slab.c 116 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890389138923893389438953896389738983899390039013902390339043905390639073908390939103911391239133914391539163917391839193920392139223923392439253926392739283929393039313932393339343935393639373938393939403941394239433944394539463947394839493950395139523953395439553956395739583959396039613962396339643965396639673968396939703971397239733974397539763977397839793980398139823983398439853986398739883989399039913992399339943995399639973998399940004001400240034004400540064007400840094010401140124013401440154016401740184019402040214022402340244025402640274028402940304031403240334034403540364037403840394040404140424043404440454046404740484049405040514052405340544055405640574058405940604061406240634064406540664067406840694070407140724073407440754076407740784079408040814082408340844085408640874088408940904091409240934094409540964097409840994100410141024103410441054106410741084109411041114112411341144115411641174118411941204121412241234124412541264127412841294130413141324133413441354136413741384139414041414142414341444145414641474148414941504151415241534154415541564157415841594160416141624163416441654166416741684169417041714172417341744175417641774178417941804181418241834184418541864187418841894190419141924193419441954196419741984199420042014202420342044205420642074208420942104211421242134214421542164217421842194220422142224223422442254226422742284229423042314232423342344235423642374238423942404241424242434244424542464247424842494250425142524253425442554256425742584259426042614262426342644265426642674268426942704271427242734274427542764277427842794280428142824283428442854286428742884289429042914292429342944295429642974298429943004301430243034304430543064307430843094310431143124313431443154316431743184319432043214322432343244325432643274328432943304331433243334334433543364337433843394340434143424343434443454346434743484349435043514352435343544355435643574358435943604361436243634364436543664367436843694370437143724373437443754376437743784379438043814382438343844385438643874388438943904391439243934394439543964397439843994400440144024403440444054406440744084409441044114412441344144415441644174418441944204421442244234424442544264427442844294430443144324433443444354436443744384439444044414442444344444445444644474448444944504451445244534454445544564457445844594460446144624463446444654466446744684469447044714472447344744475447644774478447944804481448244834484448544864487
  1. /*
  2. * linux/mm/slab.c
  3. * Written by Mark Hemment, 1996/97.
  4. * (markhe@nextd.demon.co.uk)
  5. *
  6. * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
  7. *
  8. * Major cleanup, different bufctl logic, per-cpu arrays
  9. * (c) 2000 Manfred Spraul
  10. *
  11. * Cleanup, make the head arrays unconditional, preparation for NUMA
  12. * (c) 2002 Manfred Spraul
  13. *
  14. * An implementation of the Slab Allocator as described in outline in;
  15. * UNIX Internals: The New Frontiers by Uresh Vahalia
  16. * Pub: Prentice Hall ISBN 0-13-101908-2
  17. * or with a little more detail in;
  18. * The Slab Allocator: An Object-Caching Kernel Memory Allocator
  19. * Jeff Bonwick (Sun Microsystems).
  20. * Presented at: USENIX Summer 1994 Technical Conference
  21. *
  22. * The memory is organized in caches, one cache for each object type.
  23. * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
  24. * Each cache consists out of many slabs (they are small (usually one
  25. * page long) and always contiguous), and each slab contains multiple
  26. * initialized objects.
  27. *
  28. * This means, that your constructor is used only for newly allocated
  29. * slabs and you must pass objects with the same intializations to
  30. * kmem_cache_free.
  31. *
  32. * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
  33. * normal). If you need a special memory type, then must create a new
  34. * cache for that memory type.
  35. *
  36. * In order to reduce fragmentation, the slabs are sorted in 3 groups:
  37. * full slabs with 0 free objects
  38. * partial slabs
  39. * empty slabs with no allocated objects
  40. *
  41. * If partial slabs exist, then new allocations come from these slabs,
  42. * otherwise from empty slabs or new slabs are allocated.
  43. *
  44. * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
  45. * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
  46. *
  47. * Each cache has a short per-cpu head array, most allocs
  48. * and frees go into that array, and if that array overflows, then 1/2
  49. * of the entries in the array are given back into the global cache.
  50. * The head array is strictly LIFO and should improve the cache hit rates.
  51. * On SMP, it additionally reduces the spinlock operations.
  52. *
  53. * The c_cpuarray may not be read with enabled local interrupts -
  54. * it's changed with a smp_call_function().
  55. *
  56. * SMP synchronization:
  57. * constructors and destructors are called without any locking.
  58. * Several members in struct kmem_cache and struct slab never change, they
  59. * are accessed without any locking.
  60. * The per-cpu arrays are never accessed from the wrong cpu, no locking,
  61. * and local interrupts are disabled so slab code is preempt-safe.
  62. * The non-constant members are protected with a per-cache irq spinlock.
  63. *
  64. * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
  65. * in 2000 - many ideas in the current implementation are derived from
  66. * his patch.
  67. *
  68. * Further notes from the original documentation:
  69. *
  70. * 11 April '97. Started multi-threading - markhe
  71. * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
  72. * The sem is only needed when accessing/extending the cache-chain, which
  73. * can never happen inside an interrupt (kmem_cache_create(),
  74. * kmem_cache_shrink() and kmem_cache_reap()).
  75. *
  76. * At present, each engine can be growing a cache. This should be blocked.
  77. *
  78. * 15 March 2005. NUMA slab allocator.
  79. * Shai Fultheim <shai@scalex86.org>.
  80. * Shobhit Dayal <shobhit@calsoftinc.com>
  81. * Alok N Kataria <alokk@calsoftinc.com>
  82. * Christoph Lameter <christoph@lameter.com>
  83. *
  84. * Modified the slab allocator to be node aware on NUMA systems.
  85. * Each node has its own list of partial, free and full slabs.
  86. * All object allocations for a node occur from node specific slab lists.
  87. */
  88. #include <linux/slab.h>
  89. #include <linux/mm.h>
  90. #include <linux/poison.h>
  91. #include <linux/swap.h>
  92. #include <linux/cache.h>
  93. #include <linux/interrupt.h>
  94. #include <linux/init.h>
  95. #include <linux/compiler.h>
  96. #include <linux/cpuset.h>
  97. #include <linux/seq_file.h>
  98. #include <linux/notifier.h>
  99. #include <linux/kallsyms.h>
  100. #include <linux/cpu.h>
  101. #include <linux/sysctl.h>
  102. #include <linux/module.h>
  103. #include <linux/rcupdate.h>
  104. #include <linux/string.h>
  105. #include <linux/uaccess.h>
  106. #include <linux/nodemask.h>
  107. #include <linux/mempolicy.h>
  108. #include <linux/mutex.h>
  109. #include <linux/fault-inject.h>
  110. #include <linux/rtmutex.h>
  111. #include <linux/reciprocal_div.h>
  112. #include <asm/cacheflush.h>
  113. #include <asm/tlbflush.h>
  114. #include <asm/page.h>
  115. /*
  116. * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
  117. * SLAB_RED_ZONE & SLAB_POISON.
  118. * 0 for faster, smaller code (especially in the critical paths).
  119. *
  120. * STATS - 1 to collect stats for /proc/slabinfo.
  121. * 0 for faster, smaller code (especially in the critical paths).
  122. *
  123. * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
  124. */
  125. #ifdef CONFIG_DEBUG_SLAB
  126. #define DEBUG 1
  127. #define STATS 1
  128. #define FORCED_DEBUG 1
  129. #else
  130. #define DEBUG 0
  131. #define STATS 0
  132. #define FORCED_DEBUG 0
  133. #endif
  134. /* Shouldn't this be in a header file somewhere? */
  135. #define BYTES_PER_WORD sizeof(void *)
  136. #ifndef cache_line_size
  137. #define cache_line_size() L1_CACHE_BYTES
  138. #endif
  139. #ifndef ARCH_KMALLOC_MINALIGN
  140. /*
  141. * Enforce a minimum alignment for the kmalloc caches.
  142. * Usually, the kmalloc caches are cache_line_size() aligned, except when
  143. * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
  144. * Some archs want to perform DMA into kmalloc caches and need a guaranteed
  145. * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
  146. * Note that this flag disables some debug features.
  147. */
  148. #define ARCH_KMALLOC_MINALIGN 0
  149. #endif
  150. #ifndef ARCH_SLAB_MINALIGN
  151. /*
  152. * Enforce a minimum alignment for all caches.
  153. * Intended for archs that get misalignment faults even for BYTES_PER_WORD
  154. * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
  155. * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
  156. * some debug features.
  157. */
  158. #define ARCH_SLAB_MINALIGN 0
  159. #endif
  160. #ifndef ARCH_KMALLOC_FLAGS
  161. #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
  162. #endif
  163. /* Legal flag mask for kmem_cache_create(). */
  164. #if DEBUG
  165. # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
  166. SLAB_POISON | SLAB_HWCACHE_ALIGN | \
  167. SLAB_CACHE_DMA | \
  168. SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
  169. SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
  170. SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
  171. #else
  172. # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
  173. SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
  174. SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
  175. SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
  176. #endif
  177. /*
  178. * kmem_bufctl_t:
  179. *
  180. * Bufctl's are used for linking objs within a slab
  181. * linked offsets.
  182. *
  183. * This implementation relies on "struct page" for locating the cache &
  184. * slab an object belongs to.
  185. * This allows the bufctl structure to be small (one int), but limits
  186. * the number of objects a slab (not a cache) can contain when off-slab
  187. * bufctls are used. The limit is the size of the largest general cache
  188. * that does not use off-slab slabs.
  189. * For 32bit archs with 4 kB pages, is this 56.
  190. * This is not serious, as it is only for large objects, when it is unwise
  191. * to have too many per slab.
  192. * Note: This limit can be raised by introducing a general cache whose size
  193. * is less than 512 (PAGE_SIZE<<3), but greater than 256.
  194. */
  195. typedef unsigned int kmem_bufctl_t;
  196. #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
  197. #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
  198. #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
  199. #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
  200. /*
  201. * struct slab
  202. *
  203. * Manages the objs in a slab. Placed either at the beginning of mem allocated
  204. * for a slab, or allocated from an general cache.
  205. * Slabs are chained into three list: fully used, partial, fully free slabs.
  206. */
  207. struct slab {
  208. struct list_head list;
  209. unsigned long colouroff;
  210. void *s_mem; /* including colour offset */
  211. unsigned int inuse; /* num of objs active in slab */
  212. kmem_bufctl_t free;
  213. unsigned short nodeid;
  214. };
  215. /*
  216. * struct slab_rcu
  217. *
  218. * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
  219. * arrange for kmem_freepages to be called via RCU. This is useful if
  220. * we need to approach a kernel structure obliquely, from its address
  221. * obtained without the usual locking. We can lock the structure to
  222. * stabilize it and check it's still at the given address, only if we
  223. * can be sure that the memory has not been meanwhile reused for some
  224. * other kind of object (which our subsystem's lock might corrupt).
  225. *
  226. * rcu_read_lock before reading the address, then rcu_read_unlock after
  227. * taking the spinlock within the structure expected at that address.
  228. *
  229. * We assume struct slab_rcu can overlay struct slab when destroying.
  230. */
  231. struct slab_rcu {
  232. struct rcu_head head;
  233. struct kmem_cache *cachep;
  234. void *addr;
  235. };
  236. /*
  237. * struct array_cache
  238. *
  239. * Purpose:
  240. * - LIFO ordering, to hand out cache-warm objects from _alloc
  241. * - reduce the number of linked list operations
  242. * - reduce spinlock operations
  243. *
  244. * The limit is stored in the per-cpu structure to reduce the data cache
  245. * footprint.
  246. *
  247. */
  248. struct array_cache {
  249. unsigned int avail;
  250. unsigned int limit;
  251. unsigned int batchcount;
  252. unsigned int touched;
  253. spinlock_t lock;
  254. void *entry[0]; /*
  255. * Must have this definition in here for the proper
  256. * alignment of array_cache. Also simplifies accessing
  257. * the entries.
  258. * [0] is for gcc 2.95. It should really be [].
  259. */
  260. };
  261. /*
  262. * bootstrap: The caches do not work without cpuarrays anymore, but the
  263. * cpuarrays are allocated from the generic caches...
  264. */
  265. #define BOOT_CPUCACHE_ENTRIES 1
  266. struct arraycache_init {
  267. struct array_cache cache;
  268. void *entries[BOOT_CPUCACHE_ENTRIES];
  269. };
  270. /*
  271. * The slab lists for all objects.
  272. */
  273. struct kmem_list3 {
  274. struct list_head slabs_partial; /* partial list first, better asm code */
  275. struct list_head slabs_full;
  276. struct list_head slabs_free;
  277. unsigned long free_objects;
  278. unsigned int free_limit;
  279. unsigned int colour_next; /* Per-node cache coloring */
  280. spinlock_t list_lock;
  281. struct array_cache *shared; /* shared per node */
  282. struct array_cache **alien; /* on other nodes */
  283. unsigned long next_reap; /* updated without locking */
  284. int free_touched; /* updated without locking */
  285. };
  286. /*
  287. * Need this for bootstrapping a per node allocator.
  288. */
  289. #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
  290. struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
  291. #define CACHE_CACHE 0
  292. #define SIZE_AC 1
  293. #define SIZE_L3 (1 + MAX_NUMNODES)
  294. static int drain_freelist(struct kmem_cache *cache,
  295. struct kmem_list3 *l3, int tofree);
  296. static void free_block(struct kmem_cache *cachep, void **objpp, int len,
  297. int node);
  298. static int enable_cpucache(struct kmem_cache *cachep);
  299. static void cache_reap(struct work_struct *unused);
  300. /*
  301. * This function must be completely optimized away if a constant is passed to
  302. * it. Mostly the same as what is in linux/slab.h except it returns an index.
  303. */
  304. static __always_inline int index_of(const size_t size)
  305. {
  306. extern void __bad_size(void);
  307. if (__builtin_constant_p(size)) {
  308. int i = 0;
  309. #define CACHE(x) \
  310. if (size <=x) \
  311. return i; \
  312. else \
  313. i++;
  314. #include "linux/kmalloc_sizes.h"
  315. #undef CACHE
  316. __bad_size();
  317. } else
  318. __bad_size();
  319. return 0;
  320. }
  321. static int slab_early_init = 1;
  322. #define INDEX_AC index_of(sizeof(struct arraycache_init))
  323. #define INDEX_L3 index_of(sizeof(struct kmem_list3))
  324. static void kmem_list3_init(struct kmem_list3 *parent)
  325. {
  326. INIT_LIST_HEAD(&parent->slabs_full);
  327. INIT_LIST_HEAD(&parent->slabs_partial);
  328. INIT_LIST_HEAD(&parent->slabs_free);
  329. parent->shared = NULL;
  330. parent->alien = NULL;
  331. parent->colour_next = 0;
  332. spin_lock_init(&parent->list_lock);
  333. parent->free_objects = 0;
  334. parent->free_touched = 0;
  335. }
  336. #define MAKE_LIST(cachep, listp, slab, nodeid) \
  337. do { \
  338. INIT_LIST_HEAD(listp); \
  339. list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
  340. } while (0)
  341. #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
  342. do { \
  343. MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
  344. MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
  345. MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
  346. } while (0)
  347. /*
  348. * struct kmem_cache
  349. *
  350. * manages a cache.
  351. */
  352. struct kmem_cache {
  353. /* 1) per-cpu data, touched during every alloc/free */
  354. struct array_cache *array[NR_CPUS];
  355. /* 2) Cache tunables. Protected by cache_chain_mutex */
  356. unsigned int batchcount;
  357. unsigned int limit;
  358. unsigned int shared;
  359. unsigned int buffer_size;
  360. u32 reciprocal_buffer_size;
  361. /* 3) touched by every alloc & free from the backend */
  362. struct kmem_list3 *nodelists[MAX_NUMNODES];
  363. unsigned int flags; /* constant flags */
  364. unsigned int num; /* # of objs per slab */
  365. /* 4) cache_grow/shrink */
  366. /* order of pgs per slab (2^n) */
  367. unsigned int gfporder;
  368. /* force GFP flags, e.g. GFP_DMA */
  369. gfp_t gfpflags;
  370. size_t colour; /* cache colouring range */
  371. unsigned int colour_off; /* colour offset */
  372. struct kmem_cache *slabp_cache;
  373. unsigned int slab_size;
  374. unsigned int dflags; /* dynamic flags */
  375. /* constructor func */
  376. void (*ctor) (void *, struct kmem_cache *, unsigned long);
  377. /* de-constructor func */
  378. void (*dtor) (void *, struct kmem_cache *, unsigned long);
  379. /* 5) cache creation/removal */
  380. const char *name;
  381. struct list_head next;
  382. /* 6) statistics */
  383. #if STATS
  384. unsigned long num_active;
  385. unsigned long num_allocations;
  386. unsigned long high_mark;
  387. unsigned long grown;
  388. unsigned long reaped;
  389. unsigned long errors;
  390. unsigned long max_freeable;
  391. unsigned long node_allocs;
  392. unsigned long node_frees;
  393. unsigned long node_overflow;
  394. atomic_t allochit;
  395. atomic_t allocmiss;
  396. atomic_t freehit;
  397. atomic_t freemiss;
  398. #endif
  399. #if DEBUG
  400. /*
  401. * If debugging is enabled, then the allocator can add additional
  402. * fields and/or padding to every object. buffer_size contains the total
  403. * object size including these internal fields, the following two
  404. * variables contain the offset to the user object and its size.
  405. */
  406. int obj_offset;
  407. int obj_size;
  408. #endif
  409. };
  410. #define CFLGS_OFF_SLAB (0x80000000UL)
  411. #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
  412. #define BATCHREFILL_LIMIT 16
  413. /*
  414. * Optimization question: fewer reaps means less probability for unnessary
  415. * cpucache drain/refill cycles.
  416. *
  417. * OTOH the cpuarrays can contain lots of objects,
  418. * which could lock up otherwise freeable slabs.
  419. */
  420. #define REAPTIMEOUT_CPUC (2*HZ)
  421. #define REAPTIMEOUT_LIST3 (4*HZ)
  422. #if STATS
  423. #define STATS_INC_ACTIVE(x) ((x)->num_active++)
  424. #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
  425. #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
  426. #define STATS_INC_GROWN(x) ((x)->grown++)
  427. #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
  428. #define STATS_SET_HIGH(x) \
  429. do { \
  430. if ((x)->num_active > (x)->high_mark) \
  431. (x)->high_mark = (x)->num_active; \
  432. } while (0)
  433. #define STATS_INC_ERR(x) ((x)->errors++)
  434. #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
  435. #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
  436. #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
  437. #define STATS_SET_FREEABLE(x, i) \
  438. do { \
  439. if ((x)->max_freeable < i) \
  440. (x)->max_freeable = i; \
  441. } while (0)
  442. #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
  443. #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
  444. #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
  445. #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
  446. #else
  447. #define STATS_INC_ACTIVE(x) do { } while (0)
  448. #define STATS_DEC_ACTIVE(x) do { } while (0)
  449. #define STATS_INC_ALLOCED(x) do { } while (0)
  450. #define STATS_INC_GROWN(x) do { } while (0)
  451. #define STATS_ADD_REAPED(x,y) do { } while (0)
  452. #define STATS_SET_HIGH(x) do { } while (0)
  453. #define STATS_INC_ERR(x) do { } while (0)
  454. #define STATS_INC_NODEALLOCS(x) do { } while (0)
  455. #define STATS_INC_NODEFREES(x) do { } while (0)
  456. #define STATS_INC_ACOVERFLOW(x) do { } while (0)
  457. #define STATS_SET_FREEABLE(x, i) do { } while (0)
  458. #define STATS_INC_ALLOCHIT(x) do { } while (0)
  459. #define STATS_INC_ALLOCMISS(x) do { } while (0)
  460. #define STATS_INC_FREEHIT(x) do { } while (0)
  461. #define STATS_INC_FREEMISS(x) do { } while (0)
  462. #endif
  463. #if DEBUG
  464. /*
  465. * memory layout of objects:
  466. * 0 : objp
  467. * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
  468. * the end of an object is aligned with the end of the real
  469. * allocation. Catches writes behind the end of the allocation.
