slab.c 119 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138413941404141414241434144414541464147414841494150415141524153415441554156415741584159416041614162416341644165416641674168416941704171417241734174417541764177417841794180418141824183418441854186418741884189419041914192419341944195419641974198419942004201420242034204420542064207420842094210421142124213421442154216421742184219422042214222422342244225422642274228422942304231423242334234423542364237423842394240424142424243424442454246424742484249425042514252425342544255425642574258425942604261426242634264426542664267426842694270427142724273427442754276427742784279428042814282428342844285428642874288428942904291429242934294429542964297429842994300430143024303430443054306430743084309431043114312431343144315431643174318431943204321432243234324432543264327432843294330433143324333433443354336433743384339434043414342434343444345434643474348434943504351435243534354435543564357435843594360436143624363436443654366436743684369437043714372437343744375437643774378437943804381438243834384438543864387438843894390439143924393439443954396439743984399440044014402440344044405440644074408440944104411441244134414441544164417441844194420442144224423442444254426442744284429443044314432443344344435443644374438443944404441444244434444444544464447444844494450445144524453445444554456445744584459446044614462446344644465446644674468446944704471447244734474447544764477447844794480448144824483448444854486448744884489449044914492449344944495449644974498449945004501450245034504450545064507450845094510451145124513451445154516451745184519452045214522452345244525452645274528452945304531453245334534453545364537453845394540454145424543454445454546454745484549455045514552455345544555455645574558455945604561456245634564456545664567456845694570457145724573457445754576457745784579458045814582458345844585458645874588458945904591459245934594459545964597459845994600460146024603460446054606460746084609461046114612461346144615461646174618461946204621462246234624462546264627462846294630
  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 initializations 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/proc_fs.h>
  98. #include <linux/seq_file.h>
  99. #include <linux/notifier.h>
  100. #include <linux/kallsyms.h>
  101. #include <linux/cpu.h>
  102. #include <linux/sysctl.h>
  103. #include <linux/module.h>
  104. #include <linux/rcupdate.h>
  105. #include <linux/string.h>
  106. #include <linux/uaccess.h>
  107. #include <linux/nodemask.h>
  108. #include <linux/kmemleak.h>
  109. #include <linux/mempolicy.h>
  110. #include <linux/mutex.h>
  111. #include <linux/fault-inject.h>
  112. #include <linux/rtmutex.h>
  113. #include <linux/reciprocal_div.h>
  114. #include <linux/debugobjects.h>
  115. #include <linux/kmemcheck.h>
  116. #include <linux/memory.h>
  117. #include <linux/prefetch.h>
  118. #include <asm/cacheflush.h>
  119. #include <asm/tlbflush.h>
  120. #include <asm/page.h>
  121. #include <trace/events/kmem.h>
  122. /*
  123. * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
  124. * 0 for faster, smaller code (especially in the critical paths).
  125. *
  126. * STATS - 1 to collect stats for /proc/slabinfo.
  127. * 0 for faster, smaller code (especially in the critical paths).
  128. *
  129. * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
  130. */
  131. #ifdef CONFIG_DEBUG_SLAB
  132. #define DEBUG 1
  133. #define STATS 1
  134. #define FORCED_DEBUG 1
  135. #else
  136. #define DEBUG 0
  137. #define STATS 0
  138. #define FORCED_DEBUG 0
  139. #endif
  140. /* Shouldn't this be in a header file somewhere? */
  141. #define BYTES_PER_WORD sizeof(void *)
  142. #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
  143. #ifndef ARCH_KMALLOC_FLAGS
  144. #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
  145. #endif
  146. /* Legal flag mask for kmem_cache_create(). */
  147. #if DEBUG
  148. # define CREATE_MASK (SLAB_RED_ZONE | \
  149. SLAB_POISON | SLAB_HWCACHE_ALIGN | \
  150. SLAB_CACHE_DMA | \
  151. SLAB_STORE_USER | \
  152. SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
  153. SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
  154. SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
  155. #else
  156. # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
  157. SLAB_CACHE_DMA | \
  158. SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
  159. SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
  160. SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
  161. #endif
  162. /*
  163. * kmem_bufctl_t:
  164. *
  165. * Bufctl's are used for linking objs within a slab
  166. * linked offsets.
  167. *
  168. * This implementation relies on "struct page" for locating the cache &
  169. * slab an object belongs to.
  170. * This allows the bufctl structure to be small (one int), but limits
  171. * the number of objects a slab (not a cache) can contain when off-slab
  172. * bufctls are used. The limit is the size of the largest general cache
  173. * that does not use off-slab slabs.
  174. * For 32bit archs with 4 kB pages, is this 56.
  175. * This is not serious, as it is only for large objects, when it is unwise
  176. * to have too many per slab.
  177. * Note: This limit can be raised by introducing a general cache whose size
  178. * is less than 512 (PAGE_SIZE<<3), but greater than 256.
  179. */
  180. typedef unsigned int kmem_bufctl_t;
  181. #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
  182. #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
  183. #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
  184. #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
  185. /*
  186. * struct slab_rcu
  187. *
  188. * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
  189. * arrange for kmem_freepages to be called via RCU. This is useful if
  190. * we need to approach a kernel structure obliquely, from its address
  191. * obtained without the usual locking. We can lock the structure to
  192. * stabilize it and check it's still at the given address, only if we
  193. * can be sure that the memory has not been meanwhile reused for some
  194. * other kind of object (which our subsystem's lock might corrupt).
  195. *
  196. * rcu_read_lock before reading the address, then rcu_read_unlock after
  197. * taking the spinlock within the structure expected at that address.
  198. */
  199. struct slab_rcu {
  200. struct rcu_head head;
  201. struct kmem_cache *cachep;
  202. void *addr;
  203. };
  204. /*
  205. * struct slab
  206. *
  207. * Manages the objs in a slab. Placed either at the beginning of mem allocated
  208. * for a slab, or allocated from an general cache.
  209. * Slabs are chained into three list: fully used, partial, fully free slabs.
  210. */
  211. struct slab {
  212. union {
  213. struct {
  214. struct list_head list;
  215. unsigned long colouroff;
  216. void *s_mem; /* including colour offset */
  217. unsigned int inuse; /* num of objs active in slab */
  218. kmem_bufctl_t free;
  219. unsigned short nodeid;
  220. };
  221. struct slab_rcu __slab_cover_slab_rcu;
  222. };
  223. };
  224. /*
  225. * struct array_cache
  226. *
  227. * Purpose:
  228. * - LIFO ordering, to hand out cache-warm objects from _alloc
  229. * - reduce the number of linked list operations
  230. * - reduce spinlock operations
  231. *
  232. * The limit is stored in the per-cpu structure to reduce the data cache
  233. * footprint.
  234. *
  235. */
  236. struct array_cache {
  237. unsigned int avail;
  238. unsigned int limit;
  239. unsigned int batchcount;
  240. unsigned int touched;
  241. spinlock_t lock;
  242. void *entry[]; /*
  243. * Must have this definition in here for the proper
  244. * alignment of array_cache. Also simplifies accessing
  245. * the entries.
  246. */
  247. };
  248. /*
  249. * bootstrap: The caches do not work without cpuarrays anymore, but the
  250. * cpuarrays are allocated from the generic caches...
  251. */
  252. #define BOOT_CPUCACHE_ENTRIES 1
  253. struct arraycache_init {
  254. struct array_cache cache;
  255. void *entries[BOOT_CPUCACHE_ENTRIES];
  256. };
  257. /*
  258. * The slab lists for all objects.
  259. */
  260. struct kmem_list3 {
  261. struct list_head slabs_partial; /* partial list first, better asm code */
  262. struct list_head slabs_full;
  263. struct list_head slabs_free;
  264. unsigned long free_objects;
  265. unsigned int free_limit;
  266. unsigned int colour_next; /* Per-node cache coloring */
  267. spinlock_t list_lock;
  268. struct array_cache *shared; /* shared per node */
  269. struct array_cache **alien; /* on other nodes */
  270. unsigned long next_reap; /* updated without locking */
  271. int free_touched; /* updated without locking */
  272. };
  273. /*
  274. * Need this for bootstrapping a per node allocator.
  275. */
  276. #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
  277. static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
  278. #define CACHE_CACHE 0
  279. #define SIZE_AC MAX_NUMNODES
  280. #define SIZE_L3 (2 * MAX_NUMNODES)
  281. static int drain_freelist(struct kmem_cache *cache,
  282. struct kmem_list3 *l3, int tofree);
  283. static void free_block(struct kmem_cache *cachep, void **objpp, int len,
  284. int node);
  285. static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
  286. static void cache_reap(struct work_struct *unused);
  287. /*
  288. * This function must be completely optimized away if a constant is passed to
  289. * it. Mostly the same as what is in linux/slab.h except it returns an index.
  290. */
  291. static __always_inline int index_of(const size_t size)
  292. {
  293. extern void __bad_size(void);
  294. if (__builtin_constant_p(size)) {
  295. int i = 0;
  296. #define CACHE(x) \
  297. if (size <=x) \
  298. return i; \
  299. else \
  300. i++;
  301. #include <linux/kmalloc_sizes.h>
  302. #undef CACHE
  303. __bad_size();
  304. } else
  305. __bad_size();
  306. return 0;
  307. }
  308. static int slab_early_init = 1;
  309. #define INDEX_AC index_of(sizeof(struct arraycache_init))
  310. #define INDEX_L3 index_of(sizeof(struct kmem_list3))
  311. static void kmem_list3_init(struct kmem_list3 *parent)
  312. {
  313. INIT_LIST_HEAD(&parent->slabs_full);
  314. INIT_LIST_HEAD(&parent->slabs_partial);
  315. INIT_LIST_HEAD(&parent->slabs_free);
  316. parent->shared = NULL;
  317. parent->alien = NULL;
  318. parent->colour_next = 0;
  319. spin_lock_init(&parent->list_lock);
  320. parent->free_objects = 0;
  321. parent->free_touched = 0;
  322. }
  323. #define MAKE_LIST(cachep, listp, slab, nodeid) \
  324. do { \
  325. INIT_LIST_HEAD(listp); \
  326. list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
  327. } while (0)
  328. #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
  329. do { \
  330. MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
  331. MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
  332. MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
  333. } while (0)
  334. #define CFLGS_OFF_SLAB (0x80000000UL)
  335. #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
  336. #define BATCHREFILL_LIMIT 16
  337. /*
  338. * Optimization question: fewer reaps means less probability for unnessary
  339. * cpucache drain/refill cycles.
  340. *
  341. * OTOH the cpuarrays can contain lots of objects,
  342. * which could lock up otherwise freeable slabs.
  343. */
  344. #define REAPTIMEOUT_CPUC (2*HZ)
  345. #define REAPTIMEOUT_LIST3 (4*HZ)
  346. #if STATS
  347. #define STATS_INC_ACTIVE(x) ((x)->num_active++)
  348. #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
  349. #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
  350. #define STATS_INC_GROWN(x) ((x)->grown++)
  351. #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
  352. #define STATS_SET_HIGH(x) \
  353. do { \
  354. if ((x)->num_active > (x)->high_mark) \
  355. (x)->high_mark = (x)->num_active; \
  356. } while (0)
  357. #define STATS_INC_ERR(x) ((x)->errors++)
  358. #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
  359. #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
  360. #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
  361. #define STATS_SET_FREEABLE(x, i) \
  362. do { \
  363. if ((x)->max_freeable < i) \
  364. (x)->max_freeable = i; \
  365. } while (0)
  366. #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
  367. #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
  368. #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
  369. #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
  370. #else
  371. #define STATS_INC_ACTIVE(x) do { } while (0)
  372. #define STATS_DEC_ACTIVE(x) do { } while (0)
  373. #define STATS_INC_ALLOCED(x) do { } while (0)
  374. #define STATS_INC_GROWN(x) do { } while (0)
  375. #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
  376. #define STATS_SET_HIGH(x) do { } while (0)
  377. #define STATS_INC_ERR(x) do { } while (0)
  378. #define STATS_INC_NODEALLOCS(x) do { } while (0)
  379. #define STATS_INC_NODEFREES(x) do { } while (0)
  380. #define STATS_INC_ACOVERFLOW(x) do { } while (0)
  381. #define STATS_SET_FREEABLE(x, i) do { } while (0)
  382. #define STATS_INC_ALLOCHIT(x) do { } while (0)
  383. #define STATS_INC_ALLOCMISS(x) do { } while (0)
  384. #define STATS_INC_FREEHIT(x) do { } while (0)
  385. #define STATS_INC_FREEMISS(x) do { } while (0)
  386. #endif
  387. #if DEBUG
  388. /*
  389. * memory layout of objects:
  390. * 0 : objp
  391. * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
  392. * the end of an object is aligned with the end of the real
  393. * allocation. Catches writes behind the end of the allocation.
  394. * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
  395. * redzone word.
  396. * cachep->obj_offset: The real object.
  397. * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
  398. * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
  399. * [BYTES_PER_WORD long]
  400. */
  401. static int obj_offset(struct kmem_cache *cachep)
  402. {
  403. return cachep->obj_offset;
  404. }
  405. static int obj_size(struct kmem_cache *cachep)
  406. {
  407. return cachep->obj_size;
  408. }
  409. static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
  410. {
  411. BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
  412. return (unsigned long long*) (objp + obj_offset(cachep) -
  413. sizeof(unsigned long long));
  414. }
  415. static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
  416. {
  417. BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
  418. if (cachep->flags & SLAB_STORE_USER)
  419. return (unsigned long long *)(objp + cachep->buffer_size -
  420. sizeof(unsigned long long) -
  421. REDZONE_ALIGN);
  422. return (unsigned long long *) (objp + cachep->buffer_size -
  423. sizeof(unsigned long long));
  424. }
  425. static void **dbg_userword(struct kmem_cache *cachep, void *objp)
  426. {
  427. BUG_ON(!(cachep->flags & SLAB_STORE_USER));
  428. return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
  429. }
  430. #else
  431. #define obj_offset(x) 0
  432. #define obj_size(cachep) (cachep->buffer_size)
  433. #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
  434. #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
  435. #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
  436. #endif
  437. #ifdef CONFIG_TRACING
  438. size_t slab_buffer_size(struct kmem_cache *cachep)
  439. {
  440. return cachep->buffer_size;
  441. }
  442. EXPORT_SYMBOL(slab_buffer_size);
  443. #endif
  444. /*
  445. * Do not go above this order unless 0 objects fit into the slab or
  446. * overridden on the command line.
  447. */
  448. #define SLAB_MAX_ORDER_HI 1
  449. #define SLAB_MAX_ORDER_LO 0
  450. static int slab_max_order = SLAB_MAX_ORDER_LO;
  451. static bool slab_max_order_set __initdata;
  452. /*
  453. * Functions for storing/retrieving the cachep and or slab from the page
  454. * allocator. These are used to find the slab an obj belongs to. With kfree(),
  455. * these are used to find the cache which an obj belongs to.
  456. */
  457. static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
  458. {
  459. page->lru.next = (struct list_head *)cache;
  460. }
  461. static inline struct kmem_cache *page_get_cache(struct page *page)
  462. {
  463. page = compound_head(page);
  464. BUG_ON(!PageSlab(page));
  465. return (struct kmem_cache *)page->lru.next;
  466. }
  467. static inline void page_set_slab(struct page *page, struct slab *slab)
  468. {
  469. page->lru.prev = (struct list_head *)slab;
  470. }
  471. static inline struct slab *page_get_slab(struct page *page)
  472. {
  473. BUG_ON(!PageSlab(page));
  474. return (struct slab *)page->lru.prev;
  475. }
  476. static inline struct kmem_cache *virt_to_cache(const void *obj)
  477. {
  478. struct page *page = virt_to_head_page(obj);
  479. return page_get_cache(page);
  480. }
  481. static inline struct slab *virt_to_slab(const void *obj)
  482. {
  483. struct page *page = virt_to_head_page(obj);
  484. return page_get_slab(page);
  485. }
  486. static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
  487. unsigned int idx)
  488. {
  489. return slab->s_mem + cache->buffer_size * idx;
  490. }
  491. /*
  492. * We want to avoid an expensive divide : (offset / cache->buffer_size)
  493. * Using the fact that buffer_size is a constant for a particular cache,
  494. * we can replace (offset / cache->buffer_size) by
  495. * reciprocal_divide(offset, cache->reciprocal_buffer_size)
  496. */
  497. static inline unsigned int obj_to_index(const struct kmem_cache *cache,
  498. const struct slab *slab, void *obj)
  499. {
  500. u32 offset = (obj - slab->s_mem);
  501. return reciprocal_divide(offset, cache->reciprocal_buffer_size);
  502. }
  503. /*
  504. * These are the default caches for kmalloc. Custom caches can have other sizes.
  505. */
  506. struct cache_sizes malloc_sizes[] = {
  507. #define CACHE(x) { .cs_size = (x) },
  508. #include <linux/kmalloc_sizes.h>
  509. CACHE(ULONG_MAX)
  510. #undef CACHE
  511. };
  512. EXPORT_SYMBOL(malloc_sizes);
  513. /* Must match cache_sizes above. Out of line to keep cache footprint low. */
  514. struct cache_names {
  515. char *name;
  516. char *name_dma;
  517. };
  518. static struct cache_names __initdata cache_names[] = {
  519. #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
  520. #include <linux/kmalloc_sizes.h>
  521. {NULL,}
  522. #undef CACHE
  523. };
  524. static struct arraycache_init initarray_cache __initdata =
  525. { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
  526. static struct arraycache_init initarray_generic =
  527. { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
  528. /* internal cache of cache description objs */
  529. static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES];
  530. static struct kmem_cache cache_cache = {
  531. .nodelists = cache_cache_nodelists,
  532. .batchcount = 1,
  533. .limit = BOOT_CPUCACHE_ENTRIES,
  534. .shared = 1,
  535. .buffer_size = sizeof(struct kmem_cache),
  536. .name = "kmem_cache",
  537. };
  538. #define BAD_ALIEN_MAGIC 0x01020304ul
  539. /*
  540. * chicken and egg problem: delay the per-cpu array allocation
  541. * until the general caches are up.
  542. */
  543. static enum {
  544. NONE,
  545. PARTIAL_AC,
  546. PARTIAL_L3,
  547. EARLY,
  548. LATE,
  549. FULL
  550. } g_cpucache_up;
  551. /*
  552. * used by boot code to determine if it can use slab based allocator
  553. */
  554. int slab_is_available(void)
  555. {
  556. return g_cpucache_up >= EARLY;
  557. }
  558. #ifdef CONFIG_LOCKDEP
  559. /*
  560. * Slab sometimes uses the kmalloc slabs to store the slab headers
  561. * for other slabs "off slab".
  562. * The locking for this is tricky in that it nests within the locks
  563. * of all other slabs in a few places; to deal with this special
  564. * locking we put on-slab caches into a separate lock-class.
