222. * Check whether freezing is complete. This function must be called
  3223. * between freeze_workqueues_begin() and thaw_workqueues().
  3224. *
  3225. * CONTEXT:
  3226. * Grabs and releases workqueue_lock.
  3227. *
  3228. * RETURNS:
  3229. * %true if some freezable workqueues are still busy. %false if freezing
  3230. * is complete.
  3231. */
  3232. bool freeze_workqueues_busy(void)
  3233. {
  3234. unsigned int cpu;
  3235. bool busy = false;
  3236. spin_lock(&workqueue_lock);
  3237. BUG_ON(!workqueue_freezing);
  3238. for_each_gcwq_cpu(cpu) {
  3239. struct workqueue_struct *wq;
  3240. /*
  3241. * nr_active is monotonically decreasing. It's safe
  3242. * to peek without lock.
  3243. */
  3244. list_for_each_entry(wq, &workqueues, list) {
  3245. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  3246. if (!cwq || !(wq->flags & WQ_FREEZABLE))
  3247. continue;
  3248. BUG_ON(cwq->nr_active < 0);
  3249. if (cwq->nr_active) {
  3250. busy = true;
  3251. goto out_unlock;
  3252. }
  3253. }
  3254. }
  3255. out_unlock:
  3256. spin_unlock(&workqueue_lock);
  3257. return busy;
  3258. }
  3259. /**
  3260. * thaw_workqueues - thaw workqueues
  3261. *
  3262. * Thaw workqueues. Normal queueing is restored and all collected
  3263. * frozen works are transferred to their respective gcwq worklists.
  3264. *
  3265. * CONTEXT:
  3266. * Grabs and releases workqueue_lock and gcwq->lock's.
  3267. */
  3268. void thaw_workqueues(void)
  3269. {
  3270. unsigned int cpu;
  3271. spin_lock(&workqueue_lock);
  3272. if (!workqueue_freezing)
  3273. goto out_unlock;
  3274. for_each_gcwq_cpu(cpu) {
  3275. struct global_cwq *gcwq = get_gcwq(cpu);
  3276. struct worker_pool *pool;
  3277. struct workqueue_struct *wq;
  3278. spin_lock_irq(&gcwq->lock);
  3279. BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
  3280. gcwq->flags &= ~GCWQ_FREEZING;
  3281. list_for_each_entry(wq, &workqueues, list) {
  3282. struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
  3283. if (!cwq || !(wq->flags & WQ_FREEZABLE))
  3284. continue;
  3285. /* restore max_active and repopulate worklist */
  3286. cwq_set_max_active(cwq, wq->saved_max_active);
  3287. }
  3288. for_each_worker_pool(pool, gcwq)
  3289. wake_up_worker(pool);
  3290. spin_unlock_irq(&gcwq->lock);
  3291. }
  3292. workqueue_freezing = false;
  3293. out_unlock:
  3294. spin_unlock(&workqueue_lock);
  3295. }
  3296. #endif /* CONFIG_FREEZER */
  3297. static int __init init_workqueues(void)
  3298. {
  3299. unsigned int cpu;
  3300. int i;
  3301. /* make sure we have enough bits for OFFQ CPU number */
  3302. BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
  3303. WORK_CPU_LAST);
  3304. cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
  3305. hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
  3306. /* initialize gcwqs */
  3307. for_each_gcwq_cpu(cpu) {
  3308. struct global_cwq *gcwq = get_gcwq(cpu);
  3309. struct worker_pool *pool;
  3310. spin_lock_init(&gcwq->lock);
  3311. gcwq->cpu = cpu;
  3312. gcwq->flags |= GCWQ_DISASSOCIATED;
  3313. for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
  3314. INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
  3315. for_each_worker_pool(pool, gcwq) {
  3316. pool->gcwq = gcwq;
  3317. INIT_LIST_HEAD(&pool->worklist);
  3318. INIT_LIST_HEAD(&pool->idle_list);
  3319. init_timer_deferrable(&pool->idle_timer);
  3320. pool->idle_timer.function = idle_worker_timeout;
  3321. pool->idle_timer.data = (unsigned long)pool;
  3322. setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
  3323. (unsigned long)pool);
  3324. mutex_init(&pool->assoc_mutex);
  3325. ida_init(&pool->worker_ida);
  3326. }
  3327. }
  3328. /* create the initial worker */
  3329. for_each_online_gcwq_cpu(cpu) {
  3330. struct global_cwq *gcwq = get_gcwq(cpu);
  3331. struct worker_pool *pool;
  3332. if (cpu != WORK_CPU_UNBOUND)
  3333. gcwq->flags &= ~GCWQ_DISASSOCIATED;
  3334. for_each_worker_pool(pool, gcwq) {
  3335. struct worker *worker;
  3336. worker = create_worker(pool);
  3337. BUG_ON(!worker);
  3338. spin_lock_irq(&gcwq->lock);
  3339. start_worker(worker);
  3340. spin_unlock_irq(&gcwq->lock);
  3341. }
  3342. }
  3343. system_wq = alloc_workqueue("events", 0, 0);
  3344. system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
  3345. system_long_wq = alloc_workqueue("events_long", 0, 0);
  3346. system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
  3347. WQ_UNBOUND_MAX_ACTIVE);
  3348. system_freezable_wq = alloc_workqueue("events_freezable",
  3349. WQ_FREEZABLE, 0);
  3350. BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
  3351. !system_unbound_wq || !system_freezable_wq);
  3352. return 0;
  3353. }
  3354. early_initcall(init_workqueues);