  470. * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
  471. * redzone word.
  472. * cachep->obj_offset: The real object.
  473. * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
  474. * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
  475. * [BYTES_PER_WORD long]
  476. */
  477. static int obj_offset(struct kmem_cache *cachep)
  478. {
  479. return cachep->obj_offset;
  480. }
  481. static int obj_size(struct kmem_cache *cachep)
  482. {
  483. return cachep->obj_size;
  484. }
  485. static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
  486. {
  487. BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
  488. return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
  489. }
  490. static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
  491. {
  492. BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
  493. if (cachep->flags & SLAB_STORE_USER)
  494. return (unsigned long *)(objp + cachep->buffer_size -
  495. 2 * BYTES_PER_WORD);
  496. return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
  497. }
  498. static void **dbg_userword(struct kmem_cache *cachep, void *objp)
  499. {
  500. BUG_ON(!(cachep->flags & SLAB_STORE_USER));
  501. return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
  502. }
  503. #else
  504. #define obj_offset(x) 0
  505. #define obj_size(cachep) (cachep->buffer_size)
  506. #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
  507. #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
  508. #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
  509. #endif
  510. /*
  511. * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
  512. * order.
  513. */
  514. #if defined(CONFIG_LARGE_ALLOCS)
  515. #define MAX_OBJ_ORDER 13 /* up to 32Mb */
  516. #define MAX_GFP_ORDER 13 /* up to 32Mb */
  517. #elif defined(CONFIG_MMU)
  518. #define MAX_OBJ_ORDER 5 /* 32 pages */
  519. #define MAX_GFP_ORDER 5 /* 32 pages */
  520. #else
  521. #define MAX_OBJ_ORDER 8 /* up to 1Mb */
  522. #define MAX_GFP_ORDER 8 /* up to 1Mb */
  523. #endif
  524. /*
  525. * Do not go above this order unless 0 objects fit into the slab.
  526. */
  527. #define BREAK_GFP_ORDER_HI 1
  528. #define BREAK_GFP_ORDER_LO 0
  529. static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
  530. /*
  531. * Functions for storing/retrieving the cachep and or slab from the page
  532. * allocator. These are used to find the slab an obj belongs to. With kfree(),
  533. * these are used to find the cache which an obj belongs to.
  534. */
  535. static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
  536. {
  537. page->lru.next = (struct list_head *)cache;
  538. }
  539. static inline struct kmem_cache *page_get_cache(struct page *page)
  540. {
  541. if (unlikely(PageCompound(page)))
  542. page = (struct page *)page_private(page);
  543. BUG_ON(!PageSlab(page));
  544. return (struct kmem_cache *)page->lru.next;
  545. }
  546. static inline void page_set_slab(struct page *page, struct slab *slab)
  547. {
  548. page->lru.prev = (struct list_head *)slab;
  549. }
  550. static inline struct slab *page_get_slab(struct page *page)
  551. {
  552. if (unlikely(PageCompound(page)))
  553. page = (struct page *)page_private(page);
  554. BUG_ON(!PageSlab(page));
  555. return (struct slab *)page->lru.prev;
  556. }
  557. static inline struct kmem_cache *virt_to_cache(const void *obj)
  558. {
  559. struct page *page = virt_to_page(obj);
  560. return page_get_cache(page);
  561. }
  562. static inline struct slab *virt_to_slab(const void *obj)
  563. {
  564. struct page *page = virt_to_page(obj);
  565. return page_get_slab(page);
  566. }
  567. static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
  568. unsigned int idx)
  569. {
  570. return slab->s_mem + cache->buffer_size * idx;
  571. }
  572. /*
  573. * We want to avoid an expensive divide : (offset / cache->buffer_size)
  574. * Using the fact that buffer_size is a constant for a particular cache,
  575. * we can replace (offset / cache->buffer_size) by
  576. * reciprocal_divide(offset, cache->reciprocal_buffer_size)
  577. */
  578. static inline unsigned int obj_to_index(const struct kmem_cache *cache,
  579. const struct slab *slab, void *obj)
  580. {
  581. u32 offset = (obj - slab->s_mem);
  582. return reciprocal_divide(offset, cache->reciprocal_buffer_size);
  583. }
  584. /*
  585. * These are the default caches for kmalloc. Custom caches can have other sizes.
  586. */
  587. struct cache_sizes malloc_sizes[] = {
  588. #define CACHE(x) { .cs_size = (x) },
  589. #include <linux/kmalloc_sizes.h>
  590. CACHE(ULONG_MAX)
  591. #undef CACHE
  592. };
  593. EXPORT_SYMBOL(malloc_sizes);
  594. /* Must match cache_sizes above. Out of line to keep cache footprint low. */
  595. struct cache_names {
  596. char *name;
  597. char *name_dma;
  598. };
  599. static struct cache_names __initdata cache_names[] = {
  600. #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
  601. #include <linux/kmalloc_sizes.h>
  602. {NULL,}
  603. #undef CACHE
  604. };
  605. static struct arraycache_init initarray_cache __initdata =
  606. { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
  607. static struct arraycache_init initarray_generic =
  608. { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
  609. /* internal cache of cache description objs */
  610. static struct kmem_cache cache_cache = {
  611. .batchcount = 1,
  612. .limit = BOOT_CPUCACHE_ENTRIES,
  613. .shared = 1,
  614. .buffer_size = sizeof(struct kmem_cache),
  615. .name = "kmem_cache",
  616. #if DEBUG
  617. .obj_size = sizeof(struct kmem_cache),
  618. #endif
  619. };
  620. #define BAD_ALIEN_MAGIC 0x01020304ul
  621. #ifdef CONFIG_LOCKDEP
  622. /*
  623. * Slab sometimes uses the kmalloc slabs to store the slab headers
  624. * for other slabs "off slab".
  625. * The locking for this is tricky in that it nests within the locks
  626. * of all other slabs in a few places; to deal with this special
  627. * locking we put on-slab caches into a separate lock-class.
  628. *
  629. * We set lock class for alien array caches which are up during init.
  630. * The lock annotation will be lost if all cpus of a node goes down and
  631. * then comes back up during hotplug
  632. */
  633. static struct lock_class_key on_slab_l3_key;
  634. static struct lock_class_key on_slab_alc_key;
  635. static inline void init_lock_keys(void)
  636. {
  637. int q;
  638. struct cache_sizes *s = malloc_sizes;
  639. while (s->cs_size != ULONG_MAX) {
  640. for_each_node(q) {
  641. struct array_cache **alc;
  642. int r;
  643. struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
  644. if (!l3 || OFF_SLAB(s->cs_cachep))
  645. continue;
  646. lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
  647. alc = l3->alien;
  648. /*
  649. * FIXME: This check for BAD_ALIEN_MAGIC
  650. * should go away when common slab code is taught to
  651. * work even without alien caches.
  652. * Currently, non NUMA code returns BAD_ALIEN_MAGIC
  653. * for alloc_alien_cache,
  654. */
  655. if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
  656. continue;
  657. for_each_node(r) {
  658. if (alc[r])
  659. lockdep_set_class(&alc[r]->lock,
  660. &on_slab_alc_key);
  661. }
  662. }
  663. s++;
  664. }
  665. }
  666. #else
  667. static inline void init_lock_keys(void)
  668. {
  669. }
  670. #endif
  671. /*
  672. * 1. Guard access to the cache-chain.
  673. * 2. Protect sanity of cpu_online_map against cpu hotplug events
  674. */
  675. static DEFINE_MUTEX(cache_chain_mutex);
  676. static struct list_head cache_chain;
  677. /*
  678. * chicken and egg problem: delay the per-cpu array allocation
  679. * until the general caches are up.
  680. */
  681. static enum {
  682. NONE,
  683. PARTIAL_AC,
  684. PARTIAL_L3,
  685. FULL
  686. } g_cpucache_up;
  687. /*
  688. * used by boot code to determine if it can use slab based allocator
  689. */
  690. int slab_is_available(void)
  691. {
  692. return g_cpucache_up == FULL;
  693. }
  694. static DEFINE_PER_CPU(struct delayed_work, reap_work);
  695. static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
  696. {
  697. return cachep->array[smp_processor_id()];
  698. }
  699. static inline struct kmem_cache *__find_general_cachep(size_t size,
  700. gfp_t gfpflags)
  701. {
  702. struct cache_sizes *csizep = malloc_sizes;
  703. #if DEBUG
  704. /* This happens if someone tries to call
  705. * kmem_cache_create(), or __kmalloc(), before
  706. * the generic caches are initialized.
  707. */
  708. BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
  709. #endif
  710. while (size > csizep->cs_size)
  711. csizep++;
  712. /*
  713. * Really subtle: The last entry with cs->cs_size==ULONG_MAX
  714. * has cs_{dma,}cachep==NULL. Thus no special case
  715. * for large kmalloc calls required.
  716. */
  717. #ifdef CONFIG_ZONE_DMA
  718. if (unlikely(gfpflags & GFP_DMA))
  719. return csizep->cs_dmacachep;
  720. #endif
  721. return csizep->cs_cachep;
  722. }
  723. static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
  724. {
  725. return __find_general_cachep(size, gfpflags);
  726. }
  727. static size_t slab_mgmt_size(size_t nr_objs, size_t align)
  728. {
  729. return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
  730. }
  731. /*
  732. * Calculate the number of objects and left-over bytes for a given buffer size.
  733. */
  734. static void cache_estimate(unsigned long gfporder, size_t buffer_size,
  735. size_t align, int flags, size_t *left_over,
  736. unsigned int *num)
  737. {
  738. int nr_objs;
  739. size_t mgmt_size;
  740. size_t slab_size = PAGE_SIZE << gfporder;
  741. /*
  742. * The slab management structure can be either off the slab or
  743. * on it. For the latter case, the memory allocated for a
  744. * slab is used for:
  745. *
  746. * - The struct slab
  747. * - One kmem_bufctl_t for each object
  748. * - Padding to respect alignment of @align
  749. * - @buffer_size bytes for each object
  750. *
  751. * If the slab management structure is off the slab, then the
  752. * alignment will already be calculated into the size. Because
  753. * the slabs are all pages aligned, the objects will be at the
  754. * correct alignment when allocated.
  755. */
  756. if (flags & CFLGS_OFF_SLAB) {
  757. mgmt_size = 0;
  758. nr_objs = slab_size / buffer_size;
  759. if (nr_objs > SLAB_LIMIT)
  760. nr_objs = SLAB_LIMIT;
  761. } else {
  762. /*
  763. * Ignore padding for the initial guess. The padding
  764. * is at most @align-1 bytes, and @buffer_size is at
  765. * least @align. In the worst case, this result will
  766. * be one greater than the number of objects that fit
  767. * into the memory allocation when taking the padding
  768. * into account.
  769. */
  770. nr_objs = (slab_size - sizeof(struct slab)) /
  771. (buffer_size + sizeof(kmem_bufctl_t));
  772. /*
  773. * This calculated number will be either the right
  774. * amount, or one greater than what we want.
  775. */
  776. if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
  777. > slab_size)
  778. nr_objs--;
  779. if (nr_objs > SLAB_LIMIT)
  780. nr_objs = SLAB_LIMIT;
  781. mgmt_size = slab_mgmt_size(nr_objs, align);
  782. }
  783. *num = nr_objs;
  784. *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
  785. }
  786. #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
  787. static void __slab_error(const char *function, struct kmem_cache *cachep,
  788. char *msg)
  789. {
  790. printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
  791. function, cachep->name, msg);
  792. dump_stack();
  793. }
  794. /*
  795. * By default on NUMA we use alien caches to stage the freeing of
  796. * objects allocated from other nodes. This causes massive memory
  797. * inefficiencies when using fake NUMA setup to split memory into a
  798. * large number of small nodes, so it can be disabled on the command
  799. * line
  800. */
  801. static int use_alien_caches __read_mostly = 1;
  802. static int __init noaliencache_setup(char *s)
  803. {
  804. use_alien_caches = 0;
  805. return 1;
  806. }
  807. __setup("noaliencache", noaliencache_setup);
  808. #ifdef CONFIG_NUMA
  809. /*
  810. * Special reaping functions for NUMA systems called from cache_reap().
  811. * These take care of doing round robin flushing of alien caches (containing
  812. * objects freed on different nodes from which they were allocated) and the
  813. * flushing of remote pcps by calling drain_node_pages.
  814. */
  815. static DEFINE_PER_CPU(unsigned long, reap_node);
  816. static void init_reap_node(int cpu)
  817. {
  818. int node;
  819. node = next_node(cpu_to_node(cpu), node_online_map);
  820. if (node == MAX_NUMNODES)
  821. node = first_node(node_online_map);
  822. per_cpu(reap_node, cpu) = node;
  823. }
  824. static void next_reap_node(void)
  825. {
  826. int node = __get_cpu_var(reap_node);
  827. /*
  828. * Also drain per cpu pages on remote zones
  829. */
  830. if (node != numa_node_id())
  831. drain_node_pages(node);
  832. node = next_node(node, node_online_map);
  833. if (unlikely(node >= MAX_NUMNODES))
  834. node = first_node(node_online_map);
  835. __get_cpu_var(reap_node) = node;
  836. }
  837. #else
  838. #define init_reap_node(cpu) do { } while (0)
  839. #define next_reap_node(void) do { } while (0)
  840. #endif
  841. /*
  842. * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
  843. * via the workqueue/eventd.
  844. * Add the CPU number into the expiration time to minimize the possibility of
  845. * the CPUs getting into lockstep and contending for the global cache chain
  846. * lock.
  847. */
  848. static void __devinit start_cpu_timer(int cpu)
  849. {
  850. struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
  851. /*
  852. * When this gets called from do_initcalls via cpucache_init(),
  853. * init_workqueues() has already run, so keventd will be setup
  854. * at that time.
  855. */
  856. if (keventd_up() && reap_work->work.func == NULL) {
  857. init_reap_node(cpu);
  858. INIT_DELAYED_WORK(reap_work, cache_reap);
  859. schedule_delayed_work_on(cpu, reap_work,
  860. __round_jiffies_relative(HZ, cpu));
  861. }
  862. }
  863. static struct array_cache *alloc_arraycache(int node, int entries,
  864. int batchcount)
  865. {
  866. int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
  867. struct array_cache *nc = NULL;
  868. nc = kmalloc_node(memsize, GFP_KERNEL, node);
  869. if (nc) {
  870. nc->avail = 0;
  871. nc->limit = entries;
  872. nc->batchcount = batchcount;
  873. nc->touched = 0;
  874. spin_lock_init(&nc->lock);
  875. }
  876. return nc;
  877. }
  878. /*
  879. * Transfer objects in one arraycache to another.
  880. * Locking must be handled by the caller.
  881. *
  882. * Return the number of entries transferred.
  883. */
  884. static int transfer_objects(struct array_cache *to,
  885. struct array_cache *from, unsigned int max)
  886. {
  887. /* Figure out how many entries to transfer */
  888. int nr = min(min(from->avail, max), to->limit - to->avail);
  889. if (!nr)
  890. return 0;
  891. memcpy(to->entry + to->avail, from->entry + from->avail -nr,
  892. sizeof(void *) *nr);
  893. from->avail -= nr;
  894. to->avail += nr;
  895. to->touched = 1;
  896. return nr;
  897. }
  898. #ifndef CONFIG_NUMA
  899. #define drain_alien_cache(cachep, alien) do { } while (0)
  900. #define reap_alien(cachep, l3) do { } while (0)
  901. static inline struct array_cache **alloc_alien_cache(int node, int limit)
  902. {
  903. return (struct array_cache **)BAD_ALIEN_MAGIC;
  904. }
  905. static inline void free_alien_cache(struct array_cache **ac_ptr)
  906. {
  907. }
  908. static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
  909. {
  910. return 0;
  911. }
  912. static inline void *alternate_node_alloc(struct kmem_cache *cachep,
  913. gfp_t flags)
  914. {
  915. return NULL;
  916. }
  917. static inline void *____cache_alloc_node(struct kmem_cache *cachep,
  918. gfp_t flags, int nodeid)
  919. {
  920. return NULL;
  921. }
  922. #else /* CONFIG_NUMA */
  923. static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
  924. static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
  925. static struct array_cache **alloc_alien_cache(int node, int limit)
  926. {
  927. struct array_cache **ac_ptr;
  928. int memsize = sizeof(void *) * nr_node_ids;
  929. int i;
  930. if (limit > 1)
  931. limit = 12;
  932. ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
  933. if (ac_ptr) {
  934. for_each_node(i) {
  935. if (i == node || !node_online(i)) {
  936. ac_ptr[i] = NULL;
  937. continue;
  938. }
  939. ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
  940. if (!ac_ptr[i]) {
  941. for (i--; i <= 0; i--)
  942. kfree(ac_ptr[i]);
  943. kfree(ac_ptr);
  944. return NULL;
  945. }
  946. }
  947. }
  948. return ac_ptr;
  949. }
  950. static void free_alien_cache(struct array_cache **ac_ptr)
  951. {
  952. int i;
  953. if (!ac_ptr)
  954. return;
  955. for_each_node(i)
  956. kfree(ac_ptr[i]);
  957. kfree(ac_ptr);
  958. }
  959. static void __drain_alien_cache(struct kmem_cache *cachep,
  960. struct array_cache *ac, int node)
  961. {
  962. struct kmem_list3 *rl3 = cachep->nodelists[node];
  963. if (ac->avail) {
  964. spin_lock(&rl3->list_lock);
  965. /*
  966. * Stuff objects into the remote nodes shared array first.
  967. * That way we could avoid the overhead of putting the objects
  968. * into the free lists and getting them back later.
  969. */
  970. if (rl3->shared)
  971. transfer_objects(rl3->shared, ac, ac->limit);
  972. free_block(cachep, ac->entry, ac->avail, node);
  973. ac->avail = 0;
  974. spin_unlock(&rl3->list_lock);
  975. }
  976. }
  977. /*
  978. * Called from cache_reap() to regularly drain alien caches round robin.