  565. *
  566. * We set lock class for alien array caches which are up during init.
  567. * The lock annotation will be lost if all cpus of a node goes down and
  568. * then comes back up during hotplug
  569. */
  570. static struct lock_class_key on_slab_l3_key;
  571. static struct lock_class_key on_slab_alc_key;
  572. static struct lock_class_key debugobj_l3_key;
  573. static struct lock_class_key debugobj_alc_key;
  574. static void slab_set_lock_classes(struct kmem_cache *cachep,
  575. struct lock_class_key *l3_key, struct lock_class_key *alc_key,
  576. int q)
  577. {
  578. struct array_cache **alc;
  579. struct kmem_list3 *l3;
  580. int r;
  581. l3 = cachep->nodelists[q];
  582. if (!l3)
  583. return;
  584. lockdep_set_class(&l3->list_lock, l3_key);
  585. alc = l3->alien;
  586. /*
  587. * FIXME: This check for BAD_ALIEN_MAGIC
  588. * should go away when common slab code is taught to
  589. * work even without alien caches.
  590. * Currently, non NUMA code returns BAD_ALIEN_MAGIC
  591. * for alloc_alien_cache,
  592. */
  593. if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
  594. return;
  595. for_each_node(r) {
  596. if (alc[r])
  597. lockdep_set_class(&alc[r]->lock, alc_key);
  598. }
  599. }
  600. static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
  601. {
  602. slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
  603. }
  604. static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
  605. {
  606. int node;
  607. for_each_online_node(node)
  608. slab_set_debugobj_lock_classes_node(cachep, node);
  609. }
  610. static void init_node_lock_keys(int q)
  611. {
  612. struct cache_sizes *s = malloc_sizes;
  613. if (g_cpucache_up < LATE)
  614. return;
  615. for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
  616. struct kmem_list3 *l3;
  617. l3 = s->cs_cachep->nodelists[q];
  618. if (!l3 || OFF_SLAB(s->cs_cachep))
  619. continue;
  620. slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
  621. &on_slab_alc_key, q);
  622. }
  623. }
  624. static inline void init_lock_keys(void)
  625. {
  626. int node;
  627. for_each_node(node)
  628. init_node_lock_keys(node);
  629. }
  630. #else
  631. static void init_node_lock_keys(int q)
  632. {
  633. }
  634. static inline void init_lock_keys(void)
  635. {
  636. }
  637. static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
  638. {
  639. }
  640. static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
  641. {
  642. }
  643. #endif
  644. /*
  645. * Guard access to the cache-chain.
  646. */
  647. static DEFINE_MUTEX(cache_chain_mutex);
  648. static struct list_head cache_chain;
  649. static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
  650. static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
  651. {
  652. return cachep->array[smp_processor_id()];
  653. }
  654. static inline struct kmem_cache *__find_general_cachep(size_t size,
  655. gfp_t gfpflags)
  656. {
  657. struct cache_sizes *csizep = malloc_sizes;
  658. #if DEBUG
  659. /* This happens if someone tries to call
  660. * kmem_cache_create(), or __kmalloc(), before
  661. * the generic caches are initialized.
  662. */
  663. BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
  664. #endif
  665. if (!size)
  666. return ZERO_SIZE_PTR;
  667. while (size > csizep->cs_size)
  668. csizep++;
  669. /*
  670. * Really subtle: The last entry with cs->cs_size==ULONG_MAX
  671. * has cs_{dma,}cachep==NULL. Thus no special case
  672. * for large kmalloc calls required.
  673. */
  674. #ifdef CONFIG_ZONE_DMA
  675. if (unlikely(gfpflags & GFP_DMA))
  676. return csizep->cs_dmacachep;
  677. #endif
  678. return csizep->cs_cachep;
  679. }
  680. static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
  681. {
  682. return __find_general_cachep(size, gfpflags);
  683. }
  684. static size_t slab_mgmt_size(size_t nr_objs, size_t align)
  685. {
  686. return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
  687. }
  688. /*
  689. * Calculate the number of objects and left-over bytes for a given buffer size.
  690. */
  691. static void cache_estimate(unsigned long gfporder, size_t buffer_size,
  692. size_t align, int flags, size_t *left_over,
  693. unsigned int *num)
  694. {
  695. int nr_objs;
  696. size_t mgmt_size;
  697. size_t slab_size = PAGE_SIZE << gfporder;
  698. /*
  699. * The slab management structure can be either off the slab or
  700. * on it. For the latter case, the memory allocated for a
  701. * slab is used for:
  702. *
  703. * - The struct slab
  704. * - One kmem_bufctl_t for each object
  705. * - Padding to respect alignment of @align
  706. * - @buffer_size bytes for each object
  707. *
  708. * If the slab management structure is off the slab, then the
  709. * alignment will already be calculated into the size. Because
  710. * the slabs are all pages aligned, the objects will be at the
  711. * correct alignment when allocated.
  712. */
  713. if (flags & CFLGS_OFF_SLAB) {
  714. mgmt_size = 0;
  715. nr_objs = slab_size / buffer_size;
  716. if (nr_objs > SLAB_LIMIT)
  717. nr_objs = SLAB_LIMIT;
  718. } else {
  719. /*
  720. * Ignore padding for the initial guess. The padding
  721. * is at most @align-1 bytes, and @buffer_size is at
  722. * least @align. In the worst case, this result will
  723. * be one greater than the number of objects that fit
  724. * into the memory allocation when taking the padding
  725. * into account.
  726. */
  727. nr_objs = (slab_size - sizeof(struct slab)) /
  728. (buffer_size + sizeof(kmem_bufctl_t));
  729. /*
  730. * This calculated number will be either the right
  731. * amount, or one greater than what we want.
  732. */
  733. if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
  734. > slab_size)
  735. nr_objs--;
  736. if (nr_objs > SLAB_LIMIT)
  737. nr_objs = SLAB_LIMIT;
  738. mgmt_size = slab_mgmt_size(nr_objs, align);
  739. }
  740. *num = nr_objs;
  741. *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
  742. }
  743. #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
  744. static void __slab_error(const char *function, struct kmem_cache *cachep,
  745. char *msg)
  746. {
  747. printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
  748. function, cachep->name, msg);
  749. dump_stack();
  750. }
  751. /*
  752. * By default on NUMA we use alien caches to stage the freeing of
  753. * objects allocated from other nodes. This causes massive memory
  754. * inefficiencies when using fake NUMA setup to split memory into a
  755. * large number of small nodes, so it can be disabled on the command
  756. * line
  757. */
  758. static int use_alien_caches __read_mostly = 1;
  759. static int __init noaliencache_setup(char *s)
  760. {
  761. use_alien_caches = 0;
  762. return 1;
  763. }
  764. __setup("noaliencache", noaliencache_setup);
  765. static int __init slab_max_order_setup(char *str)
  766. {
  767. get_option(&str, &slab_max_order);
  768. slab_max_order = slab_max_order < 0 ? 0 :
  769. min(slab_max_order, MAX_ORDER - 1);
  770. slab_max_order_set = true;
  771. return 1;
  772. }
  773. __setup("slab_max_order=", slab_max_order_setup);
  774. #ifdef CONFIG_NUMA
  775. /*
  776. * Special reaping functions for NUMA systems called from cache_reap().
  777. * These take care of doing round robin flushing of alien caches (containing
  778. * objects freed on different nodes from which they were allocated) and the
  779. * flushing of remote pcps by calling drain_node_pages.
  780. */
  781. static DEFINE_PER_CPU(unsigned long, slab_reap_node);
  782. static void init_reap_node(int cpu)
  783. {
  784. int node;
  785. node = next_node(cpu_to_mem(cpu), node_online_map);
  786. if (node == MAX_NUMNODES)
  787. node = first_node(node_online_map);
  788. per_cpu(slab_reap_node, cpu) = node;
  789. }
  790. static void next_reap_node(void)
  791. {
  792. int node = __this_cpu_read(slab_reap_node);
  793. node = next_node(node, node_online_map);
  794. if (unlikely(node >= MAX_NUMNODES))
  795. node = first_node(node_online_map);
  796. __this_cpu_write(slab_reap_node, node);
  797. }
  798. #else
  799. #define init_reap_node(cpu) do { } while (0)
  800. #define next_reap_node(void) do { } while (0)
  801. #endif
  802. /*
  803. * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
  804. * via the workqueue/eventd.
  805. * Add the CPU number into the expiration time to minimize the possibility of
  806. * the CPUs getting into lockstep and contending for the global cache chain
  807. * lock.
  808. */
  809. static void __cpuinit start_cpu_timer(int cpu)
  810. {
  811. struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
  812. /*
  813. * When this gets called from do_initcalls via cpucache_init(),
  814. * init_workqueues() has already run, so keventd will be setup
  815. * at that time.
  816. */
  817. if (keventd_up() && reap_work->work.func == NULL) {
  818. init_reap_node(cpu);
  819. INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
  820. schedule_delayed_work_on(cpu, reap_work,
  821. __round_jiffies_relative(HZ, cpu));
  822. }
  823. }
  824. static struct array_cache *alloc_arraycache(int node, int entries,
  825. int batchcount, gfp_t gfp)
  826. {
  827. int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
  828. struct array_cache *nc = NULL;
  829. nc = kmalloc_node(memsize, gfp, node);
  830. /*
  831. * The array_cache structures contain pointers to free object.
  832. * However, when such objects are allocated or transferred to another
  833. * cache the pointers are not cleared and they could be counted as
  834. * valid references during a kmemleak scan. Therefore, kmemleak must
  835. * not scan such objects.
  836. */
  837. kmemleak_no_scan(nc);
  838. if (nc) {
  839. nc->avail = 0;
  840. nc->limit = entries;
  841. nc->batchcount = batchcount;
  842. nc->touched = 0;
  843. spin_lock_init(&nc->lock);
  844. }
  845. return nc;
  846. }
  847. /*
  848. * Transfer objects in one arraycache to another.
  849. * Locking must be handled by the caller.
  850. *
  851. * Return the number of entries transferred.
  852. */
  853. static int transfer_objects(struct array_cache *to,
  854. struct array_cache *from, unsigned int max)
  855. {
  856. /* Figure out how many entries to transfer */
  857. int nr = min3(from->avail, max, to->limit - to->avail);
  858. if (!nr)
  859. return 0;
  860. memcpy(to->entry + to->avail, from->entry + from->avail -nr,
  861. sizeof(void *) *nr);
  862. from->avail -= nr;
  863. to->avail += nr;
  864. return nr;
  865. }
  866. #ifndef CONFIG_NUMA
  867. #define drain_alien_cache(cachep, alien) do { } while (0)
  868. #define reap_alien(cachep, l3) do { } while (0)
  869. static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
  870. {
  871. return (struct array_cache **)BAD_ALIEN_MAGIC;
  872. }
  873. static inline void free_alien_cache(struct array_cache **ac_ptr)
  874. {
  875. }
  876. static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
  877. {
  878. return 0;
  879. }
  880. static inline void *alternate_node_alloc(struct kmem_cache *cachep,
  881. gfp_t flags)
  882. {
  883. return NULL;
  884. }
  885. static inline void *____cache_alloc_node(struct kmem_cache *cachep,
  886. gfp_t flags, int nodeid)
  887. {
  888. return NULL;
  889. }
  890. #else /* CONFIG_NUMA */
  891. static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
  892. static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
  893. static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
  894. {
  895. struct array_cache **ac_ptr;
  896. int memsize = sizeof(void *) * nr_node_ids;
  897. int i;
  898. if (limit > 1)
  899. limit = 12;
  900. ac_ptr = kzalloc_node(memsize, gfp, node);
  901. if (ac_ptr) {
  902. for_each_node(i) {
  903. if (i == node || !node_online(i))
  904. continue;
  905. ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
  906. if (!ac_ptr[i]) {
  907. for (i--; i >= 0; i--)
  908. kfree(ac_ptr[i]);
  909. kfree(ac_ptr);
  910. return NULL;
  911. }
  912. }
  913. }
  914. return ac_ptr;
  915. }
  916. static void free_alien_cache(struct array_cache **ac_ptr)
  917. {
  918. int i;
  919. if (!ac_ptr)
  920. return;
  921. for_each_node(i)
  922. kfree(ac_ptr[i]);
  923. kfree(ac_ptr);
  924. }
  925. static void __drain_alien_cache(struct kmem_cache *cachep,
  926. struct array_cache *ac, int node)
  927. {
  928. struct kmem_list3 *rl3 = cachep->nodelists[node];
  929. if (ac->avail) {
  930. spin_lock(&rl3->list_lock);
  931. /*
  932. * Stuff objects into the remote nodes shared array first.
  933. * That way we could avoid the overhead of putting the objects
  934. * into the free lists and getting them back later.
  935. */
  936. if (rl3->shared)
  937. transfer_objects(rl3->shared, ac, ac->limit);
  938. free_block(cachep, ac->entry, ac->avail, node);
  939. ac->avail = 0;
  940. spin_unlock(&rl3->list_lock);
  941. }
  942. }
  943. /*
  944. * Called from cache_reap() to regularly drain alien caches round robin.
  945. */
  946. static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
  947. {
  948. int node = __this_cpu_read(slab_reap_node);
  949. if (l3->alien) {
  950. struct array_cache *ac = l3->alien[node];
  951. if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
  952. __drain_alien_cache(cachep, ac, node);
  953. spin_unlock_irq(&ac->lock);
  954. }
  955. }
  956. }
  957. static void drain_alien_cache(struct kmem_cache *cachep,
  958. struct array_cache **alien)
  959. {
  960. int i = 0;
  961. struct array_cache *ac;
  962. unsigned long flags;
  963. for_each_online_node(i) {
  964. ac = alien[i];
  965. if (ac) {
  966. spin_lock_irqsave(&ac->lock, flags);
  967. __drain_alien_cache(cachep, ac, i);
  968. spin_unlock_irqrestore(&ac->lock, flags);
  969. }
  970. }
  971. }
  972. static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
  973. {
  974. struct slab *slabp = virt_to_slab(objp);
  975. int nodeid = slabp->nodeid;
  976. struct kmem_list3 *l3;
  977. struct array_cache *alien = NULL;
  978. int node;
  979. node = numa_mem_id();
  980. /*
  981. * Make sure we are not freeing a object from another node to the array
  982. * cache on this cpu.
  983. */
  984. if (likely(slabp->nodeid == node))
  985. return 0;
  986. l3 = cachep->nodelists[node];
  987. STATS_INC_NODEFREES(cachep);
  988. if (l3->alien && l3->alien[nodeid]) {
  989. alien = l3->alien[nodeid];
  990. spin_lock(&alien->lock);
  991. if (unlikely(alien->avail == alien->limit)) {
  992. STATS_INC_ACOVERFLOW(cachep);
  993. __drain_alien_cache(cachep, alien, nodeid);
  994. }
  995. alien->entry[alien->avail++] = objp;
  996. spin_unlock(&alien->lock);
  997. } else {
  998. spin_lock(&(cachep->nodelists[nodeid])->list_lock);
  999. free_block(cachep, &objp, 1, nodeid);
  1000. spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
  1001. }
  1002. return 1;
  1003. }
  1004. #endif
  1005. /*
  1006. * Allocates and initializes nodelists for a node on each slab cache, used for
  1007. * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
  1008. * will be allocated off-node since memory is not yet online for the new node.
  1009. * When hotplugging memory or a cpu, existing nodelists are not replaced if
  1010. * already in use.
  1011. *
  1012. * Must hold cache_chain_mutex.
  1013. */
  1014. static int init_cache_nodelists_node(int node)
  1015. {
  1016. struct kmem_cache *cachep;
  1017. struct kmem_list3 *l3;
  1018. const int memsize = sizeof(struct kmem_list3);
  1019. list_for_each_entry(cachep, &cache_chain, next) {
  1020. /*
  1021. * Set up the size64 kmemlist for cpu before we can
  1022. * begin anything. Make sure some other cpu on this
  1023. * node has not already allocated this
  1024. */
  1025. if (!cachep->nodelists[node]) {
  1026. l3 = kmalloc_node(memsize, GFP_KERNEL, node);
  1027. if (!l3)
  1028. return -ENOMEM;
  1029. kmem_list3_init(l3);
  1030. l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
  1031. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1032. /*
  1033. * The l3s don't come and go as CPUs come and
  1034. * go. cache_chain_mutex is sufficient
  1035. * protection here.
  1036. */
  1037. cachep->nodelists[node] = l3;
  1038. }
  1039. spin_lock_irq(&cachep->nodelists[node]->list_lock);
  1040. cachep->nodelists[node]->free_limit =
  1041. (1 + nr_cpus_node(node)) *
  1042. cachep->batchcount + cachep->num;
  1043. spin_unlock_irq(&cachep->nodelists[node]->list_lock);
  1044. }
  1045. return 0;
  1046. }
  1047. static void __cpuinit cpuup_canceled(long cpu)
  1048. {
  1049. struct kmem_cache *cachep;
  1050. struct kmem_list3 *l3 = NULL;
  1051. int node = cpu_to_mem(cpu);
  1052. const struct cpumask *mask = cpumask_of_node(node);
  1053. list_for_each_entry(cachep, &cache_chain, next) {
  1054. struct array_cache *nc;
  1055. struct array_cache *shared;
  1056. struct array_cache **alien;
  1057. /* cpu is dead; no one can alloc from it. */
  1058. nc = cachep->array[cpu];
  1059. cachep->array[cpu] = NULL;
  1060. l3 = cachep->nodelists[node];
  1061. if (!l3)
  1062. goto free_array_cache;
  1063. spin_lock_irq(&l3->list_lock);
  1064. /* Free limit for this kmem_list3 */
  1065. l3->free_limit -= cachep->batchcount;
  1066. if (nc)
  1067. free_block(cachep, nc->entry, nc->avail, node);
  1068. if (!cpumask_empty(mask)) {
  1069. spin_unlock_irq(&l3->list_lock);
  1070. goto free_array_cache;
  1071. }
  1072. shared = l3->shared;
  1073. if (shared) {
  1074. free_block(cachep, shared->entry,
  1075. shared->avail, node);
  1076. l3->shared = NULL;
  1077. }
  1078. alien = l3->alien;
  1079. l3->alien = NULL;
  1080. spin_unlock_irq(&l3->list_lock);
  1081. kfree(shared);
  1082. if (alien) {
  1083. drain_alien_cache(cachep, alien);
  1084. free_alien_cache(alien);
  1085. }
  1086. free_array_cache:
  1087. kfree(nc);
  1088. }
  1089. /*
  1090. * In the previous loop, all the objects were freed to
  1091. * the respective cache's slabs, now we can go ahead and
  1092. * shrink each nodelist to its limit.
