perf_event.c 142 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890389138923893389438953896389738983899390039013902390339043905390639073908390939103911391239133914391539163917391839193920392139223923392439253926392739283929393039313932393339343935393639373938393939403941394239433944394539463947394839493950395139523953395439553956395739583959396039613962396339643965396639673968396939703971397239733974397539763977397839793980398139823983398439853986398739883989399039913992399339943995399639973998399940004001400240034004400540064007400840094010401140124013401440154016401740184019402040214022402340244025402640274028402940304031403240334034403540364037403840394040404140424043404440454046404740484049405040514052405340544055405640574058405940604061406240634064406540664067406840694070407140724073407440754076407740784079408040814082408340844085408640874088408940904091409240934094409540964097409840994100410141024103410441054106410741084109411041114112411341144115411641174118411941204121412241234124412541264127412841294130413141324133413441354136413741384139414041414142414341444145414641474148414941504151415241534154415541564157415841594160416141624163416441654166416741684169417041714172417341744175417641774178417941804181418241834184418541864187418841894190419141924193419441954196419741984199420042014202420342044205420642074208420942104211421242134214421542164217421842194220422142224223422442254226422742284229423042314232423342344235423642374238423942404241424242434244424542464247424842494250425142524253425442554256425742584259426042614262426342644265426642674268426942704271427242734274427542764277427842794280428142824283428442854286428742884289429042914292429342944295429642974298429943004301430243034304430543064307430843094310431143124313431443154316431743184319432043214322432343244325432643274328432943304331433243334334433543364337433843394340434143424343434443454346434743484349435043514352435343544355435643574358435943604361436243634364436543664367436843694370437143724373437443754376437743784379438043814382438343844385438643874388438943904391439243934394439543964397439843994400440144024403440444054406440744084409441044114412441344144415441644174418441944204421442244234424442544264427442844294430443144324433443444354436443744384439444044414442444344444445444644474448444944504451445244534454445544564457445844594460446144624463446444654466446744684469447044714472447344744475447644774478447944804481448244834484448544864487448844894490449144924493449444954496449744984499450045014502450345044505450645074508450945104511451245134514451545164517451845194520452145224523452445254526452745284529453045314532453345344535453645374538453945404541454245434544454545464547454845494550455145524553455445554556455745584559456045614562456345644565456645674568456945704571457245734574457545764577457845794580458145824583458445854586458745884589459045914592459345944595459645974598459946004601460246034604460546064607460846094610461146124613461446154616461746184619462046214622462346244625462646274628462946304631463246334634463546364637463846394640464146424643464446454646464746484649465046514652465346544655465646574658465946604661466246634664466546664667466846694670467146724673467446754676467746784679468046814682468346844685468646874688468946904691469246934694469546964697469846994700470147024703470447054706470747084709471047114712471347144715471647174718471947204721472247234724472547264727472847294730473147324733473447354736473747384739474047414742474347444745474647474748474947504751475247534754475547564757475847594760476147624763476447654766476747684769477047714772477347744775477647774778477947804781478247834784478547864787478847894790479147924793479447954796479747984799480048014802480348044805480648074808480948104811481248134814481548164817481848194820482148224823482448254826482748284829483048314832483348344835483648374838483948404841484248434844484548464847484848494850485148524853485448554856485748584859486048614862486348644865486648674868486948704871487248734874487548764877487848794880488148824883488448854886488748884889489048914892489348944895489648974898489949004901490249034904490549064907490849094910491149124913491449154916491749184919492049214922492349244925492649274928492949304931493249334934493549364937493849394940494149424943494449454946494749484949495049514952495349544955495649574958495949604961496249634964496549664967496849694970497149724973497449754976497749784979498049814982498349844985498649874988498949904991499249934994499549964997499849995000500150025003500450055006500750085009501050115012501350145015501650175018501950205021502250235024502550265027502850295030503150325033503450355036503750385039504050415042504350445045504650475048504950505051505250535054505550565057505850595060506150625063506450655066506750685069507050715072507350745075507650775078507950805081508250835084508550865087508850895090509150925093509450955096509750985099510051015102510351045105510651075108510951105111511251135114511551165117511851195120512151225123512451255126512751285129513051315132513351345135513651375138513951405141514251435144514551465147514851495150515151525153515451555156515751585159516051615162516351645165516651675168516951705171517251735174517551765177517851795180518151825183518451855186518751885189519051915192519351945195519651975198519952005201520252035204520552065207520852095210521152125213521452155216521752185219522052215222522352245225522652275228522952305231523252335234523552365237523852395240524152425243524452455246524752485249525052515252525352545255525652575258525952605261526252635264526552665267526852695270527152725273527452755276527752785279528052815282528352845285528652875288528952905291529252935294529552965297529852995300530153025303530453055306530753085309531053115312531353145315531653175318531953205321532253235324532553265327532853295330533153325333533453355336533753385339534053415342534353445345534653475348534953505351535253535354535553565357535853595360536153625363536453655366536753685369537053715372537353745375537653775378537953805381538253835384538553865387538853895390539153925393539453955396539753985399540054015402540354045405540654075408540954105411541254135414541554165417541854195420542154225423542454255426542754285429543054315432543354345435543654375438543954405441544254435444544554465447544854495450545154525453545454555456545754585459546054615462546354645465546654675468546954705471547254735474547554765477547854795480548154825483548454855486548754885489549054915492549354945495549654975498549955005501550255035504550555065507550855095510551155125513551455155516551755185519552055215522552355245525552655275528552955305531553255335534553555365537553855395540554155425543554455455546554755485549555055515552555355545555555655575558555955605561556255635564556555665567556855695570557155725573557455755576557755785579558055815582558355845585558655875588558955905591559255935594559555965597559855995600560156025603560456055606560756085609561056115612561356145615561656175618561956205621562256235624562556265627562856295630563156325633563456355636563756385639564056415642564356445645564656475648564956505651565256535654565556565657565856595660566156625663566456655666566756685669567056715672567356745675567656775678567956805681568256835684568556865687568856895690569156925693569456955696569756985699570057015702570357045705570657075708570957105711571257135714571557165717571857195720572157225723572457255726572757285729573057315732573357345735573657375738573957405741574257435744574557465747574857495750575157525753575457555756575757585759576057615762576357645765576657675768576957705771577257735774577557765777577857795780578157825783578457855786578757885789579057915792579357945795579657975798579958005801580258035804580558065807580858095810581158125813581458155816581758185819582058215822582358245825582658275828582958305831583258335834583558365837583858395840584158425843584458455846584758485849585058515852585358545855585658575858585958605861586258635864586558665867586858695870587158725873587458755876587758785879588058815882588358845885588658875888588958905891589258935894589558965897589858995900590159025903590459055906590759085909591059115912591359145915591659175918591959205921592259235924592559265927592859295930593159325933593459355936593759385939594059415942594359445945594659475948594959505951595259535954595559565957595859595960596159625963596459655966596759685969597059715972597359745975597659775978597959805981598259835984598559865987598859895990599159925993599459955996599759985999600060016002600360046005600660076008600960106011601260136014601560166017601860196020602160226023602460256026602760286029603060316032603360346035603660376038603960406041604260436044604560466047604860496050605160526053605460556056605760586059606060616062606360646065606660676068606960706071607260736074607560766077607860796080608160826083608460856086608760886089609060916092609360946095609660976098609961006101610261036104610561066107610861096110611161126113611461156116611761186119612061216122612361246125612661276128612961306131613261336134613561366137613861396140614161426143614461456146614761486149615061516152615361546155615661576158615961606161616261636164616561666167616861696170617161726173617461756176617761786179618061816182618361846185618661876188618961906191619261936194619561966197619861996200620162026203620462056206620762086209621062116212621362146215621662176218621962206221622262236224622562266227622862296230623162326233623462356236623762386239624062416242624362446245624662476248624962506251625262536254625562566257625862596260626162626263626462656266626762686269627062716272627362746275627662776278627962806281628262836284628562866287628862896290629162926293629462956296629762986299630063016302630363046305
  1. /*
  2. * Performance events core code:
  3. *
  4. * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
  5. * Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
  6. * Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  7. * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  8. *
  9. * For licensing details see kernel-base/COPYING
  10. */
  11. #include <linux/fs.h>
  12. #include <linux/mm.h>
  13. #include <linux/cpu.h>
  14. #include <linux/smp.h>
  15. #include <linux/file.h>
  16. #include <linux/poll.h>
  17. #include <linux/slab.h>
  18. #include <linux/hash.h>
  19. #include <linux/sysfs.h>
  20. #include <linux/dcache.h>
  21. #include <linux/percpu.h>
  22. #include <linux/ptrace.h>
  23. #include <linux/vmstat.h>
  24. #include <linux/vmalloc.h>
  25. #include <linux/hardirq.h>
  26. #include <linux/rculist.h>
  27. #include <linux/uaccess.h>
  28. #include <linux/syscalls.h>
  29. #include <linux/anon_inodes.h>
  30. #include <linux/kernel_stat.h>
  31. #include <linux/perf_event.h>
  32. #include <linux/ftrace_event.h>
  33. #include <asm/irq_regs.h>
  34. atomic_t perf_task_events __read_mostly;
  35. static atomic_t nr_mmap_events __read_mostly;
  36. static atomic_t nr_comm_events __read_mostly;
  37. static atomic_t nr_task_events __read_mostly;
  38. static LIST_HEAD(pmus);
  39. static DEFINE_MUTEX(pmus_lock);
  40. static struct srcu_struct pmus_srcu;
  41. /*
  42. * perf event paranoia level:
  43. * -1 - not paranoid at all
  44. * 0 - disallow raw tracepoint access for unpriv
  45. * 1 - disallow cpu events for unpriv
  46. * 2 - disallow kernel profiling for unpriv
  47. */
  48. int sysctl_perf_event_paranoid __read_mostly = 1;
  49. int sysctl_perf_event_mlock __read_mostly = 512; /* 'free' kb per user */
  50. /*
  51. * max perf event sample rate
  52. */
  53. int sysctl_perf_event_sample_rate __read_mostly = 100000;
  54. static atomic64_t perf_event_id;
  55. void __weak perf_event_print_debug(void) { }
  56. extern __weak const char *perf_pmu_name(void)
  57. {
  58. return "pmu";
  59. }
  60. void perf_pmu_disable(struct pmu *pmu)
  61. {
  62. int *count = this_cpu_ptr(pmu->pmu_disable_count);
  63. if (!(*count)++)
  64. pmu->pmu_disable(pmu);
  65. }
  66. void perf_pmu_enable(struct pmu *pmu)
  67. {
  68. int *count = this_cpu_ptr(pmu->pmu_disable_count);
  69. if (!--(*count))
  70. pmu->pmu_enable(pmu);
  71. }
  72. static DEFINE_PER_CPU(struct list_head, rotation_list);
  73. /*
  74. * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
  75. * because they're strictly cpu affine and rotate_start is called with IRQs
  76. * disabled, while rotate_context is called from IRQ context.
  77. */
  78. static void perf_pmu_rotate_start(struct pmu *pmu)
  79. {
  80. struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
  81. struct list_head *head = &__get_cpu_var(rotation_list);
  82. WARN_ON(!irqs_disabled());
  83. if (list_empty(&cpuctx->rotation_list))
  84. list_add(&cpuctx->rotation_list, head);
  85. }
  86. static void get_ctx(struct perf_event_context *ctx)
  87. {
  88. WARN_ON(!atomic_inc_not_zero(&ctx->refcount));
  89. }
  90. static void free_ctx(struct rcu_head *head)
  91. {
  92. struct perf_event_context *ctx;
  93. ctx = container_of(head, struct perf_event_context, rcu_head);
  94. kfree(ctx);
  95. }
  96. static void put_ctx(struct perf_event_context *ctx)
  97. {
  98. if (atomic_dec_and_test(&ctx->refcount)) {
  99. if (ctx->parent_ctx)
  100. put_ctx(ctx->parent_ctx);
  101. if (ctx->task)
  102. put_task_struct(ctx->task);
  103. call_rcu(&ctx->rcu_head, free_ctx);
  104. }
  105. }
  106. static void unclone_ctx(struct perf_event_context *ctx)
  107. {
  108. if (ctx->parent_ctx) {
  109. put_ctx(ctx->parent_ctx);
  110. ctx->parent_ctx = NULL;
  111. }
  112. }
  113. /*
  114. * If we inherit events we want to return the parent event id
  115. * to userspace.
  116. */
  117. static u64 primary_event_id(struct perf_event *event)
  118. {
  119. u64 id = event->id;
  120. if (event->parent)
  121. id = event->parent->id;
  122. return id;
  123. }
  124. /*
  125. * Get the perf_event_context for a task and lock it.
  126. * This has to cope with with the fact that until it is locked,
  127. * the context could get moved to another task.
  128. */
  129. static struct perf_event_context *
  130. perf_lock_task_context(struct task_struct *task, int ctxn, unsigned long *flags)
  131. {
  132. struct perf_event_context *ctx;
  133. rcu_read_lock();
  134. retry:
  135. ctx = rcu_dereference(task->perf_event_ctxp[ctxn]);
  136. if (ctx) {
  137. /*
  138. * If this context is a clone of another, it might
  139. * get swapped for another underneath us by
  140. * perf_event_task_sched_out, though the
  141. * rcu_read_lock() protects us from any context
  142. * getting freed. Lock the context and check if it
  143. * got swapped before we could get the lock, and retry
  144. * if so. If we locked the right context, then it
  145. * can't get swapped on us any more.
  146. */
  147. raw_spin_lock_irqsave(&ctx->lock, *flags);
  148. if (ctx != rcu_dereference(task->perf_event_ctxp[ctxn])) {
  149. raw_spin_unlock_irqrestore(&ctx->lock, *flags);
  150. goto retry;
  151. }
  152. if (!atomic_inc_not_zero(&ctx->refcount)) {
  153. raw_spin_unlock_irqrestore(&ctx->lock, *flags);
  154. ctx = NULL;
  155. }
  156. }
  157. rcu_read_unlock();
  158. return ctx;
  159. }
  160. /*
  161. * Get the context for a task and increment its pin_count so it
  162. * can't get swapped to another task. This also increments its
  163. * reference count so that the context can't get freed.
  164. */
  165. static struct perf_event_context *
  166. perf_pin_task_context(struct task_struct *task, int ctxn)
  167. {
  168. struct perf_event_context *ctx;
  169. unsigned long flags;
  170. ctx = perf_lock_task_context(task, ctxn, &flags);
  171. if (ctx) {
  172. ++ctx->pin_count;
  173. raw_spin_unlock_irqrestore(&ctx->lock, flags);
  174. }
  175. return ctx;
  176. }
  177. static void perf_unpin_context(struct perf_event_context *ctx)
  178. {
  179. unsigned long flags;
  180. raw_spin_lock_irqsave(&ctx->lock, flags);
  181. --ctx->pin_count;
  182. raw_spin_unlock_irqrestore(&ctx->lock, flags);
  183. put_ctx(ctx);
  184. }
  185. static inline u64 perf_clock(void)
  186. {
  187. return local_clock();
  188. }
  189. /*
  190. * Update the record of the current time in a context.
  191. */
  192. static void update_context_time(struct perf_event_context *ctx)
  193. {
  194. u64 now = perf_clock();
  195. ctx->time += now - ctx->timestamp;
  196. ctx->timestamp = now;
  197. }
  198. /*
  199. * Update the total_time_enabled and total_time_running fields for a event.
  200. */
  201. static void update_event_times(struct perf_event *event)
  202. {
  203. struct perf_event_context *ctx = event->ctx;
  204. u64 run_end;
  205. if (event->state < PERF_EVENT_STATE_INACTIVE ||
  206. event->group_leader->state < PERF_EVENT_STATE_INACTIVE)
  207. return;
  208. if (ctx->is_active)
  209. run_end = ctx->time;
  210. else
  211. run_end = event->tstamp_stopped;
  212. event->total_time_enabled = run_end - event->tstamp_enabled;
  213. if (event->state == PERF_EVENT_STATE_INACTIVE)
  214. run_end = event->tstamp_stopped;
  215. else
  216. run_end = ctx->time;
  217. event->total_time_running = run_end - event->tstamp_running;
  218. }
  219. /*
  220. * Update total_time_enabled and total_time_running for all events in a group.
  221. */
  222. static void update_group_times(struct perf_event *leader)
  223. {
  224. struct perf_event *event;
  225. update_event_times(leader);
  226. list_for_each_entry(event, &leader->sibling_list, group_entry)
  227. update_event_times(event);
  228. }
  229. static struct list_head *
  230. ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
  231. {
  232. if (event->attr.pinned)
  233. return &ctx->pinned_groups;
  234. else
  235. return &ctx->flexible_groups;
  236. }
  237. /*
  238. * Add a event from the lists for its context.
  239. * Must be called with ctx->mutex and ctx->lock held.
  240. */
  241. static void
  242. list_add_event(struct perf_event *event, struct perf_event_context *ctx)
  243. {
  244. WARN_ON_ONCE(event->attach_state & PERF_ATTACH_CONTEXT);
  245. event->attach_state |= PERF_ATTACH_CONTEXT;
  246. /*
  247. * If we're a stand alone event or group leader, we go to the context
  248. * list, group events are kept attached to the group so that
  249. * perf_group_detach can, at all times, locate all siblings.
  250. */
  251. if (event->group_leader == event) {
  252. struct list_head *list;
  253. if (is_software_event(event))
  254. event->group_flags |= PERF_GROUP_SOFTWARE;
  255. list = ctx_group_list(event, ctx);
  256. list_add_tail(&event->group_entry, list);
  257. }
  258. list_add_rcu(&event->event_entry, &ctx->event_list);
  259. if (!ctx->nr_events)
  260. perf_pmu_rotate_start(ctx->pmu);
  261. ctx->nr_events++;
  262. if (event->attr.inherit_stat)
  263. ctx->nr_stat++;
  264. }
  265. static void perf_group_attach(struct perf_event *event)
  266. {
  267. struct perf_event *group_leader = event->group_leader;
  268. /*
  269. * We can have double attach due to group movement in perf_event_open.
  270. */
  271. if (event->attach_state & PERF_ATTACH_GROUP)
  272. return;
  273. event->attach_state |= PERF_ATTACH_GROUP;
  274. if (group_leader == event)
  275. return;
  276. if (group_leader->group_flags & PERF_GROUP_SOFTWARE &&
  277. !is_software_event(event))
  278. group_leader->group_flags &= ~PERF_GROUP_SOFTWARE;
  279. list_add_tail(&event->group_entry, &group_leader->sibling_list);
  280. group_leader->nr_siblings++;
  281. }
  282. /*
  283. * Remove a event from the lists for its context.
  284. * Must be called with ctx->mutex and ctx->lock held.
  285. */
  286. static void
  287. list_del_event(struct perf_event *event, struct perf_event_context *ctx)
  288. {
  289. /*
  290. * We can have double detach due to exit/hot-unplug + close.
  291. */
  292. if (!(event->attach_state & PERF_ATTACH_CONTEXT))
  293. return;
  294. event->attach_state &= ~PERF_ATTACH_CONTEXT;
  295. ctx->nr_events--;
  296. if (event->attr.inherit_stat)
  297. ctx->nr_stat--;
  298. list_del_rcu(&event->event_entry);
  299. if (event->group_leader == event)
  300. list_del_init(&event->group_entry);
  301. update_group_times(event);
  302. /*
  303. * If event was in error state, then keep it
  304. * that way, otherwise bogus counts will be
  305. * returned on read(). The only way to get out
  306. * of error state is by explicit re-enabling
  307. * of the event
  308. */
  309. if (event->state > PERF_EVENT_STATE_OFF)
  310. event->state = PERF_EVENT_STATE_OFF;
  311. }
  312. static void perf_group_detach(struct perf_event *event)
  313. {
  314. struct perf_event *sibling, *tmp;
  315. struct list_head *list = NULL;
  316. /*
  317. * We can have double detach due to exit/hot-unplug + close.
  318. */
  319. if (!(event->attach_state & PERF_ATTACH_GROUP))
  320. return;
  321. event->attach_state &= ~PERF_ATTACH_GROUP;
  322. /*
  323. * If this is a sibling, remove it from its group.
  324. */
  325. if (event->group_leader != event) {
  326. list_del_init(&event->group_entry);
  327. event->group_leader->nr_siblings--;
  328. return;
  329. }
  330. if (!list_empty(&event->group_entry))
  331. list = &event->group_entry;
  332. /*
  333. * If this was a group event with sibling events then
  334. * upgrade the siblings to singleton events by adding them
  335. * to whatever list we are on.
  336. */
  337. list_for_each_entry_safe(sibling, tmp, &event->sibling_list, group_entry) {
  338. if (list)
  339. list_move_tail(&sibling->group_entry, list);
  340. sibling->group_leader = sibling;
  341. /* Inherit group flags from the previous leader */
  342. sibling->group_flags = event->group_flags;
  343. }
  344. }
  345. static inline int
  346. event_filter_match(struct perf_event *event)
  347. {
  348. return event->cpu == -1 || event->cpu == smp_processor_id();
  349. }
  350. static void
  351. event_sched_out(struct perf_event *event,
  352. struct perf_cpu_context *cpuctx,
  353. struct perf_event_context *ctx)
  354. {
  355. u64 delta;
  356. /*
  357. * An event which could not be activated because of
  358. * filter mismatch still needs to have its timings
  359. * maintained, otherwise bogus information is return
  360. * via read() for time_enabled, time_running:
  361. */
  362. if (event->state == PERF_EVENT_STATE_INACTIVE
  363. && !event_filter_match(event)) {
  364. delta = ctx->time - event->tstamp_stopped;
  365. event->tstamp_running += delta;
  366. event->tstamp_stopped = ctx->time;
  367. }
  368. if (event->state != PERF_EVENT_STATE_ACTIVE)
  369. return;
  370. event->state = PERF_EVENT_STATE_INACTIVE;
  371. if (event->pending_disable) {
  372. event->pending_disable = 0;
  373. event->state = PERF_EVENT_STATE_OFF;
  374. }
  375. event->tstamp_stopped = ctx->time;
  376. event->pmu->del(event, 0);
  377. event->oncpu = -1;
  378. if (!is_software_event(event))
  379. cpuctx->active_oncpu--;
  380. ctx->nr_active--;
  381. if (event->attr.exclusive || !cpuctx->active_oncpu)
  382. cpuctx->exclusive = 0;
  383. }
  384. static void
  385. group_sched_out(struct perf_event *group_event,
  386. struct perf_cpu_context *cpuctx,
  387. struct perf_event_context *ctx)
  388. {
  389. struct perf_event *event;
  390. int state = group_event->state;
  391. event_sched_out(group_event, cpuctx, ctx);
  392. /*
  393. * Schedule out siblings (if any):
  394. */
  395. list_for_each_entry(event, &group_event->sibling_list, group_entry)
  396. event_sched_out(event, cpuctx, ctx);
  397. if (state == PERF_EVENT_STATE_ACTIVE && group_event->attr.exclusive)
  398. cpuctx->exclusive = 0;
  399. }
  400. static inline struct perf_cpu_context *
  401. __get_cpu_context(struct perf_event_context *ctx)
  402. {
  403. return this_cpu_ptr(ctx->pmu->pmu_cpu_context);
  404. }
  405. /*
  406. * Cross CPU call to remove a performance event
  407. *
  408. * We disable the event on the hardware level first. After that we
  409. * remove it from the context list.
  410. */
  411. static void __perf_event_remove_from_context(void *info)
  412. {
  413. struct perf_event *event = info;
  414. struct perf_event_context *ctx = event->ctx;
  415. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  416. /*
  417. * If this is a task context, we need to check whether it is
  418. * the current task context of this cpu. If not it has been
  419. * scheduled out before the smp call arrived.
  420. */
  421. if (ctx->task && cpuctx->task_ctx != ctx)
  422. return;
  423. raw_spin_lock(&ctx->lock);
  424. event_sched_out(event, cpuctx, ctx);
  425. list_del_event(event, ctx);
  426. raw_spin_unlock(&ctx->lock);
  427. }
  428. /*
  429. * Remove the event from a task's (or a CPU's) list of events.
  430. *
  431. * Must be called with ctx->mutex held.
  432. *
  433. * CPU events are removed with a smp call. For task events we only
  434. * call when the task is on a CPU.
  435. *
  436. * If event->ctx is a cloned context, callers must make sure that
  437. * every task struct that event->ctx->task could possibly point to
  438. * remains valid. This is OK when called from perf_release since
  439. * that only calls us on the top-level context, which can't be a clone.
  440. * When called from perf_event_exit_task, it's OK because the
  441. * context has been detached from its task.
  442. */
  443. static void perf_event_remove_from_context(struct perf_event *event)
  444. {
  445. struct perf_event_context *ctx = event->ctx;
  446. struct task_struct *task = ctx->task;
  447. if (!task) {
  448. /*
  449. * Per cpu events are removed via an smp call and
  450. * the removal is always successful.
  451. */
  452. smp_call_function_single(event->cpu,
  453. __perf_event_remove_from_context,
  454. event, 1);
  455. return;
  456. }
  457. retry:
  458. task_oncpu_function_call(task, __perf_event_remove_from_context,
  459. event);
  460. raw_spin_lock_irq(&ctx->lock);
  461. /*
  462. * If the context is active we need to retry the smp call.
  463. */
  464. if (ctx->nr_active && !list_empty(&event->group_entry)) {
  465. raw_spin_unlock_irq(&ctx->lock);
  466. goto retry;
  467. }
  468. /*
  469. * The lock prevents that this context is scheduled in so we
  470. * can remove the event safely, if the call above did not
  471. * succeed.
  472. */
  473. if (!list_empty(&event->group_entry))
  474. list_del_event(event, ctx);
  475. raw_spin_unlock_irq(&ctx->lock);
  476. }
  477. /*
  478. * Cross CPU call to disable a performance event
  479. */
  480. static void __perf_event_disable(void *info)
  481. {
  482. struct perf_event *event = info;
  483. struct perf_event_context *ctx = event->ctx;
  484. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  485. /*
  486. * If this is a per-task event, need to check whether this
  487. * event's task is the current task on this cpu.