  979. */
  980. static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
  981. {
  982. int node = __get_cpu_var(reap_node);
  983. if (l3->alien) {
  984. struct array_cache *ac = l3->alien[node];
  985. if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
  986. __drain_alien_cache(cachep, ac, node);
  987. spin_unlock_irq(&ac->lock);
  988. }
  989. }
  990. }
  991. static void drain_alien_cache(struct kmem_cache *cachep,
  992. struct array_cache **alien)
  993. {
  994. int i = 0;
  995. struct array_cache *ac;
  996. unsigned long flags;
  997. for_each_online_node(i) {
  998. ac = alien[i];
  999. if (ac) {
  1000. spin_lock_irqsave(&ac->lock, flags);
  1001. __drain_alien_cache(cachep, ac, i);
  1002. spin_unlock_irqrestore(&ac->lock, flags);
  1003. }
  1004. }
  1005. }
  1006. static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
  1007. {
  1008. struct slab *slabp = virt_to_slab(objp);
  1009. int nodeid = slabp->nodeid;
  1010. struct kmem_list3 *l3;
  1011. struct array_cache *alien = NULL;
  1012. int node;
  1013. node = numa_node_id();
  1014. /*
  1015. * Make sure we are not freeing a object from another node to the array
  1016. * cache on this cpu.
  1017. */
  1018. if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches))
  1019. return 0;
  1020. l3 = cachep->nodelists[node];
  1021. STATS_INC_NODEFREES(cachep);
  1022. if (l3->alien && l3->alien[nodeid]) {
  1023. alien = l3->alien[nodeid];
  1024. spin_lock(&alien->lock);
  1025. if (unlikely(alien->avail == alien->limit)) {
  1026. STATS_INC_ACOVERFLOW(cachep);
  1027. __drain_alien_cache(cachep, alien, nodeid);
  1028. }
  1029. alien->entry[alien->avail++] = objp;
  1030. spin_unlock(&alien->lock);
  1031. } else {
  1032. spin_lock(&(cachep->nodelists[nodeid])->list_lock);
  1033. free_block(cachep, &objp, 1, nodeid);
  1034. spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
  1035. }
  1036. return 1;
  1037. }
  1038. #endif
  1039. static int __cpuinit cpuup_callback(struct notifier_block *nfb,
  1040. unsigned long action, void *hcpu)
  1041. {
  1042. long cpu = (long)hcpu;
  1043. struct kmem_cache *cachep;
  1044. struct kmem_list3 *l3 = NULL;
  1045. int node = cpu_to_node(cpu);
  1046. int memsize = sizeof(struct kmem_list3);
  1047. switch (action) {
  1048. case CPU_UP_PREPARE:
  1049. mutex_lock(&cache_chain_mutex);
  1050. /*
  1051. * We need to do this right in the beginning since
  1052. * alloc_arraycache's are going to use this list.
  1053. * kmalloc_node allows us to add the slab to the right
  1054. * kmem_list3 and not this cpu's kmem_list3
  1055. */
  1056. list_for_each_entry(cachep, &cache_chain, next) {
  1057. /*
  1058. * Set up the size64 kmemlist for cpu before we can
  1059. * begin anything. Make sure some other cpu on this
  1060. * node has not already allocated this
  1061. */
  1062. if (!cachep->nodelists[node]) {
  1063. l3 = kmalloc_node(memsize, GFP_KERNEL, node);
  1064. if (!l3)
  1065. goto bad;
  1066. kmem_list3_init(l3);
  1067. l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
  1068. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1069. /*
  1070. * The l3s don't come and go as CPUs come and
  1071. * go. cache_chain_mutex is sufficient
  1072. * protection here.
  1073. */
  1074. cachep->nodelists[node] = l3;
  1075. }
  1076. spin_lock_irq(&cachep->nodelists[node]->list_lock);
  1077. cachep->nodelists[node]->free_limit =
  1078. (1 + nr_cpus_node(node)) *
  1079. cachep->batchcount + cachep->num;
  1080. spin_unlock_irq(&cachep->nodelists[node]->list_lock);
  1081. }
  1082. /*
  1083. * Now we can go ahead with allocating the shared arrays and
  1084. * array caches
  1085. */
  1086. list_for_each_entry(cachep, &cache_chain, next) {
  1087. struct array_cache *nc;
  1088. struct array_cache *shared;
  1089. struct array_cache **alien = NULL;
  1090. nc = alloc_arraycache(node, cachep->limit,
  1091. cachep->batchcount);
  1092. if (!nc)
  1093. goto bad;
  1094. shared = alloc_arraycache(node,
  1095. cachep->shared * cachep->batchcount,
  1096. 0xbaadf00d);
  1097. if (!shared)
  1098. goto bad;
  1099. if (use_alien_caches) {
  1100. alien = alloc_alien_cache(node, cachep->limit);
  1101. if (!alien)
  1102. goto bad;
  1103. }
  1104. cachep->array[cpu] = nc;
  1105. l3 = cachep->nodelists[node];
  1106. BUG_ON(!l3);
  1107. spin_lock_irq(&l3->list_lock);
  1108. if (!l3->shared) {
  1109. /*
  1110. * We are serialised from CPU_DEAD or
  1111. * CPU_UP_CANCELLED by the cpucontrol lock
  1112. */
  1113. l3->shared = shared;
  1114. shared = NULL;
  1115. }
  1116. #ifdef CONFIG_NUMA
  1117. if (!l3->alien) {
  1118. l3->alien = alien;
  1119. alien = NULL;
  1120. }
  1121. #endif
  1122. spin_unlock_irq(&l3->list_lock);
  1123. kfree(shared);
  1124. free_alien_cache(alien);
  1125. }
  1126. break;
  1127. case CPU_ONLINE:
  1128. mutex_unlock(&cache_chain_mutex);
  1129. start_cpu_timer(cpu);
  1130. break;
  1131. #ifdef CONFIG_HOTPLUG_CPU
  1132. case CPU_DOWN_PREPARE:
  1133. mutex_lock(&cache_chain_mutex);
  1134. break;
  1135. case CPU_DOWN_FAILED:
  1136. mutex_unlock(&cache_chain_mutex);
  1137. break;
  1138. case CPU_DEAD:
  1139. /*
  1140. * Even if all the cpus of a node are down, we don't free the
  1141. * kmem_list3 of any cache. This to avoid a race between
  1142. * cpu_down, and a kmalloc allocation from another cpu for
  1143. * memory from the node of the cpu going down. The list3
  1144. * structure is usually allocated from kmem_cache_create() and
  1145. * gets destroyed at kmem_cache_destroy().
  1146. */
  1147. /* fall thru */
  1148. #endif
  1149. case CPU_UP_CANCELED:
  1150. list_for_each_entry(cachep, &cache_chain, next) {
  1151. struct array_cache *nc;
  1152. struct array_cache *shared;
  1153. struct array_cache **alien;
  1154. cpumask_t mask;
  1155. mask = node_to_cpumask(node);
  1156. /* cpu is dead; no one can alloc from it. */
  1157. nc = cachep->array[cpu];
  1158. cachep->array[cpu] = NULL;
  1159. l3 = cachep->nodelists[node];
  1160. if (!l3)
  1161. goto free_array_cache;
  1162. spin_lock_irq(&l3->list_lock);
  1163. /* Free limit for this kmem_list3 */
  1164. l3->free_limit -= cachep->batchcount;
  1165. if (nc)
  1166. free_block(cachep, nc->entry, nc->avail, node);
  1167. if (!cpus_empty(mask)) {
  1168. spin_unlock_irq(&l3->list_lock);
  1169. goto free_array_cache;
  1170. }
  1171. shared = l3->shared;
  1172. if (shared) {
  1173. free_block(cachep, l3->shared->entry,
  1174. l3->shared->avail, node);
  1175. l3->shared = NULL;
  1176. }
  1177. alien = l3->alien;
  1178. l3->alien = NULL;
  1179. spin_unlock_irq(&l3->list_lock);
  1180. kfree(shared);
  1181. if (alien) {
  1182. drain_alien_cache(cachep, alien);
  1183. free_alien_cache(alien);
  1184. }
  1185. free_array_cache:
  1186. kfree(nc);
  1187. }
  1188. /*
  1189. * In the previous loop, all the objects were freed to
  1190. * the respective cache's slabs, now we can go ahead and
  1191. * shrink each nodelist to its limit.
  1192. */
  1193. list_for_each_entry(cachep, &cache_chain, next) {
  1194. l3 = cachep->nodelists[node];
  1195. if (!l3)
  1196. continue;
  1197. drain_freelist(cachep, l3, l3->free_objects);
  1198. }
  1199. mutex_unlock(&cache_chain_mutex);
  1200. break;
  1201. }
  1202. return NOTIFY_OK;
  1203. bad:
  1204. return NOTIFY_BAD;
  1205. }
  1206. static struct notifier_block __cpuinitdata cpucache_notifier = {
  1207. &cpuup_callback, NULL, 0
  1208. };
  1209. /*
  1210. * swap the static kmem_list3 with kmalloced memory
  1211. */
  1212. static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
  1213. int nodeid)
  1214. {
  1215. struct kmem_list3 *ptr;
  1216. ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
  1217. BUG_ON(!ptr);
  1218. local_irq_disable();
  1219. memcpy(ptr, list, sizeof(struct kmem_list3));
  1220. /*
  1221. * Do not assume that spinlocks can be initialized via memcpy:
  1222. */
  1223. spin_lock_init(&ptr->list_lock);
  1224. MAKE_ALL_LISTS(cachep, ptr, nodeid);
  1225. cachep->nodelists[nodeid] = ptr;
  1226. local_irq_enable();
  1227. }
  1228. /*
  1229. * Initialisation. Called after the page allocator have been initialised and
  1230. * before smp_init().
  1231. */
  1232. void __init kmem_cache_init(void)
  1233. {
  1234. size_t left_over;
  1235. struct cache_sizes *sizes;
  1236. struct cache_names *names;
  1237. int i;
  1238. int order;
  1239. int node;
  1240. for (i = 0; i < NUM_INIT_LISTS; i++) {
  1241. kmem_list3_init(&initkmem_list3[i]);
  1242. if (i < MAX_NUMNODES)
  1243. cache_cache.nodelists[i] = NULL;
  1244. }
  1245. /*
  1246. * Fragmentation resistance on low memory - only use bigger
  1247. * page orders on machines with more than 32MB of memory.
  1248. */
  1249. if (num_physpages > (32 << 20) >> PAGE_SHIFT)
  1250. slab_break_gfp_order = BREAK_GFP_ORDER_HI;
  1251. /* Bootstrap is tricky, because several objects are allocated
  1252. * from caches that do not exist yet:
  1253. * 1) initialize the cache_cache cache: it contains the struct
  1254. * kmem_cache structures of all caches, except cache_cache itself:
  1255. * cache_cache is statically allocated.
  1256. * Initially an __init data area is used for the head array and the
  1257. * kmem_list3 structures, it's replaced with a kmalloc allocated
  1258. * array at the end of the bootstrap.
  1259. * 2) Create the first kmalloc cache.
  1260. * The struct kmem_cache for the new cache is allocated normally.
  1261. * An __init data area is used for the head array.
  1262. * 3) Create the remaining kmalloc caches, with minimally sized
  1263. * head arrays.
  1264. * 4) Replace the __init data head arrays for cache_cache and the first
  1265. * kmalloc cache with kmalloc allocated arrays.
  1266. * 5) Replace the __init data for kmem_list3 for cache_cache and
  1267. * the other cache's with kmalloc allocated memory.
  1268. * 6) Resize the head arrays of the kmalloc caches to their final sizes.
  1269. */
  1270. node = numa_node_id();
  1271. /* 1) create the cache_cache */
  1272. INIT_LIST_HEAD(&cache_chain);
  1273. list_add(&cache_cache.next, &cache_chain);
  1274. cache_cache.colour_off = cache_line_size();
  1275. cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
  1276. cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
  1277. cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
  1278. cache_line_size());
  1279. cache_cache.reciprocal_buffer_size =
  1280. reciprocal_value(cache_cache.buffer_size);
  1281. for (order = 0; order < MAX_ORDER; order++) {
  1282. cache_estimate(order, cache_cache.buffer_size,
  1283. cache_line_size(), 0, &left_over, &cache_cache.num);
  1284. if (cache_cache.num)
  1285. break;
  1286. }
  1287. BUG_ON(!cache_cache.num);
  1288. cache_cache.gfporder = order;
  1289. cache_cache.colour = left_over / cache_cache.colour_off;
  1290. cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
  1291. sizeof(struct slab), cache_line_size());
  1292. /* 2+3) create the kmalloc caches */
  1293. sizes = malloc_sizes;
  1294. names = cache_names;
  1295. /*
  1296. * Initialize the caches that provide memory for the array cache and the
  1297. * kmem_list3 structures first. Without this, further allocations will
  1298. * bug.
  1299. */
  1300. sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
  1301. sizes[INDEX_AC].cs_size,
  1302. ARCH_KMALLOC_MINALIGN,
  1303. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1304. NULL, NULL);
  1305. if (INDEX_AC != INDEX_L3) {
  1306. sizes[INDEX_L3].cs_cachep =
  1307. kmem_cache_create(names[INDEX_L3].name,
  1308. sizes[INDEX_L3].cs_size,
  1309. ARCH_KMALLOC_MINALIGN,
  1310. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1311. NULL, NULL);
  1312. }
  1313. slab_early_init = 0;
  1314. while (sizes->cs_size != ULONG_MAX) {
  1315. /*
  1316. * For performance, all the general caches are L1 aligned.
  1317. * This should be particularly beneficial on SMP boxes, as it
  1318. * eliminates "false sharing".
  1319. * Note for systems short on memory removing the alignment will
  1320. * allow tighter packing of the smaller caches.
  1321. */
  1322. if (!sizes->cs_cachep) {
  1323. sizes->cs_cachep = kmem_cache_create(names->name,
  1324. sizes->cs_size,
  1325. ARCH_KMALLOC_MINALIGN,
  1326. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1327. NULL, NULL);
  1328. }
  1329. #ifdef CONFIG_ZONE_DMA
  1330. sizes->cs_dmacachep = kmem_cache_create(
  1331. names->name_dma,
  1332. sizes->cs_size,
  1333. ARCH_KMALLOC_MINALIGN,
  1334. ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
  1335. SLAB_PANIC,
  1336. NULL, NULL);
  1337. #endif
  1338. sizes++;
  1339. names++;
  1340. }
  1341. /* 4) Replace the bootstrap head arrays */
  1342. {
  1343. struct array_cache *ptr;
  1344. ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
  1345. local_irq_disable();
  1346. BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
  1347. memcpy(ptr, cpu_cache_get(&cache_cache),
  1348. sizeof(struct arraycache_init));
  1349. /*
  1350. * Do not assume that spinlocks can be initialized via memcpy:
  1351. */
  1352. spin_lock_init(&ptr->lock);
  1353. cache_cache.array[smp_processor_id()] = ptr;
  1354. local_irq_enable();
  1355. ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
  1356. local_irq_disable();
  1357. BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
  1358. != &initarray_generic.cache);
  1359. memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
  1360. sizeof(struct arraycache_init));
  1361. /*
  1362. * Do not assume that spinlocks can be initialized via memcpy:
  1363. */
  1364. spin_lock_init(&ptr->lock);
  1365. malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
  1366. ptr;
  1367. local_irq_enable();
  1368. }
  1369. /* 5) Replace the bootstrap kmem_list3's */
  1370. {
  1371. int nid;
  1372. /* Replace the static kmem_list3 structures for the boot cpu */
  1373. init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
  1374. for_each_online_node(nid) {
  1375. init_list(malloc_sizes[INDEX_AC].cs_cachep,
  1376. &initkmem_list3[SIZE_AC + nid], nid);
  1377. if (INDEX_AC != INDEX_L3) {
  1378. init_list(malloc_sizes[INDEX_L3].cs_cachep,
  1379. &initkmem_list3[SIZE_L3 + nid], nid);
  1380. }
  1381. }
  1382. }
  1383. /* 6) resize the head arrays to their final sizes */
  1384. {
  1385. struct kmem_cache *cachep;
  1386. mutex_lock(&cache_chain_mutex);
  1387. list_for_each_entry(cachep, &cache_chain, next)
  1388. if (enable_cpucache(cachep))
  1389. BUG();
  1390. mutex_unlock(&cache_chain_mutex);
  1391. }
  1392. /* Annotate slab for lockdep -- annotate the malloc caches */
  1393. init_lock_keys();
  1394. /* Done! */
  1395. g_cpucache_up = FULL;
  1396. /*
  1397. * Register a cpu startup notifier callback that initializes
  1398. * cpu_cache_get for all new cpus
  1399. */
  1400. register_cpu_notifier(&cpucache_notifier);
  1401. /*
  1402. * The reap timers are started later, with a module init call: That part
  1403. * of the kernel is not yet operational.
  1404. */
  1405. }
  1406. static int __init cpucache_init(void)
  1407. {
  1408. int cpu;
  1409. /*
  1410. * Register the timers that return unneeded pages to the page allocator
  1411. */
  1412. for_each_online_cpu(cpu)
  1413. start_cpu_timer(cpu);
  1414. return 0;
  1415. }
  1416. __initcall(cpucache_init);
  1417. /*
  1418. * Interface to system's page allocator. No need to hold the cache-lock.
  1419. *
  1420. * If we requested dmaable memory, we will get it. Even if we
  1421. * did not request dmaable memory, we might get it, but that
  1422. * would be relatively rare and ignorable.
  1423. */
  1424. static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
  1425. {
  1426. struct page *page;
  1427. int nr_pages;
  1428. int i;
  1429. #ifndef CONFIG_MMU
  1430. /*
  1431. * Nommu uses slab's for process anonymous memory allocations, and thus
  1432. * requires __GFP_COMP to properly refcount higher order allocations
  1433. */
  1434. flags |= __GFP_COMP;
  1435. #endif
  1436. flags |= cachep->gfpflags;
  1437. page = alloc_pages_node(nodeid, flags, cachep->gfporder);
  1438. if (!page)
  1439. return NULL;
  1440. nr_pages = (1 << cachep->gfporder);
  1441. if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
  1442. add_zone_page_state(page_zone(page),
  1443. NR_SLAB_RECLAIMABLE, nr_pages);
  1444. else
  1445. add_zone_page_state(page_zone(page),
  1446. NR_SLAB_UNRECLAIMABLE, nr_pages);
  1447. for (i = 0; i < nr_pages; i++)
  1448. __SetPageSlab(page + i);
  1449. return page_address(page);
  1450. }
  1451. /*
  1452. * Interface to system's page release.