  1093. */
  1094. list_for_each_entry(cachep, &cache_chain, next) {
  1095. l3 = cachep->nodelists[node];
  1096. if (!l3)
  1097. continue;
  1098. drain_freelist(cachep, l3, l3->free_objects);
  1099. }
  1100. }
  1101. static int __cpuinit cpuup_prepare(long cpu)
  1102. {
  1103. struct kmem_cache *cachep;
  1104. struct kmem_list3 *l3 = NULL;
  1105. int node = cpu_to_mem(cpu);
  1106. int err;
  1107. /*
  1108. * We need to do this right in the beginning since
  1109. * alloc_arraycache's are going to use this list.
  1110. * kmalloc_node allows us to add the slab to the right
  1111. * kmem_list3 and not this cpu's kmem_list3
  1112. */
  1113. err = init_cache_nodelists_node(node);
  1114. if (err < 0)
  1115. goto bad;
  1116. /*
  1117. * Now we can go ahead with allocating the shared arrays and
  1118. * array caches
  1119. */
  1120. list_for_each_entry(cachep, &cache_chain, next) {
  1121. struct array_cache *nc;
  1122. struct array_cache *shared = NULL;
  1123. struct array_cache **alien = NULL;
  1124. nc = alloc_arraycache(node, cachep->limit,
  1125. cachep->batchcount, GFP_KERNEL);
  1126. if (!nc)
  1127. goto bad;
  1128. if (cachep->shared) {
  1129. shared = alloc_arraycache(node,
  1130. cachep->shared * cachep->batchcount,
  1131. 0xbaadf00d, GFP_KERNEL);
  1132. if (!shared) {
  1133. kfree(nc);
  1134. goto bad;
  1135. }
  1136. }
  1137. if (use_alien_caches) {
  1138. alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
  1139. if (!alien) {
  1140. kfree(shared);
  1141. kfree(nc);
  1142. goto bad;
  1143. }
  1144. }
  1145. cachep->array[cpu] = nc;
  1146. l3 = cachep->nodelists[node];
  1147. BUG_ON(!l3);
  1148. spin_lock_irq(&l3->list_lock);
  1149. if (!l3->shared) {
  1150. /*
  1151. * We are serialised from CPU_DEAD or
  1152. * CPU_UP_CANCELLED by the cpucontrol lock
  1153. */
  1154. l3->shared = shared;
  1155. shared = NULL;
  1156. }
  1157. #ifdef CONFIG_NUMA
  1158. if (!l3->alien) {
  1159. l3->alien = alien;
  1160. alien = NULL;
  1161. }
  1162. #endif
  1163. spin_unlock_irq(&l3->list_lock);
  1164. kfree(shared);
  1165. free_alien_cache(alien);
  1166. if (cachep->flags & SLAB_DEBUG_OBJECTS)
  1167. slab_set_debugobj_lock_classes_node(cachep, node);
  1168. }
  1169. init_node_lock_keys(node);
  1170. return 0;
  1171. bad:
  1172. cpuup_canceled(cpu);
  1173. return -ENOMEM;
  1174. }
  1175. static int __cpuinit cpuup_callback(struct notifier_block *nfb,
  1176. unsigned long action, void *hcpu)
  1177. {
  1178. long cpu = (long)hcpu;
  1179. int err = 0;
  1180. switch (action) {
  1181. case CPU_UP_PREPARE:
  1182. case CPU_UP_PREPARE_FROZEN:
  1183. mutex_lock(&cache_chain_mutex);
  1184. err = cpuup_prepare(cpu);
  1185. mutex_unlock(&cache_chain_mutex);
  1186. break;
  1187. case CPU_ONLINE:
  1188. case CPU_ONLINE_FROZEN:
  1189. start_cpu_timer(cpu);
  1190. break;
  1191. #ifdef CONFIG_HOTPLUG_CPU
  1192. case CPU_DOWN_PREPARE:
  1193. case CPU_DOWN_PREPARE_FROZEN:
  1194. /*
  1195. * Shutdown cache reaper. Note that the cache_chain_mutex is
  1196. * held so that if cache_reap() is invoked it cannot do
  1197. * anything expensive but will only modify reap_work
  1198. * and reschedule the timer.
  1199. */
  1200. cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
  1201. /* Now the cache_reaper is guaranteed to be not running. */
  1202. per_cpu(slab_reap_work, cpu).work.func = NULL;
  1203. break;
  1204. case CPU_DOWN_FAILED:
  1205. case CPU_DOWN_FAILED_FROZEN:
  1206. start_cpu_timer(cpu);
  1207. break;
  1208. case CPU_DEAD:
  1209. case CPU_DEAD_FROZEN:
  1210. /*
  1211. * Even if all the cpus of a node are down, we don't free the
  1212. * kmem_list3 of any cache. This to avoid a race between
  1213. * cpu_down, and a kmalloc allocation from another cpu for
  1214. * memory from the node of the cpu going down. The list3
  1215. * structure is usually allocated from kmem_cache_create() and
  1216. * gets destroyed at kmem_cache_destroy().
  1217. */
  1218. /* fall through */
  1219. #endif
  1220. case CPU_UP_CANCELED:
  1221. case CPU_UP_CANCELED_FROZEN:
  1222. mutex_lock(&cache_chain_mutex);
  1223. cpuup_canceled(cpu);
  1224. mutex_unlock(&cache_chain_mutex);
  1225. break;
  1226. }
  1227. return notifier_from_errno(err);
  1228. }
  1229. static struct notifier_block __cpuinitdata cpucache_notifier = {
  1230. &cpuup_callback, NULL, 0
  1231. };
  1232. #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
  1233. /*
  1234. * Drains freelist for a node on each slab cache, used for memory hot-remove.
  1235. * Returns -EBUSY if all objects cannot be drained so that the node is not
  1236. * removed.
  1237. *
  1238. * Must hold cache_chain_mutex.
  1239. */
  1240. static int __meminit drain_cache_nodelists_node(int node)
  1241. {
  1242. struct kmem_cache *cachep;
  1243. int ret = 0;
  1244. list_for_each_entry(cachep, &cache_chain, next) {
  1245. struct kmem_list3 *l3;
  1246. l3 = cachep->nodelists[node];
  1247. if (!l3)
  1248. continue;
  1249. drain_freelist(cachep, l3, l3->free_objects);
  1250. if (!list_empty(&l3->slabs_full) ||
  1251. !list_empty(&l3->slabs_partial)) {
  1252. ret = -EBUSY;
  1253. break;
  1254. }
  1255. }
  1256. return ret;
  1257. }
  1258. static int __meminit slab_memory_callback(struct notifier_block *self,
  1259. unsigned long action, void *arg)
  1260. {
  1261. struct memory_notify *mnb = arg;
  1262. int ret = 0;
  1263. int nid;
  1264. nid = mnb->status_change_nid;
  1265. if (nid < 0)
  1266. goto out;
  1267. switch (action) {
  1268. case MEM_GOING_ONLINE:
  1269. mutex_lock(&cache_chain_mutex);
  1270. ret = init_cache_nodelists_node(nid);
  1271. mutex_unlock(&cache_chain_mutex);
  1272. break;
  1273. case MEM_GOING_OFFLINE:
  1274. mutex_lock(&cache_chain_mutex);
  1275. ret = drain_cache_nodelists_node(nid);
  1276. mutex_unlock(&cache_chain_mutex);
  1277. break;
  1278. case MEM_ONLINE:
  1279. case MEM_OFFLINE:
  1280. case MEM_CANCEL_ONLINE:
  1281. case MEM_CANCEL_OFFLINE:
  1282. break;
  1283. }
  1284. out:
  1285. return notifier_from_errno(ret);
  1286. }
  1287. #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
  1288. /*
  1289. * swap the static kmem_list3 with kmalloced memory
  1290. */
  1291. static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
  1292. int nodeid)
  1293. {
  1294. struct kmem_list3 *ptr;
  1295. ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
  1296. BUG_ON(!ptr);
  1297. memcpy(ptr, list, sizeof(struct kmem_list3));
  1298. /*
  1299. * Do not assume that spinlocks can be initialized via memcpy:
  1300. */
  1301. spin_lock_init(&ptr->list_lock);
  1302. MAKE_ALL_LISTS(cachep, ptr, nodeid);
  1303. cachep->nodelists[nodeid] = ptr;
  1304. }
  1305. /*
  1306. * For setting up all the kmem_list3s for cache whose buffer_size is same as
  1307. * size of kmem_list3.
  1308. */
  1309. static void __init set_up_list3s(struct kmem_cache *cachep, int index)
  1310. {
  1311. int node;
  1312. for_each_online_node(node) {
  1313. cachep->nodelists[node] = &initkmem_list3[index + node];
  1314. cachep->nodelists[node]->next_reap = jiffies +
  1315. REAPTIMEOUT_LIST3 +
  1316. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1317. }
  1318. }
  1319. /*
  1320. * Initialisation. Called after the page allocator have been initialised and
  1321. * before smp_init().
  1322. */
  1323. void __init kmem_cache_init(void)
  1324. {
  1325. size_t left_over;
  1326. struct cache_sizes *sizes;
  1327. struct cache_names *names;
  1328. int i;
  1329. int order;
  1330. int node;
  1331. if (num_possible_nodes() == 1)
  1332. use_alien_caches = 0;
  1333. for (i = 0; i < NUM_INIT_LISTS; i++) {
  1334. kmem_list3_init(&initkmem_list3[i]);
  1335. if (i < MAX_NUMNODES)
  1336. cache_cache.nodelists[i] = NULL;
  1337. }
  1338. set_up_list3s(&cache_cache, CACHE_CACHE);
  1339. /*
  1340. * Fragmentation resistance on low memory - only use bigger
  1341. * page orders on machines with more than 32MB of memory if
  1342. * not overridden on the command line.
  1343. */
  1344. if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
  1345. slab_max_order = SLAB_MAX_ORDER_HI;
  1346. /* Bootstrap is tricky, because several objects are allocated
  1347. * from caches that do not exist yet:
  1348. * 1) initialize the cache_cache cache: it contains the struct
  1349. * kmem_cache structures of all caches, except cache_cache itself:
  1350. * cache_cache is statically allocated.
  1351. * Initially an __init data area is used for the head array and the
  1352. * kmem_list3 structures, it's replaced with a kmalloc allocated
  1353. * array at the end of the bootstrap.
  1354. * 2) Create the first kmalloc cache.
  1355. * The struct kmem_cache for the new cache is allocated normally.
  1356. * An __init data area is used for the head array.
  1357. * 3) Create the remaining kmalloc caches, with minimally sized
  1358. * head arrays.
  1359. * 4) Replace the __init data head arrays for cache_cache and the first
  1360. * kmalloc cache with kmalloc allocated arrays.
  1361. * 5) Replace the __init data for kmem_list3 for cache_cache and
  1362. * the other cache's with kmalloc allocated memory.
  1363. * 6) Resize the head arrays of the kmalloc caches to their final sizes.
  1364. */
  1365. node = numa_mem_id();
  1366. /* 1) create the cache_cache */
  1367. INIT_LIST_HEAD(&cache_chain);
  1368. list_add(&cache_cache.next, &cache_chain);
  1369. cache_cache.colour_off = cache_line_size();
  1370. cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
  1371. cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
  1372. /*
  1373. * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
  1374. */
  1375. cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
  1376. nr_node_ids * sizeof(struct kmem_list3 *);
  1377. #if DEBUG
  1378. cache_cache.obj_size = cache_cache.buffer_size;
  1379. #endif
  1380. cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
  1381. cache_line_size());
  1382. cache_cache.reciprocal_buffer_size =
  1383. reciprocal_value(cache_cache.buffer_size);
  1384. for (order = 0; order < MAX_ORDER; order++) {
  1385. cache_estimate(order, cache_cache.buffer_size,
  1386. cache_line_size(), 0, &left_over, &cache_cache.num);
  1387. if (cache_cache.num)
  1388. break;
  1389. }
  1390. BUG_ON(!cache_cache.num);
  1391. cache_cache.gfporder = order;
  1392. cache_cache.colour = left_over / cache_cache.colour_off;
  1393. cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
  1394. sizeof(struct slab), cache_line_size());
  1395. /* 2+3) create the kmalloc caches */
  1396. sizes = malloc_sizes;
  1397. names = cache_names;
  1398. /*
  1399. * Initialize the caches that provide memory for the array cache and the
  1400. * kmem_list3 structures first. Without this, further allocations will
  1401. * bug.
  1402. */
  1403. sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
  1404. sizes[INDEX_AC].cs_size,
  1405. ARCH_KMALLOC_MINALIGN,
  1406. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1407. NULL);
  1408. if (INDEX_AC != INDEX_L3) {
  1409. sizes[INDEX_L3].cs_cachep =
  1410. kmem_cache_create(names[INDEX_L3].name,
  1411. sizes[INDEX_L3].cs_size,
  1412. ARCH_KMALLOC_MINALIGN,
  1413. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1414. NULL);
  1415. }
  1416. slab_early_init = 0;
  1417. while (sizes->cs_size != ULONG_MAX) {
  1418. /*
  1419. * For performance, all the general caches are L1 aligned.
  1420. * This should be particularly beneficial on SMP boxes, as it
  1421. * eliminates "false sharing".
  1422. * Note for systems short on memory removing the alignment will
  1423. * allow tighter packing of the smaller caches.
  1424. */
  1425. if (!sizes->cs_cachep) {
  1426. sizes->cs_cachep = kmem_cache_create(names->name,
  1427. sizes->cs_size,
  1428. ARCH_KMALLOC_MINALIGN,
  1429. ARCH_KMALLOC_FLAGS|SLAB_PANIC,
  1430. NULL);
  1431. }
  1432. #ifdef CONFIG_ZONE_DMA
  1433. sizes->cs_dmacachep = kmem_cache_create(
  1434. names->name_dma,
  1435. sizes->cs_size,
  1436. ARCH_KMALLOC_MINALIGN,
  1437. ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
  1438. SLAB_PANIC,
  1439. NULL);
  1440. #endif
  1441. sizes++;
  1442. names++;
  1443. }
  1444. /* 4) Replace the bootstrap head arrays */
  1445. {
  1446. struct array_cache *ptr;
  1447. ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
  1448. BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
  1449. memcpy(ptr, cpu_cache_get(&cache_cache),
  1450. sizeof(struct arraycache_init));
  1451. /*
  1452. * Do not assume that spinlocks can be initialized via memcpy:
  1453. */
  1454. spin_lock_init(&ptr->lock);
  1455. cache_cache.array[smp_processor_id()] = ptr;
  1456. ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
  1457. BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
  1458. != &initarray_generic.cache);
  1459. memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
  1460. sizeof(struct arraycache_init));
  1461. /*
  1462. * Do not assume that spinlocks can be initialized via memcpy:
  1463. */
  1464. spin_lock_init(&ptr->lock);
  1465. malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
  1466. ptr;
  1467. }
  1468. /* 5) Replace the bootstrap kmem_list3's */
  1469. {
  1470. int nid;
  1471. for_each_online_node(nid) {
  1472. init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
  1473. init_list(malloc_sizes[INDEX_AC].cs_cachep,
  1474. &initkmem_list3[SIZE_AC + nid], nid);
  1475. if (INDEX_AC != INDEX_L3) {
  1476. init_list(malloc_sizes[INDEX_L3].cs_cachep,
  1477. &initkmem_list3[SIZE_L3 + nid], nid);
  1478. }
  1479. }
  1480. }
  1481. g_cpucache_up = EARLY;
  1482. }
  1483. void __init kmem_cache_init_late(void)
  1484. {
  1485. struct kmem_cache *cachep;
  1486. g_cpucache_up = LATE;
  1487. /* Annotate slab for lockdep -- annotate the malloc caches */
  1488. init_lock_keys();
  1489. /* 6) resize the head arrays to their final sizes */
  1490. mutex_lock(&cache_chain_mutex);
  1491. list_for_each_entry(cachep, &cache_chain, next)
  1492. if (enable_cpucache(cachep, GFP_NOWAIT))
  1493. BUG();
  1494. mutex_unlock(&cache_chain_mutex);
  1495. /* Done! */
  1496. g_cpucache_up = FULL;
  1497. /*
  1498. * Register a cpu startup notifier callback that initializes
  1499. * cpu_cache_get for all new cpus
  1500. */
  1501. register_cpu_notifier(&cpucache_notifier);
  1502. #ifdef CONFIG_NUMA
  1503. /*
  1504. * Register a memory hotplug callback that initializes and frees
  1505. * nodelists.
  1506. */
  1507. hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
  1508. #endif
  1509. /*
  1510. * The reap timers are started later, with a module init call: That part
  1511. * of the kernel is not yet operational.
  1512. */
  1513. }
  1514. static int __init cpucache_init(void)
  1515. {
  1516. int cpu;
  1517. /*
  1518. * Register the timers that return unneeded pages to the page allocator
  1519. */
  1520. for_each_online_cpu(cpu)
  1521. start_cpu_timer(cpu);
  1522. return 0;
  1523. }
  1524. __initcall(cpucache_init);
  1525. /*
  1526. * Interface to system's page allocator. No need to hold the cache-lock.
  1527. *
  1528. * If we requested dmaable memory, we will get it. Even if we
  1529. * did not request dmaable memory, we might get it, but that
  1530. * would be relatively rare and ignorable.
  1531. */
  1532. static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
  1533. {
  1534. struct page *page;
  1535. int nr_pages;
  1536. int i;
  1537. #ifndef CONFIG_MMU
  1538. /*
  1539. * Nommu uses slab's for process anonymous memory allocations, and thus
  1540. * requires __GFP_COMP to properly refcount higher order allocations
  1541. */
  1542. flags |= __GFP_COMP;
  1543. #endif
  1544. flags |= cachep->gfpflags;
  1545. if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
  1546. flags |= __GFP_RECLAIMABLE;
  1547. page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
  1548. if (!page)
  1549. return NULL;
  1550. nr_pages = (1 << cachep->gfporder);
  1551. if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
  1552. add_zone_page_state(page_zone(page),
  1553. NR_SLAB_RECLAIMABLE, nr_pages);
  1554. else
  1555. add_zone_page_state(page_zone(page),
  1556. NR_SLAB_UNRECLAIMABLE, nr_pages);
  1557. for (i = 0; i < nr_pages; i++)
  1558. __SetPageSlab(page + i);
  1559. if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
  1560. kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
  1561. if (cachep->ctor)
  1562. kmemcheck_mark_uninitialized_pages(page, nr_pages);
  1563. else
  1564. kmemcheck_mark_unallocated_pages(page, nr_pages);
  1565. }
  1566. return page_address(page);
  1567. }
  1568. /*
  1569. * Interface to system's page release.