  488. */
  489. if (ctx->task && cpuctx->task_ctx != ctx)
  490. return;
  491. raw_spin_lock(&ctx->lock);
  492. /*
  493. * If the event is on, turn it off.
  494. * If it is in error state, leave it in error state.
  495. */
  496. if (event->state >= PERF_EVENT_STATE_INACTIVE) {
  497. update_context_time(ctx);
  498. update_group_times(event);
  499. if (event == event->group_leader)
  500. group_sched_out(event, cpuctx, ctx);
  501. else
  502. event_sched_out(event, cpuctx, ctx);
  503. event->state = PERF_EVENT_STATE_OFF;
  504. }
  505. raw_spin_unlock(&ctx->lock);
  506. }
  507. /*
  508. * Disable a event.
  509. *
  510. * If event->ctx is a cloned context, callers must make sure that
  511. * every task struct that event->ctx->task could possibly point to
  512. * remains valid. This condition is satisifed when called through
  513. * perf_event_for_each_child or perf_event_for_each because they
  514. * hold the top-level event's child_mutex, so any descendant that
  515. * goes to exit will block in sync_child_event.
  516. * When called from perf_pending_event it's OK because event->ctx
  517. * is the current context on this CPU and preemption is disabled,
  518. * hence we can't get into perf_event_task_sched_out for this context.
  519. */
  520. void perf_event_disable(struct perf_event *event)
  521. {
  522. struct perf_event_context *ctx = event->ctx;
  523. struct task_struct *task = ctx->task;
  524. if (!task) {
  525. /*
  526. * Disable the event on the cpu that it's on
  527. */
  528. smp_call_function_single(event->cpu, __perf_event_disable,
  529. event, 1);
  530. return;
  531. }
  532. retry:
  533. task_oncpu_function_call(task, __perf_event_disable, event);
  534. raw_spin_lock_irq(&ctx->lock);
  535. /*
  536. * If the event is still active, we need to retry the cross-call.
  537. */
  538. if (event->state == PERF_EVENT_STATE_ACTIVE) {
  539. raw_spin_unlock_irq(&ctx->lock);
  540. goto retry;
  541. }
  542. /*
  543. * Since we have the lock this context can't be scheduled
  544. * in, so we can change the state safely.
  545. */
  546. if (event->state == PERF_EVENT_STATE_INACTIVE) {
  547. update_group_times(event);
  548. event->state = PERF_EVENT_STATE_OFF;
  549. }
  550. raw_spin_unlock_irq(&ctx->lock);
  551. }
  552. static int
  553. event_sched_in(struct perf_event *event,
  554. struct perf_cpu_context *cpuctx,
  555. struct perf_event_context *ctx)
  556. {
  557. if (event->state <= PERF_EVENT_STATE_OFF)
  558. return 0;
  559. event->state = PERF_EVENT_STATE_ACTIVE;
  560. event->oncpu = smp_processor_id();
  561. /*
  562. * The new state must be visible before we turn it on in the hardware:
  563. */
  564. smp_wmb();
  565. if (event->pmu->add(event, PERF_EF_START)) {
  566. event->state = PERF_EVENT_STATE_INACTIVE;
  567. event->oncpu = -1;
  568. return -EAGAIN;
  569. }
  570. event->tstamp_running += ctx->time - event->tstamp_stopped;
  571. if (!is_software_event(event))
  572. cpuctx->active_oncpu++;
  573. ctx->nr_active++;
  574. if (event->attr.exclusive)
  575. cpuctx->exclusive = 1;
  576. return 0;
  577. }
  578. static int
  579. group_sched_in(struct perf_event *group_event,
  580. struct perf_cpu_context *cpuctx,
  581. struct perf_event_context *ctx)
  582. {
  583. struct perf_event *event, *partial_group = NULL;
  584. struct pmu *pmu = group_event->pmu;
  585. u64 now = ctx->time;
  586. bool simulate = false;
  587. if (group_event->state == PERF_EVENT_STATE_OFF)
  588. return 0;
  589. pmu->start_txn(pmu);
  590. if (event_sched_in(group_event, cpuctx, ctx)) {
  591. pmu->cancel_txn(pmu);
  592. return -EAGAIN;
  593. }
  594. /*
  595. * Schedule in siblings as one group (if any):
  596. */
  597. list_for_each_entry(event, &group_event->sibling_list, group_entry) {
  598. if (event_sched_in(event, cpuctx, ctx)) {
  599. partial_group = event;
  600. goto group_error;
  601. }
  602. }
  603. if (!pmu->commit_txn(pmu))
  604. return 0;
  605. group_error:
  606. /*
  607. * Groups can be scheduled in as one unit only, so undo any
  608. * partial group before returning:
  609. * The events up to the failed event are scheduled out normally,
  610. * tstamp_stopped will be updated.
  611. *
  612. * The failed events and the remaining siblings need to have
  613. * their timings updated as if they had gone thru event_sched_in()
  614. * and event_sched_out(). This is required to get consistent timings
  615. * across the group. This also takes care of the case where the group
  616. * could never be scheduled by ensuring tstamp_stopped is set to mark
  617. * the time the event was actually stopped, such that time delta
  618. * calculation in update_event_times() is correct.
  619. */
  620. list_for_each_entry(event, &group_event->sibling_list, group_entry) {
  621. if (event == partial_group)
  622. simulate = true;
  623. if (simulate) {
  624. event->tstamp_running += now - event->tstamp_stopped;
  625. event->tstamp_stopped = now;
  626. } else {
  627. event_sched_out(event, cpuctx, ctx);
  628. }
  629. }
  630. event_sched_out(group_event, cpuctx, ctx);
  631. pmu->cancel_txn(pmu);
  632. return -EAGAIN;
  633. }
  634. /*
  635. * Work out whether we can put this event group on the CPU now.
  636. */
  637. static int group_can_go_on(struct perf_event *event,
  638. struct perf_cpu_context *cpuctx,
  639. int can_add_hw)
  640. {
  641. /*
  642. * Groups consisting entirely of software events can always go on.
  643. */
  644. if (event->group_flags & PERF_GROUP_SOFTWARE)
  645. return 1;
  646. /*
  647. * If an exclusive group is already on, no other hardware
  648. * events can go on.
  649. */
  650. if (cpuctx->exclusive)
  651. return 0;
  652. /*
  653. * If this group is exclusive and there are already
  654. * events on the CPU, it can't go on.
  655. */
  656. if (event->attr.exclusive && cpuctx->active_oncpu)
  657. return 0;
  658. /*
  659. * Otherwise, try to add it if all previous groups were able
  660. * to go on.
  661. */
  662. return can_add_hw;
  663. }
  664. static void add_event_to_ctx(struct perf_event *event,
  665. struct perf_event_context *ctx)
  666. {
  667. list_add_event(event, ctx);
  668. perf_group_attach(event);
  669. event->tstamp_enabled = ctx->time;
  670. event->tstamp_running = ctx->time;
  671. event->tstamp_stopped = ctx->time;
  672. }
  673. /*
  674. * Cross CPU call to install and enable a performance event
  675. *
  676. * Must be called with ctx->mutex held
  677. */
  678. static void __perf_install_in_context(void *info)
  679. {
  680. struct perf_event *event = info;
  681. struct perf_event_context *ctx = event->ctx;
  682. struct perf_event *leader = event->group_leader;
  683. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  684. int err;
  685. /*
  686. * If this is a task context, we need to check whether it is
  687. * the current task context of this cpu. If not it has been
  688. * scheduled out before the smp call arrived.
  689. * Or possibly this is the right context but it isn't
  690. * on this cpu because it had no events.
  691. */
  692. if (ctx->task && cpuctx->task_ctx != ctx) {
  693. if (cpuctx->task_ctx || ctx->task != current)
  694. return;
  695. cpuctx->task_ctx = ctx;
  696. }
  697. raw_spin_lock(&ctx->lock);
  698. ctx->is_active = 1;
  699. update_context_time(ctx);
  700. add_event_to_ctx(event, ctx);
  701. if (event->cpu != -1 && event->cpu != smp_processor_id())
  702. goto unlock;
  703. /*
  704. * Don't put the event on if it is disabled or if
  705. * it is in a group and the group isn't on.
  706. */
  707. if (event->state != PERF_EVENT_STATE_INACTIVE ||
  708. (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE))
  709. goto unlock;
  710. /*
  711. * An exclusive event can't go on if there are already active
  712. * hardware events, and no hardware event can go on if there
  713. * is already an exclusive event on.
  714. */
  715. if (!group_can_go_on(event, cpuctx, 1))
  716. err = -EEXIST;
  717. else
  718. err = event_sched_in(event, cpuctx, ctx);
  719. if (err) {
  720. /*
  721. * This event couldn't go on. If it is in a group
  722. * then we have to pull the whole group off.
  723. * If the event group is pinned then put it in error state.
  724. */
  725. if (leader != event)
  726. group_sched_out(leader, cpuctx, ctx);
  727. if (leader->attr.pinned) {
  728. update_group_times(leader);
  729. leader->state = PERF_EVENT_STATE_ERROR;
  730. }
  731. }
  732. unlock:
  733. raw_spin_unlock(&ctx->lock);
  734. }
  735. /*
  736. * Attach a performance event to a context
  737. *
  738. * First we add the event to the list with the hardware enable bit
  739. * in event->hw_config cleared.
  740. *
  741. * If the event is attached to a task which is on a CPU we use a smp
  742. * call to enable it in the task context. The task might have been
  743. * scheduled away, but we check this in the smp call again.
  744. *
  745. * Must be called with ctx->mutex held.
  746. */
  747. static void
  748. perf_install_in_context(struct perf_event_context *ctx,
  749. struct perf_event *event,
  750. int cpu)
  751. {
  752. struct task_struct *task = ctx->task;
  753. event->ctx = ctx;
  754. if (!task) {
  755. /*
  756. * Per cpu events are installed via an smp call and
  757. * the install is always successful.
  758. */
  759. smp_call_function_single(cpu, __perf_install_in_context,
  760. event, 1);
  761. return;
  762. }
  763. retry:
  764. task_oncpu_function_call(task, __perf_install_in_context,
  765. event);
  766. raw_spin_lock_irq(&ctx->lock);
  767. /*
  768. * we need to retry the smp call.
  769. */
  770. if (ctx->is_active && list_empty(&event->group_entry)) {
  771. raw_spin_unlock_irq(&ctx->lock);
  772. goto retry;
  773. }
  774. /*
  775. * The lock prevents that this context is scheduled in so we
  776. * can add the event safely, if it the call above did not
  777. * succeed.
  778. */
  779. if (list_empty(&event->group_entry))
  780. add_event_to_ctx(event, ctx);
  781. raw_spin_unlock_irq(&ctx->lock);
  782. }
  783. /*
  784. * Put a event into inactive state and update time fields.
  785. * Enabling the leader of a group effectively enables all
  786. * the group members that aren't explicitly disabled, so we
  787. * have to update their ->tstamp_enabled also.
  788. * Note: this works for group members as well as group leaders
  789. * since the non-leader members' sibling_lists will be empty.
  790. */
  791. static void __perf_event_mark_enabled(struct perf_event *event,
  792. struct perf_event_context *ctx)
  793. {
  794. struct perf_event *sub;
  795. event->state = PERF_EVENT_STATE_INACTIVE;
  796. event->tstamp_enabled = ctx->time - event->total_time_enabled;
  797. list_for_each_entry(sub, &event->sibling_list, group_entry) {
  798. if (sub->state >= PERF_EVENT_STATE_INACTIVE) {
  799. sub->tstamp_enabled =
  800. ctx->time - sub->total_time_enabled;
  801. }
  802. }
  803. }
  804. /*
  805. * Cross CPU call to enable a performance event
  806. */
  807. static void __perf_event_enable(void *info)
  808. {
  809. struct perf_event *event = info;
  810. struct perf_event_context *ctx = event->ctx;
  811. struct perf_event *leader = event->group_leader;
  812. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  813. int err;
  814. /*
  815. * If this is a per-task event, need to check whether this
  816. * event's task is the current task on this cpu.
  817. */
  818. if (ctx->task && cpuctx->task_ctx != ctx) {
  819. if (cpuctx->task_ctx || ctx->task != current)
  820. return;
  821. cpuctx->task_ctx = ctx;
  822. }
  823. raw_spin_lock(&ctx->lock);
  824. ctx->is_active = 1;
  825. update_context_time(ctx);
  826. if (event->state >= PERF_EVENT_STATE_INACTIVE)
  827. goto unlock;
  828. __perf_event_mark_enabled(event, ctx);
  829. if (event->cpu != -1 && event->cpu != smp_processor_id())
  830. goto unlock;
  831. /*
  832. * If the event is in a group and isn't the group leader,
  833. * then don't put it on unless the group is on.
  834. */
  835. if (leader != event && leader->state != PERF_EVENT_STATE_ACTIVE)
  836. goto unlock;
  837. if (!group_can_go_on(event, cpuctx, 1)) {
  838. err = -EEXIST;
  839. } else {
  840. if (event == leader)
  841. err = group_sched_in(event, cpuctx, ctx);
  842. else
  843. err = event_sched_in(event, cpuctx, ctx);
  844. }
  845. if (err) {
  846. /*
  847. * If this event can't go on and it's part of a
  848. * group, then the whole group has to come off.
  849. */
  850. if (leader != event)
  851. group_sched_out(leader, cpuctx, ctx);
  852. if (leader->attr.pinned) {
  853. update_group_times(leader);
  854. leader->state = PERF_EVENT_STATE_ERROR;
  855. }
  856. }
  857. unlock:
  858. raw_spin_unlock(&ctx->lock);
  859. }
  860. /*
  861. * Enable a event.
  862. *
  863. * If event->ctx is a cloned context, callers must make sure that
  864. * every task struct that event->ctx->task could possibly point to
  865. * remains valid. This condition is satisfied when called through
  866. * perf_event_for_each_child or perf_event_for_each as described
  867. * for perf_event_disable.
  868. */
  869. void perf_event_enable(struct perf_event *event)
  870. {
  871. struct perf_event_context *ctx = event->ctx;
  872. struct task_struct *task = ctx->task;
  873. if (!task) {
  874. /*
  875. * Enable the event on the cpu that it's on
  876. */
  877. smp_call_function_single(event->cpu, __perf_event_enable,
  878. event, 1);
  879. return;
  880. }
  881. raw_spin_lock_irq(&ctx->lock);
  882. if (event->state >= PERF_EVENT_STATE_INACTIVE)
  883. goto out;
  884. /*
  885. * If the event is in error state, clear that first.
  886. * That way, if we see the event in error state below, we
  887. * know that it has gone back into error state, as distinct
  888. * from the task having been scheduled away before the
  889. * cross-call arrived.
  890. */
  891. if (event->state == PERF_EVENT_STATE_ERROR)
  892. event->state = PERF_EVENT_STATE_OFF;
  893. retry:
  894. raw_spin_unlock_irq(&ctx->lock);
  895. task_oncpu_function_call(task, __perf_event_enable, event);
  896. raw_spin_lock_irq(&ctx->lock);
  897. /*
  898. * If the context is active and the event is still off,
  899. * we need to retry the cross-call.
  900. */
  901. if (ctx->is_active && event->state == PERF_EVENT_STATE_OFF)
  902. goto retry;
  903. /*
  904. * Since we have the lock this context can't be scheduled
  905. * in, so we can change the state safely.
  906. */
  907. if (event->state == PERF_EVENT_STATE_OFF)
  908. __perf_event_mark_enabled(event, ctx);
  909. out:
  910. raw_spin_unlock_irq(&ctx->lock);
  911. }
  912. static int perf_event_refresh(struct perf_event *event, int refresh)
  913. {
  914. /*
  915. * not supported on inherited events
  916. */
  917. if (event->attr.inherit)
  918. return -EINVAL;
  919. atomic_add(refresh, &event->event_limit);
  920. perf_event_enable(event);
  921. return 0;
  922. }
  923. enum event_type_t {
  924. EVENT_FLEXIBLE = 0x1,
  925. EVENT_PINNED = 0x2,
  926. EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
  927. };
  928. static void ctx_sched_out(struct perf_event_context *ctx,
  929. struct perf_cpu_context *cpuctx,
  930. enum event_type_t event_type)
  931. {
  932. struct perf_event *event;
  933. raw_spin_lock(&ctx->lock);
  934. perf_pmu_disable(ctx->pmu);
  935. ctx->is_active = 0;
  936. if (likely(!ctx->nr_events))
  937. goto out;
  938. update_context_time(ctx);
  939. if (!ctx->nr_active)
  940. goto out;
  941. if (event_type & EVENT_PINNED) {
  942. list_for_each_entry(event, &ctx->pinned_groups, group_entry)
  943. group_sched_out(event, cpuctx, ctx);
  944. }
  945. if (event_type & EVENT_FLEXIBLE) {
  946. list_for_each_entry(event, &ctx->flexible_groups, group_entry)
  947. group_sched_out(event, cpuctx, ctx);
  948. }
  949. out:
  950. perf_pmu_enable(ctx->pmu);
  951. raw_spin_unlock(&ctx->lock);
  952. }
  953. /*
  954. * Test whether two contexts are equivalent, i.e. whether they
  955. * have both been cloned from the same version of the same context
  956. * and they both have the same number of enabled events.
  957. * If the number of enabled events is the same, then the set
  958. * of enabled events should be the same, because these are both
  959. * inherited contexts, therefore we can't access individual events
  960. * in them directly with an fd; we can only enable/disable all
  961. * events via prctl, or enable/disable all events in a family
  962. * via ioctl, which will have the same effect on both contexts.
  963. */
  964. static int context_equiv(struct perf_event_context *ctx1,
  965. struct perf_event_context *ctx2)
  966. {
  967. return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
  968. && ctx1->parent_gen == ctx2->parent_gen
  969. && !ctx1->pin_count && !ctx2->pin_count;
  970. }
  971. static void __perf_event_sync_stat(struct perf_event *event,
  972. struct perf_event *next_event)
  973. {
  974. u64 value;
  975. if (!event->attr.inherit_stat)
  976. return;
  977. /*
  978. * Update the event value, we cannot use perf_event_read()
  979. * because we're in the middle of a context switch and have IRQs
  980. * disabled, which upsets smp_call_function_single(), however
  981. * we know the event must be on the current CPU, therefore we
  982. * don't need to use it.
  983. */
  984. switch (event->state) {
  985. case PERF_EVENT_STATE_ACTIVE:
  986. event->pmu->read(event);
  987. /* fall-through */
  988. case PERF_EVENT_STATE_INACTIVE:
  989. update_event_times(event);
  990. break;
  991. default:
  992. break;
  993. }
  994. /*
  995. * In order to keep per-task stats reliable we need to flip the event
  996. * values when we flip the contexts.
  997. */
  998. value = local64_read(&next_event->count);
  999. value = local64_xchg(&event->count, value);
  1000. local64_set(&next_event->count, value);
  1001. swap(event->total_time_enabled, next_event->total_time_enabled);
  1002. swap(event->total_time_running, next_event->total_time_running);
  1003. /*
  1004. * Since we swizzled the values, update the user visible data too.
  1005. */
  1006. perf_event_update_userpage(event);
  1007. perf_event_update_userpage(next_event);
  1008. }
  1009. #define list_next_entry(pos, member) \
  1010. list_entry(pos->member.next, typeof(*pos), member)
  1011. static void perf_event_sync_stat(struct perf_event_context *ctx,
  1012. struct perf_event_context *next_ctx)
  1013. {
  1014. struct perf_event *event, *next_event;
  1015. if (!ctx->nr_stat)
  1016. return;
  1017. update_context_time(ctx);
  1018. event = list_first_entry(&ctx->event_list,
  1019. struct perf_event, event_entry);
  1020. next_event = list_first_entry(&next_ctx->event_list,
  1021. struct perf_event, event_entry);
  1022. while (&event->event_entry != &ctx->event_list &&
  1023. &next_event->event_entry != &next_ctx->event_list) {
  1024. __perf_event_sync_stat(event, next_event);
  1025. event = list_next_entry(event, event_entry);
  1026. next_event = list_next_entry(next_event, event_entry);
  1027. }
  1028. }
  1029. void perf_event_context_sched_out(struct task_struct *task, int ctxn,
  1030. struct task_struct *next)
  1031. {
  1032. struct perf_event_context *ctx = task->perf_event_ctxp[ctxn];
  1033. struct perf_event_context *next_ctx;
  1034. struct perf_event_context *parent;
  1035. struct perf_cpu_context *cpuctx;
  1036. int do_switch = 1;
  1037. if (likely(!ctx))
  1038. return;
  1039. cpuctx = __get_cpu_context(ctx);
  1040. if (!cpuctx->task_ctx)
  1041. return;
  1042. rcu_read_lock();
  1043. parent = rcu_dereference(ctx->parent_ctx);
  1044. next_ctx = next->perf_event_ctxp[ctxn];
  1045. if (parent && next_ctx &&
  1046. rcu_dereference(next_ctx->parent_ctx) == parent) {
  1047. /*
  1048. * Looks like the two contexts are clones, so we might be
  1049. * able to optimize the context switch. We lock both
  1050. * contexts and check that they are clones under the
  1051. * lock (including re-checking that neither has been
  1052. * uncloned in the meantime). It doesn't matter which
  1053. * order we take the locks because no other cpu could
  1054. * be trying to lock both of these tasks.
  1055. */
  1056. raw_spin_lock(&ctx->lock);
  1057. raw_spin_lock_nested(&next_ctx->lock, SINGLE_DEPTH_NESTING);
  1058. if (context_equiv(ctx, next_ctx)) {
  1059. /*
  1060. * XXX do we need a memory barrier of sorts
  1061. * wrt to rcu_dereference() of perf_event_ctxp
  1062. */
  1063. task->perf_event_ctxp[ctxn] = next_ctx;
  1064. next->perf_event_ctxp[ctxn] = ctx;
  1065. ctx->task = next;
  1066. next_ctx->task = task;
  1067. do_switch = 0;
  1068. perf_event_sync_stat(ctx, next_ctx);
  1069. }
  1070. raw_spin_unlock(&next_ctx->lock);
  1071. raw_spin_unlock(&ctx->lock);
  1072. }
  1073. rcu_read_unlock();
  1074. if (do_switch) {
  1075. ctx_sched_out(ctx, cpuctx, EVENT_ALL);
  1076. cpuctx->task_ctx = NULL;
  1077. }
  1078. }
  1079. #define for_each_task_context_nr(ctxn) \
  1080. for ((ctxn) = 0; (ctxn) < perf_nr_task_contexts; (ctxn)++)
  1081. /*
  1082. * Called from scheduler to remove the events of the current task,
  1083. * with interrupts disabled.
  1084. *
  1085. * We stop each event and update the event value in event->count.
  1086. *
  1087. * This does not protect us against NMI, but disable()
  1088. * sets the disabled bit in the control field of event _before_
  1089. * accessing the event control register. If a NMI hits, then it will
  1090. * not restart the event.
  1091. */
  1092. void __perf_event_task_sched_out(struct task_struct *task,
  1093. struct task_struct *next)
  1094. {
  1095. int ctxn;
  1096. perf_sw_event(PERF_COUNT_SW_CONTEXT_SWITCHES, 1, 1, NULL, 0);
  1097. for_each_task_context_nr(ctxn)
  1098. perf_event_context_sched_out(task, ctxn, next);
  1099. }
  1100. static void task_ctx_sched_out(struct perf_event_context *ctx,
  1101. enum event_type_t event_type)
  1102. {
  1103. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  1104. if (!cpuctx->task_ctx)
  1105. return;
  1106. if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
  1107. return;
  1108. ctx_sched_out(ctx, cpuctx, event_type);
  1109. cpuctx->task_ctx = NULL;
  1110. }
  1111. /*
  1112. * Called with IRQs disabled
  1113. */
  1114. static void cpu_ctx_sched_out(struct perf_cpu_context *cpuctx,
  1115. enum event_type_t event_type)
  1116. {
  1117. ctx_sched_out(&cpuctx->ctx, cpuctx, event_type);
  1118. }
  1119. static void
  1120. ctx_pinned_sched_in(struct perf_event_context *ctx,
  1121. struct perf_cpu_context *cpuctx)
  1122. {
  1123. struct perf_event *event;
  1124. list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
  1125. if (event->state <= PERF_EVENT_STATE_OFF)
  1126. continue;
  1127. if (event->cpu != -1 && event->cpu != smp_processor_id())
  1128. continue;
  1129. if (group_can_go_on(event, cpuctx, 1))
  1130. group_sched_in(event, cpuctx, ctx);
  1131. /*
  1132. * If this pinned group hasn't been scheduled,
  1133. * put it in error state.
  1134. */
  1135. if (event->state == PERF_EVENT_STATE_INACTIVE) {
  1136. update_group_times(event);
  1137. event->state = PERF_EVENT_STATE_ERROR;
  1138. }
  1139. }
  1140. }
  1141. static void
  1142. ctx_flexible_sched_in(struct perf_event_context *ctx,
  1143. struct perf_cpu_context *cpuctx)
  1144. {
  1145. struct perf_event *event;
  1146. int can_add_hw = 1;
  1147. list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
  1148. /* Ignore events in OFF or ERROR state */
  1149. if (event->state <= PERF_EVENT_STATE_OFF)
  1150. continue;
  1151. /*
  1152. * Listen to the 'cpu' scheduling filter constraint
  1153. * of events:
  1154. */
  1155. if (event->cpu != -1 && event->cpu != smp_processor_id())
  1156. continue;
  1157. if (group_can_go_on(event, cpuctx, can_add_hw)) {
  1158. if (group_sched_in(event, cpuctx, ctx))
  1159. can_add_hw = 0;
  1160. }
  1161. }
  1162. }
  1163. static void
  1164. ctx_sched_in(struct perf_event_context *ctx,
  1165. struct perf_cpu_context *cpuctx,
  1166. enum event_type_t event_type)
  1167. {
  1168. raw_spin_lock(&ctx->lock);
  1169. ctx->is_active = 1;
  1170. if (likely(!ctx->nr_events))
  1171. goto out;
  1172. ctx->timestamp = perf_clock();
  1173. /*
  1174. * First go through the list and put on any pinned groups
  1175. * in order to give them the best chance of going on.