  1453. */
  1454. static void kmem_freepages(struct kmem_cache *cachep, void *addr)
  1455. {
  1456. unsigned long i = (1 << cachep->gfporder);
  1457. struct page *page = virt_to_page(addr);
  1458. const unsigned long nr_freed = i;
  1459. if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
  1460. sub_zone_page_state(page_zone(page),
  1461. NR_SLAB_RECLAIMABLE, nr_freed);
  1462. else
  1463. sub_zone_page_state(page_zone(page),
  1464. NR_SLAB_UNRECLAIMABLE, nr_freed);
  1465. while (i--) {
  1466. BUG_ON(!PageSlab(page));
  1467. __ClearPageSlab(page);
  1468. page++;
  1469. }
  1470. if (current->reclaim_state)
  1471. current->reclaim_state->reclaimed_slab += nr_freed;
  1472. free_pages((unsigned long)addr, cachep->gfporder);
  1473. }
  1474. static void kmem_rcu_free(struct rcu_head *head)
  1475. {
  1476. struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
  1477. struct kmem_cache *cachep = slab_rcu->cachep;
  1478. kmem_freepages(cachep, slab_rcu->addr);
  1479. if (OFF_SLAB(cachep))
  1480. kmem_cache_free(cachep->slabp_cache, slab_rcu);
  1481. }
  1482. #if DEBUG
  1483. #ifdef CONFIG_DEBUG_PAGEALLOC
  1484. static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
  1485. unsigned long caller)
  1486. {
  1487. int size = obj_size(cachep);
  1488. addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
  1489. if (size < 5 * sizeof(unsigned long))
  1490. return;
  1491. *addr++ = 0x12345678;
  1492. *addr++ = caller;
  1493. *addr++ = smp_processor_id();
  1494. size -= 3 * sizeof(unsigned long);
  1495. {
  1496. unsigned long *sptr = &caller;
  1497. unsigned long svalue;
  1498. while (!kstack_end(sptr)) {
  1499. svalue = *sptr++;
  1500. if (kernel_text_address(svalue)) {
  1501. *addr++ = svalue;
  1502. size -= sizeof(unsigned long);
  1503. if (size <= sizeof(unsigned long))
  1504. break;
  1505. }
  1506. }
  1507. }
  1508. *addr++ = 0x87654321;
  1509. }
  1510. #endif
  1511. static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
  1512. {
  1513. int size = obj_size(cachep);
  1514. addr = &((char *)addr)[obj_offset(cachep)];
  1515. memset(addr, val, size);
  1516. *(unsigned char *)(addr + size - 1) = POISON_END;
  1517. }
  1518. static void dump_line(char *data, int offset, int limit)
  1519. {
  1520. int i;
  1521. unsigned char error = 0;
  1522. int bad_count = 0;
  1523. printk(KERN_ERR "%03x:", offset);
  1524. for (i = 0; i < limit; i++) {
  1525. if (data[offset + i] != POISON_FREE) {
  1526. error = data[offset + i];
  1527. bad_count++;
  1528. }
  1529. printk(" %02x", (unsigned char)data[offset + i]);
  1530. }
  1531. printk("\n");
  1532. if (bad_count == 1) {
  1533. error ^= POISON_FREE;
  1534. if (!(error & (error - 1))) {
  1535. printk(KERN_ERR "Single bit error detected. Probably "
  1536. "bad RAM.\n");
  1537. #ifdef CONFIG_X86
  1538. printk(KERN_ERR "Run memtest86+ or a similar memory "
  1539. "test tool.\n");
  1540. #else
  1541. printk(KERN_ERR "Run a memory test tool.\n");
  1542. #endif
  1543. }
  1544. }
  1545. }
  1546. #endif
  1547. #if DEBUG
  1548. static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
  1549. {
  1550. int i, size;
  1551. char *realobj;
  1552. if (cachep->flags & SLAB_RED_ZONE) {
  1553. printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
  1554. *dbg_redzone1(cachep, objp),
  1555. *dbg_redzone2(cachep, objp));
  1556. }
  1557. if (cachep->flags & SLAB_STORE_USER) {
  1558. printk(KERN_ERR "Last user: [<%p>]",
  1559. *dbg_userword(cachep, objp));
  1560. print_symbol("(%s)",
  1561. (unsigned long)*dbg_userword(cachep, objp));
  1562. printk("\n");
  1563. }
  1564. realobj = (char *)objp + obj_offset(cachep);
  1565. size = obj_size(cachep);
  1566. for (i = 0; i < size && lines; i += 16, lines--) {
  1567. int limit;
  1568. limit = 16;
  1569. if (i + limit > size)
  1570. limit = size - i;
  1571. dump_line(realobj, i, limit);
  1572. }
  1573. }
  1574. static void check_poison_obj(struct kmem_cache *cachep, void *objp)
  1575. {
  1576. char *realobj;
  1577. int size, i;
  1578. int lines = 0;
  1579. realobj = (char *)objp + obj_offset(cachep);
  1580. size = obj_size(cachep);
  1581. for (i = 0; i < size; i++) {
  1582. char exp = POISON_FREE;
  1583. if (i == size - 1)
  1584. exp = POISON_END;
  1585. if (realobj[i] != exp) {
  1586. int limit;
  1587. /* Mismatch ! */
  1588. /* Print header */
  1589. if (lines == 0) {
  1590. printk(KERN_ERR
  1591. "Slab corruption: %s start=%p, len=%d\n",
  1592. cachep->name, realobj, size);
  1593. print_objinfo(cachep, objp, 0);
  1594. }
  1595. /* Hexdump the affected line */
  1596. i = (i / 16) * 16;
  1597. limit = 16;
  1598. if (i + limit > size)
  1599. limit = size - i;
  1600. dump_line(realobj, i, limit);
  1601. i += 16;
  1602. lines++;
  1603. /* Limit to 5 lines */
  1604. if (lines > 5)
  1605. break;
  1606. }
  1607. }
  1608. if (lines != 0) {
  1609. /* Print some data about the neighboring objects, if they
  1610. * exist:
  1611. */
  1612. struct slab *slabp = virt_to_slab(objp);
  1613. unsigned int objnr;
  1614. objnr = obj_to_index(cachep, slabp, objp);
  1615. if (objnr) {
  1616. objp = index_to_obj(cachep, slabp, objnr - 1);
  1617. realobj = (char *)objp + obj_offset(cachep);
  1618. printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
  1619. realobj, size);
  1620. print_objinfo(cachep, objp, 2);
  1621. }
  1622. if (objnr + 1 < cachep->num) {
  1623. objp = index_to_obj(cachep, slabp, objnr + 1);
  1624. realobj = (char *)objp + obj_offset(cachep);
  1625. printk(KERN_ERR "Next obj: start=%p, len=%d\n",
  1626. realobj, size);
  1627. print_objinfo(cachep, objp, 2);
  1628. }
  1629. }
  1630. }
  1631. #endif
  1632. #if DEBUG
  1633. /**
  1634. * slab_destroy_objs - destroy a slab and its objects
  1635. * @cachep: cache pointer being destroyed
  1636. * @slabp: slab pointer being destroyed
  1637. *
  1638. * Call the registered destructor for each object in a slab that is being
  1639. * destroyed.
  1640. */
  1641. static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
  1642. {
  1643. int i;
  1644. for (i = 0; i < cachep->num; i++) {
  1645. void *objp = index_to_obj(cachep, slabp, i);
  1646. if (cachep->flags & SLAB_POISON) {
  1647. #ifdef CONFIG_DEBUG_PAGEALLOC
  1648. if (cachep->buffer_size % PAGE_SIZE == 0 &&
  1649. OFF_SLAB(cachep))
  1650. kernel_map_pages(virt_to_page(objp),
  1651. cachep->buffer_size / PAGE_SIZE, 1);
  1652. else
  1653. check_poison_obj(cachep, objp);
  1654. #else
  1655. check_poison_obj(cachep, objp);
  1656. #endif
  1657. }
  1658. if (cachep->flags & SLAB_RED_ZONE) {
  1659. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
  1660. slab_error(cachep, "start of a freed object "
  1661. "was overwritten");
  1662. if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
  1663. slab_error(cachep, "end of a freed object "
  1664. "was overwritten");
  1665. }
  1666. if (cachep->dtor && !(cachep->flags & SLAB_POISON))
  1667. (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
  1668. }
  1669. }
  1670. #else
  1671. static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
  1672. {
  1673. if (cachep->dtor) {
  1674. int i;
  1675. for (i = 0; i < cachep->num; i++) {
  1676. void *objp = index_to_obj(cachep, slabp, i);
  1677. (cachep->dtor) (objp, cachep, 0);
  1678. }
  1679. }
  1680. }
  1681. #endif
  1682. /**
  1683. * slab_destroy - destroy and release all objects in a slab
  1684. * @cachep: cache pointer being destroyed
  1685. * @slabp: slab pointer being destroyed
  1686. *
  1687. * Destroy all the objs in a slab, and release the mem back to the system.
  1688. * Before calling the slab must have been unlinked from the cache. The
  1689. * cache-lock is not held/needed.
  1690. */
  1691. static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
  1692. {
  1693. void *addr = slabp->s_mem - slabp->colouroff;
  1694. slab_destroy_objs(cachep, slabp);
  1695. if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
  1696. struct slab_rcu *slab_rcu;
  1697. slab_rcu = (struct slab_rcu *)slabp;
  1698. slab_rcu->cachep = cachep;
  1699. slab_rcu->addr = addr;
  1700. call_rcu(&slab_rcu->head, kmem_rcu_free);
  1701. } else {
  1702. kmem_freepages(cachep, addr);
  1703. if (OFF_SLAB(cachep))
  1704. kmem_cache_free(cachep->slabp_cache, slabp);
  1705. }
  1706. }
  1707. /*
  1708. * For setting up all the kmem_list3s for cache whose buffer_size is same as
  1709. * size of kmem_list3.
  1710. */
  1711. static void set_up_list3s(struct kmem_cache *cachep, int index)
  1712. {
  1713. int node;
  1714. for_each_online_node(node) {
  1715. cachep->nodelists[node] = &initkmem_list3[index + node];
  1716. cachep->nodelists[node]->next_reap = jiffies +
  1717. REAPTIMEOUT_LIST3 +
  1718. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1719. }
  1720. }
  1721. static void __kmem_cache_destroy(struct kmem_cache *cachep)
  1722. {
  1723. int i;
  1724. struct kmem_list3 *l3;
  1725. for_each_online_cpu(i)
  1726. kfree(cachep->array[i]);
  1727. /* NUMA: free the list3 structures */
  1728. for_each_online_node(i) {
  1729. l3 = cachep->nodelists[i];
  1730. if (l3) {
  1731. kfree(l3->shared);
  1732. free_alien_cache(l3->alien);
  1733. kfree(l3);
  1734. }
  1735. }
  1736. kmem_cache_free(&cache_cache, cachep);
  1737. }
  1738. /**
  1739. * calculate_slab_order - calculate size (page order) of slabs
  1740. * @cachep: pointer to the cache that is being created
  1741. * @size: size of objects to be created in this cache.
  1742. * @align: required alignment for the objects.
  1743. * @flags: slab allocation flags
  1744. *
  1745. * Also calculates the number of objects per slab.
  1746. *
  1747. * This could be made much more intelligent. For now, try to avoid using
  1748. * high order pages for slabs. When the gfp() functions are more friendly
  1749. * towards high-order requests, this should be changed.
  1750. */
  1751. static size_t calculate_slab_order(struct kmem_cache *cachep,
  1752. size_t size, size_t align, unsigned long flags)
  1753. {
  1754. unsigned long offslab_limit;
  1755. size_t left_over = 0;
  1756. int gfporder;
  1757. for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
  1758. unsigned int num;
  1759. size_t remainder;
  1760. cache_estimate(gfporder, size, align, flags, &remainder, &num);
  1761. if (!num)
  1762. continue;
  1763. if (flags & CFLGS_OFF_SLAB) {
  1764. /*
  1765. * Max number of objs-per-slab for caches which
  1766. * use off-slab slabs. Needed to avoid a possible
  1767. * looping condition in cache_grow().
  1768. */
  1769. offslab_limit = size - sizeof(struct slab);
  1770. offslab_limit /= sizeof(kmem_bufctl_t);
  1771. if (num > offslab_limit)
  1772. break;
  1773. }
  1774. /* Found something acceptable - save it away */
  1775. cachep->num = num;
  1776. cachep->gfporder = gfporder;
  1777. left_over = remainder;
  1778. /*
  1779. * A VFS-reclaimable slab tends to have most allocations
  1780. * as GFP_NOFS and we really don't want to have to be allocating
  1781. * higher-order pages when we are unable to shrink dcache.
  1782. */
  1783. if (flags & SLAB_RECLAIM_ACCOUNT)
  1784. break;
  1785. /*
  1786. * Large number of objects is good, but very large slabs are
  1787. * currently bad for the gfp()s.
  1788. */
  1789. if (gfporder >= slab_break_gfp_order)
  1790. break;
  1791. /*
  1792. * Acceptable internal fragmentation?
  1793. */
  1794. if (left_over * 8 <= (PAGE_SIZE << gfporder))
  1795. break;
  1796. }
  1797. return left_over;
  1798. }
  1799. static int setup_cpu_cache(struct kmem_cache *cachep)
  1800. {
  1801. if (g_cpucache_up == FULL)
  1802. return enable_cpucache(cachep);
  1803. if (g_cpucache_up == NONE) {
  1804. /*
  1805. * Note: the first kmem_cache_create must create the cache
  1806. * that's used by kmalloc(24), otherwise the creation of
  1807. * further caches will BUG().
  1808. */
  1809. cachep->array[smp_processor_id()] = &initarray_generic.cache;
  1810. /*
  1811. * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
  1812. * the first cache, then we need to set up all its list3s,
  1813. * otherwise the creation of further caches will BUG().
  1814. */
  1815. set_up_list3s(cachep, SIZE_AC);
  1816. if (INDEX_AC == INDEX_L3)
  1817. g_cpucache_up = PARTIAL_L3;
  1818. else
  1819. g_cpucache_up = PARTIAL_AC;
  1820. } else {
  1821. cachep->array[smp_processor_id()] =
  1822. kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
  1823. if (g_cpucache_up == PARTIAL_AC) {
  1824. set_up_list3s(cachep, SIZE_L3);
  1825. g_cpucache_up = PARTIAL_L3;
  1826. } else {
  1827. int node;
  1828. for_each_online_node(node) {
  1829. cachep->nodelists[node] =
  1830. kmalloc_node(sizeof(struct kmem_list3),
  1831. GFP_KERNEL, node);
  1832. BUG_ON(!cachep->nodelists[node]);
  1833. kmem_list3_init(cachep->nodelists[node]);
  1834. }
  1835. }
  1836. }
  1837. cachep->nodelists[numa_node_id()]->next_reap =
  1838. jiffies + REAPTIMEOUT_LIST3 +
  1839. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1840. cpu_cache_get(cachep)->avail = 0;
  1841. cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
  1842. cpu_cache_get(cachep)->batchcount = 1;
  1843. cpu_cache_get(cachep)->touched = 0;
  1844. cachep->batchcount = 1;
  1845. cachep->limit = BOOT_CPUCACHE_ENTRIES;
  1846. return 0;
  1847. }
  1848. /**
  1849. * kmem_cache_create - Create a cache.
  1850. * @name: A string which is used in /proc/slabinfo to identify this cache.
  1851. * @size: The size of objects to be created in this cache.
  1852. * @align: The required alignment for the objects.
  1853. * @flags: SLAB flags
  1854. * @ctor: A constructor for the objects.
  1855. * @dtor: A destructor for the objects.
  1856. *
  1857. * Returns a ptr to the cache on success, NULL on failure.
  1858. * Cannot be called within a int, but can be interrupted.
  1859. * The @ctor is run when new pages are allocated by the cache
  1860. * and the @dtor is run before the pages are handed back.
  1861. *
  1862. * @name must be valid until the cache is destroyed. This implies that
  1863. * the module calling this has to destroy the cache before getting unloaded.
  1864. *
  1865. * The flags are
  1866. *
  1867. * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
  1868. * to catch references to uninitialised memory.
  1869. *
  1870. * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
  1871. * for buffer overruns.
  1872. *
  1873. * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
  1874. * cacheline. This can be beneficial if you're counting cycles as closely
  1875. * as davem.
  1876. */
  1877. struct kmem_cache *
  1878. kmem_cache_create (const char *name, size_t size, size_t align,
  1879. unsigned long flags,
  1880. void (*ctor)(void*, struct kmem_cache *, unsigned long),
  1881. void (*dtor)(void*, struct kmem_cache *, unsigned long))
  1882. {
  1883. size_t left_over, slab_size, ralign;
  1884. struct kmem_cache *cachep = NULL, *pc;
  1885. /*
  1886. * Sanity checks... these are all serious usage bugs.
  1887. */
  1888. if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
  1889. (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
  1890. printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
  1891. name);
  1892. BUG();
  1893. }
  1894. /*
  1895. * We use cache_chain_mutex to ensure a consistent view of
  1896. * cpu_online_map as well. Please see cpuup_callback
  1897. */
  1898. mutex_lock(&cache_chain_mutex);
  1899. list_for_each_entry(pc, &cache_chain, next) {
  1900. char tmp;
  1901. int res;
  1902. /*
  1903. * This happens when the module gets unloaded and doesn't
  1904. * destroy its slab cache and no-one else reuses the vmalloc
  1905. * area of the module. Print a warning.
  1906. */
  1907. res = probe_kernel_address(pc->name, tmp);
  1908. if (res) {
  1909. printk("SLAB: cache with size %d has lost its name\n",
  1910. pc->buffer_size);
  1911. continue;
  1912. }
  1913. if (!strcmp(pc->name, name)) {
  1914. printk("kmem_cache_create: duplicate cache %s\n", name);
  1915. dump_stack();
  1916. goto oops;
  1917. }
  1918. }
  1919. #if DEBUG
  1920. WARN_ON(strchr(name, ' ')); /* It confuses parsers */
  1921. if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
  1922. /* No constructor, but inital state check requested */
  1923. printk(KERN_ERR "%s: No con, but init state check "
  1924. "requested - %s\n", __FUNCTION__, name);
  1925. flags &= ~SLAB_DEBUG_INITIAL;
  1926. }
  1927. #if FORCED_DEBUG
  1928. /*
  1929. * Enable redzoning and last user accounting, except for caches with
  1930. * large objects, if the increased size would increase the object size
  1931. * above the next power of two: caches with object sizes just above a
  1932. * power of two have a significant amount of internal fragmentation.
  1933. */
  1934. if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
  1935. flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
  1936. if (!(flags & SLAB_DESTROY_BY_RCU))
  1937. flags |= SLAB_POISON;
  1938. #endif
  1939. if (flags & SLAB_DESTROY_BY_RCU)
  1940. BUG_ON(flags & SLAB_POISON);
  1941. #endif
  1942. if (flags & SLAB_DESTROY_BY_RCU)
  1943. BUG_ON(dtor);
  1944. /*
  1945. * Always checks flags, a caller might be expecting debug support which
  1946. * isn't available.