  1570. */
  1571. static void kmem_freepages(struct kmem_cache *cachep, void *addr)
  1572. {
  1573. unsigned long i = (1 << cachep->gfporder);
  1574. struct page *page = virt_to_page(addr);
  1575. const unsigned long nr_freed = i;
  1576. kmemcheck_free_shadow(page, cachep->gfporder);
  1577. if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
  1578. sub_zone_page_state(page_zone(page),
  1579. NR_SLAB_RECLAIMABLE, nr_freed);
  1580. else
  1581. sub_zone_page_state(page_zone(page),
  1582. NR_SLAB_UNRECLAIMABLE, nr_freed);
  1583. while (i--) {
  1584. BUG_ON(!PageSlab(page));
  1585. __ClearPageSlab(page);
  1586. page++;
  1587. }
  1588. if (current->reclaim_state)
  1589. current->reclaim_state->reclaimed_slab += nr_freed;
  1590. free_pages((unsigned long)addr, cachep->gfporder);
  1591. }
  1592. static void kmem_rcu_free(struct rcu_head *head)
  1593. {
  1594. struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
  1595. struct kmem_cache *cachep = slab_rcu->cachep;
  1596. kmem_freepages(cachep, slab_rcu->addr);
  1597. if (OFF_SLAB(cachep))
  1598. kmem_cache_free(cachep->slabp_cache, slab_rcu);
  1599. }
  1600. #if DEBUG
  1601. #ifdef CONFIG_DEBUG_PAGEALLOC
  1602. static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
  1603. unsigned long caller)
  1604. {
  1605. int size = obj_size(cachep);
  1606. addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
  1607. if (size < 5 * sizeof(unsigned long))
  1608. return;
  1609. *addr++ = 0x12345678;
  1610. *addr++ = caller;
  1611. *addr++ = smp_processor_id();
  1612. size -= 3 * sizeof(unsigned long);
  1613. {
  1614. unsigned long *sptr = &caller;
  1615. unsigned long svalue;
  1616. while (!kstack_end(sptr)) {
  1617. svalue = *sptr++;
  1618. if (kernel_text_address(svalue)) {
  1619. *addr++ = svalue;
  1620. size -= sizeof(unsigned long);
  1621. if (size <= sizeof(unsigned long))
  1622. break;
  1623. }
  1624. }
  1625. }
  1626. *addr++ = 0x87654321;
  1627. }
  1628. #endif
  1629. static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
  1630. {
  1631. int size = obj_size(cachep);
  1632. addr = &((char *)addr)[obj_offset(cachep)];
  1633. memset(addr, val, size);
  1634. *(unsigned char *)(addr + size - 1) = POISON_END;
  1635. }
  1636. static void dump_line(char *data, int offset, int limit)
  1637. {
  1638. int i;
  1639. unsigned char error = 0;
  1640. int bad_count = 0;
  1641. printk(KERN_ERR "%03x: ", offset);
  1642. for (i = 0; i < limit; i++) {
  1643. if (data[offset + i] != POISON_FREE) {
  1644. error = data[offset + i];
  1645. bad_count++;
  1646. }
  1647. }
  1648. print_hex_dump(KERN_CONT, "", 0, 16, 1,
  1649. &data[offset], limit, 1);
  1650. if (bad_count == 1) {
  1651. error ^= POISON_FREE;
  1652. if (!(error & (error - 1))) {
  1653. printk(KERN_ERR "Single bit error detected. Probably "
  1654. "bad RAM.\n");
  1655. #ifdef CONFIG_X86
  1656. printk(KERN_ERR "Run memtest86+ or a similar memory "
  1657. "test tool.\n");
  1658. #else
  1659. printk(KERN_ERR "Run a memory test tool.\n");
  1660. #endif
  1661. }
  1662. }
  1663. }
  1664. #endif
  1665. #if DEBUG
  1666. static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
  1667. {
  1668. int i, size;
  1669. char *realobj;
  1670. if (cachep->flags & SLAB_RED_ZONE) {
  1671. printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
  1672. *dbg_redzone1(cachep, objp),
  1673. *dbg_redzone2(cachep, objp));
  1674. }
  1675. if (cachep->flags & SLAB_STORE_USER) {
  1676. printk(KERN_ERR "Last user: [<%p>]",
  1677. *dbg_userword(cachep, objp));
  1678. print_symbol("(%s)",
  1679. (unsigned long)*dbg_userword(cachep, objp));
  1680. printk("\n");
  1681. }
  1682. realobj = (char *)objp + obj_offset(cachep);
  1683. size = obj_size(cachep);
  1684. for (i = 0; i < size && lines; i += 16, lines--) {
  1685. int limit;
  1686. limit = 16;
  1687. if (i + limit > size)
  1688. limit = size - i;
  1689. dump_line(realobj, i, limit);
  1690. }
  1691. }
  1692. static void check_poison_obj(struct kmem_cache *cachep, void *objp)
  1693. {
  1694. char *realobj;
  1695. int size, i;
  1696. int lines = 0;
  1697. realobj = (char *)objp + obj_offset(cachep);
  1698. size = obj_size(cachep);
  1699. for (i = 0; i < size; i++) {
  1700. char exp = POISON_FREE;
  1701. if (i == size - 1)
  1702. exp = POISON_END;
  1703. if (realobj[i] != exp) {
  1704. int limit;
  1705. /* Mismatch ! */
  1706. /* Print header */
  1707. if (lines == 0) {
  1708. printk(KERN_ERR
  1709. "Slab corruption (%s): %s start=%p, len=%d\n",
  1710. print_tainted(), cachep->name, realobj, size);
  1711. print_objinfo(cachep, objp, 0);
  1712. }
  1713. /* Hexdump the affected line */
  1714. i = (i / 16) * 16;
  1715. limit = 16;
  1716. if (i + limit > size)
  1717. limit = size - i;
  1718. dump_line(realobj, i, limit);
  1719. i += 16;
  1720. lines++;
  1721. /* Limit to 5 lines */
  1722. if (lines > 5)
  1723. break;
  1724. }
  1725. }
  1726. if (lines != 0) {
  1727. /* Print some data about the neighboring objects, if they
  1728. * exist:
  1729. */
  1730. struct slab *slabp = virt_to_slab(objp);
  1731. unsigned int objnr;
  1732. objnr = obj_to_index(cachep, slabp, objp);
  1733. if (objnr) {
  1734. objp = index_to_obj(cachep, slabp, objnr - 1);
  1735. realobj = (char *)objp + obj_offset(cachep);
  1736. printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
  1737. realobj, size);
  1738. print_objinfo(cachep, objp, 2);
  1739. }
  1740. if (objnr + 1 < cachep->num) {
  1741. objp = index_to_obj(cachep, slabp, objnr + 1);
  1742. realobj = (char *)objp + obj_offset(cachep);
  1743. printk(KERN_ERR "Next obj: start=%p, len=%d\n",
  1744. realobj, size);
  1745. print_objinfo(cachep, objp, 2);
  1746. }
  1747. }
  1748. }
  1749. #endif
  1750. #if DEBUG
  1751. static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
  1752. {
  1753. int i;
  1754. for (i = 0; i < cachep->num; i++) {
  1755. void *objp = index_to_obj(cachep, slabp, i);
  1756. if (cachep->flags & SLAB_POISON) {
  1757. #ifdef CONFIG_DEBUG_PAGEALLOC
  1758. if (cachep->buffer_size % PAGE_SIZE == 0 &&
  1759. OFF_SLAB(cachep))
  1760. kernel_map_pages(virt_to_page(objp),
  1761. cachep->buffer_size / PAGE_SIZE, 1);
  1762. else
  1763. check_poison_obj(cachep, objp);
  1764. #else
  1765. check_poison_obj(cachep, objp);
  1766. #endif
  1767. }
  1768. if (cachep->flags & SLAB_RED_ZONE) {
  1769. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
  1770. slab_error(cachep, "start of a freed object "
  1771. "was overwritten");
  1772. if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
  1773. slab_error(cachep, "end of a freed object "
  1774. "was overwritten");
  1775. }
  1776. }
  1777. }
  1778. #else
  1779. static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
  1780. {
  1781. }
  1782. #endif
  1783. /**
  1784. * slab_destroy - destroy and release all objects in a slab
  1785. * @cachep: cache pointer being destroyed
  1786. * @slabp: slab pointer being destroyed
  1787. *
  1788. * Destroy all the objs in a slab, and release the mem back to the system.
  1789. * Before calling the slab must have been unlinked from the cache. The
  1790. * cache-lock is not held/needed.
  1791. */
  1792. static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
  1793. {
  1794. void *addr = slabp->s_mem - slabp->colouroff;
  1795. slab_destroy_debugcheck(cachep, slabp);
  1796. if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
  1797. struct slab_rcu *slab_rcu;
  1798. slab_rcu = (struct slab_rcu *)slabp;
  1799. slab_rcu->cachep = cachep;
  1800. slab_rcu->addr = addr;
  1801. call_rcu(&slab_rcu->head, kmem_rcu_free);
  1802. } else {
  1803. kmem_freepages(cachep, addr);
  1804. if (OFF_SLAB(cachep))
  1805. kmem_cache_free(cachep->slabp_cache, slabp);
  1806. }
  1807. }
  1808. static void __kmem_cache_destroy(struct kmem_cache *cachep)
  1809. {
  1810. int i;
  1811. struct kmem_list3 *l3;
  1812. for_each_online_cpu(i)
  1813. kfree(cachep->array[i]);
  1814. /* NUMA: free the list3 structures */
  1815. for_each_online_node(i) {
  1816. l3 = cachep->nodelists[i];
  1817. if (l3) {
  1818. kfree(l3->shared);
  1819. free_alien_cache(l3->alien);
  1820. kfree(l3);
  1821. }
  1822. }
  1823. kmem_cache_free(&cache_cache, cachep);
  1824. }
  1825. /**
  1826. * calculate_slab_order - calculate size (page order) of slabs
  1827. * @cachep: pointer to the cache that is being created
  1828. * @size: size of objects to be created in this cache.
  1829. * @align: required alignment for the objects.
  1830. * @flags: slab allocation flags
  1831. *
  1832. * Also calculates the number of objects per slab.
  1833. *
  1834. * This could be made much more intelligent. For now, try to avoid using
  1835. * high order pages for slabs. When the gfp() functions are more friendly
  1836. * towards high-order requests, this should be changed.
  1837. */
  1838. static size_t calculate_slab_order(struct kmem_cache *cachep,
  1839. size_t size, size_t align, unsigned long flags)
  1840. {
  1841. unsigned long offslab_limit;
  1842. size_t left_over = 0;
  1843. int gfporder;
  1844. for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
  1845. unsigned int num;
  1846. size_t remainder;
  1847. cache_estimate(gfporder, size, align, flags, &remainder, &num);
  1848. if (!num)
  1849. continue;
  1850. if (flags & CFLGS_OFF_SLAB) {
  1851. /*
  1852. * Max number of objs-per-slab for caches which
  1853. * use off-slab slabs. Needed to avoid a possible
  1854. * looping condition in cache_grow().
  1855. */
  1856. offslab_limit = size - sizeof(struct slab);
  1857. offslab_limit /= sizeof(kmem_bufctl_t);
  1858. if (num > offslab_limit)
  1859. break;
  1860. }
  1861. /* Found something acceptable - save it away */
  1862. cachep->num = num;
  1863. cachep->gfporder = gfporder;
  1864. left_over = remainder;
  1865. /*
  1866. * A VFS-reclaimable slab tends to have most allocations
  1867. * as GFP_NOFS and we really don't want to have to be allocating
  1868. * higher-order pages when we are unable to shrink dcache.
  1869. */
  1870. if (flags & SLAB_RECLAIM_ACCOUNT)
  1871. break;
  1872. /*
  1873. * Large number of objects is good, but very large slabs are
  1874. * currently bad for the gfp()s.
  1875. */
  1876. if (gfporder >= slab_max_order)
  1877. break;
  1878. /*
  1879. * Acceptable internal fragmentation?
  1880. */
  1881. if (left_over * 8 <= (PAGE_SIZE << gfporder))
  1882. break;
  1883. }
  1884. return left_over;
  1885. }
  1886. static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
  1887. {
  1888. if (g_cpucache_up == FULL)
  1889. return enable_cpucache(cachep, gfp);
  1890. if (g_cpucache_up == NONE) {
  1891. /*
  1892. * Note: the first kmem_cache_create must create the cache
  1893. * that's used by kmalloc(24), otherwise the creation of
  1894. * further caches will BUG().
  1895. */
  1896. cachep->array[smp_processor_id()] = &initarray_generic.cache;
  1897. /*
  1898. * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
  1899. * the first cache, then we need to set up all its list3s,
  1900. * otherwise the creation of further caches will BUG().
  1901. */
  1902. set_up_list3s(cachep, SIZE_AC);
  1903. if (INDEX_AC == INDEX_L3)
  1904. g_cpucache_up = PARTIAL_L3;
  1905. else
  1906. g_cpucache_up = PARTIAL_AC;
  1907. } else {
  1908. cachep->array[smp_processor_id()] =
  1909. kmalloc(sizeof(struct arraycache_init), gfp);
  1910. if (g_cpucache_up == PARTIAL_AC) {
  1911. set_up_list3s(cachep, SIZE_L3);
  1912. g_cpucache_up = PARTIAL_L3;
  1913. } else {
  1914. int node;
  1915. for_each_online_node(node) {
  1916. cachep->nodelists[node] =
  1917. kmalloc_node(sizeof(struct kmem_list3),
  1918. gfp, node);
  1919. BUG_ON(!cachep->nodelists[node]);
  1920. kmem_list3_init(cachep->nodelists[node]);
  1921. }
  1922. }
  1923. }
  1924. cachep->nodelists[numa_mem_id()]->next_reap =
  1925. jiffies + REAPTIMEOUT_LIST3 +
  1926. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  1927. cpu_cache_get(cachep)->avail = 0;
  1928. cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
  1929. cpu_cache_get(cachep)->batchcount = 1;
  1930. cpu_cache_get(cachep)->touched = 0;
  1931. cachep->batchcount = 1;
  1932. cachep->limit = BOOT_CPUCACHE_ENTRIES;
  1933. return 0;
  1934. }
  1935. /**
  1936. * kmem_cache_create - Create a cache.
  1937. * @name: A string which is used in /proc/slabinfo to identify this cache.
  1938. * @size: The size of objects to be created in this cache.
  1939. * @align: The required alignment for the objects.
  1940. * @flags: SLAB flags
  1941. * @ctor: A constructor for the objects.
  1942. *
  1943. * Returns a ptr to the cache on success, NULL on failure.
  1944. * Cannot be called within a int, but can be interrupted.
  1945. * The @ctor is run when new pages are allocated by the cache.
  1946. *
  1947. * @name must be valid until the cache is destroyed. This implies that
  1948. * the module calling this has to destroy the cache before getting unloaded.
  1949. *
  1950. * The flags are
  1951. *
  1952. * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
  1953. * to catch references to uninitialised memory.
  1954. *
  1955. * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
  1956. * for buffer overruns.
  1957. *
  1958. * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
  1959. * cacheline. This can be beneficial if you're counting cycles as closely
  1960. * as davem.
  1961. */
  1962. struct kmem_cache *
  1963. kmem_cache_create (const char *name, size_t size, size_t align,
  1964. unsigned long flags, void (*ctor)(void *))
  1965. {
  1966. size_t left_over, slab_size, ralign;
  1967. struct kmem_cache *cachep = NULL, *pc;
  1968. gfp_t gfp;
  1969. /*
  1970. * Sanity checks... these are all serious usage bugs.
  1971. */
  1972. if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
  1973. size > KMALLOC_MAX_SIZE) {
  1974. printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
  1975. name);
  1976. BUG();
  1977. }
  1978. /*
  1979. * We use cache_chain_mutex to ensure a consistent view of
  1980. * cpu_online_mask as well. Please see cpuup_callback
  1981. */
  1982. if (slab_is_available()) {
  1983. get_online_cpus();
  1984. mutex_lock(&cache_chain_mutex);
  1985. }
  1986. list_for_each_entry(pc, &cache_chain, next) {
  1987. char tmp;
  1988. int res;
  1989. /*
  1990. * This happens when the module gets unloaded and doesn't
  1991. * destroy its slab cache and no-one else reuses the vmalloc
  1992. * area of the module. Print a warning.
  1993. */
  1994. res = probe_kernel_address(pc->name, tmp);
  1995. if (res) {
  1996. printk(KERN_ERR
  1997. "SLAB: cache with size %d has lost its name\n",
  1998. pc->buffer_size);
  1999. continue;
  2000. }
  2001. if (!strcmp(pc->name, name)) {
  2002. printk(KERN_ERR
  2003. "kmem_cache_create: duplicate cache %s\n", name);
  2004. dump_stack();
  2005. goto oops;
  2006. }
  2007. }
  2008. #if DEBUG
  2009. WARN_ON(strchr(name, ' ')); /* It confuses parsers */
  2010. #if FORCED_DEBUG
  2011. /*
  2012. * Enable redzoning and last user accounting, except for caches with
  2013. * large objects, if the increased size would increase the object size
  2014. * above the next power of two: caches with object sizes just above a
  2015. * power of two have a significant amount of internal fragmentation.
  2016. */
  2017. if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
  2018. 2 * sizeof(unsigned long long)))
  2019. flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
  2020. if (!(flags & SLAB_DESTROY_BY_RCU))
  2021. flags |= SLAB_POISON;
  2022. #endif
  2023. if (flags & SLAB_DESTROY_BY_RCU)
  2024. BUG_ON(flags & SLAB_POISON);
  2025. #endif
  2026. /*
  2027. * Always checks flags, a caller might be expecting debug support which
  2028. * isn't available.
  2029. */
  2030. BUG_ON(flags & ~CREATE_MASK);
  2031. /*
  2032. * Check that size is in terms of words. This is needed to avoid
  2033. * unaligned accesses for some archs when redzoning is used, and makes
  2034. * sure any on-slab bufctl's are also correctly aligned.
  2035. */
  2036. if (size & (BYTES_PER_WORD - 1)) {
  2037. size += (BYTES_PER_WORD - 1);
  2038. size &= ~(BYTES_PER_WORD - 1);
  2039. }
  2040. /* calculate the final buffer alignment: */
  2041. /* 1) arch recommendation: can be overridden for debug */
  2042. if (flags & SLAB_HWCACHE_ALIGN) {
  2043. /*
  2044. * Default alignment: as specified by the arch code. Except if
  2045. * an object is really small, then squeeze multiple objects into
  2046. * one cacheline.
  2047. */
  2048. ralign = cache_line_size();
  2049. while (size <= ralign / 2)
  2050. ralign /= 2;
  2051. } else {
  2052. ralign = BYTES_PER_WORD;
  2053. }
  2054. /*
  2055. * Redzoning and user store require word alignment or possibly larger.
  2056. * Note this will be overridden by architecture or caller mandated
  2057. * alignment if either is greater than BYTES_PER_WORD.