  1176. */
  1177. if (event_type & EVENT_PINNED)
  1178. ctx_pinned_sched_in(ctx, cpuctx);
  1179. /* Then walk through the lower prio flexible groups */
  1180. if (event_type & EVENT_FLEXIBLE)
  1181. ctx_flexible_sched_in(ctx, cpuctx);
  1182. out:
  1183. raw_spin_unlock(&ctx->lock);
  1184. }
  1185. static void cpu_ctx_sched_in(struct perf_cpu_context *cpuctx,
  1186. enum event_type_t event_type)
  1187. {
  1188. struct perf_event_context *ctx = &cpuctx->ctx;
  1189. ctx_sched_in(ctx, cpuctx, event_type);
  1190. }
  1191. static void task_ctx_sched_in(struct perf_event_context *ctx,
  1192. enum event_type_t event_type)
  1193. {
  1194. struct perf_cpu_context *cpuctx;
  1195. cpuctx = __get_cpu_context(ctx);
  1196. if (cpuctx->task_ctx == ctx)
  1197. return;
  1198. ctx_sched_in(ctx, cpuctx, event_type);
  1199. cpuctx->task_ctx = ctx;
  1200. }
  1201. void perf_event_context_sched_in(struct perf_event_context *ctx)
  1202. {
  1203. struct perf_cpu_context *cpuctx;
  1204. cpuctx = __get_cpu_context(ctx);
  1205. if (cpuctx->task_ctx == ctx)
  1206. return;
  1207. perf_pmu_disable(ctx->pmu);
  1208. /*
  1209. * We want to keep the following priority order:
  1210. * cpu pinned (that don't need to move), task pinned,
  1211. * cpu flexible, task flexible.
  1212. */
  1213. cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
  1214. ctx_sched_in(ctx, cpuctx, EVENT_PINNED);
  1215. cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
  1216. ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE);
  1217. cpuctx->task_ctx = ctx;
  1218. /*
  1219. * Since these rotations are per-cpu, we need to ensure the
  1220. * cpu-context we got scheduled on is actually rotating.
  1221. */
  1222. perf_pmu_rotate_start(ctx->pmu);
  1223. perf_pmu_enable(ctx->pmu);
  1224. }
  1225. /*
  1226. * Called from scheduler to add the events of the current task
  1227. * with interrupts disabled.
  1228. *
  1229. * We restore the event value and then enable it.
  1230. *
  1231. * This does not protect us against NMI, but enable()
  1232. * sets the enabled bit in the control field of event _before_
  1233. * accessing the event control register. If a NMI hits, then it will
  1234. * keep the event running.
  1235. */
  1236. void __perf_event_task_sched_in(struct task_struct *task)
  1237. {
  1238. struct perf_event_context *ctx;
  1239. int ctxn;
  1240. for_each_task_context_nr(ctxn) {
  1241. ctx = task->perf_event_ctxp[ctxn];
  1242. if (likely(!ctx))
  1243. continue;
  1244. perf_event_context_sched_in(ctx);
  1245. }
  1246. }
  1247. #define MAX_INTERRUPTS (~0ULL)
  1248. static void perf_log_throttle(struct perf_event *event, int enable);
  1249. static u64 perf_calculate_period(struct perf_event *event, u64 nsec, u64 count)
  1250. {
  1251. u64 frequency = event->attr.sample_freq;
  1252. u64 sec = NSEC_PER_SEC;
  1253. u64 divisor, dividend;
  1254. int count_fls, nsec_fls, frequency_fls, sec_fls;
  1255. count_fls = fls64(count);
  1256. nsec_fls = fls64(nsec);
  1257. frequency_fls = fls64(frequency);
  1258. sec_fls = 30;
  1259. /*
  1260. * We got @count in @nsec, with a target of sample_freq HZ
  1261. * the target period becomes:
  1262. *
  1263. * @count * 10^9
  1264. * period = -------------------
  1265. * @nsec * sample_freq
  1266. *
  1267. */
  1268. /*
  1269. * Reduce accuracy by one bit such that @a and @b converge
  1270. * to a similar magnitude.
  1271. */
  1272. #define REDUCE_FLS(a, b) \
  1273. do { \
  1274. if (a##_fls > b##_fls) { \
  1275. a >>= 1; \
  1276. a##_fls--; \
  1277. } else { \
  1278. b >>= 1; \
  1279. b##_fls--; \
  1280. } \
  1281. } while (0)
  1282. /*
  1283. * Reduce accuracy until either term fits in a u64, then proceed with
  1284. * the other, so that finally we can do a u64/u64 division.
  1285. */
  1286. while (count_fls + sec_fls > 64 && nsec_fls + frequency_fls > 64) {
  1287. REDUCE_FLS(nsec, frequency);
  1288. REDUCE_FLS(sec, count);
  1289. }
  1290. if (count_fls + sec_fls > 64) {
  1291. divisor = nsec * frequency;
  1292. while (count_fls + sec_fls > 64) {
  1293. REDUCE_FLS(count, sec);
  1294. divisor >>= 1;
  1295. }
  1296. dividend = count * sec;
  1297. } else {
  1298. dividend = count * sec;
  1299. while (nsec_fls + frequency_fls > 64) {
  1300. REDUCE_FLS(nsec, frequency);
  1301. dividend >>= 1;
  1302. }
  1303. divisor = nsec * frequency;
  1304. }
  1305. if (!divisor)
  1306. return dividend;
  1307. return div64_u64(dividend, divisor);
  1308. }
  1309. static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count)
  1310. {
  1311. struct hw_perf_event *hwc = &event->hw;
  1312. s64 period, sample_period;
  1313. s64 delta;
  1314. period = perf_calculate_period(event, nsec, count);
  1315. delta = (s64)(period - hwc->sample_period);
  1316. delta = (delta + 7) / 8; /* low pass filter */
  1317. sample_period = hwc->sample_period + delta;
  1318. if (!sample_period)
  1319. sample_period = 1;
  1320. hwc->sample_period = sample_period;
  1321. if (local64_read(&hwc->period_left) > 8*sample_period) {
  1322. event->pmu->stop(event, PERF_EF_UPDATE);
  1323. local64_set(&hwc->period_left, 0);
  1324. event->pmu->start(event, PERF_EF_RELOAD);
  1325. }
  1326. }
  1327. static void perf_ctx_adjust_freq(struct perf_event_context *ctx, u64 period)
  1328. {
  1329. struct perf_event *event;
  1330. struct hw_perf_event *hwc;
  1331. u64 interrupts, now;
  1332. s64 delta;
  1333. raw_spin_lock(&ctx->lock);
  1334. list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
  1335. if (event->state != PERF_EVENT_STATE_ACTIVE)
  1336. continue;
  1337. if (event->cpu != -1 && event->cpu != smp_processor_id())
  1338. continue;
  1339. hwc = &event->hw;
  1340. interrupts = hwc->interrupts;
  1341. hwc->interrupts = 0;
  1342. /*
  1343. * unthrottle events on the tick
  1344. */
  1345. if (interrupts == MAX_INTERRUPTS) {
  1346. perf_log_throttle(event, 1);
  1347. event->pmu->start(event, 0);
  1348. }
  1349. if (!event->attr.freq || !event->attr.sample_freq)
  1350. continue;
  1351. event->pmu->read(event);
  1352. now = local64_read(&event->count);
  1353. delta = now - hwc->freq_count_stamp;
  1354. hwc->freq_count_stamp = now;
  1355. if (delta > 0)
  1356. perf_adjust_period(event, period, delta);
  1357. }
  1358. raw_spin_unlock(&ctx->lock);
  1359. }
  1360. /*
  1361. * Round-robin a context's events:
  1362. */
  1363. static void rotate_ctx(struct perf_event_context *ctx)
  1364. {
  1365. raw_spin_lock(&ctx->lock);
  1366. /* Rotate the first entry last of non-pinned groups */
  1367. list_rotate_left(&ctx->flexible_groups);
  1368. raw_spin_unlock(&ctx->lock);
  1369. }
  1370. /*
  1371. * perf_pmu_rotate_start() and perf_rotate_context() are fully serialized
  1372. * because they're strictly cpu affine and rotate_start is called with IRQs
  1373. * disabled, while rotate_context is called from IRQ context.
  1374. */
  1375. static void perf_rotate_context(struct perf_cpu_context *cpuctx)
  1376. {
  1377. u64 interval = (u64)cpuctx->jiffies_interval * TICK_NSEC;
  1378. struct perf_event_context *ctx = NULL;
  1379. int rotate = 0, remove = 1;
  1380. if (cpuctx->ctx.nr_events) {
  1381. remove = 0;
  1382. if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
  1383. rotate = 1;
  1384. }
  1385. ctx = cpuctx->task_ctx;
  1386. if (ctx && ctx->nr_events) {
  1387. remove = 0;
  1388. if (ctx->nr_events != ctx->nr_active)
  1389. rotate = 1;
  1390. }
  1391. perf_pmu_disable(cpuctx->ctx.pmu);
  1392. perf_ctx_adjust_freq(&cpuctx->ctx, interval);
  1393. if (ctx)
  1394. perf_ctx_adjust_freq(ctx, interval);
  1395. if (!rotate)
  1396. goto done;
  1397. cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
  1398. if (ctx)
  1399. task_ctx_sched_out(ctx, EVENT_FLEXIBLE);
  1400. rotate_ctx(&cpuctx->ctx);
  1401. if (ctx)
  1402. rotate_ctx(ctx);
  1403. cpu_ctx_sched_in(cpuctx, EVENT_FLEXIBLE);
  1404. if (ctx)
  1405. task_ctx_sched_in(ctx, EVENT_FLEXIBLE);
  1406. done:
  1407. if (remove)
  1408. list_del_init(&cpuctx->rotation_list);
  1409. perf_pmu_enable(cpuctx->ctx.pmu);
  1410. }
  1411. void perf_event_task_tick(void)
  1412. {
  1413. struct list_head *head = &__get_cpu_var(rotation_list);
  1414. struct perf_cpu_context *cpuctx, *tmp;
  1415. WARN_ON(!irqs_disabled());
  1416. list_for_each_entry_safe(cpuctx, tmp, head, rotation_list) {
  1417. if (cpuctx->jiffies_interval == 1 ||
  1418. !(jiffies % cpuctx->jiffies_interval))
  1419. perf_rotate_context(cpuctx);
  1420. }
  1421. }
  1422. static int event_enable_on_exec(struct perf_event *event,
  1423. struct perf_event_context *ctx)
  1424. {
  1425. if (!event->attr.enable_on_exec)
  1426. return 0;
  1427. event->attr.enable_on_exec = 0;
  1428. if (event->state >= PERF_EVENT_STATE_INACTIVE)
  1429. return 0;
  1430. __perf_event_mark_enabled(event, ctx);
  1431. return 1;
  1432. }
  1433. /*
  1434. * Enable all of a task's events that have been marked enable-on-exec.
  1435. * This expects task == current.
  1436. */
  1437. static void perf_event_enable_on_exec(struct perf_event_context *ctx)
  1438. {
  1439. struct perf_event *event;
  1440. unsigned long flags;
  1441. int enabled = 0;
  1442. int ret;
  1443. local_irq_save(flags);
  1444. if (!ctx || !ctx->nr_events)
  1445. goto out;
  1446. task_ctx_sched_out(ctx, EVENT_ALL);
  1447. raw_spin_lock(&ctx->lock);
  1448. list_for_each_entry(event, &ctx->pinned_groups, group_entry) {
  1449. ret = event_enable_on_exec(event, ctx);
  1450. if (ret)
  1451. enabled = 1;
  1452. }
  1453. list_for_each_entry(event, &ctx->flexible_groups, group_entry) {
  1454. ret = event_enable_on_exec(event, ctx);
  1455. if (ret)
  1456. enabled = 1;
  1457. }
  1458. /*
  1459. * Unclone this context if we enabled any event.
  1460. */
  1461. if (enabled)
  1462. unclone_ctx(ctx);
  1463. raw_spin_unlock(&ctx->lock);
  1464. perf_event_context_sched_in(ctx);
  1465. out:
  1466. local_irq_restore(flags);
  1467. }
  1468. /*
  1469. * Cross CPU call to read the hardware event
  1470. */
  1471. static void __perf_event_read(void *info)
  1472. {
  1473. struct perf_event *event = info;
  1474. struct perf_event_context *ctx = event->ctx;
  1475. struct perf_cpu_context *cpuctx = __get_cpu_context(ctx);
  1476. /*
  1477. * If this is a task context, we need to check whether it is
  1478. * the current task context of this cpu. If not it has been
  1479. * scheduled out before the smp call arrived. In that case
  1480. * event->count would have been updated to a recent sample
  1481. * when the event was scheduled out.
  1482. */
  1483. if (ctx->task && cpuctx->task_ctx != ctx)
  1484. return;
  1485. raw_spin_lock(&ctx->lock);
  1486. update_context_time(ctx);
  1487. update_event_times(event);
  1488. raw_spin_unlock(&ctx->lock);
  1489. event->pmu->read(event);
  1490. }
  1491. static inline u64 perf_event_count(struct perf_event *event)
  1492. {
  1493. return local64_read(&event->count) + atomic64_read(&event->child_count);
  1494. }
  1495. static u64 perf_event_read(struct perf_event *event)
  1496. {
  1497. /*
  1498. * If event is enabled and currently active on a CPU, update the
  1499. * value in the event structure:
  1500. */
  1501. if (event->state == PERF_EVENT_STATE_ACTIVE) {
  1502. smp_call_function_single(event->oncpu,
  1503. __perf_event_read, event, 1);
  1504. } else if (event->state == PERF_EVENT_STATE_INACTIVE) {
  1505. struct perf_event_context *ctx = event->ctx;
  1506. unsigned long flags;
  1507. raw_spin_lock_irqsave(&ctx->lock, flags);
  1508. /*
  1509. * may read while context is not active
  1510. * (e.g., thread is blocked), in that case
  1511. * we cannot update context time
  1512. */
  1513. if (ctx->is_active)
  1514. update_context_time(ctx);
  1515. update_event_times(event);
  1516. raw_spin_unlock_irqrestore(&ctx->lock, flags);
  1517. }
  1518. return perf_event_count(event);
  1519. }
  1520. /*
  1521. * Callchain support
  1522. */
  1523. struct callchain_cpus_entries {
  1524. struct rcu_head rcu_head;
  1525. struct perf_callchain_entry *cpu_entries[0];
  1526. };
  1527. static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
  1528. static atomic_t nr_callchain_events;
  1529. static DEFINE_MUTEX(callchain_mutex);
  1530. struct callchain_cpus_entries *callchain_cpus_entries;
  1531. __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
  1532. struct pt_regs *regs)
  1533. {
  1534. }
  1535. __weak void perf_callchain_user(struct perf_callchain_entry *entry,
  1536. struct pt_regs *regs)
  1537. {
  1538. }
  1539. static void release_callchain_buffers_rcu(struct rcu_head *head)
  1540. {
  1541. struct callchain_cpus_entries *entries;
  1542. int cpu;
  1543. entries = container_of(head, struct callchain_cpus_entries, rcu_head);
  1544. for_each_possible_cpu(cpu)
  1545. kfree(entries->cpu_entries[cpu]);
  1546. kfree(entries);
  1547. }
  1548. static void release_callchain_buffers(void)
  1549. {
  1550. struct callchain_cpus_entries *entries;
  1551. entries = callchain_cpus_entries;
  1552. rcu_assign_pointer(callchain_cpus_entries, NULL);
  1553. call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
  1554. }
  1555. static int alloc_callchain_buffers(void)
  1556. {
  1557. int cpu;
  1558. int size;
  1559. struct callchain_cpus_entries *entries;
  1560. /*
  1561. * We can't use the percpu allocation API for data that can be
  1562. * accessed from NMI. Use a temporary manual per cpu allocation
  1563. * until that gets sorted out.
  1564. */
  1565. size = sizeof(*entries) + sizeof(struct perf_callchain_entry *) *
  1566. num_possible_cpus();
  1567. entries = kzalloc(size, GFP_KERNEL);
  1568. if (!entries)
  1569. return -ENOMEM;
  1570. size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
  1571. for_each_possible_cpu(cpu) {
  1572. entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
  1573. cpu_to_node(cpu));
  1574. if (!entries->cpu_entries[cpu])
  1575. goto fail;
  1576. }
  1577. rcu_assign_pointer(callchain_cpus_entries, entries);
  1578. return 0;
  1579. fail:
  1580. for_each_possible_cpu(cpu)
  1581. kfree(entries->cpu_entries[cpu]);
  1582. kfree(entries);
  1583. return -ENOMEM;
  1584. }
  1585. static int get_callchain_buffers(void)
  1586. {
  1587. int err = 0;
  1588. int count;
  1589. mutex_lock(&callchain_mutex);
  1590. count = atomic_inc_return(&nr_callchain_events);
  1591. if (WARN_ON_ONCE(count < 1)) {
  1592. err = -EINVAL;
  1593. goto exit;
  1594. }
  1595. if (count > 1) {
  1596. /* If the allocation failed, give up */
  1597. if (!callchain_cpus_entries)
  1598. err = -ENOMEM;
  1599. goto exit;
  1600. }
  1601. err = alloc_callchain_buffers();
  1602. if (err)
  1603. release_callchain_buffers();
  1604. exit:
  1605. mutex_unlock(&callchain_mutex);
  1606. return err;
  1607. }
  1608. static void put_callchain_buffers(void)
  1609. {
  1610. if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
  1611. release_callchain_buffers();
  1612. mutex_unlock(&callchain_mutex);
  1613. }
  1614. }
  1615. static int get_recursion_context(int *recursion)
  1616. {
  1617. int rctx;
  1618. if (in_nmi())
  1619. rctx = 3;
  1620. else if (in_irq())
  1621. rctx = 2;
  1622. else if (in_softirq())
  1623. rctx = 1;
  1624. else
  1625. rctx = 0;
  1626. if (recursion[rctx])
  1627. return -1;
  1628. recursion[rctx]++;
  1629. barrier();
  1630. return rctx;
  1631. }
  1632. static inline void put_recursion_context(int *recursion, int rctx)
  1633. {
  1634. barrier();
  1635. recursion[rctx]--;
  1636. }
  1637. static struct perf_callchain_entry *get_callchain_entry(int *rctx)
  1638. {
  1639. int cpu;
  1640. struct callchain_cpus_entries *entries;
  1641. *rctx = get_recursion_context(__get_cpu_var(callchain_recursion));
  1642. if (*rctx == -1)
  1643. return NULL;
  1644. entries = rcu_dereference(callchain_cpus_entries);
  1645. if (!entries)
  1646. return NULL;
  1647. cpu = smp_processor_id();
  1648. return &entries->cpu_entries[cpu][*rctx];
  1649. }
  1650. static void
  1651. put_callchain_entry(int rctx)
  1652. {
  1653. put_recursion_context(__get_cpu_var(callchain_recursion), rctx);
  1654. }
  1655. static struct perf_callchain_entry *perf_callchain(struct pt_regs *regs)
  1656. {
  1657. int rctx;
  1658. struct perf_callchain_entry *entry;
  1659. entry = get_callchain_entry(&rctx);
  1660. if (rctx == -1)
  1661. return NULL;
  1662. if (!entry)
  1663. goto exit_put;
  1664. entry->nr = 0;
  1665. if (!user_mode(regs)) {
  1666. perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
  1667. perf_callchain_kernel(entry, regs);
  1668. if (current->mm)
  1669. regs = task_pt_regs(current);
  1670. else
  1671. regs = NULL;
  1672. }
  1673. if (regs) {
  1674. perf_callchain_store(entry, PERF_CONTEXT_USER);
  1675. perf_callchain_user(entry, regs);
  1676. }
  1677. exit_put:
  1678. put_callchain_entry(rctx);
  1679. return entry;
  1680. }
  1681. /*
  1682. * Initialize the perf_event context in a task_struct:
  1683. */
  1684. static void __perf_event_init_context(struct perf_event_context *ctx)
  1685. {
  1686. raw_spin_lock_init(&ctx->lock);
  1687. mutex_init(&ctx->mutex);
  1688. INIT_LIST_HEAD(&ctx->pinned_groups);
  1689. INIT_LIST_HEAD(&ctx->flexible_groups);
  1690. INIT_LIST_HEAD(&ctx->event_list);
  1691. atomic_set(&ctx->refcount, 1);
  1692. }
  1693. static struct perf_event_context *
  1694. alloc_perf_context(struct pmu *pmu, struct task_struct *task)
  1695. {
  1696. struct perf_event_context *ctx;
  1697. ctx = kzalloc(sizeof(struct perf_event_context), GFP_KERNEL);
  1698. if (!ctx)
  1699. return NULL;
  1700. __perf_event_init_context(ctx);
  1701. if (task) {
  1702. ctx->task = task;
  1703. get_task_struct(task);
  1704. }
  1705. ctx->pmu = pmu;
  1706. return ctx;
  1707. }
  1708. static struct task_struct *
  1709. find_lively_task_by_vpid(pid_t vpid)
  1710. {
  1711. struct task_struct *task;
  1712. int err;
  1713. rcu_read_lock();
  1714. if (!vpid)
  1715. task = current;
  1716. else
  1717. task = find_task_by_vpid(vpid);
  1718. if (task)
  1719. get_task_struct(task);
  1720. rcu_read_unlock();
  1721. if (!task)
  1722. return ERR_PTR(-ESRCH);
  1723. /*
  1724. * Can't attach events to a dying task.
  1725. */
  1726. err = -ESRCH;
  1727. if (task->flags & PF_EXITING)
  1728. goto errout;
  1729. /* Reuse ptrace permission checks for now. */
  1730. err = -EACCES;
  1731. if (!ptrace_may_access(task, PTRACE_MODE_READ))
  1732. goto errout;
  1733. return task;
  1734. errout:
  1735. put_task_struct(task);
  1736. return ERR_PTR(err);
  1737. }
  1738. static struct perf_event_context *
  1739. find_get_context(struct pmu *pmu, struct task_struct *task, int cpu)
  1740. {
  1741. struct perf_event_context *ctx;
  1742. struct perf_cpu_context *cpuctx;
  1743. unsigned long flags;
  1744. int ctxn, err;
  1745. if (!task && cpu != -1) {
  1746. /* Must be root to operate on a CPU event: */
  1747. if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
  1748. return ERR_PTR(-EACCES);
  1749. if (cpu < 0 || cpu >= nr_cpumask_bits)
  1750. return ERR_PTR(-EINVAL);
  1751. /*
  1752. * We could be clever and allow to attach a event to an
  1753. * offline CPU and activate it when the CPU comes up, but
  1754. * that's for later.
  1755. */
  1756. if (!cpu_online(cpu))
  1757. return ERR_PTR(-ENODEV);
  1758. cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
  1759. ctx = &cpuctx->ctx;
  1760. get_ctx(ctx);
  1761. return ctx;
  1762. }
  1763. err = -EINVAL;
  1764. ctxn = pmu->task_ctx_nr;
  1765. if (ctxn < 0)
  1766. goto errout;
  1767. retry:
  1768. ctx = perf_lock_task_context(task, ctxn, &flags);
  1769. if (ctx) {
  1770. unclone_ctx(ctx);
  1771. raw_spin_unlock_irqrestore(&ctx->lock, flags);
  1772. }
  1773. if (!ctx) {
  1774. ctx = alloc_perf_context(pmu, task);
  1775. err = -ENOMEM;
  1776. if (!ctx)
  1777. goto errout;
  1778. get_ctx(ctx);
  1779. if (cmpxchg(&task->perf_event_ctxp[ctxn], NULL, ctx)) {
  1780. /*
  1781. * We raced with some other task; use
  1782. * the context they set.
  1783. */
  1784. put_task_struct(task);
  1785. kfree(ctx);
  1786. goto retry;
  1787. }
  1788. }
  1789. return ctx;
  1790. errout:
  1791. return ERR_PTR(err);
  1792. }
  1793. static void perf_event_free_filter(struct perf_event *event);
  1794. static void free_event_rcu(struct rcu_head *head)
  1795. {
  1796. struct perf_event *event;
  1797. event = container_of(head, struct perf_event, rcu_head);
  1798. if (event->ns)
  1799. put_pid_ns(event->ns);
  1800. perf_event_free_filter(event);
  1801. kfree(event);
  1802. }
  1803. static void perf_buffer_put(struct perf_buffer *buffer);
  1804. static void free_event(struct perf_event *event)
  1805. {
  1806. irq_work_sync(&event->pending);
  1807. if (!event->parent) {
  1808. if (event->attach_state & PERF_ATTACH_TASK)
  1809. jump_label_dec(&perf_task_events);
  1810. if (event->attr.mmap || event->attr.mmap_data)
  1811. atomic_dec(&nr_mmap_events);
  1812. if (event->attr.comm)
  1813. atomic_dec(&nr_comm_events);
  1814. if (event->attr.task)
  1815. atomic_dec(&nr_task_events);
  1816. if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN)
  1817. put_callchain_buffers();
  1818. }
  1819. if (event->buffer) {
  1820. perf_buffer_put(event->buffer);
  1821. event->buffer = NULL;
  1822. }
  1823. if (event->destroy)
  1824. event->destroy(event);
  1825. if (event->ctx)
  1826. put_ctx(event->ctx);
  1827. call_rcu(&event->rcu_head, free_event_rcu);
  1828. }
  1829. int perf_event_release_kernel(struct perf_event *event)
  1830. {
  1831. struct perf_event_context *ctx = event->ctx;
  1832. /*
  1833. * Remove from the PMU, can't get re-enabled since we got
  1834. * here because the last ref went.