  1947. */
  1948. BUG_ON(flags & ~CREATE_MASK);
  1949. /*
  1950. * Check that size is in terms of words. This is needed to avoid
  1951. * unaligned accesses for some archs when redzoning is used, and makes
  1952. * sure any on-slab bufctl's are also correctly aligned.
  1953. */
  1954. if (size & (BYTES_PER_WORD - 1)) {
  1955. size += (BYTES_PER_WORD - 1);
  1956. size &= ~(BYTES_PER_WORD - 1);
  1957. }
  1958. /* calculate the final buffer alignment: */
  1959. /* 1) arch recommendation: can be overridden for debug */
  1960. if (flags & SLAB_HWCACHE_ALIGN) {
  1961. /*
  1962. * Default alignment: as specified by the arch code. Except if
  1963. * an object is really small, then squeeze multiple objects into
  1964. * one cacheline.
  1965. */
  1966. ralign = cache_line_size();
  1967. while (size <= ralign / 2)
  1968. ralign /= 2;
  1969. } else {
  1970. ralign = BYTES_PER_WORD;
  1971. }
  1972. /*
  1973. * Redzoning and user store require word alignment. Note this will be
  1974. * overridden by architecture or caller mandated alignment if either
  1975. * is greater than BYTES_PER_WORD.
  1976. */
  1977. if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
  1978. ralign = BYTES_PER_WORD;
  1979. /* 2) arch mandated alignment */
  1980. if (ralign < ARCH_SLAB_MINALIGN) {
  1981. ralign = ARCH_SLAB_MINALIGN;
  1982. }
  1983. /* 3) caller mandated alignment */
  1984. if (ralign < align) {
  1985. ralign = align;
  1986. }
  1987. /* disable debug if necessary */
  1988. if (ralign > BYTES_PER_WORD)
  1989. flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
  1990. /*
  1991. * 4) Store it.
  1992. */
  1993. align = ralign;
  1994. /* Get cache's description obj. */
  1995. cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
  1996. if (!cachep)
  1997. goto oops;
  1998. #if DEBUG
  1999. cachep->obj_size = size;
  2000. /*
  2001. * Both debugging options require word-alignment which is calculated
  2002. * into align above.
  2003. */
  2004. if (flags & SLAB_RED_ZONE) {
  2005. /* add space for red zone words */
  2006. cachep->obj_offset += BYTES_PER_WORD;
  2007. size += 2 * BYTES_PER_WORD;
  2008. }
  2009. if (flags & SLAB_STORE_USER) {
  2010. /* user store requires one word storage behind the end of
  2011. * the real object.
  2012. */
  2013. size += BYTES_PER_WORD;
  2014. }
  2015. #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
  2016. if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
  2017. && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
  2018. cachep->obj_offset += PAGE_SIZE - size;
  2019. size = PAGE_SIZE;
  2020. }
  2021. #endif
  2022. #endif
  2023. /*
  2024. * Determine if the slab management is 'on' or 'off' slab.
  2025. * (bootstrapping cannot cope with offslab caches so don't do
  2026. * it too early on.)
  2027. */
  2028. if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
  2029. /*
  2030. * Size is large, assume best to place the slab management obj
  2031. * off-slab (should allow better packing of objs).
  2032. */
  2033. flags |= CFLGS_OFF_SLAB;
  2034. size = ALIGN(size, align);
  2035. left_over = calculate_slab_order(cachep, size, align, flags);
  2036. if (!cachep->num) {
  2037. printk("kmem_cache_create: couldn't create cache %s.\n", name);
  2038. kmem_cache_free(&cache_cache, cachep);
  2039. cachep = NULL;
  2040. goto oops;
  2041. }
  2042. slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
  2043. + sizeof(struct slab), align);
  2044. /*
  2045. * If the slab has been placed off-slab, and we have enough space then
  2046. * move it on-slab. This is at the expense of any extra colouring.
  2047. */
  2048. if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
  2049. flags &= ~CFLGS_OFF_SLAB;
  2050. left_over -= slab_size;
  2051. }
  2052. if (flags & CFLGS_OFF_SLAB) {
  2053. /* really off slab. No need for manual alignment */
  2054. slab_size =
  2055. cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
  2056. }
  2057. cachep->colour_off = cache_line_size();
  2058. /* Offset must be a multiple of the alignment. */
  2059. if (cachep->colour_off < align)
  2060. cachep->colour_off = align;
  2061. cachep->colour = left_over / cachep->colour_off;
  2062. cachep->slab_size = slab_size;
  2063. cachep->flags = flags;
  2064. cachep->gfpflags = 0;
  2065. if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
  2066. cachep->gfpflags |= GFP_DMA;
  2067. cachep->buffer_size = size;
  2068. cachep->reciprocal_buffer_size = reciprocal_value(size);
  2069. if (flags & CFLGS_OFF_SLAB) {
  2070. cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
  2071. /*
  2072. * This is a possibility for one of the malloc_sizes caches.
  2073. * But since we go off slab only for object size greater than
  2074. * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
  2075. * this should not happen at all.
  2076. * But leave a BUG_ON for some lucky dude.
  2077. */
  2078. BUG_ON(!cachep->slabp_cache);
  2079. }
  2080. cachep->ctor = ctor;
  2081. cachep->dtor = dtor;
  2082. cachep->name = name;
  2083. if (setup_cpu_cache(cachep)) {
  2084. __kmem_cache_destroy(cachep);
  2085. cachep = NULL;
  2086. goto oops;
  2087. }
  2088. /* cache setup completed, link it into the list */
  2089. list_add(&cachep->next, &cache_chain);
  2090. oops:
  2091. if (!cachep && (flags & SLAB_PANIC))
  2092. panic("kmem_cache_create(): failed to create slab `%s'\n",
  2093. name);
  2094. mutex_unlock(&cache_chain_mutex);
  2095. return cachep;
  2096. }
  2097. EXPORT_SYMBOL(kmem_cache_create);
  2098. #if DEBUG
  2099. static void check_irq_off(void)
  2100. {
  2101. BUG_ON(!irqs_disabled());
  2102. }
  2103. static void check_irq_on(void)
  2104. {
  2105. BUG_ON(irqs_disabled());
  2106. }
  2107. static void check_spinlock_acquired(struct kmem_cache *cachep)
  2108. {
  2109. #ifdef CONFIG_SMP
  2110. check_irq_off();
  2111. assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
  2112. #endif
  2113. }
  2114. static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
  2115. {
  2116. #ifdef CONFIG_SMP
  2117. check_irq_off();
  2118. assert_spin_locked(&cachep->nodelists[node]->list_lock);
  2119. #endif
  2120. }
  2121. #else
  2122. #define check_irq_off() do { } while(0)
  2123. #define check_irq_on() do { } while(0)
  2124. #define check_spinlock_acquired(x) do { } while(0)
  2125. #define check_spinlock_acquired_node(x, y) do { } while(0)
  2126. #endif
  2127. static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
  2128. struct array_cache *ac,
  2129. int force, int node);
  2130. static void do_drain(void *arg)
  2131. {
  2132. struct kmem_cache *cachep = arg;
  2133. struct array_cache *ac;
  2134. int node = numa_node_id();
  2135. check_irq_off();
  2136. ac = cpu_cache_get(cachep);
  2137. spin_lock(&cachep->nodelists[node]->list_lock);
  2138. free_block(cachep, ac->entry, ac->avail, node);
  2139. spin_unlock(&cachep->nodelists[node]->list_lock);
  2140. ac->avail = 0;
  2141. }
  2142. static void drain_cpu_caches(struct kmem_cache *cachep)
  2143. {
  2144. struct kmem_list3 *l3;
  2145. int node;
  2146. on_each_cpu(do_drain, cachep, 1, 1);
  2147. check_irq_on();
  2148. for_each_online_node(node) {
  2149. l3 = cachep->nodelists[node];
  2150. if (l3 && l3->alien)
  2151. drain_alien_cache(cachep, l3->alien);
  2152. }
  2153. for_each_online_node(node) {
  2154. l3 = cachep->nodelists[node];
  2155. if (l3)
  2156. drain_array(cachep, l3, l3->shared, 1, node);
  2157. }
  2158. }
  2159. /*
  2160. * Remove slabs from the list of free slabs.
  2161. * Specify the number of slabs to drain in tofree.
  2162. *
  2163. * Returns the actual number of slabs released.
  2164. */
  2165. static int drain_freelist(struct kmem_cache *cache,
  2166. struct kmem_list3 *l3, int tofree)
  2167. {
  2168. struct list_head *p;
  2169. int nr_freed;
  2170. struct slab *slabp;
  2171. nr_freed = 0;
  2172. while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
  2173. spin_lock_irq(&l3->list_lock);
  2174. p = l3->slabs_free.prev;
  2175. if (p == &l3->slabs_free) {
  2176. spin_unlock_irq(&l3->list_lock);
  2177. goto out;
  2178. }
  2179. slabp = list_entry(p, struct slab, list);
  2180. #if DEBUG
  2181. BUG_ON(slabp->inuse);
  2182. #endif
  2183. list_del(&slabp->list);
  2184. /*
  2185. * Safe to drop the lock. The slab is no longer linked
  2186. * to the cache.
  2187. */
  2188. l3->free_objects -= cache->num;
  2189. spin_unlock_irq(&l3->list_lock);
  2190. slab_destroy(cache, slabp);
  2191. nr_freed++;
  2192. }
  2193. out:
  2194. return nr_freed;
  2195. }
  2196. /* Called with cache_chain_mutex held to protect against cpu hotplug */
  2197. static int __cache_shrink(struct kmem_cache *cachep)
  2198. {
  2199. int ret = 0, i = 0;
  2200. struct kmem_list3 *l3;
  2201. drain_cpu_caches(cachep);
  2202. check_irq_on();
  2203. for_each_online_node(i) {
  2204. l3 = cachep->nodelists[i];
  2205. if (!l3)
  2206. continue;
  2207. drain_freelist(cachep, l3, l3->free_objects);
  2208. ret += !list_empty(&l3->slabs_full) ||
  2209. !list_empty(&l3->slabs_partial);
  2210. }
  2211. return (ret ? 1 : 0);
  2212. }
  2213. /**
  2214. * kmem_cache_shrink - Shrink a cache.
  2215. * @cachep: The cache to shrink.
  2216. *
  2217. * Releases as many slabs as possible for a cache.
  2218. * To help debugging, a zero exit status indicates all slabs were released.
  2219. */
  2220. int kmem_cache_shrink(struct kmem_cache *cachep)
  2221. {
  2222. int ret;
  2223. BUG_ON(!cachep || in_interrupt());
  2224. mutex_lock(&cache_chain_mutex);
  2225. ret = __cache_shrink(cachep);
  2226. mutex_unlock(&cache_chain_mutex);
  2227. return ret;
  2228. }
  2229. EXPORT_SYMBOL(kmem_cache_shrink);
  2230. /**
  2231. * kmem_cache_destroy - delete a cache
  2232. * @cachep: the cache to destroy
  2233. *
  2234. * Remove a &struct kmem_cache object from the slab cache.
  2235. *
  2236. * It is expected this function will be called by a module when it is
  2237. * unloaded. This will remove the cache completely, and avoid a duplicate
  2238. * cache being allocated each time a module is loaded and unloaded, if the
  2239. * module doesn't have persistent in-kernel storage across loads and unloads.
  2240. *
  2241. * The cache must be empty before calling this function.
  2242. *
  2243. * The caller must guarantee that noone will allocate memory from the cache
  2244. * during the kmem_cache_destroy().
  2245. */
  2246. void kmem_cache_destroy(struct kmem_cache *cachep)
  2247. {
  2248. BUG_ON(!cachep || in_interrupt());
  2249. /* Find the cache in the chain of caches. */
  2250. mutex_lock(&cache_chain_mutex);
  2251. /*
  2252. * the chain is never empty, cache_cache is never destroyed
  2253. */
  2254. list_del(&cachep->next);
  2255. if (__cache_shrink(cachep)) {
  2256. slab_error(cachep, "Can't free all objects");
  2257. list_add(&cachep->next, &cache_chain);
  2258. mutex_unlock(&cache_chain_mutex);
  2259. return;
  2260. }
  2261. if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
  2262. synchronize_rcu();
  2263. __kmem_cache_destroy(cachep);
  2264. mutex_unlock(&cache_chain_mutex);
  2265. }
  2266. EXPORT_SYMBOL(kmem_cache_destroy);
  2267. /*
  2268. * Get the memory for a slab management obj.
  2269. * For a slab cache when the slab descriptor is off-slab, slab descriptors
  2270. * always come from malloc_sizes caches. The slab descriptor cannot
  2271. * come from the same cache which is getting created because,
  2272. * when we are searching for an appropriate cache for these
  2273. * descriptors in kmem_cache_create, we search through the malloc_sizes array.
  2274. * If we are creating a malloc_sizes cache here it would not be visible to
  2275. * kmem_find_general_cachep till the initialization is complete.
  2276. * Hence we cannot have slabp_cache same as the original cache.
  2277. */
  2278. static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
  2279. int colour_off, gfp_t local_flags,
  2280. int nodeid)
  2281. {
  2282. struct slab *slabp;
  2283. if (OFF_SLAB(cachep)) {
  2284. /* Slab management obj is off-slab. */
  2285. slabp = kmem_cache_alloc_node(cachep->slabp_cache,
  2286. local_flags & ~GFP_THISNODE, nodeid);
  2287. if (!slabp)
  2288. return NULL;
  2289. } else {
  2290. slabp = objp + colour_off;
  2291. colour_off += cachep->slab_size;
  2292. }
  2293. slabp->inuse = 0;
  2294. slabp->colouroff = colour_off;
  2295. slabp->s_mem = objp + colour_off;
  2296. slabp->nodeid = nodeid;
  2297. return slabp;
  2298. }
  2299. static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
  2300. {
  2301. return (kmem_bufctl_t *) (slabp + 1);
  2302. }
  2303. static void cache_init_objs(struct kmem_cache *cachep,
  2304. struct slab *slabp, unsigned long ctor_flags)
  2305. {
  2306. int i;
  2307. for (i = 0; i < cachep->num; i++) {
  2308. void *objp = index_to_obj(cachep, slabp, i);
  2309. #if DEBUG
  2310. /* need to poison the objs? */
  2311. if (cachep->flags & SLAB_POISON)
  2312. poison_obj(cachep, objp, POISON_FREE);
  2313. if (cachep->flags & SLAB_STORE_USER)
  2314. *dbg_userword(cachep, objp) = NULL;
  2315. if (cachep->flags & SLAB_RED_ZONE) {
  2316. *dbg_redzone1(cachep, objp) = RED_INACTIVE;
  2317. *dbg_redzone2(cachep, objp) = RED_INACTIVE;
  2318. }
  2319. /*
  2320. * Constructors are not allowed to allocate memory from the same
  2321. * cache which they are a constructor for. Otherwise, deadlock.
  2322. * They must also be threaded.
  2323. */
  2324. if (cachep->ctor && !(cachep->flags & SLAB_POISON))
  2325. cachep->ctor(objp + obj_offset(cachep), cachep,
  2326. ctor_flags);
  2327. if (cachep->flags & SLAB_RED_ZONE) {
  2328. if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
  2329. slab_error(cachep, "constructor overwrote the"
  2330. " end of an object");
  2331. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
  2332. slab_error(cachep, "constructor overwrote the"
  2333. " start of an object");
  2334. }
  2335. if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
  2336. OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
  2337. kernel_map_pages(virt_to_page(objp),
  2338. cachep->buffer_size / PAGE_SIZE, 0);
  2339. #else
  2340. if (cachep->ctor)
  2341. cachep->ctor(objp, cachep, ctor_flags);
  2342. #endif
  2343. slab_bufctl(slabp)[i] = i + 1;
  2344. }
  2345. slab_bufctl(slabp)[i - 1] = BUFCTL_END;
  2346. slabp->free = 0;
  2347. }
  2348. static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
  2349. {
  2350. if (CONFIG_ZONE_DMA_FLAG) {
  2351. if (flags & GFP_DMA)
  2352. BUG_ON(!(cachep->gfpflags & GFP_DMA));
  2353. else
  2354. BUG_ON(cachep->gfpflags & GFP_DMA);
  2355. }
  2356. }
  2357. static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
  2358. int nodeid)
  2359. {
  2360. void *objp = index_to_obj(cachep, slabp, slabp->free);
  2361. kmem_bufctl_t next;
  2362. slabp->inuse++;
  2363. next = slab_bufctl(slabp)[slabp->free];
  2364. #if DEBUG
  2365. slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
  2366. WARN_ON(slabp->nodeid != nodeid);
  2367. #endif
  2368. slabp->free = next;
  2369. return objp;
  2370. }
  2371. static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
  2372. void *objp, int nodeid)
  2373. {
  2374. unsigned int objnr = obj_to_index(cachep, slabp, objp);
  2375. #if DEBUG
  2376. /* Verify that the slab belongs to the intended node */
  2377. WARN_ON(slabp->nodeid != nodeid);
  2378. if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
  2379. printk(KERN_ERR "slab: double free detected in cache "
  2380. "'%s', objp %p\n", cachep->name, objp);
  2381. BUG();
  2382. }
  2383. #endif
  2384. slab_bufctl(slabp)[objnr] = slabp->free;
  2385. slabp->free = objnr;
  2386. slabp->inuse--;
  2387. }
  2388. /*
  2389. * Map pages beginning at addr to the given cache and slab. This is required
  2390. * for the slab allocator to be able to lookup the cache and slab of a
  2391. * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
  2392. */
  2393. static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
  2394. void *addr)
  2395. {
  2396. int nr_pages;
  2397. struct page *page;
  2398. page = virt_to_page(addr);
  2399. nr_pages = 1;
  2400. if (likely(!PageCompound(page)))
  2401. nr_pages <<= cache->gfporder;
  2402. do {
  2403. page_set_cache(page, cache);
  2404. page_set_slab(page, slab);
  2405. page++;
  2406. } while (--nr_pages);
  2407. }
  2408. /*
  2409. * Grow (by 1) the number of slabs within a cache. This is called by
  2410. * kmem_cache_alloc() when there are no active objs left in a cache.
  2411. */
  2412. static int cache_grow(struct kmem_cache *cachep,
  2413. gfp_t flags, int nodeid, void *objp)
  2414. {
  2415. struct slab *slabp;
  2416. size_t offset;
  2417. gfp_t local_flags;
  2418. unsigned long ctor_flags;
  2419. struct kmem_list3 *l3;
  2420. /*
  2421. * Be lazy and only check for valid flags here, keeping it out of the
  2422. * critical path in kmem_cache_alloc().
  2423. */
  2424. BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
  2425. if (flags & __GFP_NO_GROW)
  2426. return 0;
  2427. ctor_flags = SLAB_CTOR_CONSTRUCTOR;
  2428. local_flags = (flags & GFP_LEVEL_MASK);
  2429. if (!(local_flags & __GFP_WAIT))
  2430. /*
  2431. * Not allowed to sleep. Need to tell a constructor about
  2432. * this - it might need to know...