  2058. */
  2059. if (flags & SLAB_STORE_USER)
  2060. ralign = BYTES_PER_WORD;
  2061. if (flags & SLAB_RED_ZONE) {
  2062. ralign = REDZONE_ALIGN;
  2063. /* If redzoning, ensure that the second redzone is suitably
  2064. * aligned, by adjusting the object size accordingly. */
  2065. size += REDZONE_ALIGN - 1;
  2066. size &= ~(REDZONE_ALIGN - 1);
  2067. }
  2068. /* 2) arch mandated alignment */
  2069. if (ralign < ARCH_SLAB_MINALIGN) {
  2070. ralign = ARCH_SLAB_MINALIGN;
  2071. }
  2072. /* 3) caller mandated alignment */
  2073. if (ralign < align) {
  2074. ralign = align;
  2075. }
  2076. /* disable debug if necessary */
  2077. if (ralign > __alignof__(unsigned long long))
  2078. flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
  2079. /*
  2080. * 4) Store it.
  2081. */
  2082. align = ralign;
  2083. if (slab_is_available())
  2084. gfp = GFP_KERNEL;
  2085. else
  2086. gfp = GFP_NOWAIT;
  2087. /* Get cache's description obj. */
  2088. cachep = kmem_cache_zalloc(&cache_cache, gfp);
  2089. if (!cachep)
  2090. goto oops;
  2091. cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
  2092. #if DEBUG
  2093. cachep->obj_size = size;
  2094. /*
  2095. * Both debugging options require word-alignment which is calculated
  2096. * into align above.
  2097. */
  2098. if (flags & SLAB_RED_ZONE) {
  2099. /* add space for red zone words */
  2100. cachep->obj_offset += sizeof(unsigned long long);
  2101. size += 2 * sizeof(unsigned long long);
  2102. }
  2103. if (flags & SLAB_STORE_USER) {
  2104. /* user store requires one word storage behind the end of
  2105. * the real object. But if the second red zone needs to be
  2106. * aligned to 64 bits, we must allow that much space.
  2107. */
  2108. if (flags & SLAB_RED_ZONE)
  2109. size += REDZONE_ALIGN;
  2110. else
  2111. size += BYTES_PER_WORD;
  2112. }
  2113. #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
  2114. if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
  2115. && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
  2116. cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
  2117. size = PAGE_SIZE;
  2118. }
  2119. #endif
  2120. #endif
  2121. /*
  2122. * Determine if the slab management is 'on' or 'off' slab.
  2123. * (bootstrapping cannot cope with offslab caches so don't do
  2124. * it too early on. Always use on-slab management when
  2125. * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
  2126. */
  2127. if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
  2128. !(flags & SLAB_NOLEAKTRACE))
  2129. /*
  2130. * Size is large, assume best to place the slab management obj
  2131. * off-slab (should allow better packing of objs).
  2132. */
  2133. flags |= CFLGS_OFF_SLAB;
  2134. size = ALIGN(size, align);
  2135. left_over = calculate_slab_order(cachep, size, align, flags);
  2136. if (!cachep->num) {
  2137. printk(KERN_ERR
  2138. "kmem_cache_create: couldn't create cache %s.\n", name);
  2139. kmem_cache_free(&cache_cache, cachep);
  2140. cachep = NULL;
  2141. goto oops;
  2142. }
  2143. slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
  2144. + sizeof(struct slab), align);
  2145. /*
  2146. * If the slab has been placed off-slab, and we have enough space then
  2147. * move it on-slab. This is at the expense of any extra colouring.
  2148. */
  2149. if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
  2150. flags &= ~CFLGS_OFF_SLAB;
  2151. left_over -= slab_size;
  2152. }
  2153. if (flags & CFLGS_OFF_SLAB) {
  2154. /* really off slab. No need for manual alignment */
  2155. slab_size =
  2156. cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
  2157. #ifdef CONFIG_PAGE_POISONING
  2158. /* If we're going to use the generic kernel_map_pages()
  2159. * poisoning, then it's going to smash the contents of
  2160. * the redzone and userword anyhow, so switch them off.
  2161. */
  2162. if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
  2163. flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
  2164. #endif
  2165. }
  2166. cachep->colour_off = cache_line_size();
  2167. /* Offset must be a multiple of the alignment. */
  2168. if (cachep->colour_off < align)
  2169. cachep->colour_off = align;
  2170. cachep->colour = left_over / cachep->colour_off;
  2171. cachep->slab_size = slab_size;
  2172. cachep->flags = flags;
  2173. cachep->gfpflags = 0;
  2174. if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
  2175. cachep->gfpflags |= GFP_DMA;
  2176. cachep->buffer_size = size;
  2177. cachep->reciprocal_buffer_size = reciprocal_value(size);
  2178. if (flags & CFLGS_OFF_SLAB) {
  2179. cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
  2180. /*
  2181. * This is a possibility for one of the malloc_sizes caches.
  2182. * But since we go off slab only for object size greater than
  2183. * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
  2184. * this should not happen at all.
  2185. * But leave a BUG_ON for some lucky dude.
  2186. */
  2187. BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
  2188. }
  2189. cachep->ctor = ctor;
  2190. cachep->name = name;
  2191. if (setup_cpu_cache(cachep, gfp)) {
  2192. __kmem_cache_destroy(cachep);
  2193. cachep = NULL;
  2194. goto oops;
  2195. }
  2196. if (flags & SLAB_DEBUG_OBJECTS) {
  2197. /*
  2198. * Would deadlock through slab_destroy()->call_rcu()->
  2199. * debug_object_activate()->kmem_cache_alloc().
  2200. */
  2201. WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
  2202. slab_set_debugobj_lock_classes(cachep);
  2203. }
  2204. /* cache setup completed, link it into the list */
  2205. list_add(&cachep->next, &cache_chain);
  2206. oops:
  2207. if (!cachep && (flags & SLAB_PANIC))
  2208. panic("kmem_cache_create(): failed to create slab `%s'\n",
  2209. name);
  2210. if (slab_is_available()) {
  2211. mutex_unlock(&cache_chain_mutex);
  2212. put_online_cpus();
  2213. }
  2214. return cachep;
  2215. }
  2216. EXPORT_SYMBOL(kmem_cache_create);
  2217. #if DEBUG
  2218. static void check_irq_off(void)
  2219. {
  2220. BUG_ON(!irqs_disabled());
  2221. }
  2222. static void check_irq_on(void)
  2223. {
  2224. BUG_ON(irqs_disabled());
  2225. }
  2226. static void check_spinlock_acquired(struct kmem_cache *cachep)
  2227. {
  2228. #ifdef CONFIG_SMP
  2229. check_irq_off();
  2230. assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
  2231. #endif
  2232. }
  2233. static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
  2234. {
  2235. #ifdef CONFIG_SMP
  2236. check_irq_off();
  2237. assert_spin_locked(&cachep->nodelists[node]->list_lock);
  2238. #endif
  2239. }
  2240. #else
  2241. #define check_irq_off() do { } while(0)
  2242. #define check_irq_on() do { } while(0)
  2243. #define check_spinlock_acquired(x) do { } while(0)
  2244. #define check_spinlock_acquired_node(x, y) do { } while(0)
  2245. #endif
  2246. static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
  2247. struct array_cache *ac,
  2248. int force, int node);
  2249. static void do_drain(void *arg)
  2250. {
  2251. struct kmem_cache *cachep = arg;
  2252. struct array_cache *ac;
  2253. int node = numa_mem_id();
  2254. check_irq_off();
  2255. ac = cpu_cache_get(cachep);
  2256. spin_lock(&cachep->nodelists[node]->list_lock);
  2257. free_block(cachep, ac->entry, ac->avail, node);
  2258. spin_unlock(&cachep->nodelists[node]->list_lock);
  2259. ac->avail = 0;
  2260. }
  2261. static void drain_cpu_caches(struct kmem_cache *cachep)
  2262. {
  2263. struct kmem_list3 *l3;
  2264. int node;
  2265. on_each_cpu(do_drain, cachep, 1);
  2266. check_irq_on();
  2267. for_each_online_node(node) {
  2268. l3 = cachep->nodelists[node];
  2269. if (l3 && l3->alien)
  2270. drain_alien_cache(cachep, l3->alien);
  2271. }
  2272. for_each_online_node(node) {
  2273. l3 = cachep->nodelists[node];
  2274. if (l3)
  2275. drain_array(cachep, l3, l3->shared, 1, node);
  2276. }
  2277. }
  2278. /*
  2279. * Remove slabs from the list of free slabs.
  2280. * Specify the number of slabs to drain in tofree.
  2281. *
  2282. * Returns the actual number of slabs released.
  2283. */
  2284. static int drain_freelist(struct kmem_cache *cache,
  2285. struct kmem_list3 *l3, int tofree)
  2286. {
  2287. struct list_head *p;
  2288. int nr_freed;
  2289. struct slab *slabp;
  2290. nr_freed = 0;
  2291. while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
  2292. spin_lock_irq(&l3->list_lock);
  2293. p = l3->slabs_free.prev;
  2294. if (p == &l3->slabs_free) {
  2295. spin_unlock_irq(&l3->list_lock);
  2296. goto out;
  2297. }
  2298. slabp = list_entry(p, struct slab, list);
  2299. #if DEBUG
  2300. BUG_ON(slabp->inuse);
  2301. #endif
  2302. list_del(&slabp->list);
  2303. /*
  2304. * Safe to drop the lock. The slab is no longer linked
  2305. * to the cache.
  2306. */
  2307. l3->free_objects -= cache->num;
  2308. spin_unlock_irq(&l3->list_lock);
  2309. slab_destroy(cache, slabp);
  2310. nr_freed++;
  2311. }
  2312. out:
  2313. return nr_freed;
  2314. }
  2315. /* Called with cache_chain_mutex held to protect against cpu hotplug */
  2316. static int __cache_shrink(struct kmem_cache *cachep)
  2317. {
  2318. int ret = 0, i = 0;
  2319. struct kmem_list3 *l3;
  2320. drain_cpu_caches(cachep);
  2321. check_irq_on();
  2322. for_each_online_node(i) {
  2323. l3 = cachep->nodelists[i];
  2324. if (!l3)
  2325. continue;
  2326. drain_freelist(cachep, l3, l3->free_objects);
  2327. ret += !list_empty(&l3->slabs_full) ||
  2328. !list_empty(&l3->slabs_partial);
  2329. }
  2330. return (ret ? 1 : 0);
  2331. }
  2332. /**
  2333. * kmem_cache_shrink - Shrink a cache.
  2334. * @cachep: The cache to shrink.
  2335. *
  2336. * Releases as many slabs as possible for a cache.
  2337. * To help debugging, a zero exit status indicates all slabs were released.
  2338. */
  2339. int kmem_cache_shrink(struct kmem_cache *cachep)
  2340. {
  2341. int ret;
  2342. BUG_ON(!cachep || in_interrupt());
  2343. get_online_cpus();
  2344. mutex_lock(&cache_chain_mutex);
  2345. ret = __cache_shrink(cachep);
  2346. mutex_unlock(&cache_chain_mutex);
  2347. put_online_cpus();
  2348. return ret;
  2349. }
  2350. EXPORT_SYMBOL(kmem_cache_shrink);
  2351. /**
  2352. * kmem_cache_destroy - delete a cache
  2353. * @cachep: the cache to destroy
  2354. *
  2355. * Remove a &struct kmem_cache object from the slab cache.
  2356. *
  2357. * It is expected this function will be called by a module when it is
  2358. * unloaded. This will remove the cache completely, and avoid a duplicate
  2359. * cache being allocated each time a module is loaded and unloaded, if the
  2360. * module doesn't have persistent in-kernel storage across loads and unloads.
  2361. *
  2362. * The cache must be empty before calling this function.
  2363. *
  2364. * The caller must guarantee that no one will allocate memory from the cache
  2365. * during the kmem_cache_destroy().
  2366. */
  2367. void kmem_cache_destroy(struct kmem_cache *cachep)
  2368. {
  2369. BUG_ON(!cachep || in_interrupt());
  2370. /* Find the cache in the chain of caches. */
  2371. get_online_cpus();
  2372. mutex_lock(&cache_chain_mutex);
  2373. /*
  2374. * the chain is never empty, cache_cache is never destroyed
  2375. */
  2376. list_del(&cachep->next);
  2377. if (__cache_shrink(cachep)) {
  2378. slab_error(cachep, "Can't free all objects");
  2379. list_add(&cachep->next, &cache_chain);
  2380. mutex_unlock(&cache_chain_mutex);
  2381. put_online_cpus();
  2382. return;
  2383. }
  2384. if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
  2385. rcu_barrier();
  2386. __kmem_cache_destroy(cachep);
  2387. mutex_unlock(&cache_chain_mutex);
  2388. put_online_cpus();
  2389. }
  2390. EXPORT_SYMBOL(kmem_cache_destroy);
  2391. /*
  2392. * Get the memory for a slab management obj.
  2393. * For a slab cache when the slab descriptor is off-slab, slab descriptors
  2394. * always come from malloc_sizes caches. The slab descriptor cannot
  2395. * come from the same cache which is getting created because,
  2396. * when we are searching for an appropriate cache for these
  2397. * descriptors in kmem_cache_create, we search through the malloc_sizes array.
  2398. * If we are creating a malloc_sizes cache here it would not be visible to
  2399. * kmem_find_general_cachep till the initialization is complete.
  2400. * Hence we cannot have slabp_cache same as the original cache.
  2401. */
  2402. static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
  2403. int colour_off, gfp_t local_flags,
  2404. int nodeid)
  2405. {
  2406. struct slab *slabp;
  2407. if (OFF_SLAB(cachep)) {
  2408. /* Slab management obj is off-slab. */
  2409. slabp = kmem_cache_alloc_node(cachep->slabp_cache,
  2410. local_flags, nodeid);
  2411. /*
  2412. * If the first object in the slab is leaked (it's allocated
  2413. * but no one has a reference to it), we want to make sure
  2414. * kmemleak does not treat the ->s_mem pointer as a reference
  2415. * to the object. Otherwise we will not report the leak.
  2416. */
  2417. kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
  2418. local_flags);
  2419. if (!slabp)
  2420. return NULL;
  2421. } else {
  2422. slabp = objp + colour_off;
  2423. colour_off += cachep->slab_size;
  2424. }
  2425. slabp->inuse = 0;
  2426. slabp->colouroff = colour_off;
  2427. slabp->s_mem = objp + colour_off;
  2428. slabp->nodeid = nodeid;
  2429. slabp->free = 0;
  2430. return slabp;
  2431. }
  2432. static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
  2433. {
  2434. return (kmem_bufctl_t *) (slabp + 1);
  2435. }
  2436. static void cache_init_objs(struct kmem_cache *cachep,
  2437. struct slab *slabp)
  2438. {
  2439. int i;
  2440. for (i = 0; i < cachep->num; i++) {
  2441. void *objp = index_to_obj(cachep, slabp, i);
  2442. #if DEBUG
  2443. /* need to poison the objs? */
  2444. if (cachep->flags & SLAB_POISON)
  2445. poison_obj(cachep, objp, POISON_FREE);
  2446. if (cachep->flags & SLAB_STORE_USER)
  2447. *dbg_userword(cachep, objp) = NULL;
  2448. if (cachep->flags & SLAB_RED_ZONE) {
  2449. *dbg_redzone1(cachep, objp) = RED_INACTIVE;
  2450. *dbg_redzone2(cachep, objp) = RED_INACTIVE;
  2451. }
  2452. /*
  2453. * Constructors are not allowed to allocate memory from the same
  2454. * cache which they are a constructor for. Otherwise, deadlock.
  2455. * They must also be threaded.
  2456. */
  2457. if (cachep->ctor && !(cachep->flags & SLAB_POISON))
  2458. cachep->ctor(objp + obj_offset(cachep));
  2459. if (cachep->flags & SLAB_RED_ZONE) {
  2460. if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
  2461. slab_error(cachep, "constructor overwrote the"
  2462. " end of an object");
  2463. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
  2464. slab_error(cachep, "constructor overwrote the"
  2465. " start of an object");
  2466. }
  2467. if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
  2468. OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
  2469. kernel_map_pages(virt_to_page(objp),
  2470. cachep->buffer_size / PAGE_SIZE, 0);
  2471. #else
  2472. if (cachep->ctor)
  2473. cachep->ctor(objp);
  2474. #endif
  2475. slab_bufctl(slabp)[i] = i + 1;
  2476. }
  2477. slab_bufctl(slabp)[i - 1] = BUFCTL_END;
  2478. }
  2479. static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
  2480. {
  2481. if (CONFIG_ZONE_DMA_FLAG) {
  2482. if (flags & GFP_DMA)
  2483. BUG_ON(!(cachep->gfpflags & GFP_DMA));
  2484. else
  2485. BUG_ON(cachep->gfpflags & GFP_DMA);
  2486. }
  2487. }
  2488. static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
  2489. int nodeid)
  2490. {
  2491. void *objp = index_to_obj(cachep, slabp, slabp->free);
  2492. kmem_bufctl_t next;
  2493. slabp->inuse++;
  2494. next = slab_bufctl(slabp)[slabp->free];
  2495. #if DEBUG
  2496. slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
  2497. WARN_ON(slabp->nodeid != nodeid);
  2498. #endif
  2499. slabp->free = next;
  2500. return objp;
  2501. }
  2502. static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
  2503. void *objp, int nodeid)
  2504. {
  2505. unsigned int objnr = obj_to_index(cachep, slabp, objp);
  2506. #if DEBUG
  2507. /* Verify that the slab belongs to the intended node */
  2508. WARN_ON(slabp->nodeid != nodeid);
  2509. if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
  2510. printk(KERN_ERR "slab: double free detected in cache "
  2511. "'%s', objp %p\n", cachep->name, objp);
  2512. BUG();
  2513. }
  2514. #endif
  2515. slab_bufctl(slabp)[objnr] = slabp->free;
  2516. slabp->free = objnr;
  2517. slabp->inuse--;
  2518. }
  2519. /*
  2520. * Map pages beginning at addr to the given cache and slab. This is required
  2521. * for the slab allocator to be able to lookup the cache and slab of a
  2522. * virtual address for kfree, ksize, and slab debugging.
  2523. */
  2524. static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
  2525. void *addr)
  2526. {
  2527. int nr_pages;
  2528. struct page *page;
  2529. page = virt_to_page(addr);
  2530. nr_pages = 1;
  2531. if (likely(!PageCompound(page)))
  2532. nr_pages <<= cache->gfporder;
  2533. do {
  2534. page_set_cache(page, cache);
  2535. page_set_slab(page, slab);
  2536. page++;
  2537. } while (--nr_pages);
  2538. }
  2539. /*
  2540. * Grow (by 1) the number of slabs within a cache. This is called by
  2541. * kmem_cache_alloc() when there are no active objs left in a cache.
  2542. */
  2543. static int cache_grow(struct kmem_cache *cachep,
  2544. gfp_t flags, int nodeid, void *objp)
  2545. {
  2546. struct slab *slabp;
  2547. size_t offset;
  2548. gfp_t local_flags;
  2549. struct kmem_list3 *l3;
  2550. /*
  2551. * Be lazy and only check for valid flags here, keeping it out of the
  2552. * critical path in kmem_cache_alloc().