  1835. */
  1836. perf_event_disable(event);
  1837. WARN_ON_ONCE(ctx->parent_ctx);
  1838. /*
  1839. * There are two ways this annotation is useful:
  1840. *
  1841. * 1) there is a lock recursion from perf_event_exit_task
  1842. * see the comment there.
  1843. *
  1844. * 2) there is a lock-inversion with mmap_sem through
  1845. * perf_event_read_group(), which takes faults while
  1846. * holding ctx->mutex, however this is called after
  1847. * the last filedesc died, so there is no possibility
  1848. * to trigger the AB-BA case.
  1849. */
  1850. mutex_lock_nested(&ctx->mutex, SINGLE_DEPTH_NESTING);
  1851. raw_spin_lock_irq(&ctx->lock);
  1852. perf_group_detach(event);
  1853. list_del_event(event, ctx);
  1854. raw_spin_unlock_irq(&ctx->lock);
  1855. mutex_unlock(&ctx->mutex);
  1856. mutex_lock(&event->owner->perf_event_mutex);
  1857. list_del_init(&event->owner_entry);
  1858. mutex_unlock(&event->owner->perf_event_mutex);
  1859. put_task_struct(event->owner);
  1860. free_event(event);
  1861. return 0;
  1862. }
  1863. EXPORT_SYMBOL_GPL(perf_event_release_kernel);
  1864. /*
  1865. * Called when the last reference to the file is gone.
  1866. */
  1867. static int perf_release(struct inode *inode, struct file *file)
  1868. {
  1869. struct perf_event *event = file->private_data;
  1870. file->private_data = NULL;
  1871. return perf_event_release_kernel(event);
  1872. }
  1873. static int perf_event_read_size(struct perf_event *event)
  1874. {
  1875. int entry = sizeof(u64); /* value */
  1876. int size = 0;
  1877. int nr = 1;
  1878. if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  1879. size += sizeof(u64);
  1880. if (event->attr.read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  1881. size += sizeof(u64);
  1882. if (event->attr.read_format & PERF_FORMAT_ID)
  1883. entry += sizeof(u64);
  1884. if (event->attr.read_format & PERF_FORMAT_GROUP) {
  1885. nr += event->group_leader->nr_siblings;
  1886. size += sizeof(u64);
  1887. }
  1888. size += entry * nr;
  1889. return size;
  1890. }
  1891. u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
  1892. {
  1893. struct perf_event *child;
  1894. u64 total = 0;
  1895. *enabled = 0;
  1896. *running = 0;
  1897. mutex_lock(&event->child_mutex);
  1898. total += perf_event_read(event);
  1899. *enabled += event->total_time_enabled +
  1900. atomic64_read(&event->child_total_time_enabled);
  1901. *running += event->total_time_running +
  1902. atomic64_read(&event->child_total_time_running);
  1903. list_for_each_entry(child, &event->child_list, child_list) {
  1904. total += perf_event_read(child);
  1905. *enabled += child->total_time_enabled;
  1906. *running += child->total_time_running;
  1907. }
  1908. mutex_unlock(&event->child_mutex);
  1909. return total;
  1910. }
  1911. EXPORT_SYMBOL_GPL(perf_event_read_value);
  1912. static int perf_event_read_group(struct perf_event *event,
  1913. u64 read_format, char __user *buf)
  1914. {
  1915. struct perf_event *leader = event->group_leader, *sub;
  1916. int n = 0, size = 0, ret = -EFAULT;
  1917. struct perf_event_context *ctx = leader->ctx;
  1918. u64 values[5];
  1919. u64 count, enabled, running;
  1920. mutex_lock(&ctx->mutex);
  1921. count = perf_event_read_value(leader, &enabled, &running);
  1922. values[n++] = 1 + leader->nr_siblings;
  1923. if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  1924. values[n++] = enabled;
  1925. if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  1926. values[n++] = running;
  1927. values[n++] = count;
  1928. if (read_format & PERF_FORMAT_ID)
  1929. values[n++] = primary_event_id(leader);
  1930. size = n * sizeof(u64);
  1931. if (copy_to_user(buf, values, size))
  1932. goto unlock;
  1933. ret = size;
  1934. list_for_each_entry(sub, &leader->sibling_list, group_entry) {
  1935. n = 0;
  1936. values[n++] = perf_event_read_value(sub, &enabled, &running);
  1937. if (read_format & PERF_FORMAT_ID)
  1938. values[n++] = primary_event_id(sub);
  1939. size = n * sizeof(u64);
  1940. if (copy_to_user(buf + ret, values, size)) {
  1941. ret = -EFAULT;
  1942. goto unlock;
  1943. }
  1944. ret += size;
  1945. }
  1946. unlock:
  1947. mutex_unlock(&ctx->mutex);
  1948. return ret;
  1949. }
  1950. static int perf_event_read_one(struct perf_event *event,
  1951. u64 read_format, char __user *buf)
  1952. {
  1953. u64 enabled, running;
  1954. u64 values[4];
  1955. int n = 0;
  1956. values[n++] = perf_event_read_value(event, &enabled, &running);
  1957. if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  1958. values[n++] = enabled;
  1959. if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  1960. values[n++] = running;
  1961. if (read_format & PERF_FORMAT_ID)
  1962. values[n++] = primary_event_id(event);
  1963. if (copy_to_user(buf, values, n * sizeof(u64)))
  1964. return -EFAULT;
  1965. return n * sizeof(u64);
  1966. }
  1967. /*
  1968. * Read the performance event - simple non blocking version for now
  1969. */
  1970. static ssize_t
  1971. perf_read_hw(struct perf_event *event, char __user *buf, size_t count)
  1972. {
  1973. u64 read_format = event->attr.read_format;
  1974. int ret;
  1975. /*
  1976. * Return end-of-file for a read on a event that is in
  1977. * error state (i.e. because it was pinned but it couldn't be
  1978. * scheduled on to the CPU at some point).
  1979. */
  1980. if (event->state == PERF_EVENT_STATE_ERROR)
  1981. return 0;
  1982. if (count < perf_event_read_size(event))
  1983. return -ENOSPC;
  1984. WARN_ON_ONCE(event->ctx->parent_ctx);
  1985. if (read_format & PERF_FORMAT_GROUP)
  1986. ret = perf_event_read_group(event, read_format, buf);
  1987. else
  1988. ret = perf_event_read_one(event, read_format, buf);
  1989. return ret;
  1990. }
  1991. static ssize_t
  1992. perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
  1993. {
  1994. struct perf_event *event = file->private_data;
  1995. return perf_read_hw(event, buf, count);
  1996. }
  1997. static unsigned int perf_poll(struct file *file, poll_table *wait)
  1998. {
  1999. struct perf_event *event = file->private_data;
  2000. struct perf_buffer *buffer;
  2001. unsigned int events = POLL_HUP;
  2002. rcu_read_lock();
  2003. buffer = rcu_dereference(event->buffer);
  2004. if (buffer)
  2005. events = atomic_xchg(&buffer->poll, 0);
  2006. rcu_read_unlock();
  2007. poll_wait(file, &event->waitq, wait);
  2008. return events;
  2009. }
  2010. static void perf_event_reset(struct perf_event *event)
  2011. {
  2012. (void)perf_event_read(event);
  2013. local64_set(&event->count, 0);
  2014. perf_event_update_userpage(event);
  2015. }
  2016. /*
  2017. * Holding the top-level event's child_mutex means that any
  2018. * descendant process that has inherited this event will block
  2019. * in sync_child_event if it goes to exit, thus satisfying the
  2020. * task existence requirements of perf_event_enable/disable.
  2021. */
  2022. static void perf_event_for_each_child(struct perf_event *event,
  2023. void (*func)(struct perf_event *))
  2024. {
  2025. struct perf_event *child;
  2026. WARN_ON_ONCE(event->ctx->parent_ctx);
  2027. mutex_lock(&event->child_mutex);
  2028. func(event);
  2029. list_for_each_entry(child, &event->child_list, child_list)
  2030. func(child);
  2031. mutex_unlock(&event->child_mutex);
  2032. }
  2033. static void perf_event_for_each(struct perf_event *event,
  2034. void (*func)(struct perf_event *))
  2035. {
  2036. struct perf_event_context *ctx = event->ctx;
  2037. struct perf_event *sibling;
  2038. WARN_ON_ONCE(ctx->parent_ctx);
  2039. mutex_lock(&ctx->mutex);
  2040. event = event->group_leader;
  2041. perf_event_for_each_child(event, func);
  2042. func(event);
  2043. list_for_each_entry(sibling, &event->sibling_list, group_entry)
  2044. perf_event_for_each_child(event, func);
  2045. mutex_unlock(&ctx->mutex);
  2046. }
  2047. static int perf_event_period(struct perf_event *event, u64 __user *arg)
  2048. {
  2049. struct perf_event_context *ctx = event->ctx;
  2050. int ret = 0;
  2051. u64 value;
  2052. if (!event->attr.sample_period)
  2053. return -EINVAL;
  2054. if (copy_from_user(&value, arg, sizeof(value)))
  2055. return -EFAULT;
  2056. if (!value)
  2057. return -EINVAL;
  2058. raw_spin_lock_irq(&ctx->lock);
  2059. if (event->attr.freq) {
  2060. if (value > sysctl_perf_event_sample_rate) {
  2061. ret = -EINVAL;
  2062. goto unlock;
  2063. }
  2064. event->attr.sample_freq = value;
  2065. } else {
  2066. event->attr.sample_period = value;
  2067. event->hw.sample_period = value;
  2068. }
  2069. unlock:
  2070. raw_spin_unlock_irq(&ctx->lock);
  2071. return ret;
  2072. }
  2073. static const struct file_operations perf_fops;
  2074. static struct perf_event *perf_fget_light(int fd, int *fput_needed)
  2075. {
  2076. struct file *file;
  2077. file = fget_light(fd, fput_needed);
  2078. if (!file)
  2079. return ERR_PTR(-EBADF);
  2080. if (file->f_op != &perf_fops) {
  2081. fput_light(file, *fput_needed);
  2082. *fput_needed = 0;
  2083. return ERR_PTR(-EBADF);
  2084. }
  2085. return file->private_data;
  2086. }
  2087. static int perf_event_set_output(struct perf_event *event,
  2088. struct perf_event *output_event);
  2089. static int perf_event_set_filter(struct perf_event *event, void __user *arg);
  2090. static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
  2091. {
  2092. struct perf_event *event = file->private_data;
  2093. void (*func)(struct perf_event *);
  2094. u32 flags = arg;
  2095. switch (cmd) {
  2096. case PERF_EVENT_IOC_ENABLE:
  2097. func = perf_event_enable;
  2098. break;
  2099. case PERF_EVENT_IOC_DISABLE:
  2100. func = perf_event_disable;
  2101. break;
  2102. case PERF_EVENT_IOC_RESET:
  2103. func = perf_event_reset;
  2104. break;
  2105. case PERF_EVENT_IOC_REFRESH:
  2106. return perf_event_refresh(event, arg);
  2107. case PERF_EVENT_IOC_PERIOD:
  2108. return perf_event_period(event, (u64 __user *)arg);
  2109. case PERF_EVENT_IOC_SET_OUTPUT:
  2110. {
  2111. struct perf_event *output_event = NULL;
  2112. int fput_needed = 0;
  2113. int ret;
  2114. if (arg != -1) {
  2115. output_event = perf_fget_light(arg, &fput_needed);
  2116. if (IS_ERR(output_event))
  2117. return PTR_ERR(output_event);
  2118. }
  2119. ret = perf_event_set_output(event, output_event);
  2120. if (output_event)
  2121. fput_light(output_event->filp, fput_needed);
  2122. return ret;
  2123. }
  2124. case PERF_EVENT_IOC_SET_FILTER:
  2125. return perf_event_set_filter(event, (void __user *)arg);
  2126. default:
  2127. return -ENOTTY;
  2128. }
  2129. if (flags & PERF_IOC_FLAG_GROUP)
  2130. perf_event_for_each(event, func);
  2131. else
  2132. perf_event_for_each_child(event, func);
  2133. return 0;
  2134. }
  2135. int perf_event_task_enable(void)
  2136. {
  2137. struct perf_event *event;
  2138. mutex_lock(&current->perf_event_mutex);
  2139. list_for_each_entry(event, &current->perf_event_list, owner_entry)
  2140. perf_event_for_each_child(event, perf_event_enable);
  2141. mutex_unlock(&current->perf_event_mutex);
  2142. return 0;
  2143. }
  2144. int perf_event_task_disable(void)
  2145. {
  2146. struct perf_event *event;
  2147. mutex_lock(&current->perf_event_mutex);
  2148. list_for_each_entry(event, &current->perf_event_list, owner_entry)
  2149. perf_event_for_each_child(event, perf_event_disable);
  2150. mutex_unlock(&current->perf_event_mutex);
  2151. return 0;
  2152. }
  2153. #ifndef PERF_EVENT_INDEX_OFFSET
  2154. # define PERF_EVENT_INDEX_OFFSET 0
  2155. #endif
  2156. static int perf_event_index(struct perf_event *event)
  2157. {
  2158. if (event->hw.state & PERF_HES_STOPPED)
  2159. return 0;
  2160. if (event->state != PERF_EVENT_STATE_ACTIVE)
  2161. return 0;
  2162. return event->hw.idx + 1 - PERF_EVENT_INDEX_OFFSET;
  2163. }
  2164. /*
  2165. * Callers need to ensure there can be no nesting of this function, otherwise
  2166. * the seqlock logic goes bad. We can not serialize this because the arch
  2167. * code calls this from NMI context.
  2168. */
  2169. void perf_event_update_userpage(struct perf_event *event)
  2170. {
  2171. struct perf_event_mmap_page *userpg;
  2172. struct perf_buffer *buffer;
  2173. rcu_read_lock();
  2174. buffer = rcu_dereference(event->buffer);
  2175. if (!buffer)
  2176. goto unlock;
  2177. userpg = buffer->user_page;
  2178. /*
  2179. * Disable preemption so as to not let the corresponding user-space
  2180. * spin too long if we get preempted.
  2181. */
  2182. preempt_disable();
  2183. ++userpg->lock;
  2184. barrier();
  2185. userpg->index = perf_event_index(event);
  2186. userpg->offset = perf_event_count(event);
  2187. if (event->state == PERF_EVENT_STATE_ACTIVE)
  2188. userpg->offset -= local64_read(&event->hw.prev_count);
  2189. userpg->time_enabled = event->total_time_enabled +
  2190. atomic64_read(&event->child_total_time_enabled);
  2191. userpg->time_running = event->total_time_running +
  2192. atomic64_read(&event->child_total_time_running);
  2193. barrier();
  2194. ++userpg->lock;
  2195. preempt_enable();
  2196. unlock:
  2197. rcu_read_unlock();
  2198. }
  2199. static unsigned long perf_data_size(struct perf_buffer *buffer);
  2200. static void
  2201. perf_buffer_init(struct perf_buffer *buffer, long watermark, int flags)
  2202. {
  2203. long max_size = perf_data_size(buffer);
  2204. if (watermark)
  2205. buffer->watermark = min(max_size, watermark);
  2206. if (!buffer->watermark)
  2207. buffer->watermark = max_size / 2;
  2208. if (flags & PERF_BUFFER_WRITABLE)
  2209. buffer->writable = 1;
  2210. atomic_set(&buffer->refcount, 1);
  2211. }
  2212. #ifndef CONFIG_PERF_USE_VMALLOC
  2213. /*
  2214. * Back perf_mmap() with regular GFP_KERNEL-0 pages.
  2215. */
  2216. static struct page *
  2217. perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
  2218. {
  2219. if (pgoff > buffer->nr_pages)
  2220. return NULL;
  2221. if (pgoff == 0)
  2222. return virt_to_page(buffer->user_page);
  2223. return virt_to_page(buffer->data_pages[pgoff - 1]);
  2224. }
  2225. static void *perf_mmap_alloc_page(int cpu)
  2226. {
  2227. struct page *page;
  2228. int node;
  2229. node = (cpu == -1) ? cpu : cpu_to_node(cpu);
  2230. page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
  2231. if (!page)
  2232. return NULL;
  2233. return page_address(page);
  2234. }
  2235. static struct perf_buffer *
  2236. perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
  2237. {
  2238. struct perf_buffer *buffer;
  2239. unsigned long size;
  2240. int i;
  2241. size = sizeof(struct perf_buffer);
  2242. size += nr_pages * sizeof(void *);
  2243. buffer = kzalloc(size, GFP_KERNEL);
  2244. if (!buffer)
  2245. goto fail;
  2246. buffer->user_page = perf_mmap_alloc_page(cpu);
  2247. if (!buffer->user_page)
  2248. goto fail_user_page;
  2249. for (i = 0; i < nr_pages; i++) {
  2250. buffer->data_pages[i] = perf_mmap_alloc_page(cpu);
  2251. if (!buffer->data_pages[i])
  2252. goto fail_data_pages;
  2253. }
  2254. buffer->nr_pages = nr_pages;
  2255. perf_buffer_init(buffer, watermark, flags);
  2256. return buffer;
  2257. fail_data_pages:
  2258. for (i--; i >= 0; i--)
  2259. free_page((unsigned long)buffer->data_pages[i]);
  2260. free_page((unsigned long)buffer->user_page);
  2261. fail_user_page:
  2262. kfree(buffer);
  2263. fail:
  2264. return NULL;
  2265. }
  2266. static void perf_mmap_free_page(unsigned long addr)
  2267. {
  2268. struct page *page = virt_to_page((void *)addr);
  2269. page->mapping = NULL;
  2270. __free_page(page);
  2271. }
  2272. static void perf_buffer_free(struct perf_buffer *buffer)
  2273. {
  2274. int i;
  2275. perf_mmap_free_page((unsigned long)buffer->user_page);
  2276. for (i = 0; i < buffer->nr_pages; i++)
  2277. perf_mmap_free_page((unsigned long)buffer->data_pages[i]);
  2278. kfree(buffer);
  2279. }
  2280. static inline int page_order(struct perf_buffer *buffer)
  2281. {
  2282. return 0;
  2283. }
  2284. #else
  2285. /*
  2286. * Back perf_mmap() with vmalloc memory.
  2287. *
  2288. * Required for architectures that have d-cache aliasing issues.
  2289. */
  2290. static inline int page_order(struct perf_buffer *buffer)
  2291. {
  2292. return buffer->page_order;
  2293. }
  2294. static struct page *
  2295. perf_mmap_to_page(struct perf_buffer *buffer, unsigned long pgoff)
  2296. {
  2297. if (pgoff > (1UL << page_order(buffer)))
  2298. return NULL;
  2299. return vmalloc_to_page((void *)buffer->user_page + pgoff * PAGE_SIZE);
  2300. }
  2301. static void perf_mmap_unmark_page(void *addr)
  2302. {
  2303. struct page *page = vmalloc_to_page(addr);
  2304. page->mapping = NULL;
  2305. }
  2306. static void perf_buffer_free_work(struct work_struct *work)
  2307. {
  2308. struct perf_buffer *buffer;
  2309. void *base;
  2310. int i, nr;
  2311. buffer = container_of(work, struct perf_buffer, work);
  2312. nr = 1 << page_order(buffer);
  2313. base = buffer->user_page;
  2314. for (i = 0; i < nr + 1; i++)
  2315. perf_mmap_unmark_page(base + (i * PAGE_SIZE));
  2316. vfree(base);
  2317. kfree(buffer);
  2318. }
  2319. static void perf_buffer_free(struct perf_buffer *buffer)
  2320. {
  2321. schedule_work(&buffer->work);
  2322. }
  2323. static struct perf_buffer *
  2324. perf_buffer_alloc(int nr_pages, long watermark, int cpu, int flags)
  2325. {
  2326. struct perf_buffer *buffer;
  2327. unsigned long size;
  2328. void *all_buf;
  2329. size = sizeof(struct perf_buffer);
  2330. size += sizeof(void *);
  2331. buffer = kzalloc(size, GFP_KERNEL);
  2332. if (!buffer)
  2333. goto fail;
  2334. INIT_WORK(&buffer->work, perf_buffer_free_work);
  2335. all_buf = vmalloc_user((nr_pages + 1) * PAGE_SIZE);
  2336. if (!all_buf)
  2337. goto fail_all_buf;
  2338. buffer->user_page = all_buf;
  2339. buffer->data_pages[0] = all_buf + PAGE_SIZE;
  2340. buffer->page_order = ilog2(nr_pages);
  2341. buffer->nr_pages = 1;
  2342. perf_buffer_init(buffer, watermark, flags);
  2343. return buffer;
  2344. fail_all_buf:
  2345. kfree(buffer);
  2346. fail:
  2347. return NULL;
  2348. }
  2349. #endif
  2350. static unsigned long perf_data_size(struct perf_buffer *buffer)
  2351. {
  2352. return buffer->nr_pages << (PAGE_SHIFT + page_order(buffer));
  2353. }
  2354. static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
  2355. {
  2356. struct perf_event *event = vma->vm_file->private_data;
  2357. struct perf_buffer *buffer;
  2358. int ret = VM_FAULT_SIGBUS;
  2359. if (vmf->flags & FAULT_FLAG_MKWRITE) {
  2360. if (vmf->pgoff == 0)
  2361. ret = 0;
  2362. return ret;
  2363. }
  2364. rcu_read_lock();
  2365. buffer = rcu_dereference(event->buffer);
  2366. if (!buffer)
  2367. goto unlock;
  2368. if (vmf->pgoff && (vmf->flags & FAULT_FLAG_WRITE))
  2369. goto unlock;
  2370. vmf->page = perf_mmap_to_page(buffer, vmf->pgoff);
  2371. if (!vmf->page)
  2372. goto unlock;
  2373. get_page(vmf->page);
  2374. vmf->page->mapping = vma->vm_file->f_mapping;
  2375. vmf->page->index = vmf->pgoff;
  2376. ret = 0;
  2377. unlock:
  2378. rcu_read_unlock();
  2379. return ret;
  2380. }
  2381. static void perf_buffer_free_rcu(struct rcu_head *rcu_head)
  2382. {
  2383. struct perf_buffer *buffer;
  2384. buffer = container_of(rcu_head, struct perf_buffer, rcu_head);
  2385. perf_buffer_free(buffer);
  2386. }
  2387. static struct perf_buffer *perf_buffer_get(struct perf_event *event)
  2388. {
  2389. struct perf_buffer *buffer;
  2390. rcu_read_lock();
  2391. buffer = rcu_dereference(event->buffer);
  2392. if (buffer) {
  2393. if (!atomic_inc_not_zero(&buffer->refcount))
  2394. buffer = NULL;
  2395. }
  2396. rcu_read_unlock();
  2397. return buffer;
  2398. }
  2399. static void perf_buffer_put(struct perf_buffer *buffer)
  2400. {
  2401. if (!atomic_dec_and_test(&buffer->refcount))
  2402. return;
  2403. call_rcu(&buffer->rcu_head, perf_buffer_free_rcu);
  2404. }
  2405. static void perf_mmap_open(struct vm_area_struct *vma)
  2406. {
  2407. struct perf_event *event = vma->vm_file->private_data;
  2408. atomic_inc(&event->mmap_count);
  2409. }
  2410. static void perf_mmap_close(struct vm_area_struct *vma)
  2411. {
  2412. struct perf_event *event = vma->vm_file->private_data;
  2413. if (atomic_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex)) {
  2414. unsigned long size = perf_data_size(event->buffer);
  2415. struct user_struct *user = event->mmap_user;
  2416. struct perf_buffer *buffer = event->buffer;
  2417. atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
  2418. vma->vm_mm->locked_vm -= event->mmap_locked;
  2419. rcu_assign_pointer(event->buffer, NULL);
  2420. mutex_unlock(&event->mmap_mutex);
  2421. perf_buffer_put(buffer);
  2422. free_uid(user);
  2423. }
  2424. }
  2425. static const struct vm_operations_struct perf_mmap_vmops = {
  2426. .open = perf_mmap_open,
  2427. .close = perf_mmap_close,
  2428. .fault = perf_mmap_fault,
  2429. .page_mkwrite = perf_mmap_fault,
  2430. };
  2431. static int perf_mmap(struct file *file, struct vm_area_struct *vma)
  2432. {
  2433. struct perf_event *event = file->private_data;
  2434. unsigned long user_locked, user_lock_limit;
  2435. struct user_struct *user = current_user();
  2436. unsigned long locked, lock_limit;
  2437. struct perf_buffer *buffer;
  2438. unsigned long vma_size;
  2439. unsigned long nr_pages;
  2440. long user_extra, extra;
  2441. int ret = 0, flags = 0;
  2442. /*
  2443. * Don't allow mmap() of inherited per-task counters. This would
  2444. * create a performance issue due to all children writing to the
  2445. * same buffer.
  2446. */
  2447. if (event->cpu == -1 && event->attr.inherit)
  2448. return -EINVAL;
  2449. if (!(vma->vm_flags & VM_SHARED))
  2450. return -EINVAL;
  2451. vma_size = vma->vm_end - vma->vm_start;
  2452. nr_pages = (vma_size / PAGE_SIZE) - 1;
  2453. /*
  2454. * If we have buffer pages ensure they're a power-of-two number, so we
  2455. * can do bitmasks instead of modulo.