  2433. */
  2434. ctor_flags |= SLAB_CTOR_ATOMIC;
  2435. /* Take the l3 list lock to change the colour_next on this node */
  2436. check_irq_off();
  2437. l3 = cachep->nodelists[nodeid];
  2438. spin_lock(&l3->list_lock);
  2439. /* Get colour for the slab, and cal the next value. */
  2440. offset = l3->colour_next;
  2441. l3->colour_next++;
  2442. if (l3->colour_next >= cachep->colour)
  2443. l3->colour_next = 0;
  2444. spin_unlock(&l3->list_lock);
  2445. offset *= cachep->colour_off;
  2446. if (local_flags & __GFP_WAIT)
  2447. local_irq_enable();
  2448. /*
  2449. * The test for missing atomic flag is performed here, rather than
  2450. * the more obvious place, simply to reduce the critical path length
  2451. * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
  2452. * will eventually be caught here (where it matters).
  2453. */
  2454. kmem_flagcheck(cachep, flags);
  2455. /*
  2456. * Get mem for the objs. Attempt to allocate a physical page from
  2457. * 'nodeid'.
  2458. */
  2459. if (!objp)
  2460. objp = kmem_getpages(cachep, flags, nodeid);
  2461. if (!objp)
  2462. goto failed;
  2463. /* Get slab management. */
  2464. slabp = alloc_slabmgmt(cachep, objp, offset,
  2465. local_flags & ~GFP_THISNODE, nodeid);
  2466. if (!slabp)
  2467. goto opps1;
  2468. slabp->nodeid = nodeid;
  2469. slab_map_pages(cachep, slabp, objp);
  2470. cache_init_objs(cachep, slabp, ctor_flags);
  2471. if (local_flags & __GFP_WAIT)
  2472. local_irq_disable();
  2473. check_irq_off();
  2474. spin_lock(&l3->list_lock);
  2475. /* Make slab active. */
  2476. list_add_tail(&slabp->list, &(l3->slabs_free));
  2477. STATS_INC_GROWN(cachep);
  2478. l3->free_objects += cachep->num;
  2479. spin_unlock(&l3->list_lock);
  2480. return 1;
  2481. opps1:
  2482. kmem_freepages(cachep, objp);
  2483. failed:
  2484. if (local_flags & __GFP_WAIT)
  2485. local_irq_disable();
  2486. return 0;
  2487. }
  2488. #if DEBUG
  2489. /*
  2490. * Perform extra freeing checks:
  2491. * - detect bad pointers.
  2492. * - POISON/RED_ZONE checking
  2493. * - destructor calls, for caches with POISON+dtor
  2494. */
  2495. static void kfree_debugcheck(const void *objp)
  2496. {
  2497. if (!virt_addr_valid(objp)) {
  2498. printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
  2499. (unsigned long)objp);
  2500. BUG();
  2501. }
  2502. }
  2503. static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
  2504. {
  2505. unsigned long redzone1, redzone2;
  2506. redzone1 = *dbg_redzone1(cache, obj);
  2507. redzone2 = *dbg_redzone2(cache, obj);
  2508. /*
  2509. * Redzone is ok.
  2510. */
  2511. if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
  2512. return;
  2513. if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
  2514. slab_error(cache, "double free detected");
  2515. else
  2516. slab_error(cache, "memory outside object was overwritten");
  2517. printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
  2518. obj, redzone1, redzone2);
  2519. }
  2520. static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
  2521. void *caller)
  2522. {
  2523. struct page *page;
  2524. unsigned int objnr;
  2525. struct slab *slabp;
  2526. objp -= obj_offset(cachep);
  2527. kfree_debugcheck(objp);
  2528. page = virt_to_page(objp);
  2529. slabp = page_get_slab(page);
  2530. if (cachep->flags & SLAB_RED_ZONE) {
  2531. verify_redzone_free(cachep, objp);
  2532. *dbg_redzone1(cachep, objp) = RED_INACTIVE;
  2533. *dbg_redzone2(cachep, objp) = RED_INACTIVE;
  2534. }
  2535. if (cachep->flags & SLAB_STORE_USER)
  2536. *dbg_userword(cachep, objp) = caller;
  2537. objnr = obj_to_index(cachep, slabp, objp);
  2538. BUG_ON(objnr >= cachep->num);
  2539. BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
  2540. if (cachep->flags & SLAB_DEBUG_INITIAL) {
  2541. /*
  2542. * Need to call the slab's constructor so the caller can
  2543. * perform a verify of its state (debugging). Called without
  2544. * the cache-lock held.
  2545. */
  2546. cachep->ctor(objp + obj_offset(cachep),
  2547. cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
  2548. }
  2549. if (cachep->flags & SLAB_POISON && cachep->dtor) {
  2550. /* we want to cache poison the object,
  2551. * call the destruction callback
  2552. */
  2553. cachep->dtor(objp + obj_offset(cachep), cachep, 0);
  2554. }
  2555. #ifdef CONFIG_DEBUG_SLAB_LEAK
  2556. slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
  2557. #endif
  2558. if (cachep->flags & SLAB_POISON) {
  2559. #ifdef CONFIG_DEBUG_PAGEALLOC
  2560. if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
  2561. store_stackinfo(cachep, objp, (unsigned long)caller);
  2562. kernel_map_pages(virt_to_page(objp),
  2563. cachep->buffer_size / PAGE_SIZE, 0);
  2564. } else {
  2565. poison_obj(cachep, objp, POISON_FREE);
  2566. }
  2567. #else
  2568. poison_obj(cachep, objp, POISON_FREE);
  2569. #endif
  2570. }
  2571. return objp;
  2572. }
  2573. static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
  2574. {
  2575. kmem_bufctl_t i;
  2576. int entries = 0;
  2577. /* Check slab's freelist to see if this obj is there. */
  2578. for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
  2579. entries++;
  2580. if (entries > cachep->num || i >= cachep->num)
  2581. goto bad;
  2582. }
  2583. if (entries != cachep->num - slabp->inuse) {
  2584. bad:
  2585. printk(KERN_ERR "slab: Internal list corruption detected in "
  2586. "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
  2587. cachep->name, cachep->num, slabp, slabp->inuse);
  2588. for (i = 0;
  2589. i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
  2590. i++) {
  2591. if (i % 16 == 0)
  2592. printk("\n%03x:", i);
  2593. printk(" %02x", ((unsigned char *)slabp)[i]);
  2594. }
  2595. printk("\n");
  2596. BUG();
  2597. }
  2598. }
  2599. #else
  2600. #define kfree_debugcheck(x) do { } while(0)
  2601. #define cache_free_debugcheck(x,objp,z) (objp)
  2602. #define check_slabp(x,y) do { } while(0)
  2603. #endif
  2604. static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
  2605. {
  2606. int batchcount;
  2607. struct kmem_list3 *l3;
  2608. struct array_cache *ac;
  2609. int node;
  2610. node = numa_node_id();
  2611. check_irq_off();
  2612. ac = cpu_cache_get(cachep);
  2613. retry:
  2614. batchcount = ac->batchcount;
  2615. if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
  2616. /*
  2617. * If there was little recent activity on this cache, then
  2618. * perform only a partial refill. Otherwise we could generate
  2619. * refill bouncing.
  2620. */
  2621. batchcount = BATCHREFILL_LIMIT;
  2622. }
  2623. l3 = cachep->nodelists[node];
  2624. BUG_ON(ac->avail > 0 || !l3);
  2625. spin_lock(&l3->list_lock);
  2626. /* See if we can refill from the shared array */
  2627. if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
  2628. goto alloc_done;
  2629. while (batchcount > 0) {
  2630. struct list_head *entry;
  2631. struct slab *slabp;
  2632. /* Get slab alloc is to come from. */
  2633. entry = l3->slabs_partial.next;
  2634. if (entry == &l3->slabs_partial) {
  2635. l3->free_touched = 1;
  2636. entry = l3->slabs_free.next;
  2637. if (entry == &l3->slabs_free)
  2638. goto must_grow;
  2639. }
  2640. slabp = list_entry(entry, struct slab, list);
  2641. check_slabp(cachep, slabp);
  2642. check_spinlock_acquired(cachep);
  2643. while (slabp->inuse < cachep->num && batchcount--) {
  2644. STATS_INC_ALLOCED(cachep);
  2645. STATS_INC_ACTIVE(cachep);
  2646. STATS_SET_HIGH(cachep);
  2647. ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
  2648. node);
  2649. }
  2650. check_slabp(cachep, slabp);
  2651. /* move slabp to correct slabp list: */
  2652. list_del(&slabp->list);
  2653. if (slabp->free == BUFCTL_END)
  2654. list_add(&slabp->list, &l3->slabs_full);
  2655. else
  2656. list_add(&slabp->list, &l3->slabs_partial);
  2657. }
  2658. must_grow:
  2659. l3->free_objects -= ac->avail;
  2660. alloc_done:
  2661. spin_unlock(&l3->list_lock);
  2662. if (unlikely(!ac->avail)) {
  2663. int x;
  2664. x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
  2665. /* cache_grow can reenable interrupts, then ac could change. */
  2666. ac = cpu_cache_get(cachep);
  2667. if (!x && ac->avail == 0) /* no objects in sight? abort */
  2668. return NULL;
  2669. if (!ac->avail) /* objects refilled by interrupt? */
  2670. goto retry;
  2671. }
  2672. ac->touched = 1;
  2673. return ac->entry[--ac->avail];
  2674. }
  2675. static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
  2676. gfp_t flags)
  2677. {
  2678. might_sleep_if(flags & __GFP_WAIT);
  2679. #if DEBUG
  2680. kmem_flagcheck(cachep, flags);
  2681. #endif
  2682. }
  2683. #if DEBUG
  2684. static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
  2685. gfp_t flags, void *objp, void *caller)
  2686. {
  2687. if (!objp)
  2688. return objp;
  2689. if (cachep->flags & SLAB_POISON) {
  2690. #ifdef CONFIG_DEBUG_PAGEALLOC
  2691. if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
  2692. kernel_map_pages(virt_to_page(objp),
  2693. cachep->buffer_size / PAGE_SIZE, 1);
  2694. else
  2695. check_poison_obj(cachep, objp);
  2696. #else
  2697. check_poison_obj(cachep, objp);
  2698. #endif
  2699. poison_obj(cachep, objp, POISON_INUSE);
  2700. }
  2701. if (cachep->flags & SLAB_STORE_USER)
  2702. *dbg_userword(cachep, objp) = caller;
  2703. if (cachep->flags & SLAB_RED_ZONE) {
  2704. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
  2705. *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
  2706. slab_error(cachep, "double free, or memory outside"
  2707. " object was overwritten");
  2708. printk(KERN_ERR
  2709. "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
  2710. objp, *dbg_redzone1(cachep, objp),
  2711. *dbg_redzone2(cachep, objp));
  2712. }
  2713. *dbg_redzone1(cachep, objp) = RED_ACTIVE;
  2714. *dbg_redzone2(cachep, objp) = RED_ACTIVE;
  2715. }
  2716. #ifdef CONFIG_DEBUG_SLAB_LEAK
  2717. {
  2718. struct slab *slabp;
  2719. unsigned objnr;
  2720. slabp = page_get_slab(virt_to_page(objp));
  2721. objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
  2722. slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
  2723. }
  2724. #endif
  2725. objp += obj_offset(cachep);
  2726. if (cachep->ctor && cachep->flags & SLAB_POISON) {
  2727. unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
  2728. if (!(flags & __GFP_WAIT))
  2729. ctor_flags |= SLAB_CTOR_ATOMIC;
  2730. cachep->ctor(objp, cachep, ctor_flags);
  2731. }
  2732. #if ARCH_SLAB_MINALIGN
  2733. if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
  2734. printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
  2735. objp, ARCH_SLAB_MINALIGN);
  2736. }
  2737. #endif
  2738. return objp;
  2739. }
  2740. #else
  2741. #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
  2742. #endif
  2743. #ifdef CONFIG_FAILSLAB
  2744. static struct failslab_attr {
  2745. struct fault_attr attr;
  2746. u32 ignore_gfp_wait;
  2747. #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
  2748. struct dentry *ignore_gfp_wait_file;
  2749. #endif
  2750. } failslab = {
  2751. .attr = FAULT_ATTR_INITIALIZER,
  2752. .ignore_gfp_wait = 1,
  2753. };
  2754. static int __init setup_failslab(char *str)
  2755. {
  2756. return setup_fault_attr(&failslab.attr, str);
  2757. }
  2758. __setup("failslab=", setup_failslab);
  2759. static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
  2760. {
  2761. if (cachep == &cache_cache)
  2762. return 0;
  2763. if (flags & __GFP_NOFAIL)
  2764. return 0;
  2765. if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
  2766. return 0;
  2767. return should_fail(&failslab.attr, obj_size(cachep));
  2768. }
  2769. #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
  2770. static int __init failslab_debugfs(void)
  2771. {
  2772. mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
  2773. struct dentry *dir;
  2774. int err;
  2775. err = init_fault_attr_dentries(&failslab.attr, "failslab");
  2776. if (err)
  2777. return err;
  2778. dir = failslab.attr.dentries.dir;
  2779. failslab.ignore_gfp_wait_file =
  2780. debugfs_create_bool("ignore-gfp-wait", mode, dir,
  2781. &failslab.ignore_gfp_wait);
  2782. if (!failslab.ignore_gfp_wait_file) {
  2783. err = -ENOMEM;
  2784. debugfs_remove(failslab.ignore_gfp_wait_file);
  2785. cleanup_fault_attr_dentries(&failslab.attr);
  2786. }
  2787. return err;
  2788. }
  2789. late_initcall(failslab_debugfs);
  2790. #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
  2791. #else /* CONFIG_FAILSLAB */
  2792. static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
  2793. {
  2794. return 0;
  2795. }
  2796. #endif /* CONFIG_FAILSLAB */
  2797. static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  2798. {
  2799. void *objp;
  2800. struct array_cache *ac;
  2801. check_irq_off();
  2802. if (should_failslab(cachep, flags))
  2803. return NULL;
  2804. ac = cpu_cache_get(cachep);
  2805. if (likely(ac->avail)) {
  2806. STATS_INC_ALLOCHIT(cachep);
  2807. ac->touched = 1;
  2808. objp = ac->entry[--ac->avail];
  2809. } else {
  2810. STATS_INC_ALLOCMISS(cachep);
  2811. objp = cache_alloc_refill(cachep, flags);
  2812. }
  2813. return objp;
  2814. }
  2815. #ifdef CONFIG_NUMA
  2816. /*
  2817. * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
  2818. *
  2819. * If we are in_interrupt, then process context, including cpusets and
  2820. * mempolicy, may not apply and should not be used for allocation policy.
  2821. */
  2822. static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
  2823. {
  2824. int nid_alloc, nid_here;
  2825. if (in_interrupt() || (flags & __GFP_THISNODE))
  2826. return NULL;
  2827. nid_alloc = nid_here = numa_node_id();
  2828. if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
  2829. nid_alloc = cpuset_mem_spread_node();
  2830. else if (current->mempolicy)
  2831. nid_alloc = slab_node(current->mempolicy);
  2832. if (nid_alloc != nid_here)
  2833. return ____cache_alloc_node(cachep, flags, nid_alloc);
  2834. return NULL;
  2835. }
  2836. /*
  2837. * Fallback function if there was no memory available and no objects on a
  2838. * certain node and fall back is permitted. First we scan all the
  2839. * available nodelists for available objects. If that fails then we
  2840. * perform an allocation without specifying a node. This allows the page
  2841. * allocator to do its reclaim / fallback magic. We then insert the
  2842. * slab into the proper nodelist and then allocate from it.
  2843. */
  2844. static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
  2845. {
  2846. struct zonelist *zonelist;
  2847. gfp_t local_flags;
  2848. struct zone **z;
  2849. void *obj = NULL;
  2850. int nid;
  2851. if (flags & __GFP_THISNODE)
  2852. return NULL;
  2853. zonelist = &NODE_DATA(slab_node(current->mempolicy))
  2854. ->node_zonelists[gfp_zone(flags)];
  2855. local_flags = (flags & GFP_LEVEL_MASK);
  2856. retry:
  2857. /*
  2858. * Look through allowed nodes for objects available
  2859. * from existing per node queues.
  2860. */
  2861. for (z = zonelist->zones; *z && !obj; z++) {
  2862. nid = zone_to_nid(*z);
  2863. if (cpuset_zone_allowed_hardwall(*z, flags) &&
  2864. cache->nodelists[nid] &&
  2865. cache->nodelists[nid]->free_objects)
  2866. obj = ____cache_alloc_node(cache,
  2867. flags | GFP_THISNODE, nid);
  2868. }
  2869. if (!obj && !(flags & __GFP_NO_GROW)) {
  2870. /*
  2871. * This allocation will be performed within the constraints
  2872. * of the current cpuset / memory policy requirements.
  2873. * We may trigger various forms of reclaim on the allowed
  2874. * set and go into memory reserves if necessary.
  2875. */
  2876. if (local_flags & __GFP_WAIT)
  2877. local_irq_enable();
  2878. kmem_flagcheck(cache, flags);
  2879. obj = kmem_getpages(cache, flags, -1);
  2880. if (local_flags & __GFP_WAIT)
  2881. local_irq_disable();
  2882. if (obj) {
  2883. /*
  2884. * Insert into the appropriate per node queues
  2885. */
  2886. nid = page_to_nid(virt_to_page(obj));
  2887. if (cache_grow(cache, flags, nid, obj)) {
  2888. obj = ____cache_alloc_node(cache,
  2889. flags | GFP_THISNODE, nid);
  2890. if (!obj)
  2891. /*
  2892. * Another processor may allocate the
  2893. * objects in the slab since we are
  2894. * not holding any locks.
  2895. */
  2896. goto retry;
  2897. } else {
  2898. /* cache_grow already freed obj */
  2899. obj = NULL;
  2900. }
  2901. }
  2902. }
  2903. return obj;
  2904. }
  2905. /*
  2906. * A interface to enable slab creation on nodeid
  2907. */
  2908. static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
  2909. int nodeid)
  2910. {
  2911. struct list_head *entry;
  2912. struct slab *slabp;
  2913. struct kmem_list3 *l3;
  2914. void *obj;
  2915. int x;
  2916. l3 = cachep->nodelists[nodeid];
  2917. BUG_ON(!l3);
  2918. retry:
  2919. check_irq_off();
  2920. spin_lock(&l3->list_lock);
  2921. entry = l3->slabs_partial.next;
  2922. if (entry == &l3->slabs_partial) {
  2923. l3->free_touched = 1;
  2924. entry = l3->slabs_free.next;
  2925. if (entry == &l3->slabs_free)
  2926. goto must_grow;
  2927. }
  2928. slabp = list_entry(entry, struct slab, list);
  2929. check_spinlock_acquired_node(cachep, nodeid);
  2930. check_slabp(cachep, slabp);
  2931. STATS_INC_NODEALLOCS(cachep);
  2932. STATS_INC_ACTIVE(cachep);
  2933. STATS_SET_HIGH(cachep);
  2934. BUG_ON(slabp->inuse == cachep->num);
  2935. obj = slab_get_obj(cachep, slabp, nodeid);
  2936. check_slabp(cachep, slabp);
  2937. l3->free_objects--;
  2938. /* move slabp to correct slabp list: */
  2939. list_del(&slabp->list);
  2940. if (slabp->free == BUFCTL_END)
  2941. list_add(&slabp->list, &l3->slabs_full);
  2942. else
  2943. list_add(&slabp->list, &l3->slabs_partial);
  2944. spin_unlock(&l3->list_lock);
  2945. goto done;
  2946. must_grow:
  2947. spin_unlock(&l3->list_lock);
  2948. x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
  2949. if (x)
  2950. goto retry;
  2951. return fallback_alloc(cachep, flags);
  2952. done:
  2953. return obj;
  2954. }
  2955. /**
  2956. * kmem_cache_alloc_node - Allocate an object on the specified node
  2957. * @cachep: The cache to allocate from.