  2553. */
  2554. BUG_ON(flags & GFP_SLAB_BUG_MASK);
  2555. local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
  2556. /* Take the l3 list lock to change the colour_next on this node */
  2557. check_irq_off();
  2558. l3 = cachep->nodelists[nodeid];
  2559. spin_lock(&l3->list_lock);
  2560. /* Get colour for the slab, and cal the next value. */
  2561. offset = l3->colour_next;
  2562. l3->colour_next++;
  2563. if (l3->colour_next >= cachep->colour)
  2564. l3->colour_next = 0;
  2565. spin_unlock(&l3->list_lock);
  2566. offset *= cachep->colour_off;
  2567. if (local_flags & __GFP_WAIT)
  2568. local_irq_enable();
  2569. /*
  2570. * The test for missing atomic flag is performed here, rather than
  2571. * the more obvious place, simply to reduce the critical path length
  2572. * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
  2573. * will eventually be caught here (where it matters).
  2574. */
  2575. kmem_flagcheck(cachep, flags);
  2576. /*
  2577. * Get mem for the objs. Attempt to allocate a physical page from
  2578. * 'nodeid'.
  2579. */
  2580. if (!objp)
  2581. objp = kmem_getpages(cachep, local_flags, nodeid);
  2582. if (!objp)
  2583. goto failed;
  2584. /* Get slab management. */
  2585. slabp = alloc_slabmgmt(cachep, objp, offset,
  2586. local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
  2587. if (!slabp)
  2588. goto opps1;
  2589. slab_map_pages(cachep, slabp, objp);
  2590. cache_init_objs(cachep, slabp);
  2591. if (local_flags & __GFP_WAIT)
  2592. local_irq_disable();
  2593. check_irq_off();
  2594. spin_lock(&l3->list_lock);
  2595. /* Make slab active. */
  2596. list_add_tail(&slabp->list, &(l3->slabs_free));
  2597. STATS_INC_GROWN(cachep);
  2598. l3->free_objects += cachep->num;
  2599. spin_unlock(&l3->list_lock);
  2600. return 1;
  2601. opps1:
  2602. kmem_freepages(cachep, objp);
  2603. failed:
  2604. if (local_flags & __GFP_WAIT)
  2605. local_irq_disable();
  2606. return 0;
  2607. }
  2608. #if DEBUG
  2609. /*
  2610. * Perform extra freeing checks:
  2611. * - detect bad pointers.
  2612. * - POISON/RED_ZONE checking
  2613. */
  2614. static void kfree_debugcheck(const void *objp)
  2615. {
  2616. if (!virt_addr_valid(objp)) {
  2617. printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
  2618. (unsigned long)objp);
  2619. BUG();
  2620. }
  2621. }
  2622. static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
  2623. {
  2624. unsigned long long redzone1, redzone2;
  2625. redzone1 = *dbg_redzone1(cache, obj);
  2626. redzone2 = *dbg_redzone2(cache, obj);
  2627. /*
  2628. * Redzone is ok.
  2629. */
  2630. if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
  2631. return;
  2632. if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
  2633. slab_error(cache, "double free detected");
  2634. else
  2635. slab_error(cache, "memory outside object was overwritten");
  2636. printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
  2637. obj, redzone1, redzone2);
  2638. }
  2639. static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
  2640. void *caller)
  2641. {
  2642. struct page *page;
  2643. unsigned int objnr;
  2644. struct slab *slabp;
  2645. BUG_ON(virt_to_cache(objp) != cachep);
  2646. objp -= obj_offset(cachep);
  2647. kfree_debugcheck(objp);
  2648. page = virt_to_head_page(objp);
  2649. slabp = page_get_slab(page);
  2650. if (cachep->flags & SLAB_RED_ZONE) {
  2651. verify_redzone_free(cachep, objp);
  2652. *dbg_redzone1(cachep, objp) = RED_INACTIVE;
  2653. *dbg_redzone2(cachep, objp) = RED_INACTIVE;
  2654. }
  2655. if (cachep->flags & SLAB_STORE_USER)
  2656. *dbg_userword(cachep, objp) = caller;
  2657. objnr = obj_to_index(cachep, slabp, objp);
  2658. BUG_ON(objnr >= cachep->num);
  2659. BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
  2660. #ifdef CONFIG_DEBUG_SLAB_LEAK
  2661. slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
  2662. #endif
  2663. if (cachep->flags & SLAB_POISON) {
  2664. #ifdef CONFIG_DEBUG_PAGEALLOC
  2665. if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
  2666. store_stackinfo(cachep, objp, (unsigned long)caller);
  2667. kernel_map_pages(virt_to_page(objp),
  2668. cachep->buffer_size / PAGE_SIZE, 0);
  2669. } else {
  2670. poison_obj(cachep, objp, POISON_FREE);
  2671. }
  2672. #else
  2673. poison_obj(cachep, objp, POISON_FREE);
  2674. #endif
  2675. }
  2676. return objp;
  2677. }
  2678. static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
  2679. {
  2680. kmem_bufctl_t i;
  2681. int entries = 0;
  2682. /* Check slab's freelist to see if this obj is there. */
  2683. for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
  2684. entries++;
  2685. if (entries > cachep->num || i >= cachep->num)
  2686. goto bad;
  2687. }
  2688. if (entries != cachep->num - slabp->inuse) {
  2689. bad:
  2690. printk(KERN_ERR "slab: Internal list corruption detected in "
  2691. "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
  2692. cachep->name, cachep->num, slabp, slabp->inuse,
  2693. print_tainted());
  2694. print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
  2695. sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
  2696. 1);
  2697. BUG();
  2698. }
  2699. }
  2700. #else
  2701. #define kfree_debugcheck(x) do { } while(0)
  2702. #define cache_free_debugcheck(x,objp,z) (objp)
  2703. #define check_slabp(x,y) do { } while(0)
  2704. #endif
  2705. static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
  2706. {
  2707. int batchcount;
  2708. struct kmem_list3 *l3;
  2709. struct array_cache *ac;
  2710. int node;
  2711. retry:
  2712. check_irq_off();
  2713. node = numa_mem_id();
  2714. ac = cpu_cache_get(cachep);
  2715. batchcount = ac->batchcount;
  2716. if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
  2717. /*
  2718. * If there was little recent activity on this cache, then
  2719. * perform only a partial refill. Otherwise we could generate
  2720. * refill bouncing.
  2721. */
  2722. batchcount = BATCHREFILL_LIMIT;
  2723. }
  2724. l3 = cachep->nodelists[node];
  2725. BUG_ON(ac->avail > 0 || !l3);
  2726. spin_lock(&l3->list_lock);
  2727. /* See if we can refill from the shared array */
  2728. if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
  2729. l3->shared->touched = 1;
  2730. goto alloc_done;
  2731. }
  2732. while (batchcount > 0) {
  2733. struct list_head *entry;
  2734. struct slab *slabp;
  2735. /* Get slab alloc is to come from. */
  2736. entry = l3->slabs_partial.next;
  2737. if (entry == &l3->slabs_partial) {
  2738. l3->free_touched = 1;
  2739. entry = l3->slabs_free.next;
  2740. if (entry == &l3->slabs_free)
  2741. goto must_grow;
  2742. }
  2743. slabp = list_entry(entry, struct slab, list);
  2744. check_slabp(cachep, slabp);
  2745. check_spinlock_acquired(cachep);
  2746. /*
  2747. * The slab was either on partial or free list so
  2748. * there must be at least one object available for
  2749. * allocation.
  2750. */
  2751. BUG_ON(slabp->inuse >= cachep->num);
  2752. while (slabp->inuse < cachep->num && batchcount--) {
  2753. STATS_INC_ALLOCED(cachep);
  2754. STATS_INC_ACTIVE(cachep);
  2755. STATS_SET_HIGH(cachep);
  2756. ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
  2757. node);
  2758. }
  2759. check_slabp(cachep, slabp);
  2760. /* move slabp to correct slabp list: */
  2761. list_del(&slabp->list);
  2762. if (slabp->free == BUFCTL_END)
  2763. list_add(&slabp->list, &l3->slabs_full);
  2764. else
  2765. list_add(&slabp->list, &l3->slabs_partial);
  2766. }
  2767. must_grow:
  2768. l3->free_objects -= ac->avail;
  2769. alloc_done:
  2770. spin_unlock(&l3->list_lock);
  2771. if (unlikely(!ac->avail)) {
  2772. int x;
  2773. x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
  2774. /* cache_grow can reenable interrupts, then ac could change. */
  2775. ac = cpu_cache_get(cachep);
  2776. if (!x && ac->avail == 0) /* no objects in sight? abort */
  2777. return NULL;
  2778. if (!ac->avail) /* objects refilled by interrupt? */
  2779. goto retry;
  2780. }
  2781. ac->touched = 1;
  2782. return ac->entry[--ac->avail];
  2783. }
  2784. static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
  2785. gfp_t flags)
  2786. {
  2787. might_sleep_if(flags & __GFP_WAIT);
  2788. #if DEBUG
  2789. kmem_flagcheck(cachep, flags);
  2790. #endif
  2791. }
  2792. #if DEBUG
  2793. static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
  2794. gfp_t flags, void *objp, void *caller)
  2795. {
  2796. if (!objp)
  2797. return objp;
  2798. if (cachep->flags & SLAB_POISON) {
  2799. #ifdef CONFIG_DEBUG_PAGEALLOC
  2800. if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
  2801. kernel_map_pages(virt_to_page(objp),
  2802. cachep->buffer_size / PAGE_SIZE, 1);
  2803. else
  2804. check_poison_obj(cachep, objp);
  2805. #else
  2806. check_poison_obj(cachep, objp);
  2807. #endif
  2808. poison_obj(cachep, objp, POISON_INUSE);
  2809. }
  2810. if (cachep->flags & SLAB_STORE_USER)
  2811. *dbg_userword(cachep, objp) = caller;
  2812. if (cachep->flags & SLAB_RED_ZONE) {
  2813. if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
  2814. *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
  2815. slab_error(cachep, "double free, or memory outside"
  2816. " object was overwritten");
  2817. printk(KERN_ERR
  2818. "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
  2819. objp, *dbg_redzone1(cachep, objp),
  2820. *dbg_redzone2(cachep, objp));
  2821. }
  2822. *dbg_redzone1(cachep, objp) = RED_ACTIVE;
  2823. *dbg_redzone2(cachep, objp) = RED_ACTIVE;
  2824. }
  2825. #ifdef CONFIG_DEBUG_SLAB_LEAK
  2826. {
  2827. struct slab *slabp;
  2828. unsigned objnr;
  2829. slabp = page_get_slab(virt_to_head_page(objp));
  2830. objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
  2831. slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
  2832. }
  2833. #endif
  2834. objp += obj_offset(cachep);
  2835. if (cachep->ctor && cachep->flags & SLAB_POISON)
  2836. cachep->ctor(objp);
  2837. if (ARCH_SLAB_MINALIGN &&
  2838. ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
  2839. printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
  2840. objp, (int)ARCH_SLAB_MINALIGN);
  2841. }
  2842. return objp;
  2843. }
  2844. #else
  2845. #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
  2846. #endif
  2847. static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
  2848. {
  2849. if (cachep == &cache_cache)
  2850. return false;
  2851. return should_failslab(obj_size(cachep), flags, cachep->flags);
  2852. }
  2853. static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  2854. {
  2855. void *objp;
  2856. struct array_cache *ac;
  2857. check_irq_off();
  2858. ac = cpu_cache_get(cachep);
  2859. if (likely(ac->avail)) {
  2860. STATS_INC_ALLOCHIT(cachep);
  2861. ac->touched = 1;
  2862. objp = ac->entry[--ac->avail];
  2863. } else {
  2864. STATS_INC_ALLOCMISS(cachep);
  2865. objp = cache_alloc_refill(cachep, flags);
  2866. /*
  2867. * the 'ac' may be updated by cache_alloc_refill(),
  2868. * and kmemleak_erase() requires its correct value.
  2869. */
  2870. ac = cpu_cache_get(cachep);
  2871. }
  2872. /*
  2873. * To avoid a false negative, if an object that is in one of the
  2874. * per-CPU caches is leaked, we need to make sure kmemleak doesn't
  2875. * treat the array pointers as a reference to the object.
  2876. */
  2877. if (objp)
  2878. kmemleak_erase(&ac->entry[ac->avail]);
  2879. return objp;
  2880. }
  2881. #ifdef CONFIG_NUMA
  2882. /*
  2883. * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
  2884. *
  2885. * If we are in_interrupt, then process context, including cpusets and
  2886. * mempolicy, may not apply and should not be used for allocation policy.
  2887. */
  2888. static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
  2889. {
  2890. int nid_alloc, nid_here;
  2891. if (in_interrupt() || (flags & __GFP_THISNODE))
  2892. return NULL;
  2893. nid_alloc = nid_here = numa_mem_id();
  2894. get_mems_allowed();
  2895. if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
  2896. nid_alloc = cpuset_slab_spread_node();
  2897. else if (current->mempolicy)
  2898. nid_alloc = slab_node(current->mempolicy);
  2899. put_mems_allowed();
  2900. if (nid_alloc != nid_here)
  2901. return ____cache_alloc_node(cachep, flags, nid_alloc);
  2902. return NULL;
  2903. }
  2904. /*
  2905. * Fallback function if there was no memory available and no objects on a
  2906. * certain node and fall back is permitted. First we scan all the
  2907. * available nodelists for available objects. If that fails then we
  2908. * perform an allocation without specifying a node. This allows the page
  2909. * allocator to do its reclaim / fallback magic. We then insert the
  2910. * slab into the proper nodelist and then allocate from it.
  2911. */
  2912. static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
  2913. {
  2914. struct zonelist *zonelist;
  2915. gfp_t local_flags;
  2916. struct zoneref *z;
  2917. struct zone *zone;
  2918. enum zone_type high_zoneidx = gfp_zone(flags);
  2919. void *obj = NULL;
  2920. int nid;
  2921. if (flags & __GFP_THISNODE)
  2922. return NULL;
  2923. get_mems_allowed();
  2924. zonelist = node_zonelist(slab_node(current->mempolicy), flags);
  2925. local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
  2926. retry:
  2927. /*
  2928. * Look through allowed nodes for objects available
  2929. * from existing per node queues.
  2930. */
  2931. for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
  2932. nid = zone_to_nid(zone);
  2933. if (cpuset_zone_allowed_hardwall(zone, flags) &&
  2934. cache->nodelists[nid] &&
  2935. cache->nodelists[nid]->free_objects) {
  2936. obj = ____cache_alloc_node(cache,
  2937. flags | GFP_THISNODE, nid);
  2938. if (obj)
  2939. break;
  2940. }
  2941. }
  2942. if (!obj) {
  2943. /*
  2944. * This allocation will be performed within the constraints
  2945. * of the current cpuset / memory policy requirements.
  2946. * We may trigger various forms of reclaim on the allowed
  2947. * set and go into memory reserves if necessary.
  2948. */
  2949. if (local_flags & __GFP_WAIT)
  2950. local_irq_enable();
  2951. kmem_flagcheck(cache, flags);
  2952. obj = kmem_getpages(cache, local_flags, numa_mem_id());
  2953. if (local_flags & __GFP_WAIT)
  2954. local_irq_disable();
  2955. if (obj) {
  2956. /*
  2957. * Insert into the appropriate per node queues
  2958. */
  2959. nid = page_to_nid(virt_to_page(obj));
  2960. if (cache_grow(cache, flags, nid, obj)) {
  2961. obj = ____cache_alloc_node(cache,
  2962. flags | GFP_THISNODE, nid);
  2963. if (!obj)
  2964. /*
  2965. * Another processor may allocate the
  2966. * objects in the slab since we are
  2967. * not holding any locks.
  2968. */
  2969. goto retry;
  2970. } else {
  2971. /* cache_grow already freed obj */
  2972. obj = NULL;
  2973. }
  2974. }
  2975. }
  2976. put_mems_allowed();
  2977. return obj;
  2978. }
  2979. /*
  2980. * A interface to enable slab creation on nodeid
  2981. */
  2982. static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
  2983. int nodeid)
  2984. {
  2985. struct list_head *entry;
  2986. struct slab *slabp;
  2987. struct kmem_list3 *l3;
  2988. void *obj;
  2989. int x;
  2990. l3 = cachep->nodelists[nodeid];
  2991. BUG_ON(!l3);
  2992. retry:
  2993. check_irq_off();
  2994. spin_lock(&l3->list_lock);
  2995. entry = l3->slabs_partial.next;
  2996. if (entry == &l3->slabs_partial) {
  2997. l3->free_touched = 1;
  2998. entry = l3->slabs_free.next;
  2999. if (entry == &l3->slabs_free)
  3000. goto must_grow;
  3001. }
  3002. slabp = list_entry(entry, struct slab, list);
  3003. check_spinlock_acquired_node(cachep, nodeid);
  3004. check_slabp(cachep, slabp);
  3005. STATS_INC_NODEALLOCS(cachep);
  3006. STATS_INC_ACTIVE(cachep);
  3007. STATS_SET_HIGH(cachep);
  3008. BUG_ON(slabp->inuse == cachep->num);
  3009. obj = slab_get_obj(cachep, slabp, nodeid);
  3010. check_slabp(cachep, slabp);
  3011. l3->free_objects--;
  3012. /* move slabp to correct slabp list: */
  3013. list_del(&slabp->list);
  3014. if (slabp->free == BUFCTL_END)
  3015. list_add(&slabp->list, &l3->slabs_full);
  3016. else
  3017. list_add(&slabp->list, &l3->slabs_partial);
  3018. spin_unlock(&l3->list_lock);
  3019. goto done;
  3020. must_grow:
  3021. spin_unlock(&l3->list_lock);
  3022. x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
  3023. if (x)
  3024. goto retry;
  3025. return fallback_alloc(cachep, flags);
  3026. done:
  3027. return obj;
  3028. }
  3029. /**
  3030. * kmem_cache_alloc_node - Allocate an object on the specified node
  3031. * @cachep: The cache to allocate from.
  3032. * @flags: See kmalloc().
  3033. * @nodeid: node number of the target node.
  3034. * @caller: return address of caller, used for debug information
  3035. *
  3036. * Identical to kmem_cache_alloc but it will allocate memory on the given
  3037. * node, which can improve the performance for cpu bound structures.
  3038. *
  3039. * Fallback to other node is possible if __GFP_THISNODE is not set.