  2456. */
  2457. if (nr_pages != 0 && !is_power_of_2(nr_pages))
  2458. return -EINVAL;
  2459. if (vma_size != PAGE_SIZE * (1 + nr_pages))
  2460. return -EINVAL;
  2461. if (vma->vm_pgoff != 0)
  2462. return -EINVAL;
  2463. WARN_ON_ONCE(event->ctx->parent_ctx);
  2464. mutex_lock(&event->mmap_mutex);
  2465. if (event->buffer) {
  2466. if (event->buffer->nr_pages == nr_pages)
  2467. atomic_inc(&event->buffer->refcount);
  2468. else
  2469. ret = -EINVAL;
  2470. goto unlock;
  2471. }
  2472. user_extra = nr_pages + 1;
  2473. user_lock_limit = sysctl_perf_event_mlock >> (PAGE_SHIFT - 10);
  2474. /*
  2475. * Increase the limit linearly with more CPUs:
  2476. */
  2477. user_lock_limit *= num_online_cpus();
  2478. user_locked = atomic_long_read(&user->locked_vm) + user_extra;
  2479. extra = 0;
  2480. if (user_locked > user_lock_limit)
  2481. extra = user_locked - user_lock_limit;
  2482. lock_limit = rlimit(RLIMIT_MEMLOCK);
  2483. lock_limit >>= PAGE_SHIFT;
  2484. locked = vma->vm_mm->locked_vm + extra;
  2485. if ((locked > lock_limit) && perf_paranoid_tracepoint_raw() &&
  2486. !capable(CAP_IPC_LOCK)) {
  2487. ret = -EPERM;
  2488. goto unlock;
  2489. }
  2490. WARN_ON(event->buffer);
  2491. if (vma->vm_flags & VM_WRITE)
  2492. flags |= PERF_BUFFER_WRITABLE;
  2493. buffer = perf_buffer_alloc(nr_pages, event->attr.wakeup_watermark,
  2494. event->cpu, flags);
  2495. if (!buffer) {
  2496. ret = -ENOMEM;
  2497. goto unlock;
  2498. }
  2499. rcu_assign_pointer(event->buffer, buffer);
  2500. atomic_long_add(user_extra, &user->locked_vm);
  2501. event->mmap_locked = extra;
  2502. event->mmap_user = get_current_user();
  2503. vma->vm_mm->locked_vm += event->mmap_locked;
  2504. unlock:
  2505. if (!ret)
  2506. atomic_inc(&event->mmap_count);
  2507. mutex_unlock(&event->mmap_mutex);
  2508. vma->vm_flags |= VM_RESERVED;
  2509. vma->vm_ops = &perf_mmap_vmops;
  2510. return ret;
  2511. }
  2512. static int perf_fasync(int fd, struct file *filp, int on)
  2513. {
  2514. struct inode *inode = filp->f_path.dentry->d_inode;
  2515. struct perf_event *event = filp->private_data;
  2516. int retval;
  2517. mutex_lock(&inode->i_mutex);
  2518. retval = fasync_helper(fd, filp, on, &event->fasync);
  2519. mutex_unlock(&inode->i_mutex);
  2520. if (retval < 0)
  2521. return retval;
  2522. return 0;
  2523. }
  2524. static const struct file_operations perf_fops = {
  2525. .llseek = no_llseek,
  2526. .release = perf_release,
  2527. .read = perf_read,
  2528. .poll = perf_poll,
  2529. .unlocked_ioctl = perf_ioctl,
  2530. .compat_ioctl = perf_ioctl,
  2531. .mmap = perf_mmap,
  2532. .fasync = perf_fasync,
  2533. };
  2534. /*
  2535. * Perf event wakeup
  2536. *
  2537. * If there's data, ensure we set the poll() state and publish everything
  2538. * to user-space before waking everybody up.
  2539. */
  2540. void perf_event_wakeup(struct perf_event *event)
  2541. {
  2542. wake_up_all(&event->waitq);
  2543. if (event->pending_kill) {
  2544. kill_fasync(&event->fasync, SIGIO, event->pending_kill);
  2545. event->pending_kill = 0;
  2546. }
  2547. }
  2548. static void perf_pending_event(struct irq_work *entry)
  2549. {
  2550. struct perf_event *event = container_of(entry,
  2551. struct perf_event, pending);
  2552. if (event->pending_disable) {
  2553. event->pending_disable = 0;
  2554. __perf_event_disable(event);
  2555. }
  2556. if (event->pending_wakeup) {
  2557. event->pending_wakeup = 0;
  2558. perf_event_wakeup(event);
  2559. }
  2560. }
  2561. /*
  2562. * We assume there is only KVM supporting the callbacks.
  2563. * Later on, we might change it to a list if there is
  2564. * another virtualization implementation supporting the callbacks.
  2565. */
  2566. struct perf_guest_info_callbacks *perf_guest_cbs;
  2567. int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
  2568. {
  2569. perf_guest_cbs = cbs;
  2570. return 0;
  2571. }
  2572. EXPORT_SYMBOL_GPL(perf_register_guest_info_callbacks);
  2573. int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *cbs)
  2574. {
  2575. perf_guest_cbs = NULL;
  2576. return 0;
  2577. }
  2578. EXPORT_SYMBOL_GPL(perf_unregister_guest_info_callbacks);
  2579. /*
  2580. * Output
  2581. */
  2582. static bool perf_output_space(struct perf_buffer *buffer, unsigned long tail,
  2583. unsigned long offset, unsigned long head)
  2584. {
  2585. unsigned long mask;
  2586. if (!buffer->writable)
  2587. return true;
  2588. mask = perf_data_size(buffer) - 1;
  2589. offset = (offset - tail) & mask;
  2590. head = (head - tail) & mask;
  2591. if ((int)(head - offset) < 0)
  2592. return false;
  2593. return true;
  2594. }
  2595. static void perf_output_wakeup(struct perf_output_handle *handle)
  2596. {
  2597. atomic_set(&handle->buffer->poll, POLL_IN);
  2598. if (handle->nmi) {
  2599. handle->event->pending_wakeup = 1;
  2600. irq_work_queue(&handle->event->pending);
  2601. } else
  2602. perf_event_wakeup(handle->event);
  2603. }
  2604. /*
  2605. * We need to ensure a later event_id doesn't publish a head when a former
  2606. * event isn't done writing. However since we need to deal with NMIs we
  2607. * cannot fully serialize things.
  2608. *
  2609. * We only publish the head (and generate a wakeup) when the outer-most
  2610. * event completes.
  2611. */
  2612. static void perf_output_get_handle(struct perf_output_handle *handle)
  2613. {
  2614. struct perf_buffer *buffer = handle->buffer;
  2615. preempt_disable();
  2616. local_inc(&buffer->nest);
  2617. handle->wakeup = local_read(&buffer->wakeup);
  2618. }
  2619. static void perf_output_put_handle(struct perf_output_handle *handle)
  2620. {
  2621. struct perf_buffer *buffer = handle->buffer;
  2622. unsigned long head;
  2623. again:
  2624. head = local_read(&buffer->head);
  2625. /*
  2626. * IRQ/NMI can happen here, which means we can miss a head update.
  2627. */
  2628. if (!local_dec_and_test(&buffer->nest))
  2629. goto out;
  2630. /*
  2631. * Publish the known good head. Rely on the full barrier implied
  2632. * by atomic_dec_and_test() order the buffer->head read and this
  2633. * write.
  2634. */
  2635. buffer->user_page->data_head = head;
  2636. /*
  2637. * Now check if we missed an update, rely on the (compiler)
  2638. * barrier in atomic_dec_and_test() to re-read buffer->head.
  2639. */
  2640. if (unlikely(head != local_read(&buffer->head))) {
  2641. local_inc(&buffer->nest);
  2642. goto again;
  2643. }
  2644. if (handle->wakeup != local_read(&buffer->wakeup))
  2645. perf_output_wakeup(handle);
  2646. out:
  2647. preempt_enable();
  2648. }
  2649. __always_inline void perf_output_copy(struct perf_output_handle *handle,
  2650. const void *buf, unsigned int len)
  2651. {
  2652. do {
  2653. unsigned long size = min_t(unsigned long, handle->size, len);
  2654. memcpy(handle->addr, buf, size);
  2655. len -= size;
  2656. handle->addr += size;
  2657. buf += size;
  2658. handle->size -= size;
  2659. if (!handle->size) {
  2660. struct perf_buffer *buffer = handle->buffer;
  2661. handle->page++;
  2662. handle->page &= buffer->nr_pages - 1;
  2663. handle->addr = buffer->data_pages[handle->page];
  2664. handle->size = PAGE_SIZE << page_order(buffer);
  2665. }
  2666. } while (len);
  2667. }
  2668. int perf_output_begin(struct perf_output_handle *handle,
  2669. struct perf_event *event, unsigned int size,
  2670. int nmi, int sample)
  2671. {
  2672. struct perf_buffer *buffer;
  2673. unsigned long tail, offset, head;
  2674. int have_lost;
  2675. struct {
  2676. struct perf_event_header header;
  2677. u64 id;
  2678. u64 lost;
  2679. } lost_event;
  2680. rcu_read_lock();
  2681. /*
  2682. * For inherited events we send all the output towards the parent.
  2683. */
  2684. if (event->parent)
  2685. event = event->parent;
  2686. buffer = rcu_dereference(event->buffer);
  2687. if (!buffer)
  2688. goto out;
  2689. handle->buffer = buffer;
  2690. handle->event = event;
  2691. handle->nmi = nmi;
  2692. handle->sample = sample;
  2693. if (!buffer->nr_pages)
  2694. goto out;
  2695. have_lost = local_read(&buffer->lost);
  2696. if (have_lost)
  2697. size += sizeof(lost_event);
  2698. perf_output_get_handle(handle);
  2699. do {
  2700. /*
  2701. * Userspace could choose to issue a mb() before updating the
  2702. * tail pointer. So that all reads will be completed before the
  2703. * write is issued.
  2704. */
  2705. tail = ACCESS_ONCE(buffer->user_page->data_tail);
  2706. smp_rmb();
  2707. offset = head = local_read(&buffer->head);
  2708. head += size;
  2709. if (unlikely(!perf_output_space(buffer, tail, offset, head)))
  2710. goto fail;
  2711. } while (local_cmpxchg(&buffer->head, offset, head) != offset);
  2712. if (head - local_read(&buffer->wakeup) > buffer->watermark)
  2713. local_add(buffer->watermark, &buffer->wakeup);
  2714. handle->page = offset >> (PAGE_SHIFT + page_order(buffer));
  2715. handle->page &= buffer->nr_pages - 1;
  2716. handle->size = offset & ((PAGE_SIZE << page_order(buffer)) - 1);
  2717. handle->addr = buffer->data_pages[handle->page];
  2718. handle->addr += handle->size;
  2719. handle->size = (PAGE_SIZE << page_order(buffer)) - handle->size;
  2720. if (have_lost) {
  2721. lost_event.header.type = PERF_RECORD_LOST;
  2722. lost_event.header.misc = 0;
  2723. lost_event.header.size = sizeof(lost_event);
  2724. lost_event.id = event->id;
  2725. lost_event.lost = local_xchg(&buffer->lost, 0);
  2726. perf_output_put(handle, lost_event);
  2727. }
  2728. return 0;
  2729. fail:
  2730. local_inc(&buffer->lost);
  2731. perf_output_put_handle(handle);
  2732. out:
  2733. rcu_read_unlock();
  2734. return -ENOSPC;
  2735. }
  2736. void perf_output_end(struct perf_output_handle *handle)
  2737. {
  2738. struct perf_event *event = handle->event;
  2739. struct perf_buffer *buffer = handle->buffer;
  2740. int wakeup_events = event->attr.wakeup_events;
  2741. if (handle->sample && wakeup_events) {
  2742. int events = local_inc_return(&buffer->events);
  2743. if (events >= wakeup_events) {
  2744. local_sub(wakeup_events, &buffer->events);
  2745. local_inc(&buffer->wakeup);
  2746. }
  2747. }
  2748. perf_output_put_handle(handle);
  2749. rcu_read_unlock();
  2750. }
  2751. static u32 perf_event_pid(struct perf_event *event, struct task_struct *p)
  2752. {
  2753. /*
  2754. * only top level events have the pid namespace they were created in
  2755. */
  2756. if (event->parent)
  2757. event = event->parent;
  2758. return task_tgid_nr_ns(p, event->ns);
  2759. }
  2760. static u32 perf_event_tid(struct perf_event *event, struct task_struct *p)
  2761. {
  2762. /*
  2763. * only top level events have the pid namespace they were created in
  2764. */
  2765. if (event->parent)
  2766. event = event->parent;
  2767. return task_pid_nr_ns(p, event->ns);
  2768. }
  2769. static void perf_output_read_one(struct perf_output_handle *handle,
  2770. struct perf_event *event)
  2771. {
  2772. u64 read_format = event->attr.read_format;
  2773. u64 values[4];
  2774. int n = 0;
  2775. values[n++] = perf_event_count(event);
  2776. if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
  2777. values[n++] = event->total_time_enabled +
  2778. atomic64_read(&event->child_total_time_enabled);
  2779. }
  2780. if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
  2781. values[n++] = event->total_time_running +
  2782. atomic64_read(&event->child_total_time_running);
  2783. }
  2784. if (read_format & PERF_FORMAT_ID)
  2785. values[n++] = primary_event_id(event);
  2786. perf_output_copy(handle, values, n * sizeof(u64));
  2787. }
  2788. /*
  2789. * XXX PERF_FORMAT_GROUP vs inherited events seems difficult.
  2790. */
  2791. static void perf_output_read_group(struct perf_output_handle *handle,
  2792. struct perf_event *event)
  2793. {
  2794. struct perf_event *leader = event->group_leader, *sub;
  2795. u64 read_format = event->attr.read_format;
  2796. u64 values[5];
  2797. int n = 0;
  2798. values[n++] = 1 + leader->nr_siblings;
  2799. if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
  2800. values[n++] = leader->total_time_enabled;
  2801. if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
  2802. values[n++] = leader->total_time_running;
  2803. if (leader != event)
  2804. leader->pmu->read(leader);
  2805. values[n++] = perf_event_count(leader);
  2806. if (read_format & PERF_FORMAT_ID)
  2807. values[n++] = primary_event_id(leader);
  2808. perf_output_copy(handle, values, n * sizeof(u64));
  2809. list_for_each_entry(sub, &leader->sibling_list, group_entry) {
  2810. n = 0;
  2811. if (sub != event)
  2812. sub->pmu->read(sub);
  2813. values[n++] = perf_event_count(sub);
  2814. if (read_format & PERF_FORMAT_ID)
  2815. values[n++] = primary_event_id(sub);
  2816. perf_output_copy(handle, values, n * sizeof(u64));
  2817. }
  2818. }
  2819. static void perf_output_read(struct perf_output_handle *handle,
  2820. struct perf_event *event)
  2821. {
  2822. if (event->attr.read_format & PERF_FORMAT_GROUP)
  2823. perf_output_read_group(handle, event);
  2824. else
  2825. perf_output_read_one(handle, event);
  2826. }
  2827. void perf_output_sample(struct perf_output_handle *handle,
  2828. struct perf_event_header *header,
  2829. struct perf_sample_data *data,
  2830. struct perf_event *event)
  2831. {
  2832. u64 sample_type = data->type;
  2833. perf_output_put(handle, *header);
  2834. if (sample_type & PERF_SAMPLE_IP)
  2835. perf_output_put(handle, data->ip);
  2836. if (sample_type & PERF_SAMPLE_TID)
  2837. perf_output_put(handle, data->tid_entry);
  2838. if (sample_type & PERF_SAMPLE_TIME)
  2839. perf_output_put(handle, data->time);
  2840. if (sample_type & PERF_SAMPLE_ADDR)
  2841. perf_output_put(handle, data->addr);
  2842. if (sample_type & PERF_SAMPLE_ID)
  2843. perf_output_put(handle, data->id);
  2844. if (sample_type & PERF_SAMPLE_STREAM_ID)
  2845. perf_output_put(handle, data->stream_id);
  2846. if (sample_type & PERF_SAMPLE_CPU)
  2847. perf_output_put(handle, data->cpu_entry);
  2848. if (sample_type & PERF_SAMPLE_PERIOD)
  2849. perf_output_put(handle, data->period);
  2850. if (sample_type & PERF_SAMPLE_READ)
  2851. perf_output_read(handle, event);
  2852. if (sample_type & PERF_SAMPLE_CALLCHAIN) {
  2853. if (data->callchain) {
  2854. int size = 1;
  2855. if (data->callchain)
  2856. size += data->callchain->nr;
  2857. size *= sizeof(u64);
  2858. perf_output_copy(handle, data->callchain, size);
  2859. } else {
  2860. u64 nr = 0;
  2861. perf_output_put(handle, nr);
  2862. }
  2863. }
  2864. if (sample_type & PERF_SAMPLE_RAW) {
  2865. if (data->raw) {
  2866. perf_output_put(handle, data->raw->size);
  2867. perf_output_copy(handle, data->raw->data,
  2868. data->raw->size);
  2869. } else {
  2870. struct {
  2871. u32 size;
  2872. u32 data;
  2873. } raw = {
  2874. .size = sizeof(u32),
  2875. .data = 0,
  2876. };
  2877. perf_output_put(handle, raw);
  2878. }
  2879. }
  2880. }
  2881. void perf_prepare_sample(struct perf_event_header *header,
  2882. struct perf_sample_data *data,
  2883. struct perf_event *event,
  2884. struct pt_regs *regs)
  2885. {
  2886. u64 sample_type = event->attr.sample_type;
  2887. data->type = sample_type;
  2888. header->type = PERF_RECORD_SAMPLE;
  2889. header->size = sizeof(*header);
  2890. header->misc = 0;
  2891. header->misc |= perf_misc_flags(regs);
  2892. if (sample_type & PERF_SAMPLE_IP) {
  2893. data->ip = perf_instruction_pointer(regs);
  2894. header->size += sizeof(data->ip);
  2895. }
  2896. if (sample_type & PERF_SAMPLE_TID) {
  2897. /* namespace issues */
  2898. data->tid_entry.pid = perf_event_pid(event, current);
  2899. data->tid_entry.tid = perf_event_tid(event, current);
  2900. header->size += sizeof(data->tid_entry);
  2901. }
  2902. if (sample_type & PERF_SAMPLE_TIME) {
  2903. data->time = perf_clock();
  2904. header->size += sizeof(data->time);
  2905. }
  2906. if (sample_type & PERF_SAMPLE_ADDR)
  2907. header->size += sizeof(data->addr);
  2908. if (sample_type & PERF_SAMPLE_ID) {
  2909. data->id = primary_event_id(event);
  2910. header->size += sizeof(data->id);
  2911. }
  2912. if (sample_type & PERF_SAMPLE_STREAM_ID) {
  2913. data->stream_id = event->id;
  2914. header->size += sizeof(data->stream_id);
  2915. }
  2916. if (sample_type & PERF_SAMPLE_CPU) {
  2917. data->cpu_entry.cpu = raw_smp_processor_id();
  2918. data->cpu_entry.reserved = 0;
  2919. header->size += sizeof(data->cpu_entry);
  2920. }
  2921. if (sample_type & PERF_SAMPLE_PERIOD)
  2922. header->size += sizeof(data->period);
  2923. if (sample_type & PERF_SAMPLE_READ)
  2924. header->size += perf_event_read_size(event);
  2925. if (sample_type & PERF_SAMPLE_CALLCHAIN) {
  2926. int size = 1;
  2927. data->callchain = perf_callchain(regs);
  2928. if (data->callchain)
  2929. size += data->callchain->nr;
  2930. header->size += size * sizeof(u64);
  2931. }
  2932. if (sample_type & PERF_SAMPLE_RAW) {
  2933. int size = sizeof(u32);
  2934. if (data->raw)
  2935. size += data->raw->size;
  2936. else
  2937. size += sizeof(u32);
  2938. WARN_ON_ONCE(size & (sizeof(u64)-1));
  2939. header->size += size;
  2940. }
  2941. }
  2942. static void perf_event_output(struct perf_event *event, int nmi,
  2943. struct perf_sample_data *data,
  2944. struct pt_regs *regs)
  2945. {
  2946. struct perf_output_handle handle;
  2947. struct perf_event_header header;
  2948. /* protect the callchain buffers */
  2949. rcu_read_lock();
  2950. perf_prepare_sample(&header, data, event, regs);
  2951. if (perf_output_begin(&handle, event, header.size, nmi, 1))
  2952. goto exit;
  2953. perf_output_sample(&handle, &header, data, event);
  2954. perf_output_end(&handle);
  2955. exit:
  2956. rcu_read_unlock();
  2957. }
  2958. /*
  2959. * read event_id
  2960. */
  2961. struct perf_read_event {
  2962. struct perf_event_header header;
  2963. u32 pid;
  2964. u32 tid;
  2965. };
  2966. static void
  2967. perf_event_read_event(struct perf_event *event,
  2968. struct task_struct *task)
  2969. {
  2970. struct perf_output_handle handle;
  2971. struct perf_read_event read_event = {
  2972. .header = {
  2973. .type = PERF_RECORD_READ,
  2974. .misc = 0,
  2975. .size = sizeof(read_event) + perf_event_read_size(event),
  2976. },
  2977. .pid = perf_event_pid(event, task),
  2978. .tid = perf_event_tid(event, task),
  2979. };
  2980. int ret;
  2981. ret = perf_output_begin(&handle, event, read_event.header.size, 0, 0);
  2982. if (ret)
  2983. return;
  2984. perf_output_put(&handle, read_event);
  2985. perf_output_read(&handle, event);
  2986. perf_output_end(&handle);
  2987. }
  2988. /*
  2989. * task tracking -- fork/exit
  2990. *
  2991. * enabled by: attr.comm | attr.mmap | attr.mmap_data | attr.task
  2992. */
  2993. struct perf_task_event {
  2994. struct task_struct *task;
  2995. struct perf_event_context *task_ctx;
  2996. struct {
  2997. struct perf_event_header header;
  2998. u32 pid;
  2999. u32 ppid;
  3000. u32 tid;
  3001. u32 ptid;
  3002. u64 time;
  3003. } event_id;
  3004. };
  3005. static void perf_event_task_output(struct perf_event *event,
  3006. struct perf_task_event *task_event)
  3007. {
  3008. struct perf_output_handle handle;
  3009. struct task_struct *task = task_event->task;
  3010. int size, ret;
  3011. size = task_event->event_id.header.size;
  3012. ret = perf_output_begin(&handle, event, size, 0, 0);
  3013. if (ret)
  3014. return;
  3015. task_event->event_id.pid = perf_event_pid(event, task);
  3016. task_event->event_id.ppid = perf_event_pid(event, current);
  3017. task_event->event_id.tid = perf_event_tid(event, task);
  3018. task_event->event_id.ptid = perf_event_tid(event, current);
  3019. perf_output_put(&handle, task_event->event_id);
  3020. perf_output_end(&handle);
  3021. }
  3022. static int perf_event_task_match(struct perf_event *event)
  3023. {
  3024. if (event->state < PERF_EVENT_STATE_INACTIVE)
  3025. return 0;
  3026. if (event->cpu != -1 && event->cpu != smp_processor_id())
  3027. return 0;
  3028. if (event->attr.comm || event->attr.mmap ||
  3029. event->attr.mmap_data || event->attr.task)
  3030. return 1;
  3031. return 0;
  3032. }
  3033. static void perf_event_task_ctx(struct perf_event_context *ctx,
  3034. struct perf_task_event *task_event)
  3035. {
  3036. struct perf_event *event;
  3037. list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
  3038. if (perf_event_task_match(event))
  3039. perf_event_task_output(event, task_event);
  3040. }
  3041. }
  3042. static void perf_event_task_event(struct perf_task_event *task_event)
  3043. {
  3044. struct perf_cpu_context *cpuctx;
  3045. struct perf_event_context *ctx;
  3046. struct pmu *pmu;
  3047. int ctxn;
  3048. rcu_read_lock();
  3049. list_for_each_entry_rcu(pmu, &pmus, entry) {
  3050. cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
  3051. perf_event_task_ctx(&cpuctx->ctx, task_event);
  3052. ctx = task_event->task_ctx;
  3053. if (!ctx) {
  3054. ctxn = pmu->task_ctx_nr;
  3055. if (ctxn < 0)
  3056. goto next;
  3057. ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
  3058. }
  3059. if (ctx)
  3060. perf_event_task_ctx(ctx, task_event);
  3061. next:
  3062. put_cpu_ptr(pmu->pmu_cpu_context);
  3063. }
  3064. rcu_read_unlock();
  3065. }
  3066. static void perf_event_task(struct task_struct *task,
  3067. struct perf_event_context *task_ctx,
  3068. int new)
  3069. {
  3070. struct perf_task_event task_event;
  3071. if (!atomic_read(&nr_comm_events) &&
  3072. !atomic_read(&nr_mmap_events) &&
  3073. !atomic_read(&nr_task_events))
  3074. return;
  3075. task_event = (struct perf_task_event){
  3076. .task = task,
  3077. .task_ctx = task_ctx,
  3078. .event_id = {
  3079. .header = {
  3080. .type = new ? PERF_RECORD_FORK : PERF_RECORD_EXIT,
  3081. .misc = 0,
  3082. .size = sizeof(task_event.event_id),
  3083. },
  3084. /* .pid */
  3085. /* .ppid */
  3086. /* .tid */
  3087. /* .ptid */
  3088. .time = perf_clock(),
  3089. },
  3090. };
  3091. perf_event_task_event(&task_event);
  3092. }
  3093. void perf_event_fork(struct task_struct *task)
  3094. {
  3095. perf_event_task(task, NULL, 1);
  3096. }
  3097. /*
  3098. * comm tracking
  3099. */
  3100. struct perf_comm_event {
  3101. struct task_struct *task;
  3102. char *comm;
  3103. int comm_size;
  3104. struct {
  3105. struct perf_event_header header;
  3106. u32 pid;
  3107. u32 tid;
  3108. } event_id;
  3109. };
  3110. static void perf_event_comm_output(struct perf_event *event,
  3111. struct perf_comm_event *comm_event)
  3112. {
  3113. struct perf_output_handle handle;
  3114. int size = comm_event->event_id.header.size;
  3115. int ret = perf_output_begin(&handle, event, size, 0, 0);
  3116. if (ret)
  3117. return;
  3118. comm_event->event_id.pid = perf_event_pid(event, comm_event->task);
  3119. comm_event->event_id.tid = perf_event_tid(event, comm_event->task);
  3120. perf_output_put(&handle, comm_event->event_id);
  3121. perf_output_copy(&handle, comm_event->comm,
  3122. comm_event->comm_size);
  3123. perf_output_end(&handle);
  3124. }
  3125. static int perf_event_comm_match(struct perf_event *event)
  3126. {
  3127. if (event->state < PERF_EVENT_STATE_INACTIVE)
  3128. return 0;
  3129. if (event->cpu != -1 && event->cpu != smp_processor_id())
  3130. return 0;
  3131. if (event->attr.comm)
  3132. return 1;
  3133. return 0;
  3134. }
  3135. static void perf_event_comm_ctx(struct perf_event_context *ctx,
  3136. struct perf_comm_event *comm_event)
  3137. {
  3138. struct perf_event *event;
  3139. list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
  3140. if (perf_event_comm_match(event))
  3141. perf_event_comm_output(event, comm_event);
  3142. }
  3143. }
  3144. static void perf_event_comm_event(struct perf_comm_event *comm_event)
  3145. {
  3146. struct perf_cpu_context *cpuctx;
  3147. struct perf_event_context *ctx;
  3148. char comm[TASK_COMM_LEN];
  3149. unsigned int size;
  3150. struct pmu *pmu;
  3151. int ctxn;
  3152. memset(comm, 0, sizeof(comm));
  3153. strlcpy(comm, comm_event->task->comm, sizeof(comm));
  3154. size = ALIGN(strlen(comm)+1, sizeof(u64));
  3155. comm_event->comm = comm;
  3156. comm_event->comm_size = size;
  3157. comm_event->event_id.header.size = sizeof(comm_event->event_id) + size;
  3158. rcu_read_lock();
  3159. list_for_each_entry_rcu(pmu, &pmus, entry) {
  3160. cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
  3161. perf_event_comm_ctx(&cpuctx->ctx, comm_event);
  3162. ctxn = pmu->task_ctx_nr;
  3163. if (ctxn < 0)
  3164. goto next;
  3165. ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
  3166. if (ctx)
  3167. perf_event_comm_ctx(ctx, comm_event);
  3168. next:
  3169. put_cpu_ptr(pmu->pmu_cpu_context);
  3170. }
  3171. rcu_read_unlock();
  3172. }
  3173. void perf_event_comm(struct task_struct *task)
  3174. {
  3175. struct perf_comm_event comm_event;
  3176. struct perf_event_context *ctx;
  3177. int ctxn;
  3178. for_each_task_context_nr(ctxn) {
  3179. ctx = task->perf_event_ctxp[ctxn];
  3180. if (!ctx)
  3181. continue;
  3182. perf_event_enable_on_exec(ctx);
  3183. }
  3184. if (!atomic_read(&nr_comm_events))
  3185. return;
  3186. comm_event = (struct perf_comm_event){
  3187. .task = task,
  3188. /* .comm */
  3189. /* .comm_size */
  3190. .event_id = {
  3191. .header = {
  3192. .type = PERF_RECORD_COMM,
  3193. .misc = 0,
  3194. /* .size */
  3195. },
  3196. /* .pid */
  3197. /* .tid */
  3198. },
  3199. };
  3200. perf_event_comm_event(&comm_event);
  3201. }
  3202. /*
  3203. * mmap tracking
  3204. */
  3205. struct perf_mmap_event {
  3206. struct vm_area_struct *vma;
  3207. const char *file_name;
  3208. int file_size;
  3209. struct {
  3210. struct perf_event_header header;
  3211. u32 pid;
  3212. u32 tid;
  3213. u64 start;
  3214. u64 len;
  3215. u64 pgoff;
  3216. } event_id;
  3217. };
  3218. static void perf_event_mmap_output(struct perf_event *event,
  3219. struct perf_mmap_event *mmap_event)
  3220. {
  3221. struct perf_output_handle handle;
  3222. int size = mmap_event->event_id.header.size;
  3223. int ret = perf_output_begin(&handle, event, size, 0, 0);
  3224. if (ret)
  3225. return;
  3226. mmap_event->event_id.pid = perf_event_pid(event, current);
  3227. mmap_event->event_id.tid = perf_event_tid(event, current);
  3228. perf_output_put(&handle, mmap_event->event_id);
  3229. perf_output_copy(&handle, mmap_event->file_name,
  3230. mmap_event->file_size);
  3231. perf_output_end(&handle);
  3232. }
  3233. static int perf_event_mmap_match(struct perf_event *event,
  3234. struct perf_mmap_event *mmap_event,
  3235. int executable)
  3236. {
  3237. if (event->state < PERF_EVENT_STATE_INACTIVE)
  3238. return 0;
  3239. if (event->cpu != -1 && event->cpu != smp_processor_id())
  3240. return 0;
  3241. if ((!executable && event->attr.mmap_data) ||
  3242. (executable && event->attr.mmap))
  3243. return 1;
  3244. return 0;
  3245. }
  3246. static void perf_event_mmap_ctx(struct perf_event_context *ctx,
  3247. struct perf_mmap_event *mmap_event,
  3248. int executable)
  3249. {
  3250. struct perf_event *event;
  3251. list_for_each_entry_rcu(event, &ctx->event_list, event_entry) {
  3252. if (perf_event_mmap_match(event, mmap_event, executable))
  3253. perf_event_mmap_output(event, mmap_event);
  3254. }
  3255. }
  3256. static void perf_event_mmap_event(struct perf_mmap_event *mmap_event)
  3257. {
  3258. struct perf_cpu_context *cpuctx;
  3259. struct perf_event_context *ctx;
  3260. struct vm_area_struct *vma = mmap_event->vma;
  3261. struct file *file = vma->vm_file;
  3262. unsigned int size;
  3263. char tmp[16];
  3264. char *buf = NULL;
  3265. const char *name;
  3266. struct pmu *pmu;
  3267. int ctxn;
  3268. memset(tmp, 0, sizeof(tmp));
  3269. if (file) {
  3270. /*
  3271. * d_path works from the end of the buffer backwards, so we
  3272. * need to add enough zero bytes after the string to handle
  3273. * the 64bit alignment we do later.