  2958. * @flags: See kmalloc().
  2959. * @nodeid: node number of the target node.
  2960. * @caller: return address of caller, used for debug information
  2961. *
  2962. * Identical to kmem_cache_alloc but it will allocate memory on the given
  2963. * node, which can improve the performance for cpu bound structures.
  2964. *
  2965. * Fallback to other node is possible if __GFP_THISNODE is not set.
  2966. */
  2967. static __always_inline void *
  2968. __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
  2969. void *caller)
  2970. {
  2971. unsigned long save_flags;
  2972. void *ptr;
  2973. cache_alloc_debugcheck_before(cachep, flags);
  2974. local_irq_save(save_flags);
  2975. if (unlikely(nodeid == -1))
  2976. nodeid = numa_node_id();
  2977. if (unlikely(!cachep->nodelists[nodeid])) {
  2978. /* Node not bootstrapped yet */
  2979. ptr = fallback_alloc(cachep, flags);
  2980. goto out;
  2981. }
  2982. if (nodeid == numa_node_id()) {
  2983. /*
  2984. * Use the locally cached objects if possible.
  2985. * However ____cache_alloc does not allow fallback
  2986. * to other nodes. It may fail while we still have
  2987. * objects on other nodes available.
  2988. */
  2989. ptr = ____cache_alloc(cachep, flags);
  2990. if (ptr)
  2991. goto out;
  2992. }
  2993. /* ___cache_alloc_node can fall back to other nodes */
  2994. ptr = ____cache_alloc_node(cachep, flags, nodeid);
  2995. out:
  2996. local_irq_restore(save_flags);
  2997. ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
  2998. return ptr;
  2999. }
  3000. static __always_inline void *
  3001. __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
  3002. {
  3003. void *objp;
  3004. if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
  3005. objp = alternate_node_alloc(cache, flags);
  3006. if (objp)
  3007. goto out;
  3008. }
  3009. objp = ____cache_alloc(cache, flags);
  3010. /*
  3011. * We may just have run out of memory on the local node.
  3012. * ____cache_alloc_node() knows how to locate memory on other nodes
  3013. */
  3014. if (!objp)
  3015. objp = ____cache_alloc_node(cache, flags, numa_node_id());
  3016. out:
  3017. return objp;
  3018. }
  3019. #else
  3020. static __always_inline void *
  3021. __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  3022. {
  3023. return ____cache_alloc(cachep, flags);
  3024. }
  3025. #endif /* CONFIG_NUMA */
  3026. static __always_inline void *
  3027. __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
  3028. {
  3029. unsigned long save_flags;
  3030. void *objp;
  3031. cache_alloc_debugcheck_before(cachep, flags);
  3032. local_irq_save(save_flags);
  3033. objp = __do_cache_alloc(cachep, flags);
  3034. local_irq_restore(save_flags);
  3035. objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
  3036. prefetchw(objp);
  3037. return objp;
  3038. }
  3039. /*
  3040. * Caller needs to acquire correct kmem_list's list_lock
  3041. */
  3042. static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
  3043. int node)
  3044. {
  3045. int i;
  3046. struct kmem_list3 *l3;
  3047. for (i = 0; i < nr_objects; i++) {
  3048. void *objp = objpp[i];
  3049. struct slab *slabp;
  3050. slabp = virt_to_slab(objp);
  3051. l3 = cachep->nodelists[node];
  3052. list_del(&slabp->list);
  3053. check_spinlock_acquired_node(cachep, node);
  3054. check_slabp(cachep, slabp);
  3055. slab_put_obj(cachep, slabp, objp, node);
  3056. STATS_DEC_ACTIVE(cachep);
  3057. l3->free_objects++;
  3058. check_slabp(cachep, slabp);
  3059. /* fixup slab chains */
  3060. if (slabp->inuse == 0) {
  3061. if (l3->free_objects > l3->free_limit) {
  3062. l3->free_objects -= cachep->num;
  3063. /* No need to drop any previously held
  3064. * lock here, even if we have a off-slab slab
  3065. * descriptor it is guaranteed to come from
  3066. * a different cache, refer to comments before
  3067. * alloc_slabmgmt.
  3068. */
  3069. slab_destroy(cachep, slabp);
  3070. } else {
  3071. list_add(&slabp->list, &l3->slabs_free);
  3072. }
  3073. } else {
  3074. /* Unconditionally move a slab to the end of the
  3075. * partial list on free - maximum time for the
  3076. * other objects to be freed, too.
  3077. */
  3078. list_add_tail(&slabp->list, &l3->slabs_partial);
  3079. }
  3080. }
  3081. }
  3082. static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
  3083. {
  3084. int batchcount;
  3085. struct kmem_list3 *l3;
  3086. int node = numa_node_id();
  3087. batchcount = ac->batchcount;
  3088. #if DEBUG
  3089. BUG_ON(!batchcount || batchcount > ac->avail);
  3090. #endif
  3091. check_irq_off();
  3092. l3 = cachep->nodelists[node];
  3093. spin_lock(&l3->list_lock);
  3094. if (l3->shared) {
  3095. struct array_cache *shared_array = l3->shared;
  3096. int max = shared_array->limit - shared_array->avail;
  3097. if (max) {
  3098. if (batchcount > max)
  3099. batchcount = max;
  3100. memcpy(&(shared_array->entry[shared_array->avail]),
  3101. ac->entry, sizeof(void *) * batchcount);
  3102. shared_array->avail += batchcount;
  3103. goto free_done;
  3104. }
  3105. }
  3106. free_block(cachep, ac->entry, batchcount, node);
  3107. free_done:
  3108. #if STATS
  3109. {
  3110. int i = 0;
  3111. struct list_head *p;
  3112. p = l3->slabs_free.next;
  3113. while (p != &(l3->slabs_free)) {
  3114. struct slab *slabp;
  3115. slabp = list_entry(p, struct slab, list);
  3116. BUG_ON(slabp->inuse);
  3117. i++;
  3118. p = p->next;
  3119. }
  3120. STATS_SET_FREEABLE(cachep, i);
  3121. }
  3122. #endif
  3123. spin_unlock(&l3->list_lock);
  3124. ac->avail -= batchcount;
  3125. memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
  3126. }
  3127. /*
  3128. * Release an obj back to its cache. If the obj has a constructed state, it must
  3129. * be in this state _before_ it is released. Called with disabled ints.
  3130. */
  3131. static inline void __cache_free(struct kmem_cache *cachep, void *objp)
  3132. {
  3133. struct array_cache *ac = cpu_cache_get(cachep);
  3134. check_irq_off();
  3135. objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
  3136. if (cache_free_alien(cachep, objp))
  3137. return;
  3138. if (likely(ac->avail < ac->limit)) {
  3139. STATS_INC_FREEHIT(cachep);
  3140. ac->entry[ac->avail++] = objp;
  3141. return;
  3142. } else {
  3143. STATS_INC_FREEMISS(cachep);
  3144. cache_flusharray(cachep, ac);
  3145. ac->entry[ac->avail++] = objp;
  3146. }
  3147. }
  3148. /**
  3149. * kmem_cache_alloc - Allocate an object
  3150. * @cachep: The cache to allocate from.
  3151. * @flags: See kmalloc().
  3152. *
  3153. * Allocate an object from this cache. The flags are only relevant
  3154. * if the cache has no available objects.
  3155. */
  3156. void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  3157. {
  3158. return __cache_alloc(cachep, flags, __builtin_return_address(0));
  3159. }
  3160. EXPORT_SYMBOL(kmem_cache_alloc);
  3161. /**
  3162. * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
  3163. * @cache: The cache to allocate from.
  3164. * @flags: See kmalloc().
  3165. *
  3166. * Allocate an object from this cache and set the allocated memory to zero.
  3167. * The flags are only relevant if the cache has no available objects.
  3168. */
  3169. void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
  3170. {
  3171. void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
  3172. if (ret)
  3173. memset(ret, 0, obj_size(cache));
  3174. return ret;
  3175. }
  3176. EXPORT_SYMBOL(kmem_cache_zalloc);
  3177. /**
  3178. * kmem_ptr_validate - check if an untrusted pointer might
  3179. * be a slab entry.
  3180. * @cachep: the cache we're checking against
  3181. * @ptr: pointer to validate
  3182. *
  3183. * This verifies that the untrusted pointer looks sane:
  3184. * it is _not_ a guarantee that the pointer is actually
  3185. * part of the slab cache in question, but it at least
  3186. * validates that the pointer can be dereferenced and
  3187. * looks half-way sane.
  3188. *
  3189. * Currently only used for dentry validation.
  3190. */
  3191. int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
  3192. {
  3193. unsigned long addr = (unsigned long)ptr;
  3194. unsigned long min_addr = PAGE_OFFSET;
  3195. unsigned long align_mask = BYTES_PER_WORD - 1;
  3196. unsigned long size = cachep->buffer_size;
  3197. struct page *page;
  3198. if (unlikely(addr < min_addr))
  3199. goto out;
  3200. if (unlikely(addr > (unsigned long)high_memory - size))
  3201. goto out;
  3202. if (unlikely(addr & align_mask))
  3203. goto out;
  3204. if (unlikely(!kern_addr_valid(addr)))
  3205. goto out;
  3206. if (unlikely(!kern_addr_valid(addr + size - 1)))
  3207. goto out;
  3208. page = virt_to_page(ptr);
  3209. if (unlikely(!PageSlab(page)))
  3210. goto out;
  3211. if (unlikely(page_get_cache(page) != cachep))
  3212. goto out;
  3213. return 1;
  3214. out:
  3215. return 0;
  3216. }
  3217. #ifdef CONFIG_NUMA
  3218. void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
  3219. {
  3220. return __cache_alloc_node(cachep, flags, nodeid,
  3221. __builtin_return_address(0));
  3222. }
  3223. EXPORT_SYMBOL(kmem_cache_alloc_node);
  3224. static __always_inline void *
  3225. __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
  3226. {
  3227. struct kmem_cache *cachep;
  3228. cachep = kmem_find_general_cachep(size, flags);
  3229. if (unlikely(cachep == NULL))
  3230. return NULL;
  3231. return kmem_cache_alloc_node(cachep, flags, node);
  3232. }
  3233. #ifdef CONFIG_DEBUG_SLAB
  3234. void *__kmalloc_node(size_t size, gfp_t flags, int node)
  3235. {
  3236. return __do_kmalloc_node(size, flags, node,
  3237. __builtin_return_address(0));
  3238. }
  3239. EXPORT_SYMBOL(__kmalloc_node);
  3240. void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
  3241. int node, void *caller)
  3242. {
  3243. return __do_kmalloc_node(size, flags, node, caller);
  3244. }
  3245. EXPORT_SYMBOL(__kmalloc_node_track_caller);
  3246. #else
  3247. void *__kmalloc_node(size_t size, gfp_t flags, int node)
  3248. {
  3249. return __do_kmalloc_node(size, flags, node, NULL);
  3250. }
  3251. EXPORT_SYMBOL(__kmalloc_node);
  3252. #endif /* CONFIG_DEBUG_SLAB */
  3253. #endif /* CONFIG_NUMA */
  3254. /**
  3255. * __do_kmalloc - allocate memory
  3256. * @size: how many bytes of memory are required.
  3257. * @flags: the type of memory to allocate (see kmalloc).
  3258. * @caller: function caller for debug tracking of the caller
  3259. */
  3260. static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
  3261. void *caller)
  3262. {
  3263. struct kmem_cache *cachep;
  3264. /* If you want to save a few bytes .text space: replace
  3265. * __ with kmem_.
  3266. * Then kmalloc uses the uninlined functions instead of the inline
  3267. * functions.
  3268. */
  3269. cachep = __find_general_cachep(size, flags);
  3270. if (unlikely(cachep == NULL))
  3271. return NULL;
  3272. return __cache_alloc(cachep, flags, caller);
  3273. }
  3274. #ifdef CONFIG_DEBUG_SLAB
  3275. void *__kmalloc(size_t size, gfp_t flags)
  3276. {
  3277. return __do_kmalloc(size, flags, __builtin_return_address(0));
  3278. }
  3279. EXPORT_SYMBOL(__kmalloc);
  3280. void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
  3281. {
  3282. return __do_kmalloc(size, flags, caller);
  3283. }
  3284. EXPORT_SYMBOL(__kmalloc_track_caller);
  3285. #else
  3286. void *__kmalloc(size_t size, gfp_t flags)
  3287. {
  3288. return __do_kmalloc(size, flags, NULL);
  3289. }
  3290. EXPORT_SYMBOL(__kmalloc);
  3291. #endif
  3292. /**
  3293. * kmem_cache_free - Deallocate an object
  3294. * @cachep: The cache the allocation was from.
  3295. * @objp: The previously allocated object.
  3296. *
  3297. * Free an object which was previously allocated from this
  3298. * cache.
  3299. */
  3300. void kmem_cache_free(struct kmem_cache *cachep, void *objp)
  3301. {
  3302. unsigned long flags;
  3303. BUG_ON(virt_to_cache(objp) != cachep);
  3304. local_irq_save(flags);
  3305. debug_check_no_locks_freed(objp, obj_size(cachep));
  3306. __cache_free(cachep, objp);
  3307. local_irq_restore(flags);
  3308. }
  3309. EXPORT_SYMBOL(kmem_cache_free);
  3310. /**
  3311. * kfree - free previously allocated memory
  3312. * @objp: pointer returned by kmalloc.
  3313. *
  3314. * If @objp is NULL, no operation is performed.
  3315. *
  3316. * Don't free memory not originally allocated by kmalloc()
  3317. * or you will run into trouble.
  3318. */
  3319. void kfree(const void *objp)
  3320. {
  3321. struct kmem_cache *c;
  3322. unsigned long flags;
  3323. if (unlikely(!objp))
  3324. return;
  3325. local_irq_save(flags);
  3326. kfree_debugcheck(objp);
  3327. c = virt_to_cache(objp);
  3328. debug_check_no_locks_freed(objp, obj_size(c));
  3329. __cache_free(c, (void *)objp);
  3330. local_irq_restore(flags);
  3331. }
  3332. EXPORT_SYMBOL(kfree);
  3333. unsigned int kmem_cache_size(struct kmem_cache *cachep)
  3334. {
  3335. return obj_size(cachep);
  3336. }
  3337. EXPORT_SYMBOL(kmem_cache_size);
  3338. const char *kmem_cache_name(struct kmem_cache *cachep)
  3339. {
  3340. return cachep->name;
  3341. }
  3342. EXPORT_SYMBOL_GPL(kmem_cache_name);
  3343. /*
  3344. * This initializes kmem_list3 or resizes varioius caches for all nodes.
  3345. */
  3346. static int alloc_kmemlist(struct kmem_cache *cachep)
  3347. {
  3348. int node;
  3349. struct kmem_list3 *l3;
  3350. struct array_cache *new_shared;
  3351. struct array_cache **new_alien = NULL;
  3352. for_each_online_node(node) {
  3353. if (use_alien_caches) {
  3354. new_alien = alloc_alien_cache(node, cachep->limit);
  3355. if (!new_alien)
  3356. goto fail;
  3357. }
  3358. new_shared = alloc_arraycache(node,
  3359. cachep->shared*cachep->batchcount,
  3360. 0xbaadf00d);
  3361. if (!new_shared) {
  3362. free_alien_cache(new_alien);
  3363. goto fail;
  3364. }
  3365. l3 = cachep->nodelists[node];
  3366. if (l3) {
  3367. struct array_cache *shared = l3->shared;
  3368. spin_lock_irq(&l3->list_lock);
  3369. if (shared)
  3370. free_block(cachep, shared->entry,
  3371. shared->avail, node);
  3372. l3->shared = new_shared;
  3373. if (!l3->alien) {
  3374. l3->alien = new_alien;
  3375. new_alien = NULL;
  3376. }
  3377. l3->free_limit = (1 + nr_cpus_node(node)) *
  3378. cachep->batchcount + cachep->num;
  3379. spin_unlock_irq(&l3->list_lock);
  3380. kfree(shared);
  3381. free_alien_cache(new_alien);
  3382. continue;
  3383. }
  3384. l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
  3385. if (!l3) {
  3386. free_alien_cache(new_alien);
  3387. kfree(new_shared);
  3388. goto fail;
  3389. }
  3390. kmem_list3_init(l3);
  3391. l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
  3392. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  3393. l3->shared = new_shared;
  3394. l3->alien = new_alien;
  3395. l3->free_limit = (1 + nr_cpus_node(node)) *
  3396. cachep->batchcount + cachep->num;
  3397. cachep->nodelists[node] = l3;
  3398. }
  3399. return 0;
  3400. fail:
  3401. if (!cachep->next.next) {
  3402. /* Cache is not active yet. Roll back what we did */
  3403. node--;
  3404. while (node >= 0) {
  3405. if (cachep->nodelists[node]) {
  3406. l3 = cachep->nodelists[node];
  3407. kfree(l3->shared);
  3408. free_alien_cache(l3->alien);
  3409. kfree(l3);
  3410. cachep->nodelists[node] = NULL;
  3411. }
  3412. node--;
  3413. }
  3414. }
  3415. return -ENOMEM;
  3416. }
  3417. struct ccupdate_struct {
  3418. struct kmem_cache *cachep;
  3419. struct array_cache *new[NR_CPUS];
  3420. };
  3421. static void do_ccupdate_local(void *info)
  3422. {
  3423. struct ccupdate_struct *new = info;
  3424. struct array_cache *old;
  3425. check_irq_off();
  3426. old = cpu_cache_get(new->cachep);
  3427. new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
  3428. new->new[smp_processor_id()] = old;
  3429. }
  3430. /* Always called with the cache_chain_mutex held */
  3431. static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
  3432. int batchcount, int shared)
  3433. {
  3434. struct ccupdate_struct *new;
  3435. int i;
  3436. new = kzalloc(sizeof(*new), GFP_KERNEL);
  3437. if (!new)
  3438. return -ENOMEM;
  3439. for_each_online_cpu(i) {
  3440. new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
  3441. batchcount);
  3442. if (!new->new[i]) {
  3443. for (i--; i >= 0; i--)
  3444. kfree(new->new[i]);
  3445. kfree(new);
  3446. return -ENOMEM;
  3447. }
  3448. }
  3449. new->cachep = cachep;
  3450. on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
  3451. check_irq_on();
  3452. cachep->batchcount = batchcount;
  3453. cachep->limit = limit;
  3454. cachep->shared = shared;
  3455. for_each_online_cpu(i) {
  3456. struct array_cache *ccold = new->new[i];
  3457. if (!ccold)
  3458. continue;
  3459. spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
  3460. free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
  3461. spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
  3462. kfree(ccold);
  3463. }
  3464. kfree(new);
  3465. return alloc_kmemlist(cachep);
  3466. }
  3467. /* Called with cache_chain_mutex held always */
  3468. static int enable_cpucache(struct kmem_cache *cachep)
  3469. {
  3470. int err;
  3471. int limit, shared;
  3472. /*
  3473. * The head array serves three purposes:
  3474. * - create a LIFO ordering, i.e. return objects that are cache-warm
  3475. * - reduce the number of spinlock operations.