  3040. */
  3041. static __always_inline void *
  3042. __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
  3043. void *caller)
  3044. {
  3045. unsigned long save_flags;
  3046. void *ptr;
  3047. int slab_node = numa_mem_id();
  3048. flags &= gfp_allowed_mask;
  3049. lockdep_trace_alloc(flags);
  3050. if (slab_should_failslab(cachep, flags))
  3051. return NULL;
  3052. cache_alloc_debugcheck_before(cachep, flags);
  3053. local_irq_save(save_flags);
  3054. if (nodeid == NUMA_NO_NODE)
  3055. nodeid = slab_node;
  3056. if (unlikely(!cachep->nodelists[nodeid])) {
  3057. /* Node not bootstrapped yet */
  3058. ptr = fallback_alloc(cachep, flags);
  3059. goto out;
  3060. }
  3061. if (nodeid == slab_node) {
  3062. /*
  3063. * Use the locally cached objects if possible.
  3064. * However ____cache_alloc does not allow fallback
  3065. * to other nodes. It may fail while we still have
  3066. * objects on other nodes available.
  3067. */
  3068. ptr = ____cache_alloc(cachep, flags);
  3069. if (ptr)
  3070. goto out;
  3071. }
  3072. /* ___cache_alloc_node can fall back to other nodes */
  3073. ptr = ____cache_alloc_node(cachep, flags, nodeid);
  3074. out:
  3075. local_irq_restore(save_flags);
  3076. ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
  3077. kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
  3078. flags);
  3079. if (likely(ptr))
  3080. kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
  3081. if (unlikely((flags & __GFP_ZERO) && ptr))
  3082. memset(ptr, 0, obj_size(cachep));
  3083. return ptr;
  3084. }
  3085. static __always_inline void *
  3086. __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
  3087. {
  3088. void *objp;
  3089. if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
  3090. objp = alternate_node_alloc(cache, flags);
  3091. if (objp)
  3092. goto out;
  3093. }
  3094. objp = ____cache_alloc(cache, flags);
  3095. /*
  3096. * We may just have run out of memory on the local node.
  3097. * ____cache_alloc_node() knows how to locate memory on other nodes
  3098. */
  3099. if (!objp)
  3100. objp = ____cache_alloc_node(cache, flags, numa_mem_id());
  3101. out:
  3102. return objp;
  3103. }
  3104. #else
  3105. static __always_inline void *
  3106. __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  3107. {
  3108. return ____cache_alloc(cachep, flags);
  3109. }
  3110. #endif /* CONFIG_NUMA */
  3111. static __always_inline void *
  3112. __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
  3113. {
  3114. unsigned long save_flags;
  3115. void *objp;
  3116. flags &= gfp_allowed_mask;
  3117. lockdep_trace_alloc(flags);
  3118. if (slab_should_failslab(cachep, flags))
  3119. return NULL;
  3120. cache_alloc_debugcheck_before(cachep, flags);
  3121. local_irq_save(save_flags);
  3122. objp = __do_cache_alloc(cachep, flags);
  3123. local_irq_restore(save_flags);
  3124. objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
  3125. kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
  3126. flags);
  3127. prefetchw(objp);
  3128. if (likely(objp))
  3129. kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
  3130. if (unlikely((flags & __GFP_ZERO) && objp))
  3131. memset(objp, 0, obj_size(cachep));
  3132. return objp;
  3133. }
  3134. /*
  3135. * Caller needs to acquire correct kmem_list's list_lock
  3136. */
  3137. static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
  3138. int node)
  3139. {
  3140. int i;
  3141. struct kmem_list3 *l3;
  3142. for (i = 0; i < nr_objects; i++) {
  3143. void *objp = objpp[i];
  3144. struct slab *slabp;
  3145. slabp = virt_to_slab(objp);
  3146. l3 = cachep->nodelists[node];
  3147. list_del(&slabp->list);
  3148. check_spinlock_acquired_node(cachep, node);
  3149. check_slabp(cachep, slabp);
  3150. slab_put_obj(cachep, slabp, objp, node);
  3151. STATS_DEC_ACTIVE(cachep);
  3152. l3->free_objects++;
  3153. check_slabp(cachep, slabp);
  3154. /* fixup slab chains */
  3155. if (slabp->inuse == 0) {
  3156. if (l3->free_objects > l3->free_limit) {
  3157. l3->free_objects -= cachep->num;
  3158. /* No need to drop any previously held
  3159. * lock here, even if we have a off-slab slab
  3160. * descriptor it is guaranteed to come from
  3161. * a different cache, refer to comments before
  3162. * alloc_slabmgmt.
  3163. */
  3164. slab_destroy(cachep, slabp);
  3165. } else {
  3166. list_add(&slabp->list, &l3->slabs_free);
  3167. }
  3168. } else {
  3169. /* Unconditionally move a slab to the end of the
  3170. * partial list on free - maximum time for the
  3171. * other objects to be freed, too.
  3172. */
  3173. list_add_tail(&slabp->list, &l3->slabs_partial);
  3174. }
  3175. }
  3176. }
  3177. static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
  3178. {
  3179. int batchcount;
  3180. struct kmem_list3 *l3;
  3181. int node = numa_mem_id();
  3182. batchcount = ac->batchcount;
  3183. #if DEBUG
  3184. BUG_ON(!batchcount || batchcount > ac->avail);
  3185. #endif
  3186. check_irq_off();
  3187. l3 = cachep->nodelists[node];
  3188. spin_lock(&l3->list_lock);
  3189. if (l3->shared) {
  3190. struct array_cache *shared_array = l3->shared;
  3191. int max = shared_array->limit - shared_array->avail;
  3192. if (max) {
  3193. if (batchcount > max)
  3194. batchcount = max;
  3195. memcpy(&(shared_array->entry[shared_array->avail]),
  3196. ac->entry, sizeof(void *) * batchcount);
  3197. shared_array->avail += batchcount;
  3198. goto free_done;
  3199. }
  3200. }
  3201. free_block(cachep, ac->entry, batchcount, node);
  3202. free_done:
  3203. #if STATS
  3204. {
  3205. int i = 0;
  3206. struct list_head *p;
  3207. p = l3->slabs_free.next;
  3208. while (p != &(l3->slabs_free)) {
  3209. struct slab *slabp;
  3210. slabp = list_entry(p, struct slab, list);
  3211. BUG_ON(slabp->inuse);
  3212. i++;
  3213. p = p->next;
  3214. }
  3215. STATS_SET_FREEABLE(cachep, i);
  3216. }
  3217. #endif
  3218. spin_unlock(&l3->list_lock);
  3219. ac->avail -= batchcount;
  3220. memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
  3221. }
  3222. /*
  3223. * Release an obj back to its cache. If the obj has a constructed state, it must
  3224. * be in this state _before_ it is released. Called with disabled ints.
  3225. */
  3226. static inline void __cache_free(struct kmem_cache *cachep, void *objp,
  3227. void *caller)
  3228. {
  3229. struct array_cache *ac = cpu_cache_get(cachep);
  3230. check_irq_off();
  3231. kmemleak_free_recursive(objp, cachep->flags);
  3232. objp = cache_free_debugcheck(cachep, objp, caller);
  3233. kmemcheck_slab_free(cachep, objp, obj_size(cachep));
  3234. /*
  3235. * Skip calling cache_free_alien() when the platform is not numa.
  3236. * This will avoid cache misses that happen while accessing slabp (which
  3237. * is per page memory reference) to get nodeid. Instead use a global
  3238. * variable to skip the call, which is mostly likely to be present in
  3239. * the cache.
  3240. */
  3241. if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
  3242. return;
  3243. if (likely(ac->avail < ac->limit)) {
  3244. STATS_INC_FREEHIT(cachep);
  3245. } else {
  3246. STATS_INC_FREEMISS(cachep);
  3247. cache_flusharray(cachep, ac);
  3248. }
  3249. ac->entry[ac->avail++] = objp;
  3250. }
  3251. /**
  3252. * kmem_cache_alloc - Allocate an object
  3253. * @cachep: The cache to allocate from.
  3254. * @flags: See kmalloc().
  3255. *
  3256. * Allocate an object from this cache. The flags are only relevant
  3257. * if the cache has no available objects.
  3258. */
  3259. void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
  3260. {
  3261. void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
  3262. trace_kmem_cache_alloc(_RET_IP_, ret,
  3263. obj_size(cachep), cachep->buffer_size, flags);
  3264. return ret;
  3265. }
  3266. EXPORT_SYMBOL(kmem_cache_alloc);
  3267. #ifdef CONFIG_TRACING
  3268. void *
  3269. kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
  3270. {
  3271. void *ret;
  3272. ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
  3273. trace_kmalloc(_RET_IP_, ret,
  3274. size, slab_buffer_size(cachep), flags);
  3275. return ret;
  3276. }
  3277. EXPORT_SYMBOL(kmem_cache_alloc_trace);
  3278. #endif
  3279. #ifdef CONFIG_NUMA
  3280. void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
  3281. {
  3282. void *ret = __cache_alloc_node(cachep, flags, nodeid,
  3283. __builtin_return_address(0));
  3284. trace_kmem_cache_alloc_node(_RET_IP_, ret,
  3285. obj_size(cachep), cachep->buffer_size,
  3286. flags, nodeid);
  3287. return ret;
  3288. }
  3289. EXPORT_SYMBOL(kmem_cache_alloc_node);
  3290. #ifdef CONFIG_TRACING
  3291. void *kmem_cache_alloc_node_trace(size_t size,
  3292. struct kmem_cache *cachep,
  3293. gfp_t flags,
  3294. int nodeid)
  3295. {
  3296. void *ret;
  3297. ret = __cache_alloc_node(cachep, flags, nodeid,
  3298. __builtin_return_address(0));
  3299. trace_kmalloc_node(_RET_IP_, ret,
  3300. size, slab_buffer_size(cachep),
  3301. flags, nodeid);
  3302. return ret;
  3303. }
  3304. EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
  3305. #endif
  3306. static __always_inline void *
  3307. __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
  3308. {
  3309. struct kmem_cache *cachep;
  3310. cachep = kmem_find_general_cachep(size, flags);
  3311. if (unlikely(ZERO_OR_NULL_PTR(cachep)))
  3312. return cachep;
  3313. return kmem_cache_alloc_node_trace(size, cachep, flags, node);
  3314. }
  3315. #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
  3316. void *__kmalloc_node(size_t size, gfp_t flags, int node)
  3317. {
  3318. return __do_kmalloc_node(size, flags, node,
  3319. __builtin_return_address(0));
  3320. }
  3321. EXPORT_SYMBOL(__kmalloc_node);
  3322. void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
  3323. int node, unsigned long caller)
  3324. {
  3325. return __do_kmalloc_node(size, flags, node, (void *)caller);
  3326. }
  3327. EXPORT_SYMBOL(__kmalloc_node_track_caller);
  3328. #else
  3329. void *__kmalloc_node(size_t size, gfp_t flags, int node)
  3330. {
  3331. return __do_kmalloc_node(size, flags, node, NULL);
  3332. }
  3333. EXPORT_SYMBOL(__kmalloc_node);
  3334. #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
  3335. #endif /* CONFIG_NUMA */
  3336. /**
  3337. * __do_kmalloc - allocate memory
  3338. * @size: how many bytes of memory are required.
  3339. * @flags: the type of memory to allocate (see kmalloc).
  3340. * @caller: function caller for debug tracking of the caller
  3341. */
  3342. static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
  3343. void *caller)
  3344. {
  3345. struct kmem_cache *cachep;
  3346. void *ret;
  3347. /* If you want to save a few bytes .text space: replace
  3348. * __ with kmem_.
  3349. * Then kmalloc uses the uninlined functions instead of the inline
  3350. * functions.
  3351. */
  3352. cachep = __find_general_cachep(size, flags);
  3353. if (unlikely(ZERO_OR_NULL_PTR(cachep)))
  3354. return cachep;
  3355. ret = __cache_alloc(cachep, flags, caller);
  3356. trace_kmalloc((unsigned long) caller, ret,
  3357. size, cachep->buffer_size, flags);
  3358. return ret;
  3359. }
  3360. #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
  3361. void *__kmalloc(size_t size, gfp_t flags)
  3362. {
  3363. return __do_kmalloc(size, flags, __builtin_return_address(0));
  3364. }
  3365. EXPORT_SYMBOL(__kmalloc);
  3366. void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
  3367. {
  3368. return __do_kmalloc(size, flags, (void *)caller);
  3369. }
  3370. EXPORT_SYMBOL(__kmalloc_track_caller);
  3371. #else
  3372. void *__kmalloc(size_t size, gfp_t flags)
  3373. {
  3374. return __do_kmalloc(size, flags, NULL);
  3375. }
  3376. EXPORT_SYMBOL(__kmalloc);
  3377. #endif
  3378. /**
  3379. * kmem_cache_free - Deallocate an object
  3380. * @cachep: The cache the allocation was from.
  3381. * @objp: The previously allocated object.
  3382. *
  3383. * Free an object which was previously allocated from this
  3384. * cache.
  3385. */
  3386. void kmem_cache_free(struct kmem_cache *cachep, void *objp)
  3387. {
  3388. unsigned long flags;
  3389. local_irq_save(flags);
  3390. debug_check_no_locks_freed(objp, obj_size(cachep));
  3391. if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
  3392. debug_check_no_obj_freed(objp, obj_size(cachep));
  3393. __cache_free(cachep, objp, __builtin_return_address(0));
  3394. local_irq_restore(flags);
  3395. trace_kmem_cache_free(_RET_IP_, objp);
  3396. }
  3397. EXPORT_SYMBOL(kmem_cache_free);
  3398. /**
  3399. * kfree - free previously allocated memory
  3400. * @objp: pointer returned by kmalloc.
  3401. *
  3402. * If @objp is NULL, no operation is performed.
  3403. *
  3404. * Don't free memory not originally allocated by kmalloc()
  3405. * or you will run into trouble.
  3406. */
  3407. void kfree(const void *objp)
  3408. {
  3409. struct kmem_cache *c;
  3410. unsigned long flags;
  3411. trace_kfree(_RET_IP_, objp);
  3412. if (unlikely(ZERO_OR_NULL_PTR(objp)))
  3413. return;
  3414. local_irq_save(flags);
  3415. kfree_debugcheck(objp);
  3416. c = virt_to_cache(objp);
  3417. debug_check_no_locks_freed(objp, obj_size(c));
  3418. debug_check_no_obj_freed(objp, obj_size(c));
  3419. __cache_free(c, (void *)objp, __builtin_return_address(0));
  3420. local_irq_restore(flags);
  3421. }
  3422. EXPORT_SYMBOL(kfree);
  3423. unsigned int kmem_cache_size(struct kmem_cache *cachep)
  3424. {
  3425. return obj_size(cachep);
  3426. }
  3427. EXPORT_SYMBOL(kmem_cache_size);
  3428. /*
  3429. * This initializes kmem_list3 or resizes various caches for all nodes.
  3430. */
  3431. static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
  3432. {
  3433. int node;
  3434. struct kmem_list3 *l3;
  3435. struct array_cache *new_shared;
  3436. struct array_cache **new_alien = NULL;
  3437. for_each_online_node(node) {
  3438. if (use_alien_caches) {
  3439. new_alien = alloc_alien_cache(node, cachep->limit, gfp);
  3440. if (!new_alien)
  3441. goto fail;
  3442. }
  3443. new_shared = NULL;
  3444. if (cachep->shared) {
  3445. new_shared = alloc_arraycache(node,
  3446. cachep->shared*cachep->batchcount,
  3447. 0xbaadf00d, gfp);
  3448. if (!new_shared) {
  3449. free_alien_cache(new_alien);
  3450. goto fail;
  3451. }
  3452. }
  3453. l3 = cachep->nodelists[node];
  3454. if (l3) {
  3455. struct array_cache *shared = l3->shared;
  3456. spin_lock_irq(&l3->list_lock);
  3457. if (shared)
  3458. free_block(cachep, shared->entry,
  3459. shared->avail, node);
  3460. l3->shared = new_shared;
  3461. if (!l3->alien) {
  3462. l3->alien = new_alien;
  3463. new_alien = NULL;
  3464. }
  3465. l3->free_limit = (1 + nr_cpus_node(node)) *
  3466. cachep->batchcount + cachep->num;
  3467. spin_unlock_irq(&l3->list_lock);
  3468. kfree(shared);
  3469. free_alien_cache(new_alien);
  3470. continue;
  3471. }
  3472. l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
  3473. if (!l3) {
  3474. free_alien_cache(new_alien);
  3475. kfree(new_shared);
  3476. goto fail;
  3477. }
  3478. kmem_list3_init(l3);
  3479. l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
  3480. ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
  3481. l3->shared = new_shared;
  3482. l3->alien = new_alien;
  3483. l3->free_limit = (1 + nr_cpus_node(node)) *
  3484. cachep->batchcount + cachep->num;
  3485. cachep->nodelists[node] = l3;
  3486. }
  3487. return 0;
  3488. fail:
  3489. if (!cachep->next.next) {
  3490. /* Cache is not active yet. Roll back what we did */
  3491. node--;
  3492. while (node >= 0) {
  3493. if (cachep->nodelists[node]) {
  3494. l3 = cachep->nodelists[node];
  3495. kfree(l3->shared);
  3496. free_alien_cache(l3->alien);
  3497. kfree(l3);
  3498. cachep->nodelists[node] = NULL;
  3499. }
  3500. node--;
  3501. }
  3502. }
  3503. return -ENOMEM;
  3504. }
  3505. struct ccupdate_struct {
  3506. struct kmem_cache *cachep;
  3507. struct array_cache *new[0];
  3508. };
  3509. static void do_ccupdate_local(void *info)
  3510. {
  3511. struct ccupdate_struct *new = info;
  3512. struct array_cache *old;
  3513. check_irq_off();
  3514. old = cpu_cache_get(new->cachep);
  3515. new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
  3516. new->new[smp_processor_id()] = old;
  3517. }
  3518. /* Always called with the cache_chain_mutex held */
  3519. static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
  3520. int batchcount, int shared, gfp_t gfp)
  3521. {
  3522. struct ccupdate_struct *new;
  3523. int i;
  3524. new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
  3525. gfp);
  3526. if (!new)
  3527. return -ENOMEM;
  3528. for_each_online_cpu(i) {
  3529. new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
  3530. batchcount, gfp);
  3531. if (!new->new[i]) {
  3532. for (i--; i >= 0; i--)
  3533. kfree(new->new[i]);
  3534. kfree(new);
  3535. return -ENOMEM;
  3536. }
  3537. }
  3538. new->cachep = cachep;
  3539. on_each_cpu(do_ccupdate_local, (void *)new, 1);
  3540. check_irq_on();
  3541. cachep->batchcount = batchcount;
  3542. cachep->limit = limit;
  3543. cachep->shared = shared;
  3544. for_each_online_cpu(i) {
  3545. struct array_cache *ccold = new->new[i];
  3546. if (!ccold)
  3547. continue;
  3548. spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
  3549. free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
  3550. spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
  3551. kfree(ccold);
  3552. }
  3553. kfree(new);
  3554. return alloc_kmemlist(cachep, gfp);
  3555. }
  3556. /* Called with cache_chain_mutex held always */
  3557. static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
  3558. {
  3559. int err;
  3560. int limit, shared;
  3561. /*
  3562. * The head array serves three purposes:
  3563. * - create a LIFO ordering, i.e. return objects that are cache-warm
  3564. * - reduce the number of spinlock operations.