  3274. */
  3275. buf = kzalloc(PATH_MAX + sizeof(u64), GFP_KERNEL);
  3276. if (!buf) {
  3277. name = strncpy(tmp, "//enomem", sizeof(tmp));
  3278. goto got_name;
  3279. }
  3280. name = d_path(&file->f_path, buf, PATH_MAX);
  3281. if (IS_ERR(name)) {
  3282. name = strncpy(tmp, "//toolong", sizeof(tmp));
  3283. goto got_name;
  3284. }
  3285. } else {
  3286. if (arch_vma_name(mmap_event->vma)) {
  3287. name = strncpy(tmp, arch_vma_name(mmap_event->vma),
  3288. sizeof(tmp));
  3289. goto got_name;
  3290. }
  3291. if (!vma->vm_mm) {
  3292. name = strncpy(tmp, "[vdso]", sizeof(tmp));
  3293. goto got_name;
  3294. } else if (vma->vm_start <= vma->vm_mm->start_brk &&
  3295. vma->vm_end >= vma->vm_mm->brk) {
  3296. name = strncpy(tmp, "[heap]", sizeof(tmp));
  3297. goto got_name;
  3298. } else if (vma->vm_start <= vma->vm_mm->start_stack &&
  3299. vma->vm_end >= vma->vm_mm->start_stack) {
  3300. name = strncpy(tmp, "[stack]", sizeof(tmp));
  3301. goto got_name;
  3302. }
  3303. name = strncpy(tmp, "//anon", sizeof(tmp));
  3304. goto got_name;
  3305. }
  3306. got_name:
  3307. size = ALIGN(strlen(name)+1, sizeof(u64));
  3308. mmap_event->file_name = name;
  3309. mmap_event->file_size = size;
  3310. mmap_event->event_id.header.size = sizeof(mmap_event->event_id) + size;
  3311. rcu_read_lock();
  3312. list_for_each_entry_rcu(pmu, &pmus, entry) {
  3313. cpuctx = get_cpu_ptr(pmu->pmu_cpu_context);
  3314. perf_event_mmap_ctx(&cpuctx->ctx, mmap_event,
  3315. vma->vm_flags & VM_EXEC);
  3316. ctxn = pmu->task_ctx_nr;
  3317. if (ctxn < 0)
  3318. goto next;
  3319. ctx = rcu_dereference(current->perf_event_ctxp[ctxn]);
  3320. if (ctx) {
  3321. perf_event_mmap_ctx(ctx, mmap_event,
  3322. vma->vm_flags & VM_EXEC);
  3323. }
  3324. next:
  3325. put_cpu_ptr(pmu->pmu_cpu_context);
  3326. }
  3327. rcu_read_unlock();
  3328. kfree(buf);
  3329. }
  3330. void perf_event_mmap(struct vm_area_struct *vma)
  3331. {
  3332. struct perf_mmap_event mmap_event;
  3333. if (!atomic_read(&nr_mmap_events))
  3334. return;
  3335. mmap_event = (struct perf_mmap_event){
  3336. .vma = vma,
  3337. /* .file_name */
  3338. /* .file_size */
  3339. .event_id = {
  3340. .header = {
  3341. .type = PERF_RECORD_MMAP,
  3342. .misc = PERF_RECORD_MISC_USER,
  3343. /* .size */
  3344. },
  3345. /* .pid */
  3346. /* .tid */
  3347. .start = vma->vm_start,
  3348. .len = vma->vm_end - vma->vm_start,
  3349. .pgoff = (u64)vma->vm_pgoff << PAGE_SHIFT,
  3350. },
  3351. };
  3352. perf_event_mmap_event(&mmap_event);
  3353. }
  3354. /*
  3355. * IRQ throttle logging
  3356. */
  3357. static void perf_log_throttle(struct perf_event *event, int enable)
  3358. {
  3359. struct perf_output_handle handle;
  3360. int ret;
  3361. struct {
  3362. struct perf_event_header header;
  3363. u64 time;
  3364. u64 id;
  3365. u64 stream_id;
  3366. } throttle_event = {
  3367. .header = {
  3368. .type = PERF_RECORD_THROTTLE,
  3369. .misc = 0,
  3370. .size = sizeof(throttle_event),
  3371. },
  3372. .time = perf_clock(),
  3373. .id = primary_event_id(event),
  3374. .stream_id = event->id,
  3375. };
  3376. if (enable)
  3377. throttle_event.header.type = PERF_RECORD_UNTHROTTLE;
  3378. ret = perf_output_begin(&handle, event, sizeof(throttle_event), 1, 0);
  3379. if (ret)
  3380. return;
  3381. perf_output_put(&handle, throttle_event);
  3382. perf_output_end(&handle);
  3383. }
  3384. /*
  3385. * Generic event overflow handling, sampling.
  3386. */
  3387. static int __perf_event_overflow(struct perf_event *event, int nmi,
  3388. int throttle, struct perf_sample_data *data,
  3389. struct pt_regs *regs)
  3390. {
  3391. int events = atomic_read(&event->event_limit);
  3392. struct hw_perf_event *hwc = &event->hw;
  3393. int ret = 0;
  3394. if (!throttle) {
  3395. hwc->interrupts++;
  3396. } else {
  3397. if (hwc->interrupts != MAX_INTERRUPTS) {
  3398. hwc->interrupts++;
  3399. if (HZ * hwc->interrupts >
  3400. (u64)sysctl_perf_event_sample_rate) {
  3401. hwc->interrupts = MAX_INTERRUPTS;
  3402. perf_log_throttle(event, 0);
  3403. ret = 1;
  3404. }
  3405. } else {
  3406. /*
  3407. * Keep re-disabling events even though on the previous
  3408. * pass we disabled it - just in case we raced with a
  3409. * sched-in and the event got enabled again:
  3410. */
  3411. ret = 1;
  3412. }
  3413. }
  3414. if (event->attr.freq) {
  3415. u64 now = perf_clock();
  3416. s64 delta = now - hwc->freq_time_stamp;
  3417. hwc->freq_time_stamp = now;
  3418. if (delta > 0 && delta < 2*TICK_NSEC)
  3419. perf_adjust_period(event, delta, hwc->last_period);
  3420. }
  3421. /*
  3422. * XXX event_limit might not quite work as expected on inherited
  3423. * events
  3424. */
  3425. event->pending_kill = POLL_IN;
  3426. if (events && atomic_dec_and_test(&event->event_limit)) {
  3427. ret = 1;
  3428. event->pending_kill = POLL_HUP;
  3429. if (nmi) {
  3430. event->pending_disable = 1;
  3431. irq_work_queue(&event->pending);
  3432. } else
  3433. perf_event_disable(event);
  3434. }
  3435. if (event->overflow_handler)
  3436. event->overflow_handler(event, nmi, data, regs);
  3437. else
  3438. perf_event_output(event, nmi, data, regs);
  3439. return ret;
  3440. }
  3441. int perf_event_overflow(struct perf_event *event, int nmi,
  3442. struct perf_sample_data *data,
  3443. struct pt_regs *regs)
  3444. {
  3445. return __perf_event_overflow(event, nmi, 1, data, regs);
  3446. }
  3447. /*
  3448. * Generic software event infrastructure
  3449. */
  3450. struct swevent_htable {
  3451. struct swevent_hlist *swevent_hlist;
  3452. struct mutex hlist_mutex;
  3453. int hlist_refcount;
  3454. /* Recursion avoidance in each contexts */
  3455. int recursion[PERF_NR_CONTEXTS];
  3456. };
  3457. static DEFINE_PER_CPU(struct swevent_htable, swevent_htable);
  3458. /*
  3459. * We directly increment event->count and keep a second value in
  3460. * event->hw.period_left to count intervals. This period event
  3461. * is kept in the range [-sample_period, 0] so that we can use the
  3462. * sign as trigger.
  3463. */
  3464. static u64 perf_swevent_set_period(struct perf_event *event)
  3465. {
  3466. struct hw_perf_event *hwc = &event->hw;
  3467. u64 period = hwc->last_period;
  3468. u64 nr, offset;
  3469. s64 old, val;
  3470. hwc->last_period = hwc->sample_period;
  3471. again:
  3472. old = val = local64_read(&hwc->period_left);
  3473. if (val < 0)
  3474. return 0;
  3475. nr = div64_u64(period + val, period);
  3476. offset = nr * period;
  3477. val -= offset;
  3478. if (local64_cmpxchg(&hwc->period_left, old, val) != old)
  3479. goto again;
  3480. return nr;
  3481. }
  3482. static void perf_swevent_overflow(struct perf_event *event, u64 overflow,
  3483. int nmi, struct perf_sample_data *data,
  3484. struct pt_regs *regs)
  3485. {
  3486. struct hw_perf_event *hwc = &event->hw;
  3487. int throttle = 0;
  3488. data->period = event->hw.last_period;
  3489. if (!overflow)
  3490. overflow = perf_swevent_set_period(event);
  3491. if (hwc->interrupts == MAX_INTERRUPTS)
  3492. return;
  3493. for (; overflow; overflow--) {
  3494. if (__perf_event_overflow(event, nmi, throttle,
  3495. data, regs)) {
  3496. /*
  3497. * We inhibit the overflow from happening when
  3498. * hwc->interrupts == MAX_INTERRUPTS.
  3499. */
  3500. break;
  3501. }
  3502. throttle = 1;
  3503. }
  3504. }
  3505. static void perf_swevent_event(struct perf_event *event, u64 nr,
  3506. int nmi, struct perf_sample_data *data,
  3507. struct pt_regs *regs)
  3508. {
  3509. struct hw_perf_event *hwc = &event->hw;
  3510. local64_add(nr, &event->count);
  3511. if (!regs)
  3512. return;
  3513. if (!hwc->sample_period)
  3514. return;
  3515. if (nr == 1 && hwc->sample_period == 1 && !event->attr.freq)
  3516. return perf_swevent_overflow(event, 1, nmi, data, regs);
  3517. if (local64_add_negative(nr, &hwc->period_left))
  3518. return;
  3519. perf_swevent_overflow(event, 0, nmi, data, regs);
  3520. }
  3521. static int perf_exclude_event(struct perf_event *event,
  3522. struct pt_regs *regs)
  3523. {
  3524. if (event->hw.state & PERF_HES_STOPPED)
  3525. return 0;
  3526. if (regs) {
  3527. if (event->attr.exclude_user && user_mode(regs))
  3528. return 1;
  3529. if (event->attr.exclude_kernel && !user_mode(regs))
  3530. return 1;
  3531. }
  3532. return 0;
  3533. }
  3534. static int perf_swevent_match(struct perf_event *event,
  3535. enum perf_type_id type,
  3536. u32 event_id,
  3537. struct perf_sample_data *data,
  3538. struct pt_regs *regs)
  3539. {
  3540. if (event->attr.type != type)
  3541. return 0;
  3542. if (event->attr.config != event_id)
  3543. return 0;
  3544. if (perf_exclude_event(event, regs))
  3545. return 0;
  3546. return 1;
  3547. }
  3548. static inline u64 swevent_hash(u64 type, u32 event_id)
  3549. {
  3550. u64 val = event_id | (type << 32);
  3551. return hash_64(val, SWEVENT_HLIST_BITS);
  3552. }
  3553. static inline struct hlist_head *
  3554. __find_swevent_head(struct swevent_hlist *hlist, u64 type, u32 event_id)
  3555. {
  3556. u64 hash = swevent_hash(type, event_id);
  3557. return &hlist->heads[hash];
  3558. }
  3559. /* For the read side: events when they trigger */
  3560. static inline struct hlist_head *
  3561. find_swevent_head_rcu(struct swevent_htable *swhash, u64 type, u32 event_id)
  3562. {
  3563. struct swevent_hlist *hlist;
  3564. hlist = rcu_dereference(swhash->swevent_hlist);
  3565. if (!hlist)
  3566. return NULL;
  3567. return __find_swevent_head(hlist, type, event_id);
  3568. }
  3569. /* For the event head insertion and removal in the hlist */
  3570. static inline struct hlist_head *
  3571. find_swevent_head(struct swevent_htable *swhash, struct perf_event *event)
  3572. {
  3573. struct swevent_hlist *hlist;
  3574. u32 event_id = event->attr.config;
  3575. u64 type = event->attr.type;
  3576. /*
  3577. * Event scheduling is always serialized against hlist allocation
  3578. * and release. Which makes the protected version suitable here.
  3579. * The context lock guarantees that.
  3580. */
  3581. hlist = rcu_dereference_protected(swhash->swevent_hlist,
  3582. lockdep_is_held(&event->ctx->lock));
  3583. if (!hlist)
  3584. return NULL;
  3585. return __find_swevent_head(hlist, type, event_id);
  3586. }
  3587. static void do_perf_sw_event(enum perf_type_id type, u32 event_id,
  3588. u64 nr, int nmi,
  3589. struct perf_sample_data *data,
  3590. struct pt_regs *regs)
  3591. {
  3592. struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
  3593. struct perf_event *event;
  3594. struct hlist_node *node;
  3595. struct hlist_head *head;
  3596. rcu_read_lock();
  3597. head = find_swevent_head_rcu(swhash, type, event_id);
  3598. if (!head)
  3599. goto end;
  3600. hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
  3601. if (perf_swevent_match(event, type, event_id, data, regs))
  3602. perf_swevent_event(event, nr, nmi, data, regs);
  3603. }
  3604. end:
  3605. rcu_read_unlock();
  3606. }
  3607. int perf_swevent_get_recursion_context(void)
  3608. {
  3609. struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
  3610. return get_recursion_context(swhash->recursion);
  3611. }
  3612. EXPORT_SYMBOL_GPL(perf_swevent_get_recursion_context);
  3613. void inline perf_swevent_put_recursion_context(int rctx)
  3614. {
  3615. struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
  3616. put_recursion_context(swhash->recursion, rctx);
  3617. }
  3618. void __perf_sw_event(u32 event_id, u64 nr, int nmi,
  3619. struct pt_regs *regs, u64 addr)
  3620. {
  3621. struct perf_sample_data data;
  3622. int rctx;
  3623. preempt_disable_notrace();
  3624. rctx = perf_swevent_get_recursion_context();
  3625. if (rctx < 0)
  3626. return;
  3627. perf_sample_data_init(&data, addr);
  3628. do_perf_sw_event(PERF_TYPE_SOFTWARE, event_id, nr, nmi, &data, regs);
  3629. perf_swevent_put_recursion_context(rctx);
  3630. preempt_enable_notrace();
  3631. }
  3632. static void perf_swevent_read(struct perf_event *event)
  3633. {
  3634. }
  3635. static int perf_swevent_add(struct perf_event *event, int flags)
  3636. {
  3637. struct swevent_htable *swhash = &__get_cpu_var(swevent_htable);
  3638. struct hw_perf_event *hwc = &event->hw;
  3639. struct hlist_head *head;
  3640. if (hwc->sample_period) {
  3641. hwc->last_period = hwc->sample_period;
  3642. perf_swevent_set_period(event);
  3643. }
  3644. hwc->state = !(flags & PERF_EF_START);
  3645. head = find_swevent_head(swhash, event);
  3646. if (WARN_ON_ONCE(!head))
  3647. return -EINVAL;
  3648. hlist_add_head_rcu(&event->hlist_entry, head);
  3649. return 0;
  3650. }
  3651. static void perf_swevent_del(struct perf_event *event, int flags)
  3652. {
  3653. hlist_del_rcu(&event->hlist_entry);
  3654. }
  3655. static void perf_swevent_start(struct perf_event *event, int flags)
  3656. {
  3657. event->hw.state = 0;
  3658. }
  3659. static void perf_swevent_stop(struct perf_event *event, int flags)
  3660. {
  3661. event->hw.state = PERF_HES_STOPPED;
  3662. }
  3663. /* Deref the hlist from the update side */
  3664. static inline struct swevent_hlist *
  3665. swevent_hlist_deref(struct swevent_htable *swhash)
  3666. {
  3667. return rcu_dereference_protected(swhash->swevent_hlist,
  3668. lockdep_is_held(&swhash->hlist_mutex));
  3669. }
  3670. static void swevent_hlist_release_rcu(struct rcu_head *rcu_head)
  3671. {
  3672. struct swevent_hlist *hlist;
  3673. hlist = container_of(rcu_head, struct swevent_hlist, rcu_head);
  3674. kfree(hlist);
  3675. }
  3676. static void swevent_hlist_release(struct swevent_htable *swhash)
  3677. {
  3678. struct swevent_hlist *hlist = swevent_hlist_deref(swhash);
  3679. if (!hlist)
  3680. return;
  3681. rcu_assign_pointer(swhash->swevent_hlist, NULL);
  3682. call_rcu(&hlist->rcu_head, swevent_hlist_release_rcu);
  3683. }
  3684. static void swevent_hlist_put_cpu(struct perf_event *event, int cpu)
  3685. {
  3686. struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
  3687. mutex_lock(&swhash->hlist_mutex);
  3688. if (!--swhash->hlist_refcount)
  3689. swevent_hlist_release(swhash);
  3690. mutex_unlock(&swhash->hlist_mutex);
  3691. }
  3692. static void swevent_hlist_put(struct perf_event *event)
  3693. {
  3694. int cpu;
  3695. if (event->cpu != -1) {
  3696. swevent_hlist_put_cpu(event, event->cpu);
  3697. return;
  3698. }
  3699. for_each_possible_cpu(cpu)
  3700. swevent_hlist_put_cpu(event, cpu);
  3701. }
  3702. static int swevent_hlist_get_cpu(struct perf_event *event, int cpu)
  3703. {
  3704. struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
  3705. int err = 0;
  3706. mutex_lock(&swhash->hlist_mutex);
  3707. if (!swevent_hlist_deref(swhash) && cpu_online(cpu)) {
  3708. struct swevent_hlist *hlist;
  3709. hlist = kzalloc(sizeof(*hlist), GFP_KERNEL);
  3710. if (!hlist) {
  3711. err = -ENOMEM;
  3712. goto exit;
  3713. }
  3714. rcu_assign_pointer(swhash->swevent_hlist, hlist);
  3715. }
  3716. swhash->hlist_refcount++;
  3717. exit:
  3718. mutex_unlock(&swhash->hlist_mutex);
  3719. return err;
  3720. }
  3721. static int swevent_hlist_get(struct perf_event *event)
  3722. {
  3723. int err;
  3724. int cpu, failed_cpu;
  3725. if (event->cpu != -1)
  3726. return swevent_hlist_get_cpu(event, event->cpu);
  3727. get_online_cpus();
  3728. for_each_possible_cpu(cpu) {
  3729. err = swevent_hlist_get_cpu(event, cpu);
  3730. if (err) {
  3731. failed_cpu = cpu;
  3732. goto fail;
  3733. }
  3734. }
  3735. put_online_cpus();
  3736. return 0;
  3737. fail:
  3738. for_each_possible_cpu(cpu) {
  3739. if (cpu == failed_cpu)
  3740. break;
  3741. swevent_hlist_put_cpu(event, cpu);
  3742. }
  3743. put_online_cpus();
  3744. return err;
  3745. }
  3746. atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
  3747. static void sw_perf_event_destroy(struct perf_event *event)
  3748. {
  3749. u64 event_id = event->attr.config;
  3750. WARN_ON(event->parent);
  3751. jump_label_dec(&perf_swevent_enabled[event_id]);
  3752. swevent_hlist_put(event);
  3753. }
  3754. static int perf_swevent_init(struct perf_event *event)
  3755. {
  3756. int event_id = event->attr.config;
  3757. if (event->attr.type != PERF_TYPE_SOFTWARE)
  3758. return -ENOENT;
  3759. switch (event_id) {
  3760. case PERF_COUNT_SW_CPU_CLOCK:
  3761. case PERF_COUNT_SW_TASK_CLOCK:
  3762. return -ENOENT;
  3763. default:
  3764. break;
  3765. }
  3766. if (event_id > PERF_COUNT_SW_MAX)
  3767. return -ENOENT;
  3768. if (!event->parent) {
  3769. int err;
  3770. err = swevent_hlist_get(event);
  3771. if (err)
  3772. return err;
  3773. jump_label_inc(&perf_swevent_enabled[event_id]);
  3774. event->destroy = sw_perf_event_destroy;
  3775. }
  3776. return 0;
  3777. }
  3778. static struct pmu perf_swevent = {
  3779. .task_ctx_nr = perf_sw_context,
  3780. .event_init = perf_swevent_init,
  3781. .add = perf_swevent_add,
  3782. .del = perf_swevent_del,
  3783. .start = perf_swevent_start,
  3784. .stop = perf_swevent_stop,
  3785. .read = perf_swevent_read,
  3786. };
  3787. #ifdef CONFIG_EVENT_TRACING
  3788. static int perf_tp_filter_match(struct perf_event *event,
  3789. struct perf_sample_data *data)
  3790. {
  3791. void *record = data->raw->data;
  3792. if (likely(!event->filter) || filter_match_preds(event->filter, record))
  3793. return 1;
  3794. return 0;
  3795. }
  3796. static int perf_tp_event_match(struct perf_event *event,
  3797. struct perf_sample_data *data,
  3798. struct pt_regs *regs)
  3799. {
  3800. /*
  3801. * All tracepoints are from kernel-space.