  3476. * - reduce the number of linked list operations on the slab and
  3477. * bufctl chains: array operations are cheaper.
  3478. * The numbers are guessed, we should auto-tune as described by
  3479. * Bonwick.
  3480. */
  3481. if (cachep->buffer_size > 131072)
  3482. limit = 1;
  3483. else if (cachep->buffer_size > PAGE_SIZE)
  3484. limit = 8;
  3485. else if (cachep->buffer_size > 1024)
  3486. limit = 24;
  3487. else if (cachep->buffer_size > 256)
  3488. limit = 54;
  3489. else
  3490. limit = 120;
  3491. /*
  3492. * CPU bound tasks (e.g. network routing) can exhibit cpu bound
  3493. * allocation behaviour: Most allocs on one cpu, most free operations
  3494. * on another cpu. For these cases, an efficient object passing between
  3495. * cpus is necessary. This is provided by a shared array. The array
  3496. * replaces Bonwick's magazine layer.
  3497. * On uniprocessor, it's functionally equivalent (but less efficient)
  3498. * to a larger limit. Thus disabled by default.
  3499. */
  3500. shared = 0;
  3501. #ifdef CONFIG_SMP
  3502. if (cachep->buffer_size <= PAGE_SIZE)
  3503. shared = 8;
  3504. #endif
  3505. #if DEBUG
  3506. /*
  3507. * With debugging enabled, large batchcount lead to excessively long
  3508. * periods with disabled local interrupts. Limit the batchcount
  3509. */
  3510. if (limit > 32)
  3511. limit = 32;
  3512. #endif
  3513. err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
  3514. if (err)
  3515. printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
  3516. cachep->name, -err);
  3517. return err;
  3518. }
  3519. /*
  3520. * Drain an array if it contains any elements taking the l3 lock only if
  3521. * necessary. Note that the l3 listlock also protects the array_cache
  3522. * if drain_array() is used on the shared array.
  3523. */
  3524. void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
  3525. struct array_cache *ac, int force, int node)
  3526. {
  3527. int tofree;
  3528. if (!ac || !ac->avail)
  3529. return;
  3530. if (ac->touched && !force) {
  3531. ac->touched = 0;
  3532. } else {
  3533. spin_lock_irq(&l3->list_lock);
  3534. if (ac->avail) {
  3535. tofree = force ? ac->avail : (ac->limit + 4) / 5;
  3536. if (tofree > ac->avail)
  3537. tofree = (ac->avail + 1) / 2;
  3538. free_block(cachep, ac->entry, tofree, node);
  3539. ac->avail -= tofree;
  3540. memmove(ac->entry, &(ac->entry[tofree]),
  3541. sizeof(void *) * ac->avail);
  3542. }
  3543. spin_unlock_irq(&l3->list_lock);
  3544. }
  3545. }
  3546. /**
  3547. * cache_reap - Reclaim memory from caches.
  3548. * @w: work descriptor
  3549. *
  3550. * Called from workqueue/eventd every few seconds.
  3551. * Purpose:
  3552. * - clear the per-cpu caches for this CPU.
  3553. * - return freeable pages to the main free memory pool.
  3554. *
  3555. * If we cannot acquire the cache chain mutex then just give up - we'll try
  3556. * again on the next iteration.
  3557. */
  3558. static void cache_reap(struct work_struct *w)
  3559. {
  3560. struct kmem_cache *searchp;
  3561. struct kmem_list3 *l3;
  3562. int node = numa_node_id();
  3563. struct delayed_work *work =
  3564. container_of(w, struct delayed_work, work);
  3565. if (!mutex_trylock(&cache_chain_mutex))
  3566. /* Give up. Setup the next iteration. */
  3567. goto out;
  3568. list_for_each_entry(searchp, &cache_chain, next) {
  3569. check_irq_on();
  3570. /*
  3571. * We only take the l3 lock if absolutely necessary and we
  3572. * have established with reasonable certainty that
  3573. * we can do some work if the lock was obtained.
  3574. */
  3575. l3 = searchp->nodelists[node];
  3576. reap_alien(searchp, l3);
  3577. drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
  3578. /*
  3579. * These are racy checks but it does not matter
  3580. * if we skip one check or scan twice.
  3581. */
  3582. if (time_after(l3->next_reap, jiffies))
  3583. goto next;
  3584. l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
  3585. drain_array(searchp, l3, l3->shared, 0, node);
  3586. if (l3->free_touched)
  3587. l3->free_touched = 0;
  3588. else {
  3589. int freed;
  3590. freed = drain_freelist(searchp, l3, (l3->free_limit +
  3591. 5 * searchp->num - 1) / (5 * searchp->num));
  3592. STATS_ADD_REAPED(searchp, freed);
  3593. }
  3594. next:
  3595. cond_resched();
  3596. }
  3597. check_irq_on();
  3598. mutex_unlock(&cache_chain_mutex);
  3599. next_reap_node();
  3600. refresh_cpu_vm_stats(smp_processor_id());
  3601. out:
  3602. /* Set up the next iteration */
  3603. schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
  3604. }
  3605. #ifdef CONFIG_PROC_FS
  3606. static void print_slabinfo_header(struct seq_file *m)
  3607. {
  3608. /*
  3609. * Output format version, so at least we can change it
  3610. * without _too_ many complaints.
  3611. */
  3612. #if STATS
  3613. seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
  3614. #else
  3615. seq_puts(m, "slabinfo - version: 2.1\n");
  3616. #endif
  3617. seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
  3618. "<objperslab> <pagesperslab>");
  3619. seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
  3620. seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
  3621. #if STATS
  3622. seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
  3623. "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
  3624. seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
  3625. #endif
  3626. seq_putc(m, '\n');
  3627. }
  3628. static void *s_start(struct seq_file *m, loff_t *pos)
  3629. {
  3630. loff_t n = *pos;
  3631. struct list_head *p;
  3632. mutex_lock(&cache_chain_mutex);
  3633. if (!n)
  3634. print_slabinfo_header(m);
  3635. p = cache_chain.next;
  3636. while (n--) {
  3637. p = p->next;
  3638. if (p == &cache_chain)
  3639. return NULL;
  3640. }
  3641. return list_entry(p, struct kmem_cache, next);
  3642. }
  3643. static void *s_next(struct seq_file *m, void *p, loff_t *pos)
  3644. {
  3645. struct kmem_cache *cachep = p;
  3646. ++*pos;
  3647. return cachep->next.next == &cache_chain ?
  3648. NULL : list_entry(cachep->next.next, struct kmem_cache, next);
  3649. }
  3650. static void s_stop(struct seq_file *m, void *p)
  3651. {
  3652. mutex_unlock(&cache_chain_mutex);
  3653. }
  3654. static int s_show(struct seq_file *m, void *p)
  3655. {
  3656. struct kmem_cache *cachep = p;
  3657. struct slab *slabp;
  3658. unsigned long active_objs;
  3659. unsigned long num_objs;
  3660. unsigned long active_slabs = 0;
  3661. unsigned long num_slabs, free_objects = 0, shared_avail = 0;
  3662. const char *name;
  3663. char *error = NULL;
  3664. int node;
  3665. struct kmem_list3 *l3;
  3666. active_objs = 0;
  3667. num_slabs = 0;
  3668. for_each_online_node(node) {
  3669. l3 = cachep->nodelists[node];
  3670. if (!l3)
  3671. continue;
  3672. check_irq_on();
  3673. spin_lock_irq(&l3->list_lock);
  3674. list_for_each_entry(slabp, &l3->slabs_full, list) {
  3675. if (slabp->inuse != cachep->num && !error)
  3676. error = "slabs_full accounting error";
  3677. active_objs += cachep->num;
  3678. active_slabs++;
  3679. }
  3680. list_for_each_entry(slabp, &l3->slabs_partial, list) {
  3681. if (slabp->inuse == cachep->num && !error)
  3682. error = "slabs_partial inuse accounting error";
  3683. if (!slabp->inuse && !error)
  3684. error = "slabs_partial/inuse accounting error";
  3685. active_objs += slabp->inuse;
  3686. active_slabs++;
  3687. }
  3688. list_for_each_entry(slabp, &l3->slabs_free, list) {
  3689. if (slabp->inuse && !error)
  3690. error = "slabs_free/inuse accounting error";
  3691. num_slabs++;
  3692. }
  3693. free_objects += l3->free_objects;
  3694. if (l3->shared)
  3695. shared_avail += l3->shared->avail;
  3696. spin_unlock_irq(&l3->list_lock);
  3697. }
  3698. num_slabs += active_slabs;
  3699. num_objs = num_slabs * cachep->num;
  3700. if (num_objs - active_objs != free_objects && !error)
  3701. error = "free_objects accounting error";
  3702. name = cachep->name;
  3703. if (error)
  3704. printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
  3705. seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
  3706. name, active_objs, num_objs, cachep->buffer_size,
  3707. cachep->num, (1 << cachep->gfporder));
  3708. seq_printf(m, " : tunables %4u %4u %4u",
  3709. cachep->limit, cachep->batchcount, cachep->shared);
  3710. seq_printf(m, " : slabdata %6lu %6lu %6lu",
  3711. active_slabs, num_slabs, shared_avail);
  3712. #if STATS
  3713. { /* list3 stats */
  3714. unsigned long high = cachep->high_mark;
  3715. unsigned long allocs = cachep->num_allocations;
  3716. unsigned long grown = cachep->grown;
  3717. unsigned long reaped = cachep->reaped;
  3718. unsigned long errors = cachep->errors;
  3719. unsigned long max_freeable = cachep->max_freeable;
  3720. unsigned long node_allocs = cachep->node_allocs;
  3721. unsigned long node_frees = cachep->node_frees;
  3722. unsigned long overflows = cachep->node_overflow;
  3723. seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
  3724. %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
  3725. reaped, errors, max_freeable, node_allocs,
  3726. node_frees, overflows);
  3727. }
  3728. /* cpu stats */
  3729. {
  3730. unsigned long allochit = atomic_read(&cachep->allochit);
  3731. unsigned long allocmiss = atomic_read(&cachep->allocmiss);
  3732. unsigned long freehit = atomic_read(&cachep->freehit);
  3733. unsigned long freemiss = atomic_read(&cachep->freemiss);
  3734. seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
  3735. allochit, allocmiss, freehit, freemiss);
  3736. }
  3737. #endif
  3738. seq_putc(m, '\n');
  3739. return 0;
  3740. }
  3741. /*
  3742. * slabinfo_op - iterator that generates /proc/slabinfo
  3743. *
  3744. * Output layout:
  3745. * cache-name
  3746. * num-active-objs
  3747. * total-objs
  3748. * object size
  3749. * num-active-slabs
  3750. * total-slabs
  3751. * num-pages-per-slab
  3752. * + further values on SMP and with statistics enabled
  3753. */
  3754. const struct seq_operations slabinfo_op = {
  3755. .start = s_start,
  3756. .next = s_next,
  3757. .stop = s_stop,
  3758. .show = s_show,
  3759. };
  3760. #define MAX_SLABINFO_WRITE 128
  3761. /**
  3762. * slabinfo_write - Tuning for the slab allocator
  3763. * @file: unused
  3764. * @buffer: user buffer
  3765. * @count: data length
  3766. * @ppos: unused
  3767. */
  3768. ssize_t slabinfo_write(struct file *file, const char __user * buffer,
  3769. size_t count, loff_t *ppos)
  3770. {
  3771. char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
  3772. int limit, batchcount, shared, res;
  3773. struct kmem_cache *cachep;
  3774. if (count > MAX_SLABINFO_WRITE)
  3775. return -EINVAL;
  3776. if (copy_from_user(&kbuf, buffer, count))
  3777. return -EFAULT;
  3778. kbuf[MAX_SLABINFO_WRITE] = '\0';
  3779. tmp = strchr(kbuf, ' ');
  3780. if (!tmp)
  3781. return -EINVAL;
  3782. *tmp = '\0';
  3783. tmp++;
  3784. if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
  3785. return -EINVAL;
  3786. /* Find the cache in the chain of caches. */
  3787. mutex_lock(&cache_chain_mutex);
  3788. res = -EINVAL;
  3789. list_for_each_entry(cachep, &cache_chain, next) {
  3790. if (!strcmp(cachep->name, kbuf)) {
  3791. if (limit < 1 || batchcount < 1 ||
  3792. batchcount > limit || shared < 0) {
  3793. res = 0;
  3794. } else {
  3795. res = do_tune_cpucache(cachep, limit,
  3796. batchcount, shared);
  3797. }
  3798. break;
  3799. }
  3800. }
  3801. mutex_unlock(&cache_chain_mutex);
  3802. if (res >= 0)
  3803. res = count;
  3804. return res;
  3805. }
  3806. #ifdef CONFIG_DEBUG_SLAB_LEAK
  3807. static void *leaks_start(struct seq_file *m, loff_t *pos)
  3808. {
  3809. loff_t n = *pos;
  3810. struct list_head *p;
  3811. mutex_lock(&cache_chain_mutex);
  3812. p = cache_chain.next;
  3813. while (n--) {
  3814. p = p->next;
  3815. if (p == &cache_chain)
  3816. return NULL;
  3817. }
  3818. return list_entry(p, struct kmem_cache, next);
  3819. }
  3820. static inline int add_caller(unsigned long *n, unsigned long v)
  3821. {
  3822. unsigned long *p;
  3823. int l;
  3824. if (!v)
  3825. return 1;
  3826. l = n[1];
  3827. p = n + 2;
  3828. while (l) {
  3829. int i = l/2;
  3830. unsigned long *q = p + 2 * i;
  3831. if (*q == v) {
  3832. q[1]++;
  3833. return 1;
  3834. }
  3835. if (*q > v) {
  3836. l = i;
  3837. } else {
  3838. p = q + 2;
  3839. l -= i + 1;
  3840. }
  3841. }
  3842. if (++n[1] == n[0])
  3843. return 0;
  3844. memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
  3845. p[0] = v;
  3846. p[1] = 1;
  3847. return 1;
  3848. }
  3849. static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
  3850. {
  3851. void *p;
  3852. int i;
  3853. if (n[0] == n[1])
  3854. return;
  3855. for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
  3856. if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
  3857. continue;
  3858. if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
  3859. return;
  3860. }
  3861. }
  3862. static void show_symbol(struct seq_file *m, unsigned long address)
  3863. {
  3864. #ifdef CONFIG_KALLSYMS
  3865. char *modname;
  3866. const char *name;
  3867. unsigned long offset, size;
  3868. char namebuf[KSYM_NAME_LEN+1];
  3869. name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
  3870. if (name) {
  3871. seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
  3872. if (modname)
  3873. seq_printf(m, " [%s]", modname);
  3874. return;
  3875. }
  3876. #endif
  3877. seq_printf(m, "%p", (void *)address);
  3878. }
  3879. static int leaks_show(struct seq_file *m, void *p)
  3880. {
  3881. struct kmem_cache *cachep = p;
  3882. struct slab *slabp;
  3883. struct kmem_list3 *l3;
  3884. const char *name;
  3885. unsigned long *n = m->private;
  3886. int node;
  3887. int i;
  3888. if (!(cachep->flags & SLAB_STORE_USER))
  3889. return 0;
  3890. if (!(cachep->flags & SLAB_RED_ZONE))
  3891. return 0;
  3892. /* OK, we can do it */
  3893. n[1] = 0;
  3894. for_each_online_node(node) {
  3895. l3 = cachep->nodelists[node];
  3896. if (!l3)
  3897. continue;
  3898. check_irq_on();
  3899. spin_lock_irq(&l3->list_lock);
  3900. list_for_each_entry(slabp, &l3->slabs_full, list)
  3901. handle_slab(n, cachep, slabp);
  3902. list_for_each_entry(slabp, &l3->slabs_partial, list)
  3903. handle_slab(n, cachep, slabp);
  3904. spin_unlock_irq(&l3->list_lock);
  3905. }
  3906. name = cachep->name;
  3907. if (n[0] == n[1]) {
  3908. /* Increase the buffer size */
  3909. mutex_unlock(&cache_chain_mutex);
  3910. m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
  3911. if (!m->private) {
  3912. /* Too bad, we are really out */
  3913. m->private = n;
  3914. mutex_lock(&cache_chain_mutex);
  3915. return -ENOMEM;
  3916. }
  3917. *(unsigned long *)m->private = n[0] * 2;
  3918. kfree(n);
  3919. mutex_lock(&cache_chain_mutex);
  3920. /* Now make sure this entry will be retried */
  3921. m->count = m->size;
  3922. return 0;
  3923. }
  3924. for (i = 0; i < n[1]; i++) {
  3925. seq_printf(m, "%s: %lu ", name, n[2*i+3]);
  3926. show_symbol(m, n[2*i+2]);
  3927. seq_putc(m, '\n');
  3928. }
  3929. return 0;
  3930. }
  3931. const struct seq_operations slabstats_op = {
  3932. .start = leaks_start,
  3933. .next = s_next,
  3934. .stop = s_stop,
  3935. .show = leaks_show,
  3936. };
  3937. #endif
  3938. #endif
  3939. /**
  3940. * ksize - get the actual amount of memory allocated for a given object
  3941. * @objp: Pointer to the object
  3942. *
  3943. * kmalloc may internally round up allocations and return more memory
  3944. * than requested. ksize() can be used to determine the actual amount of
  3945. * memory allocated. The caller may use this additional memory, even though
  3946. * a smaller amount of memory was initially specified with the kmalloc call.
  3947. * The caller must guarantee that objp points to a valid object previously
  3948. * allocated with either kmalloc() or kmem_cache_alloc(). The object
  3949. * must not be freed during the duration of the call.
  3950. */
  3951. unsigned int ksize(const void *objp)
  3952. {
  3953. if (unlikely(objp == NULL))
  3954. return 0;
  3955. return obj_size(virt_to_cache(objp));
  3956. }