  3565. * - reduce the number of linked list operations on the slab and
  3566. * bufctl chains: array operations are cheaper.
  3567. * The numbers are guessed, we should auto-tune as described by
  3568. * Bonwick.
  3569. */
  3570. if (cachep->buffer_size > 131072)
  3571. limit = 1;
  3572. else if (cachep->buffer_size > PAGE_SIZE)
  3573. limit = 8;
  3574. else if (cachep->buffer_size > 1024)
  3575. limit = 24;
  3576. else if (cachep->buffer_size > 256)
  3577. limit = 54;
  3578. else
  3579. limit = 120;
  3580. /*
  3581. * CPU bound tasks (e.g. network routing) can exhibit cpu bound
  3582. * allocation behaviour: Most allocs on one cpu, most free operations
  3583. * on another cpu. For these cases, an efficient object passing between
  3584. * cpus is necessary. This is provided by a shared array. The array
  3585. * replaces Bonwick's magazine layer.
  3586. * On uniprocessor, it's functionally equivalent (but less efficient)
  3587. * to a larger limit. Thus disabled by default.
  3588. */
  3589. shared = 0;
  3590. if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
  3591. shared = 8;
  3592. #if DEBUG
  3593. /*
  3594. * With debugging enabled, large batchcount lead to excessively long
  3595. * periods with disabled local interrupts. Limit the batchcount
  3596. */
  3597. if (limit > 32)
  3598. limit = 32;
  3599. #endif
  3600. err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
  3601. if (err)
  3602. printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
  3603. cachep->name, -err);
  3604. return err;
  3605. }
  3606. /*
  3607. * Drain an array if it contains any elements taking the l3 lock only if
  3608. * necessary. Note that the l3 listlock also protects the array_cache
  3609. * if drain_array() is used on the shared array.
  3610. */
  3611. static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
  3612. struct array_cache *ac, int force, int node)
  3613. {
  3614. int tofree;
  3615. if (!ac || !ac->avail)
  3616. return;
  3617. if (ac->touched && !force) {
  3618. ac->touched = 0;
  3619. } else {
  3620. spin_lock_irq(&l3->list_lock);
  3621. if (ac->avail) {
  3622. tofree = force ? ac->avail : (ac->limit + 4) / 5;
  3623. if (tofree > ac->avail)
  3624. tofree = (ac->avail + 1) / 2;
  3625. free_block(cachep, ac->entry, tofree, node);
  3626. ac->avail -= tofree;
  3627. memmove(ac->entry, &(ac->entry[tofree]),
  3628. sizeof(void *) * ac->avail);
  3629. }
  3630. spin_unlock_irq(&l3->list_lock);
  3631. }
  3632. }
  3633. /**
  3634. * cache_reap - Reclaim memory from caches.
  3635. * @w: work descriptor
  3636. *
  3637. * Called from workqueue/eventd every few seconds.
  3638. * Purpose:
  3639. * - clear the per-cpu caches for this CPU.
  3640. * - return freeable pages to the main free memory pool.
  3641. *
  3642. * If we cannot acquire the cache chain mutex then just give up - we'll try
  3643. * again on the next iteration.
  3644. */
  3645. static void cache_reap(struct work_struct *w)
  3646. {
  3647. struct kmem_cache *searchp;
  3648. struct kmem_list3 *l3;
  3649. int node = numa_mem_id();
  3650. struct delayed_work *work = to_delayed_work(w);
  3651. if (!mutex_trylock(&cache_chain_mutex))
  3652. /* Give up. Setup the next iteration. */
  3653. goto out;
  3654. list_for_each_entry(searchp, &cache_chain, next) {
  3655. check_irq_on();
  3656. /*
  3657. * We only take the l3 lock if absolutely necessary and we
  3658. * have established with reasonable certainty that
  3659. * we can do some work if the lock was obtained.
  3660. */
  3661. l3 = searchp->nodelists[node];
  3662. reap_alien(searchp, l3);
  3663. drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
  3664. /*
  3665. * These are racy checks but it does not matter
  3666. * if we skip one check or scan twice.
  3667. */
  3668. if (time_after(l3->next_reap, jiffies))
  3669. goto next;
  3670. l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
  3671. drain_array(searchp, l3, l3->shared, 0, node);
  3672. if (l3->free_touched)
  3673. l3->free_touched = 0;
  3674. else {
  3675. int freed;
  3676. freed = drain_freelist(searchp, l3, (l3->free_limit +
  3677. 5 * searchp->num - 1) / (5 * searchp->num));
  3678. STATS_ADD_REAPED(searchp, freed);
  3679. }
  3680. next:
  3681. cond_resched();
  3682. }
  3683. check_irq_on();
  3684. mutex_unlock(&cache_chain_mutex);
  3685. next_reap_node();
  3686. out:
  3687. /* Set up the next iteration */
  3688. schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
  3689. }
  3690. #ifdef CONFIG_SLABINFO
  3691. static void print_slabinfo_header(struct seq_file *m)
  3692. {
  3693. /*
  3694. * Output format version, so at least we can change it
  3695. * without _too_ many complaints.
  3696. */
  3697. #if STATS
  3698. seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
  3699. #else
  3700. seq_puts(m, "slabinfo - version: 2.1\n");
  3701. #endif
  3702. seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
  3703. "<objperslab> <pagesperslab>");
  3704. seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
  3705. seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
  3706. #if STATS
  3707. seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
  3708. "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
  3709. seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
  3710. #endif
  3711. seq_putc(m, '\n');
  3712. }
  3713. static void *s_start(struct seq_file *m, loff_t *pos)
  3714. {
  3715. loff_t n = *pos;
  3716. mutex_lock(&cache_chain_mutex);
  3717. if (!n)
  3718. print_slabinfo_header(m);
  3719. return seq_list_start(&cache_chain, *pos);
  3720. }
  3721. static void *s_next(struct seq_file *m, void *p, loff_t *pos)
  3722. {
  3723. return seq_list_next(p, &cache_chain, pos);
  3724. }
  3725. static void s_stop(struct seq_file *m, void *p)
  3726. {
  3727. mutex_unlock(&cache_chain_mutex);
  3728. }
  3729. static int s_show(struct seq_file *m, void *p)
  3730. {
  3731. struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
  3732. struct slab *slabp;
  3733. unsigned long active_objs;
  3734. unsigned long num_objs;
  3735. unsigned long active_slabs = 0;
  3736. unsigned long num_slabs, free_objects = 0, shared_avail = 0;
  3737. const char *name;
  3738. char *error = NULL;
  3739. int node;
  3740. struct kmem_list3 *l3;
  3741. active_objs = 0;
  3742. num_slabs = 0;
  3743. for_each_online_node(node) {
  3744. l3 = cachep->nodelists[node];
  3745. if (!l3)
  3746. continue;
  3747. check_irq_on();
  3748. spin_lock_irq(&l3->list_lock);
  3749. list_for_each_entry(slabp, &l3->slabs_full, list) {
  3750. if (slabp->inuse != cachep->num && !error)
  3751. error = "slabs_full accounting error";
  3752. active_objs += cachep->num;
  3753. active_slabs++;
  3754. }
  3755. list_for_each_entry(slabp, &l3->slabs_partial, list) {
  3756. if (slabp->inuse == cachep->num && !error)
  3757. error = "slabs_partial inuse accounting error";
  3758. if (!slabp->inuse && !error)
  3759. error = "slabs_partial/inuse accounting error";
  3760. active_objs += slabp->inuse;
  3761. active_slabs++;
  3762. }
  3763. list_for_each_entry(slabp, &l3->slabs_free, list) {
  3764. if (slabp->inuse && !error)
  3765. error = "slabs_free/inuse accounting error";
  3766. num_slabs++;
  3767. }
  3768. free_objects += l3->free_objects;
  3769. if (l3->shared)
  3770. shared_avail += l3->shared->avail;
  3771. spin_unlock_irq(&l3->list_lock);
  3772. }
  3773. num_slabs += active_slabs;
  3774. num_objs = num_slabs * cachep->num;
  3775. if (num_objs - active_objs != free_objects && !error)
  3776. error = "free_objects accounting error";
  3777. name = cachep->name;
  3778. if (error)
  3779. printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
  3780. seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
  3781. name, active_objs, num_objs, cachep->buffer_size,
  3782. cachep->num, (1 << cachep->gfporder));
  3783. seq_printf(m, " : tunables %4u %4u %4u",
  3784. cachep->limit, cachep->batchcount, cachep->shared);
  3785. seq_printf(m, " : slabdata %6lu %6lu %6lu",
  3786. active_slabs, num_slabs, shared_avail);
  3787. #if STATS
  3788. { /* list3 stats */
  3789. unsigned long high = cachep->high_mark;
  3790. unsigned long allocs = cachep->num_allocations;
  3791. unsigned long grown = cachep->grown;
  3792. unsigned long reaped = cachep->reaped;
  3793. unsigned long errors = cachep->errors;
  3794. unsigned long max_freeable = cachep->max_freeable;
  3795. unsigned long node_allocs = cachep->node_allocs;
  3796. unsigned long node_frees = cachep->node_frees;
  3797. unsigned long overflows = cachep->node_overflow;
  3798. seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
  3799. "%4lu %4lu %4lu %4lu %4lu",
  3800. allocs, high, grown,
  3801. reaped, errors, max_freeable, node_allocs,
  3802. node_frees, overflows);
  3803. }
  3804. /* cpu stats */
  3805. {
  3806. unsigned long allochit = atomic_read(&cachep->allochit);
  3807. unsigned long allocmiss = atomic_read(&cachep->allocmiss);
  3808. unsigned long freehit = atomic_read(&cachep->freehit);
  3809. unsigned long freemiss = atomic_read(&cachep->freemiss);
  3810. seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
  3811. allochit, allocmiss, freehit, freemiss);
  3812. }
  3813. #endif
  3814. seq_putc(m, '\n');
  3815. return 0;
  3816. }
  3817. /*
  3818. * slabinfo_op - iterator that generates /proc/slabinfo
  3819. *
  3820. * Output layout:
  3821. * cache-name
  3822. * num-active-objs
  3823. * total-objs
  3824. * object size
  3825. * num-active-slabs
  3826. * total-slabs
  3827. * num-pages-per-slab
  3828. * + further values on SMP and with statistics enabled
  3829. */
  3830. static const struct seq_operations slabinfo_op = {
  3831. .start = s_start,
  3832. .next = s_next,
  3833. .stop = s_stop,
  3834. .show = s_show,
  3835. };
  3836. #define MAX_SLABINFO_WRITE 128
  3837. /**
  3838. * slabinfo_write - Tuning for the slab allocator
  3839. * @file: unused
  3840. * @buffer: user buffer
  3841. * @count: data length
  3842. * @ppos: unused
  3843. */
  3844. static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
  3845. size_t count, loff_t *ppos)
  3846. {
  3847. char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
  3848. int limit, batchcount, shared, res;
  3849. struct kmem_cache *cachep;
  3850. if (count > MAX_SLABINFO_WRITE)
  3851. return -EINVAL;
  3852. if (copy_from_user(&kbuf, buffer, count))
  3853. return -EFAULT;
  3854. kbuf[MAX_SLABINFO_WRITE] = '\0';
  3855. tmp = strchr(kbuf, ' ');
  3856. if (!tmp)
  3857. return -EINVAL;
  3858. *tmp = '\0';
  3859. tmp++;
  3860. if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
  3861. return -EINVAL;
  3862. /* Find the cache in the chain of caches. */
  3863. mutex_lock(&cache_chain_mutex);
  3864. res = -EINVAL;
  3865. list_for_each_entry(cachep, &cache_chain, next) {
  3866. if (!strcmp(cachep->name, kbuf)) {
  3867. if (limit < 1 || batchcount < 1 ||
  3868. batchcount > limit || shared < 0) {
  3869. res = 0;
  3870. } else {
  3871. res = do_tune_cpucache(cachep, limit,
  3872. batchcount, shared,
  3873. GFP_KERNEL);
  3874. }
  3875. break;
  3876. }
  3877. }
  3878. mutex_unlock(&cache_chain_mutex);
  3879. if (res >= 0)
  3880. res = count;
  3881. return res;
  3882. }
  3883. static int slabinfo_open(struct inode *inode, struct file *file)
  3884. {
  3885. return seq_open(file, &slabinfo_op);
  3886. }
  3887. static const struct file_operations proc_slabinfo_operations = {
  3888. .open = slabinfo_open,
  3889. .read = seq_read,
  3890. .write = slabinfo_write,
  3891. .llseek = seq_lseek,
  3892. .release = seq_release,
  3893. };
  3894. #ifdef CONFIG_DEBUG_SLAB_LEAK
  3895. static void *leaks_start(struct seq_file *m, loff_t *pos)
  3896. {
  3897. mutex_lock(&cache_chain_mutex);
  3898. return seq_list_start(&cache_chain, *pos);
  3899. }
  3900. static inline int add_caller(unsigned long *n, unsigned long v)
  3901. {
  3902. unsigned long *p;
  3903. int l;
  3904. if (!v)
  3905. return 1;
  3906. l = n[1];
  3907. p = n + 2;
  3908. while (l) {
  3909. int i = l/2;
  3910. unsigned long *q = p + 2 * i;
  3911. if (*q == v) {
  3912. q[1]++;
  3913. return 1;
  3914. }
  3915. if (*q > v) {
  3916. l = i;
  3917. } else {
  3918. p = q + 2;
  3919. l -= i + 1;
  3920. }
  3921. }
  3922. if (++n[1] == n[0])
  3923. return 0;
  3924. memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
  3925. p[0] = v;
  3926. p[1] = 1;
  3927. return 1;
  3928. }
  3929. static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
  3930. {
  3931. void *p;
  3932. int i;
  3933. if (n[0] == n[1])
  3934. return;
  3935. for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
  3936. if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
  3937. continue;
  3938. if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
  3939. return;
  3940. }
  3941. }
  3942. static void show_symbol(struct seq_file *m, unsigned long address)
  3943. {
  3944. #ifdef CONFIG_KALLSYMS
  3945. unsigned long offset, size;
  3946. char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
  3947. if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
  3948. seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
  3949. if (modname[0])
  3950. seq_printf(m, " [%s]", modname);
  3951. return;
  3952. }
  3953. #endif
  3954. seq_printf(m, "%p", (void *)address);
  3955. }
  3956. static int leaks_show(struct seq_file *m, void *p)
  3957. {
  3958. struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
  3959. struct slab *slabp;
  3960. struct kmem_list3 *l3;
  3961. const char *name;
  3962. unsigned long *n = m->private;
  3963. int node;
  3964. int i;
  3965. if (!(cachep->flags & SLAB_STORE_USER))
  3966. return 0;
  3967. if (!(cachep->flags & SLAB_RED_ZONE))
  3968. return 0;
  3969. /* OK, we can do it */
  3970. n[1] = 0;
  3971. for_each_online_node(node) {
  3972. l3 = cachep->nodelists[node];
  3973. if (!l3)
  3974. continue;
  3975. check_irq_on();
  3976. spin_lock_irq(&l3->list_lock);
  3977. list_for_each_entry(slabp, &l3->slabs_full, list)
  3978. handle_slab(n, cachep, slabp);
  3979. list_for_each_entry(slabp, &l3->slabs_partial, list)
  3980. handle_slab(n, cachep, slabp);
  3981. spin_unlock_irq(&l3->list_lock);
  3982. }
  3983. name = cachep->name;
  3984. if (n[0] == n[1]) {
  3985. /* Increase the buffer size */
  3986. mutex_unlock(&cache_chain_mutex);
  3987. m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
  3988. if (!m->private) {
  3989. /* Too bad, we are really out */
  3990. m->private = n;
  3991. mutex_lock(&cache_chain_mutex);
  3992. return -ENOMEM;
  3993. }
  3994. *(unsigned long *)m->private = n[0] * 2;
  3995. kfree(n);
  3996. mutex_lock(&cache_chain_mutex);
  3997. /* Now make sure this entry will be retried */
  3998. m->count = m->size;
  3999. return 0;
  4000. }
  4001. for (i = 0; i < n[1]; i++) {
  4002. seq_printf(m, "%s: %lu ", name, n[2*i+3]);
  4003. show_symbol(m, n[2*i+2]);
  4004. seq_putc(m, '\n');
  4005. }
  4006. return 0;
  4007. }
  4008. static const struct seq_operations slabstats_op = {
  4009. .start = leaks_start,
  4010. .next = s_next,
  4011. .stop = s_stop,
  4012. .show = leaks_show,
  4013. };
  4014. static int slabstats_open(struct inode *inode, struct file *file)
  4015. {
  4016. unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
  4017. int ret = -ENOMEM;
  4018. if (n) {
  4019. ret = seq_open(file, &slabstats_op);
  4020. if (!ret) {
  4021. struct seq_file *m = file->private_data;
  4022. *n = PAGE_SIZE / (2 * sizeof(unsigned long));
  4023. m->private = n;
  4024. n = NULL;
  4025. }
  4026. kfree(n);
  4027. }
  4028. return ret;
  4029. }
  4030. static const struct file_operations proc_slabstats_operations = {
  4031. .open = slabstats_open,
  4032. .read = seq_read,
  4033. .llseek = seq_lseek,
  4034. .release = seq_release_private,
  4035. };
  4036. #endif
  4037. static int __init slab_proc_init(void)
  4038. {
  4039. proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
  4040. #ifdef CONFIG_DEBUG_SLAB_LEAK
  4041. proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
  4042. #endif
  4043. return 0;
  4044. }
  4045. module_init(slab_proc_init);
  4046. #endif
  4047. /**
  4048. * ksize - get the actual amount of memory allocated for a given object
  4049. * @objp: Pointer to the object
  4050. *
  4051. * kmalloc may internally round up allocations and return more memory
  4052. * than requested. ksize() can be used to determine the actual amount of
  4053. * memory allocated. The caller may use this additional memory, even though
  4054. * a smaller amount of memory was initially specified with the kmalloc call.
  4055. * The caller must guarantee that objp points to a valid object previously
  4056. * allocated with either kmalloc() or kmem_cache_alloc(). The object
  4057. * must not be freed during the duration of the call.
  4058. */
  4059. size_t ksize(const void *objp)
  4060. {
  4061. BUG_ON(!objp);
  4062. if (unlikely(objp == ZERO_SIZE_PTR))
  4063. return 0;
  4064. return obj_size(virt_to_cache(objp));
  4065. }
  4066. EXPORT_SYMBOL(ksize);