  3802. */
  3803. if (event->attr.exclude_kernel)
  3804. return 0;
  3805. if (!perf_tp_filter_match(event, data))
  3806. return 0;
  3807. return 1;
  3808. }
  3809. void perf_tp_event(u64 addr, u64 count, void *record, int entry_size,
  3810. struct pt_regs *regs, struct hlist_head *head, int rctx)
  3811. {
  3812. struct perf_sample_data data;
  3813. struct perf_event *event;
  3814. struct hlist_node *node;
  3815. struct perf_raw_record raw = {
  3816. .size = entry_size,
  3817. .data = record,
  3818. };
  3819. perf_sample_data_init(&data, addr);
  3820. data.raw = &raw;
  3821. hlist_for_each_entry_rcu(event, node, head, hlist_entry) {
  3822. if (perf_tp_event_match(event, &data, regs))
  3823. perf_swevent_event(event, count, 1, &data, regs);
  3824. }
  3825. perf_swevent_put_recursion_context(rctx);
  3826. }
  3827. EXPORT_SYMBOL_GPL(perf_tp_event);
  3828. static void tp_perf_event_destroy(struct perf_event *event)
  3829. {
  3830. perf_trace_destroy(event);
  3831. }
  3832. static int perf_tp_event_init(struct perf_event *event)
  3833. {
  3834. int err;
  3835. if (event->attr.type != PERF_TYPE_TRACEPOINT)
  3836. return -ENOENT;
  3837. /*
  3838. * Raw tracepoint data is a severe data leak, only allow root to
  3839. * have these.
  3840. */
  3841. if ((event->attr.sample_type & PERF_SAMPLE_RAW) &&
  3842. perf_paranoid_tracepoint_raw() &&
  3843. !capable(CAP_SYS_ADMIN))
  3844. return -EPERM;
  3845. err = perf_trace_init(event);
  3846. if (err)
  3847. return err;
  3848. event->destroy = tp_perf_event_destroy;
  3849. return 0;
  3850. }
  3851. static struct pmu perf_tracepoint = {
  3852. .task_ctx_nr = perf_sw_context,
  3853. .event_init = perf_tp_event_init,
  3854. .add = perf_trace_add,
  3855. .del = perf_trace_del,
  3856. .start = perf_swevent_start,
  3857. .stop = perf_swevent_stop,
  3858. .read = perf_swevent_read,
  3859. };
  3860. static inline void perf_tp_register(void)
  3861. {
  3862. perf_pmu_register(&perf_tracepoint);
  3863. }
  3864. static int perf_event_set_filter(struct perf_event *event, void __user *arg)
  3865. {
  3866. char *filter_str;
  3867. int ret;
  3868. if (event->attr.type != PERF_TYPE_TRACEPOINT)
  3869. return -EINVAL;
  3870. filter_str = strndup_user(arg, PAGE_SIZE);
  3871. if (IS_ERR(filter_str))
  3872. return PTR_ERR(filter_str);
  3873. ret = ftrace_profile_set_filter(event, event->attr.config, filter_str);
  3874. kfree(filter_str);
  3875. return ret;
  3876. }
  3877. static void perf_event_free_filter(struct perf_event *event)
  3878. {
  3879. ftrace_profile_free_filter(event);
  3880. }
  3881. #else
  3882. static inline void perf_tp_register(void)
  3883. {
  3884. }
  3885. static int perf_event_set_filter(struct perf_event *event, void __user *arg)
  3886. {
  3887. return -ENOENT;
  3888. }
  3889. static void perf_event_free_filter(struct perf_event *event)
  3890. {
  3891. }
  3892. #endif /* CONFIG_EVENT_TRACING */
  3893. #ifdef CONFIG_HAVE_HW_BREAKPOINT
  3894. void perf_bp_event(struct perf_event *bp, void *data)
  3895. {
  3896. struct perf_sample_data sample;
  3897. struct pt_regs *regs = data;
  3898. perf_sample_data_init(&sample, bp->attr.bp_addr);
  3899. if (!bp->hw.state && !perf_exclude_event(bp, regs))
  3900. perf_swevent_event(bp, 1, 1, &sample, regs);
  3901. }
  3902. #endif
  3903. /*
  3904. * hrtimer based swevent callback
  3905. */
  3906. static enum hrtimer_restart perf_swevent_hrtimer(struct hrtimer *hrtimer)
  3907. {
  3908. enum hrtimer_restart ret = HRTIMER_RESTART;
  3909. struct perf_sample_data data;
  3910. struct pt_regs *regs;
  3911. struct perf_event *event;
  3912. u64 period;
  3913. event = container_of(hrtimer, struct perf_event, hw.hrtimer);
  3914. event->pmu->read(event);
  3915. perf_sample_data_init(&data, 0);
  3916. data.period = event->hw.last_period;
  3917. regs = get_irq_regs();
  3918. if (regs && !perf_exclude_event(event, regs)) {
  3919. if (!(event->attr.exclude_idle && current->pid == 0))
  3920. if (perf_event_overflow(event, 0, &data, regs))
  3921. ret = HRTIMER_NORESTART;
  3922. }
  3923. period = max_t(u64, 10000, event->hw.sample_period);
  3924. hrtimer_forward_now(hrtimer, ns_to_ktime(period));
  3925. return ret;
  3926. }
  3927. static void perf_swevent_start_hrtimer(struct perf_event *event)
  3928. {
  3929. struct hw_perf_event *hwc = &event->hw;
  3930. hrtimer_init(&hwc->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
  3931. hwc->hrtimer.function = perf_swevent_hrtimer;
  3932. if (hwc->sample_period) {
  3933. s64 period = local64_read(&hwc->period_left);
  3934. if (period) {
  3935. if (period < 0)
  3936. period = 10000;
  3937. local64_set(&hwc->period_left, 0);
  3938. } else {
  3939. period = max_t(u64, 10000, hwc->sample_period);
  3940. }
  3941. __hrtimer_start_range_ns(&hwc->hrtimer,
  3942. ns_to_ktime(period), 0,
  3943. HRTIMER_MODE_REL_PINNED, 0);
  3944. }
  3945. }
  3946. static void perf_swevent_cancel_hrtimer(struct perf_event *event)
  3947. {
  3948. struct hw_perf_event *hwc = &event->hw;
  3949. if (hwc->sample_period) {
  3950. ktime_t remaining = hrtimer_get_remaining(&hwc->hrtimer);
  3951. local64_set(&hwc->period_left, ktime_to_ns(remaining));
  3952. hrtimer_cancel(&hwc->hrtimer);
  3953. }
  3954. }
  3955. /*
  3956. * Software event: cpu wall time clock
  3957. */
  3958. static void cpu_clock_event_update(struct perf_event *event)
  3959. {
  3960. s64 prev;
  3961. u64 now;
  3962. now = local_clock();
  3963. prev = local64_xchg(&event->hw.prev_count, now);
  3964. local64_add(now - prev, &event->count);
  3965. }
  3966. static void cpu_clock_event_start(struct perf_event *event, int flags)
  3967. {
  3968. local64_set(&event->hw.prev_count, local_clock());
  3969. perf_swevent_start_hrtimer(event);
  3970. }
  3971. static void cpu_clock_event_stop(struct perf_event *event, int flags)
  3972. {
  3973. perf_swevent_cancel_hrtimer(event);
  3974. cpu_clock_event_update(event);
  3975. }
  3976. static int cpu_clock_event_add(struct perf_event *event, int flags)
  3977. {
  3978. if (flags & PERF_EF_START)
  3979. cpu_clock_event_start(event, flags);
  3980. return 0;
  3981. }
  3982. static void cpu_clock_event_del(struct perf_event *event, int flags)
  3983. {
  3984. cpu_clock_event_stop(event, flags);
  3985. }
  3986. static void cpu_clock_event_read(struct perf_event *event)
  3987. {
  3988. cpu_clock_event_update(event);
  3989. }
  3990. static int cpu_clock_event_init(struct perf_event *event)
  3991. {
  3992. if (event->attr.type != PERF_TYPE_SOFTWARE)
  3993. return -ENOENT;
  3994. if (event->attr.config != PERF_COUNT_SW_CPU_CLOCK)
  3995. return -ENOENT;
  3996. return 0;
  3997. }
  3998. static struct pmu perf_cpu_clock = {
  3999. .task_ctx_nr = perf_sw_context,
  4000. .event_init = cpu_clock_event_init,
  4001. .add = cpu_clock_event_add,
  4002. .del = cpu_clock_event_del,
  4003. .start = cpu_clock_event_start,
  4004. .stop = cpu_clock_event_stop,
  4005. .read = cpu_clock_event_read,
  4006. };
  4007. /*
  4008. * Software event: task time clock
  4009. */
  4010. static void task_clock_event_update(struct perf_event *event, u64 now)
  4011. {
  4012. u64 prev;
  4013. s64 delta;
  4014. prev = local64_xchg(&event->hw.prev_count, now);
  4015. delta = now - prev;
  4016. local64_add(delta, &event->count);
  4017. }
  4018. static void task_clock_event_start(struct perf_event *event, int flags)
  4019. {
  4020. local64_set(&event->hw.prev_count, event->ctx->time);
  4021. perf_swevent_start_hrtimer(event);
  4022. }
  4023. static void task_clock_event_stop(struct perf_event *event, int flags)
  4024. {
  4025. perf_swevent_cancel_hrtimer(event);
  4026. task_clock_event_update(event, event->ctx->time);
  4027. }
  4028. static int task_clock_event_add(struct perf_event *event, int flags)
  4029. {
  4030. if (flags & PERF_EF_START)
  4031. task_clock_event_start(event, flags);
  4032. return 0;
  4033. }
  4034. static void task_clock_event_del(struct perf_event *event, int flags)
  4035. {
  4036. task_clock_event_stop(event, PERF_EF_UPDATE);
  4037. }
  4038. static void task_clock_event_read(struct perf_event *event)
  4039. {
  4040. u64 time;
  4041. if (!in_nmi()) {
  4042. update_context_time(event->ctx);
  4043. time = event->ctx->time;
  4044. } else {
  4045. u64 now = perf_clock();
  4046. u64 delta = now - event->ctx->timestamp;
  4047. time = event->ctx->time + delta;
  4048. }
  4049. task_clock_event_update(event, time);
  4050. }
  4051. static int task_clock_event_init(struct perf_event *event)
  4052. {
  4053. if (event->attr.type != PERF_TYPE_SOFTWARE)
  4054. return -ENOENT;
  4055. if (event->attr.config != PERF_COUNT_SW_TASK_CLOCK)
  4056. return -ENOENT;
  4057. return 0;
  4058. }
  4059. static struct pmu perf_task_clock = {
  4060. .task_ctx_nr = perf_sw_context,
  4061. .event_init = task_clock_event_init,
  4062. .add = task_clock_event_add,
  4063. .del = task_clock_event_del,
  4064. .start = task_clock_event_start,
  4065. .stop = task_clock_event_stop,
  4066. .read = task_clock_event_read,
  4067. };
  4068. static void perf_pmu_nop_void(struct pmu *pmu)
  4069. {
  4070. }
  4071. static int perf_pmu_nop_int(struct pmu *pmu)
  4072. {
  4073. return 0;
  4074. }
  4075. static void perf_pmu_start_txn(struct pmu *pmu)
  4076. {
  4077. perf_pmu_disable(pmu);
  4078. }
  4079. static int perf_pmu_commit_txn(struct pmu *pmu)
  4080. {
  4081. perf_pmu_enable(pmu);
  4082. return 0;
  4083. }
  4084. static void perf_pmu_cancel_txn(struct pmu *pmu)
  4085. {
  4086. perf_pmu_enable(pmu);
  4087. }
  4088. /*
  4089. * Ensures all contexts with the same task_ctx_nr have the same
  4090. * pmu_cpu_context too.
  4091. */
  4092. static void *find_pmu_context(int ctxn)
  4093. {
  4094. struct pmu *pmu;
  4095. if (ctxn < 0)
  4096. return NULL;
  4097. list_for_each_entry(pmu, &pmus, entry) {
  4098. if (pmu->task_ctx_nr == ctxn)
  4099. return pmu->pmu_cpu_context;
  4100. }
  4101. return NULL;
  4102. }
  4103. static void free_pmu_context(void * __percpu cpu_context)
  4104. {
  4105. struct pmu *pmu;
  4106. mutex_lock(&pmus_lock);
  4107. /*
  4108. * Like a real lame refcount.
  4109. */
  4110. list_for_each_entry(pmu, &pmus, entry) {
  4111. if (pmu->pmu_cpu_context == cpu_context)
  4112. goto out;
  4113. }
  4114. free_percpu(cpu_context);
  4115. out:
  4116. mutex_unlock(&pmus_lock);
  4117. }
  4118. int perf_pmu_register(struct pmu *pmu)
  4119. {
  4120. int cpu, ret;
  4121. mutex_lock(&pmus_lock);
  4122. ret = -ENOMEM;
  4123. pmu->pmu_disable_count = alloc_percpu(int);
  4124. if (!pmu->pmu_disable_count)
  4125. goto unlock;
  4126. pmu->pmu_cpu_context = find_pmu_context(pmu->task_ctx_nr);
  4127. if (pmu->pmu_cpu_context)
  4128. goto got_cpu_context;
  4129. pmu->pmu_cpu_context = alloc_percpu(struct perf_cpu_context);
  4130. if (!pmu->pmu_cpu_context)
  4131. goto free_pdc;
  4132. for_each_possible_cpu(cpu) {
  4133. struct perf_cpu_context *cpuctx;
  4134. cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
  4135. __perf_event_init_context(&cpuctx->ctx);
  4136. cpuctx->ctx.type = cpu_context;
  4137. cpuctx->ctx.pmu = pmu;
  4138. cpuctx->jiffies_interval = 1;
  4139. INIT_LIST_HEAD(&cpuctx->rotation_list);
  4140. }
  4141. got_cpu_context:
  4142. if (!pmu->start_txn) {
  4143. if (pmu->pmu_enable) {
  4144. /*
  4145. * If we have pmu_enable/pmu_disable calls, install
  4146. * transaction stubs that use that to try and batch
  4147. * hardware accesses.
  4148. */
  4149. pmu->start_txn = perf_pmu_start_txn;
  4150. pmu->commit_txn = perf_pmu_commit_txn;
  4151. pmu->cancel_txn = perf_pmu_cancel_txn;
  4152. } else {
  4153. pmu->start_txn = perf_pmu_nop_void;
  4154. pmu->commit_txn = perf_pmu_nop_int;
  4155. pmu->cancel_txn = perf_pmu_nop_void;
  4156. }
  4157. }
  4158. if (!pmu->pmu_enable) {
  4159. pmu->pmu_enable = perf_pmu_nop_void;
  4160. pmu->pmu_disable = perf_pmu_nop_void;
  4161. }
  4162. list_add_rcu(&pmu->entry, &pmus);
  4163. ret = 0;
  4164. unlock:
  4165. mutex_unlock(&pmus_lock);
  4166. return ret;
  4167. free_pdc:
  4168. free_percpu(pmu->pmu_disable_count);
  4169. goto unlock;
  4170. }
  4171. void perf_pmu_unregister(struct pmu *pmu)
  4172. {
  4173. mutex_lock(&pmus_lock);
  4174. list_del_rcu(&pmu->entry);
  4175. mutex_unlock(&pmus_lock);
  4176. /*
  4177. * We dereference the pmu list under both SRCU and regular RCU, so
  4178. * synchronize against both of those.
  4179. */
  4180. synchronize_srcu(&pmus_srcu);
  4181. synchronize_rcu();
  4182. free_percpu(pmu->pmu_disable_count);
  4183. free_pmu_context(pmu->pmu_cpu_context);
  4184. }
  4185. struct pmu *perf_init_event(struct perf_event *event)
  4186. {
  4187. struct pmu *pmu = NULL;
  4188. int idx;
  4189. idx = srcu_read_lock(&pmus_srcu);
  4190. list_for_each_entry_rcu(pmu, &pmus, entry) {
  4191. int ret = pmu->event_init(event);
  4192. if (!ret)
  4193. goto unlock;
  4194. if (ret != -ENOENT) {
  4195. pmu = ERR_PTR(ret);
  4196. goto unlock;
  4197. }
  4198. }
  4199. pmu = ERR_PTR(-ENOENT);
  4200. unlock:
  4201. srcu_read_unlock(&pmus_srcu, idx);
  4202. return pmu;
  4203. }
  4204. /*
  4205. * Allocate and initialize a event structure
  4206. */
  4207. static struct perf_event *
  4208. perf_event_alloc(struct perf_event_attr *attr, int cpu,
  4209. struct task_struct *task,
  4210. struct perf_event *group_leader,
  4211. struct perf_event *parent_event,
  4212. perf_overflow_handler_t overflow_handler)
  4213. {
  4214. struct pmu *pmu;
  4215. struct perf_event *event;
  4216. struct hw_perf_event *hwc;
  4217. long err;
  4218. event = kzalloc(sizeof(*event), GFP_KERNEL);
  4219. if (!event)
  4220. return ERR_PTR(-ENOMEM);
  4221. /*
  4222. * Single events are their own group leaders, with an
  4223. * empty sibling list:
  4224. */
  4225. if (!group_leader)
  4226. group_leader = event;
  4227. mutex_init(&event->child_mutex);
  4228. INIT_LIST_HEAD(&event->child_list);
  4229. INIT_LIST_HEAD(&event->group_entry);
  4230. INIT_LIST_HEAD(&event->event_entry);
  4231. INIT_LIST_HEAD(&event->sibling_list);
  4232. init_waitqueue_head(&event->waitq);
  4233. init_irq_work(&event->pending, perf_pending_event);
  4234. mutex_init(&event->mmap_mutex);
  4235. event->cpu = cpu;
  4236. event->attr = *attr;
  4237. event->group_leader = group_leader;
  4238. event->pmu = NULL;
  4239. event->oncpu = -1;
  4240. event->parent = parent_event;
  4241. event->ns = get_pid_ns(current->nsproxy->pid_ns);
  4242. event->id = atomic64_inc_return(&perf_event_id);
  4243. event->state = PERF_EVENT_STATE_INACTIVE;
  4244. if (task) {
  4245. event->attach_state = PERF_ATTACH_TASK;
  4246. #ifdef CONFIG_HAVE_HW_BREAKPOINT
  4247. /*
  4248. * hw_breakpoint is a bit difficult here..
  4249. */
  4250. if (attr->type == PERF_TYPE_BREAKPOINT)
  4251. event->hw.bp_target = task;
  4252. #endif
  4253. }
  4254. if (!overflow_handler && parent_event)
  4255. overflow_handler = parent_event->overflow_handler;
  4256. event->overflow_handler = overflow_handler;
  4257. if (attr->disabled)
  4258. event->state = PERF_EVENT_STATE_OFF;
  4259. pmu = NULL;
  4260. hwc = &event->hw;
  4261. hwc->sample_period = attr->sample_period;
  4262. if (attr->freq && attr->sample_freq)
  4263. hwc->sample_period = 1;
  4264. hwc->last_period = hwc->sample_period;
  4265. local64_set(&hwc->period_left, hwc->sample_period);
  4266. /*
  4267. * we currently do not support PERF_FORMAT_GROUP on inherited events
  4268. */
  4269. if (attr->inherit && (attr->read_format & PERF_FORMAT_GROUP))
  4270. goto done;
  4271. pmu = perf_init_event(event);
  4272. done:
  4273. err = 0;
  4274. if (!pmu)
  4275. err = -EINVAL;
  4276. else if (IS_ERR(pmu))
  4277. err = PTR_ERR(pmu);
  4278. if (err) {
  4279. if (event->ns)
  4280. put_pid_ns(event->ns);
  4281. kfree(event);
  4282. return ERR_PTR(err);
  4283. }
  4284. event->pmu = pmu;
  4285. if (!event->parent) {
  4286. if (event->attach_state & PERF_ATTACH_TASK)
  4287. jump_label_inc(&perf_task_events);
  4288. if (event->attr.mmap || event->attr.mmap_data)
  4289. atomic_inc(&nr_mmap_events);
  4290. if (event->attr.comm)
  4291. atomic_inc(&nr_comm_events);
  4292. if (event->attr.task)
  4293. atomic_inc(&nr_task_events);
  4294. if (event->attr.sample_type & PERF_SAMPLE_CALLCHAIN) {
  4295. err = get_callchain_buffers();
  4296. if (err) {
  4297. free_event(event);
  4298. return ERR_PTR(err);
  4299. }
  4300. }
  4301. }
  4302. return event;
  4303. }
  4304. static int perf_copy_attr(struct perf_event_attr __user *uattr,
  4305. struct perf_event_attr *attr)
  4306. {
  4307. u32 size;
  4308. int ret;
  4309. if (!access_ok(VERIFY_WRITE, uattr, PERF_ATTR_SIZE_VER0))
  4310. return -EFAULT;
  4311. /*
  4312. * zero the full structure, so that a short copy will be nice.
  4313. */
  4314. memset(attr, 0, sizeof(*attr));
  4315. ret = get_user(size, &uattr->size);
  4316. if (ret)
  4317. return ret;
  4318. if (size > PAGE_SIZE) /* silly large */
  4319. goto err_size;
  4320. if (!size) /* abi compat */
  4321. size = PERF_ATTR_SIZE_VER0;
  4322. if (size < PERF_ATTR_SIZE_VER0)
  4323. goto err_size;
  4324. /*
  4325. * If we're handed a bigger struct than we know of,
  4326. * ensure all the unknown bits are 0 - i.e. new
  4327. * user-space does not rely on any kernel feature
  4328. * extensions we dont know about yet.
  4329. */
  4330. if (size > sizeof(*attr)) {
  4331. unsigned char __user *addr;
  4332. unsigned char __user *end;
  4333. unsigned char val;
  4334. addr = (void __user *)uattr + sizeof(*attr);
  4335. end = (void __user *)uattr + size;
  4336. for (; addr < end; addr++) {
  4337. ret = get_user(val, addr);
  4338. if (ret)
  4339. return ret;
  4340. if (val)
  4341. goto err_size;
  4342. }
  4343. size = sizeof(*attr);
  4344. }
  4345. ret = copy_from_user(attr, uattr, size);
  4346. if (ret)
  4347. return -EFAULT;
  4348. /*
  4349. * If the type exists, the corresponding creation will verify
  4350. * the attr->config.
  4351. */
  4352. if (attr->type >= PERF_TYPE_MAX)
  4353. return -EINVAL;
  4354. if (attr->__reserved_1)
  4355. return -EINVAL;
  4356. if (attr->sample_type & ~(PERF_SAMPLE_MAX-1))
  4357. return -EINVAL;
  4358. if (attr->read_format & ~(PERF_FORMAT_MAX-1))
  4359. return -EINVAL;
  4360. out:
  4361. return ret;
  4362. err_size:
  4363. put_user(sizeof(*attr), &uattr->size);
  4364. ret = -E2BIG;
  4365. goto out;
  4366. }
  4367. static int
  4368. perf_event_set_output(struct perf_event *event, struct perf_event *output_event)
  4369. {
  4370. struct perf_buffer *buffer = NULL, *old_buffer = NULL;
  4371. int ret = -EINVAL;
  4372. if (!output_event)
  4373. goto set;
  4374. /* don't allow circular references */
  4375. if (event == output_event)
  4376. goto out;
  4377. /*
  4378. * Don't allow cross-cpu buffers
  4379. */
  4380. if (output_event->cpu != event->cpu)
  4381. goto out;
  4382. /*
  4383. * If its not a per-cpu buffer, it must be the same task.
  4384. */
  4385. if (output_event->cpu == -1 && output_event->ctx != event->ctx)
  4386. goto out;
  4387. set:
  4388. mutex_lock(&event->mmap_mutex);
  4389. /* Can't redirect output if we've got an active mmap() */
  4390. if (atomic_read(&event->mmap_count))
  4391. goto unlock;
  4392. if (output_event) {
  4393. /* get the buffer we want to redirect to */
  4394. buffer = perf_buffer_get(output_event);
  4395. if (!buffer)
  4396. goto unlock;
  4397. }
  4398. old_buffer = event->buffer;
  4399. rcu_assign_pointer(event->buffer, buffer);
  4400. ret = 0;
  4401. unlock:
  4402. mutex_unlock(&event->mmap_mutex);
  4403. if (old_buffer)
  4404. perf_buffer_put(old_buffer);
  4405. out:
  4406. return ret;
  4407. }
  4408. /**
  4409. * sys_perf_event_open - open a performance event, associate it to a task/cpu
  4410. *
  4411. * @attr_uptr: event_id type attributes for monitoring/sampling
  4412. * @pid: target pid
  4413. * @cpu: target cpu
  4414. * @group_fd: group leader event fd
  4415. */
  4416. SYSCALL_DEFINE5(perf_event_open,
  4417. struct perf_event_attr __user *, attr_uptr,
  4418. pid_t, pid, int, cpu, int, group_fd, unsigned long, flags)
  4419. {
  4420. struct perf_event *group_leader = NULL, *output_event = NULL;
  4421. struct perf_event *event, *sibling;
  4422. struct perf_event_attr attr;
  4423. struct perf_event_context *ctx;
  4424. struct file *event_file = NULL;
  4425. struct file *group_file = NULL;
  4426. struct task_struct *task = NULL;
  4427. struct pmu *pmu;
  4428. int event_fd;
  4429. int move_group = 0;
  4430. int fput_needed = 0;
  4431. int err;
  4432. /* for future expandability... */
  4433. if (flags & ~(PERF_FLAG_FD_NO_GROUP | PERF_FLAG_FD_OUTPUT))
  4434. return -EINVAL;
  4435. err = perf_copy_attr(attr_uptr, &attr);
  4436. if (err)
  4437. return err;
  4438. if (!attr.exclude_kernel) {
  4439. if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
  4440. return -EACCES;
  4441. }
  4442. if (attr.freq) {
  4443. if (attr.sample_freq > sysctl_perf_event_sample_rate)
  4444. return -EINVAL;
  4445. }
  4446. event_fd = get_unused_fd_flags(O_RDWR);
  4447. if (event_fd < 0)
  4448. return event_fd;
  4449. if (group_fd != -1) {
  4450. group_leader = perf_fget_light(group_fd, &fput_needed);
  4451. if (IS_ERR(group_leader)) {
  4452. err = PTR_ERR(group_leader);
  4453. goto err_fd;
  4454. }
  4455. group_file = group_leader->filp;
  4456. if (flags & PERF_FLAG_FD_OUTPUT)
  4457. output_event = group_leader;
  4458. if (flags & PERF_FLAG_FD_NO_GROUP)
  4459. group_leader = NULL;
  4460. }
  4461. if (pid != -1) {
  4462. task = find_lively_task_by_vpid(pid);
  4463. if (IS_ERR(task)) {
  4464. err = PTR_ERR(task);
  4465. goto err_group_fd;
  4466. }
  4467. }
  4468. event = perf_event_alloc(&attr, cpu, task, group_leader, NULL, NULL);
  4469. if (IS_ERR(event)) {
  4470. err = PTR_ERR(event);
  4471. goto err_task;
  4472. }
  4473. /*
  4474. * Special case software events and allow them to be part of
  4475. * any hardware group.
  4476. */
  4477. pmu = event->pmu;
  4478. if (group_leader &&
  4479. (is_software_event(event) != is_software_event(group_leader))) {
  4480. if (is_software_event(event)) {
  4481. /*
  4482. * If event and group_leader are not both a software
  4483. * event, and event is, then group leader is not.
  4484. *
  4485. * Allow the addition of software events to !software
  4486. * groups, this is safe because software events never
  4487. * fail to schedule.
  4488. */
  4489. pmu = group_leader->pmu;
  4490. } else if (is_software_event(group_leader) &&
  4491. (group_leader->group_flags & PERF_GROUP_SOFTWARE)) {
  4492. /*
  4493. * In case the group is a pure software group, and we
  4494. * try to add a hardware event, move the whole group to
  4495. * the hardware context.
  4496. */
  4497. move_group = 1;
  4498. }
  4499. }
  4500. /*
  4501. * Get the target context (task or percpu):
  4502. */
  4503. ctx = find_get_context(pmu, task, cpu);
  4504. if (IS_ERR(ctx)) {
  4505. err = PTR_ERR(ctx);
  4506. goto err_alloc;
  4507. }
  4508. /*
  4509. * Look up the group leader (we will attach this event to it):
  4510. */
  4511. if (group_leader) {
  4512. err = -EINVAL;
  4513. /*
  4514. * Do not allow a recursive hierarchy (this new sibling
  4515. * becoming part of another group-sibling):
  4516. */
  4517. if (group_leader->group_leader != group_leader)
  4518. goto err_context;
  4519. /*
  4520. * Do not allow to attach to a group in a different
  4521. * task or CPU context:
  4522. */
  4523. if (move_group) {
  4524. if (group_leader->ctx->type != ctx->type)
  4525. goto err_context;
  4526. } else {
  4527. if (group_leader->ctx != ctx)
  4528. goto err_context;
  4529. }
  4530. /*
  4531. * Only a group leader can be exclusive or pinned
  4532. */
  4533. if (attr.exclusive || attr.pinned)
  4534. goto err_context;
  4535. }
  4536. if (output_event) {
  4537. err = perf_event_set_output(event, output_event);
  4538. if (err)
  4539. goto err_context;
  4540. }
  4541. event_file = anon_inode_getfile("[perf_event]", &perf_fops, event, O_RDWR);
  4542. if (IS_ERR(event_file)) {
  4543. err = PTR_ERR(event_file);
  4544. goto err_context;
  4545. }
  4546. if (move_group) {
  4547. struct perf_event_context *gctx = group_leader->ctx;
  4548. mutex_lock(&gctx->mutex);
  4549. perf_event_remove_from_context(group_leader);
  4550. list_for_each_entry(sibling, &group_leader->sibling_list,
  4551. group_entry) {
  4552. perf_event_remove_from_context(sibling);
  4553. put_ctx(gctx);
  4554. }
  4555. mutex_unlock(&gctx->mutex);
  4556. put_ctx(gctx);
  4557. }
  4558. event->filp = event_file;
  4559. WARN_ON_ONCE(ctx->parent_ctx);
  4560. mutex_lock(&ctx->mutex);
  4561. if (move_group) {
  4562. perf_install_in_context(ctx, group_leader, cpu);
  4563. get_ctx(ctx);
  4564. list_for_each_entry(sibling, &group_leader->sibling_list,
  4565. group_entry) {
  4566. perf_install_in_context(ctx, sibling, cpu);
  4567. get_ctx(ctx);
  4568. }
  4569. }
  4570. perf_install_in_context(ctx, event, cpu);
  4571. ++ctx->generation;
  4572. mutex_unlock(&ctx->mutex);
  4573. event->owner = current;
  4574. get_task_struct(current);
  4575. mutex_lock(&current->perf_event_mutex);
  4576. list_add_tail(&event->owner_entry, &current->perf_event_list);
  4577. mutex_unlock(&current->perf_event_mutex);
  4578. /*
  4579. * Drop the reference on the group_event after placing the
  4580. * new event on the sibling_list. This ensures destruction
  4581. * of the group leader will find the pointer to itself in
  4582. * perf_group_detach().
  4583. */
  4584. fput_light(group_file, fput_needed);
  4585. fd_install(event_fd, event_file);
  4586. return event_fd;
  4587. err_context:
  4588. put_ctx(ctx);
  4589. err_alloc:
  4590. free_event(event);
  4591. err_task:
  4592. if (task)
  4593. put_task_struct(task);
  4594. err_group_fd:
  4595. fput_light(group_file, fput_needed);
  4596. err_fd:
  4597. put_unused_fd(event_fd);
  4598. return err;
  4599. }
  4600. /**
  4601. * perf_event_create_kernel_counter
  4602. *
  4603. * @attr: attributes of the counter to create
  4604. * @cpu: cpu in which the counter is bound
  4605. * @task: task to profile (NULL for percpu)
  4606. */
  4607. struct perf_event *
  4608. perf_event_create_kernel_counter(struct perf_event_attr *attr, int cpu,
  4609. struct task_struct *task,
  4610. perf_overflow_handler_t overflow_handler)
  4611. {
  4612. struct perf_event_context *ctx;
  4613. struct perf_event *event;
  4614. int err;
  4615. /*
  4616. * Get the target context (task or percpu):
  4617. */
  4618. event = perf_event_alloc(attr, cpu, task, NULL, NULL, overflow_handler);
  4619. if (IS_ERR(event)) {
  4620. err = PTR_ERR(event);
  4621. goto err;
  4622. }
  4623. ctx = find_get_context(event->pmu, task, cpu);
  4624. if (IS_ERR(ctx)) {
  4625. err = PTR_ERR(ctx);
  4626. goto err_free;
  4627. }
  4628. event->filp = NULL;
  4629. WARN_ON_ONCE(ctx->parent_ctx);
  4630. mutex_lock(&ctx->mutex);
  4631. perf_install_in_context(ctx, event, cpu);
  4632. ++ctx->generation;
  4633. mutex_unlock(&ctx->mutex);
  4634. event->owner = current;
  4635. get_task_struct(current);
  4636. mutex_lock(&current->perf_event_mutex);
  4637. list_add_tail(&event->owner_entry, &current->perf_event_list);
  4638. mutex_unlock(&current->perf_event_mutex);
  4639. return event;
  4640. err_free:
  4641. free_event(event);
  4642. err:
  4643. return ERR_PTR(err);
  4644. }
  4645. EXPORT_SYMBOL_GPL(perf_event_create_kernel_counter);
  4646. static void sync_child_event(struct perf_event *child_event,
  4647. struct task_struct *child)
  4648. {
  4649. struct perf_event *parent_event = child_event->parent;
  4650. u64 child_val;
  4651. if (child_event->attr.inherit_stat)
  4652. perf_event_read_event(child_event, child);
  4653. child_val = perf_event_count(child_event);
  4654. /*
  4655. * Add back the child's count to the parent's count:
  4656. */
  4657. atomic64_add(child_val, &parent_event->child_count);
  4658. atomic64_add(child_event->total_time_enabled,
  4659. &parent_event->child_total_time_enabled);
  4660. atomic64_add(child_event->total_time_running,
  4661. &parent_event->child_total_time_running);
  4662. /*
  4663. * Remove this event from the parent's list
  4664. */
  4665. WARN_ON_ONCE(parent_event->ctx->parent_ctx);
  4666. mutex_lock(&parent_event->child_mutex);
  4667. list_del_init(&child_event->child_list);
  4668. mutex_unlock(&parent_event->child_mutex);
  4669. /*
  4670. * Release the parent event, if this was the last
  4671. * reference to it.
  4672. */
  4673. fput(parent_event->filp);
  4674. }
  4675. static void
  4676. __perf_event_exit_task(struct perf_event *child_event,
  4677. struct perf_event_context *child_ctx,
  4678. struct task_struct *child)
  4679. {
  4680. struct perf_event *parent_event;
  4681. perf_event_remove_from_context(child_event);
  4682. parent_event = child_event->parent;
  4683. /*
  4684. * It can happen that parent exits first, and has events
  4685. * that are still around due to the child reference. These
  4686. * events need to be zapped - but otherwise linger.
  4687. */
  4688. if (parent_event) {
  4689. sync_child_event(child_event, child);
  4690. free_event(child_event);
  4691. }
  4692. }
  4693. static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
  4694. {
  4695. struct perf_event *child_event, *tmp;
  4696. struct perf_event_context *child_ctx;
  4697. unsigned long flags;
  4698. if (likely(!child->perf_event_ctxp[ctxn])) {
  4699. perf_event_task(child, NULL, 0);
  4700. return;
  4701. }
  4702. local_irq_save(flags);
  4703. /*
  4704. * We can't reschedule here because interrupts are disabled,
  4705. * and either child is current or it is a task that can't be
  4706. * scheduled, so we are now safe from rescheduling changing
  4707. * our context.
  4708. */
  4709. child_ctx = child->perf_event_ctxp[ctxn];
  4710. task_ctx_sched_out(child_ctx, EVENT_ALL);
  4711. /*
  4712. * Take the context lock here so that if find_get_context is
  4713. * reading child->perf_event_ctxp, we wait until it has
  4714. * incremented the context's refcount before we do put_ctx below.
  4715. */
  4716. raw_spin_lock(&child_ctx->lock);
  4717. child->perf_event_ctxp[ctxn] = NULL;
  4718. /*
  4719. * If this context is a clone; unclone it so it can't get
  4720. * swapped to another process while we're removing all
  4721. * the events from it.
  4722. */
  4723. unclone_ctx(child_ctx);
  4724. update_context_time(child_ctx);
  4725. raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
  4726. /*
  4727. * Report the task dead after unscheduling the events so that we
  4728. * won't get any samples after PERF_RECORD_EXIT. We can however still
  4729. * get a few PERF_RECORD_READ events.
  4730. */
  4731. perf_event_task(child, child_ctx, 0);
  4732. /*
  4733. * We can recurse on the same lock type through:
  4734. *
  4735. * __perf_event_exit_task()
  4736. * sync_child_event()
  4737. * fput(parent_event->filp)
  4738. * perf_release()
  4739. * mutex_lock(&ctx->mutex)
  4740. *
  4741. * But since its the parent context it won't be the same instance.
  4742. */
  4743. mutex_lock(&child_ctx->mutex);
  4744. again:
  4745. list_for_each_entry_safe(child_event, tmp, &child_ctx->pinned_groups,
  4746. group_entry)
  4747. __perf_event_exit_task(child_event, child_ctx, child);
  4748. list_for_each_entry_safe(child_event, tmp, &child_ctx->flexible_groups,
  4749. group_entry)
  4750. __perf_event_exit_task(child_event, child_ctx, child);
  4751. /*
  4752. * If the last event was a group event, it will have appended all
  4753. * its siblings to the list, but we obtained 'tmp' before that which
  4754. * will still point to the list head terminating the iteration.
  4755. */
  4756. if (!list_empty(&child_ctx->pinned_groups) ||
  4757. !list_empty(&child_ctx->flexible_groups))
  4758. goto again;
  4759. mutex_unlock(&child_ctx->mutex);
  4760. put_ctx(child_ctx);
  4761. }
  4762. /*
  4763. * When a child task exits, feed back event values to parent events.
  4764. */
  4765. void perf_event_exit_task(struct task_struct *child)
  4766. {
  4767. int ctxn;
  4768. for_each_task_context_nr(ctxn)
  4769. perf_event_exit_task_context(child, ctxn);
  4770. }
  4771. static void perf_free_event(struct perf_event *event,
  4772. struct perf_event_context *ctx)
  4773. {
  4774. struct perf_event *parent = event->parent;
  4775. if (WARN_ON_ONCE(!parent))
  4776. return;
  4777. mutex_lock(&parent->child_mutex);
  4778. list_del_init(&event->child_list);
  4779. mutex_unlock(&parent->child_mutex);
  4780. fput(parent->filp);
  4781. perf_group_detach(event);
  4782. list_del_event(event, ctx);
  4783. free_event(event);
  4784. }
  4785. /*
  4786. * free an unexposed, unused context as created by inheritance by
  4787. * perf_event_init_task below, used by fork() in case of fail.
  4788. */
  4789. void perf_event_free_task(struct task_struct *task)
  4790. {
  4791. struct perf_event_context *ctx;
  4792. struct perf_event *event, *tmp;
  4793. int ctxn;
  4794. for_each_task_context_nr(ctxn) {
  4795. ctx = task->perf_event_ctxp[ctxn];
  4796. if (!ctx)
  4797. continue;
  4798. mutex_lock(&ctx->mutex);
  4799. again:
  4800. list_for_each_entry_safe(event, tmp, &ctx->pinned_groups,
  4801. group_entry)
  4802. perf_free_event(event, ctx);
  4803. list_for_each_entry_safe(event, tmp, &ctx->flexible_groups,
  4804. group_entry)
  4805. perf_free_event(event, ctx);
  4806. if (!list_empty(&ctx->pinned_groups) ||
  4807. !list_empty(&ctx->flexible_groups))
  4808. goto again;
  4809. mutex_unlock(&ctx->mutex);
  4810. put_ctx(ctx);
  4811. }
  4812. }
  4813. void perf_event_delayed_put(struct task_struct *task)
  4814. {
  4815. int ctxn;
  4816. for_each_task_context_nr(ctxn)
  4817. WARN_ON_ONCE(task->perf_event_ctxp[ctxn]);
  4818. }
  4819. /*
  4820. * inherit a event from parent task to child task:
  4821. */
  4822. static struct perf_event *
  4823. inherit_event(struct perf_event *parent_event,
  4824. struct task_struct *parent,
  4825. struct perf_event_context *parent_ctx,
  4826. struct task_struct *child,
  4827. struct perf_event *group_leader,
  4828. struct perf_event_context *child_ctx)
  4829. {
  4830. struct perf_event *child_event;
  4831. unsigned long flags;
  4832. /*
  4833. * Instead of creating recursive hierarchies of events,
  4834. * we link inherited events back to the original parent,
  4835. * which has a filp for sure, which we use as the reference
  4836. * count:
  4837. */
  4838. if (parent_event->parent)
  4839. parent_event = parent_event->parent;
  4840. child_event = perf_event_alloc(&parent_event->attr,
  4841. parent_event->cpu,
  4842. child,
  4843. group_leader, parent_event,
  4844. NULL);
  4845. if (IS_ERR(child_event))
  4846. return child_event;
  4847. get_ctx(child_ctx);
  4848. /*
  4849. * Make the child state follow the state of the parent event,
  4850. * not its attr.disabled bit. We hold the parent's mutex,
  4851. * so we won't race with perf_event_{en, dis}able_family.
  4852. */
  4853. if (parent_event->state >= PERF_EVENT_STATE_INACTIVE)
  4854. child_event->state = PERF_EVENT_STATE_INACTIVE;
  4855. else
  4856. child_event->state = PERF_EVENT_STATE_OFF;
  4857. if (parent_event->attr.freq) {
  4858. u64 sample_period = parent_event->hw.sample_period;
  4859. struct hw_perf_event *hwc = &child_event->hw;
  4860. hwc->sample_period = sample_period;
  4861. hwc->last_period = sample_period;
  4862. local64_set(&hwc->period_left, sample_period);
  4863. }
  4864. child_event->ctx = child_ctx;
  4865. child_event->overflow_handler = parent_event->overflow_handler;
  4866. /*
  4867. * Link it up in the child's context:
  4868. */
  4869. raw_spin_lock_irqsave(&child_ctx->lock, flags);
  4870. add_event_to_ctx(child_event, child_ctx);
  4871. raw_spin_unlock_irqrestore(&child_ctx->lock, flags);
  4872. /*
  4873. * Get a reference to the parent filp - we will fput it
  4874. * when the child event exits. This is safe to do because
  4875. * we are in the parent and we know that the filp still
  4876. * exists and has a nonzero count:
  4877. */
  4878. atomic_long_inc(&parent_event->filp->f_count);
  4879. /*
  4880. * Link this into the parent event's child list
  4881. */
  4882. WARN_ON_ONCE(parent_event->ctx->parent_ctx);
  4883. mutex_lock(&parent_event->child_mutex);
  4884. list_add_tail(&child_event->child_list, &parent_event->child_list);
  4885. mutex_unlock(&parent_event->child_mutex);
  4886. return child_event;
  4887. }
  4888. static int inherit_group(struct perf_event *parent_event,
  4889. struct task_struct *parent,
  4890. struct perf_event_context *parent_ctx,
  4891. struct task_struct *child,
  4892. struct perf_event_context *child_ctx)
  4893. {
  4894. struct perf_event *leader;
  4895. struct perf_event *sub;
  4896. struct perf_event *child_ctr;
  4897. leader = inherit_event(parent_event, parent, parent_ctx,
  4898. child, NULL, child_ctx);
  4899. if (IS_ERR(leader))
  4900. return PTR_ERR(leader);
  4901. list_for_each_entry(sub, &parent_event->sibling_list, group_entry) {
  4902. child_ctr = inherit_event(sub, parent, parent_ctx,
  4903. child, leader, child_ctx);
  4904. if (IS_ERR(child_ctr))
  4905. return PTR_ERR(child_ctr);
  4906. }
  4907. return 0;
  4908. }
  4909. static int
  4910. inherit_task_group(struct perf_event *event, struct task_struct *parent,
  4911. struct perf_event_context *parent_ctx,
  4912. struct task_struct *child, int ctxn,
  4913. int *inherited_all)
  4914. {
  4915. int ret;
  4916. struct perf_event_context *child_ctx;
  4917. if (!event->attr.inherit) {
  4918. *inherited_all = 0;
  4919. return 0;
  4920. }
  4921. child_ctx = child->perf_event_ctxp[ctxn];
  4922. if (!child_ctx) {
  4923. /*
  4924. * This is executed from the parent task context, so
  4925. * inherit events that have been marked for cloning.
  4926. * First allocate and initialize a context for the
  4927. * child.
  4928. */
  4929. child_ctx = alloc_perf_context(event->pmu, child);
  4930. if (!child_ctx)
  4931. return -ENOMEM;
  4932. child->perf_event_ctxp[ctxn] = child_ctx;
  4933. }
  4934. ret = inherit_group(event, parent, parent_ctx,
  4935. child, child_ctx);
  4936. if (ret)
  4937. *inherited_all = 0;
  4938. return ret;
  4939. }
  4940. /*
  4941. * Initialize the perf_event context in task_struct
  4942. */
  4943. int perf_event_init_context(struct task_struct *child, int ctxn)
  4944. {
  4945. struct perf_event_context *child_ctx, *parent_ctx;
  4946. struct perf_event_context *cloned_ctx;
  4947. struct perf_event *event;
  4948. struct task_struct *parent = current;
  4949. int inherited_all = 1;
  4950. int ret = 0;
  4951. child->perf_event_ctxp[ctxn] = NULL;
  4952. mutex_init(&child->perf_event_mutex);
  4953. INIT_LIST_HEAD(&child->perf_event_list);
  4954. if (likely(!parent->perf_event_ctxp[ctxn]))
  4955. return 0;
  4956. /*
  4957. * If the parent's context is a clone, pin it so it won't get
  4958. * swapped under us.
  4959. */
  4960. parent_ctx = perf_pin_task_context(parent, ctxn);
  4961. /*
  4962. * No need to check if parent_ctx != NULL here; since we saw
  4963. * it non-NULL earlier, the only reason for it to become NULL
  4964. * is if we exit, and since we're currently in the middle of
  4965. * a fork we can't be exiting at the same time.
  4966. */
  4967. /*
  4968. * Lock the parent list. No need to lock the child - not PID
  4969. * hashed yet and not running, so nobody can access it.
  4970. */
  4971. mutex_lock(&parent_ctx->mutex);
  4972. /*
  4973. * We dont have to disable NMIs - we are only looking at
  4974. * the list, not manipulating it:
  4975. */
  4976. list_for_each_entry(event, &parent_ctx->pinned_groups, group_entry) {
  4977. ret = inherit_task_group(event, parent, parent_ctx,
  4978. child, ctxn, &inherited_all);
  4979. if (ret)
  4980. break;
  4981. }
  4982. list_for_each_entry(event, &parent_ctx->flexible_groups, group_entry) {
  4983. ret = inherit_task_group(event, parent, parent_ctx,
  4984. child, ctxn, &inherited_all);
  4985. if (ret)
  4986. break;
  4987. }
  4988. child_ctx = child->perf_event_ctxp[ctxn];
  4989. if (child_ctx && inherited_all) {
  4990. /*
  4991. * Mark the child context as a clone of the parent
  4992. * context, or of whatever the parent is a clone of.
  4993. * Note that if the parent is a clone, it could get
  4994. * uncloned at any point, but that doesn't matter
  4995. * because the list of events and the generation
  4996. * count can't have changed since we took the mutex.
  4997. */
  4998. cloned_ctx = rcu_dereference(parent_ctx->parent_ctx);
  4999. if (cloned_ctx) {
  5000. child_ctx->parent_ctx = cloned_ctx;
  5001. child_ctx->parent_gen = parent_ctx->parent_gen;
  5002. } else {
  5003. child_ctx->parent_ctx = parent_ctx;
  5004. child_ctx->parent_gen = parent_ctx->generation;
  5005. }
  5006. get_ctx(child_ctx->parent_ctx);
  5007. }
  5008. mutex_unlock(&parent_ctx->mutex);
  5009. perf_unpin_context(parent_ctx);
  5010. return ret;
  5011. }
  5012. /*
  5013. * Initialize the perf_event context in task_struct
  5014. */
  5015. int perf_event_init_task(struct task_struct *child)
  5016. {
  5017. int ctxn, ret;
  5018. for_each_task_context_nr(ctxn) {
  5019. ret = perf_event_init_context(child, ctxn);
  5020. if (ret)
  5021. return ret;
  5022. }
  5023. return 0;
  5024. }
  5025. static void __init perf_event_init_all_cpus(void)
  5026. {
  5027. struct swevent_htable *swhash;
  5028. int cpu;
  5029. for_each_possible_cpu(cpu) {
  5030. swhash = &per_cpu(swevent_htable, cpu);
  5031. mutex_init(&swhash->hlist_mutex);
  5032. INIT_LIST_HEAD(&per_cpu(rotation_list, cpu));
  5033. }
  5034. }
  5035. static void __cpuinit perf_event_init_cpu(int cpu)
  5036. {
  5037. struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
  5038. mutex_lock(&swhash->hlist_mutex);
  5039. if (swhash->hlist_refcount > 0) {
  5040. struct swevent_hlist *hlist;
  5041. hlist = kzalloc_node(sizeof(*hlist), GFP_KERNEL, cpu_to_node(cpu));
  5042. WARN_ON(!hlist);
  5043. rcu_assign_pointer(swhash->swevent_hlist, hlist);
  5044. }
  5045. mutex_unlock(&swhash->hlist_mutex);
  5046. }
  5047. #ifdef CONFIG_HOTPLUG_CPU
  5048. static void perf_pmu_rotate_stop(struct pmu *pmu)
  5049. {
  5050. struct perf_cpu_context *cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
  5051. WARN_ON(!irqs_disabled());
  5052. list_del_init(&cpuctx->rotation_list);
  5053. }
  5054. static void __perf_event_exit_context(void *__info)
  5055. {
  5056. struct perf_event_context *ctx = __info;
  5057. struct perf_event *event, *tmp;
  5058. perf_pmu_rotate_stop(ctx->pmu);
  5059. list_for_each_entry_safe(event, tmp, &ctx->pinned_groups, group_entry)
  5060. __perf_event_remove_from_context(event);
  5061. list_for_each_entry_safe(event, tmp, &ctx->flexible_groups, group_entry)
  5062. __perf_event_remove_from_context(event);
  5063. }
  5064. static void perf_event_exit_cpu_context(int cpu)
  5065. {
  5066. struct perf_event_context *ctx;
  5067. struct pmu *pmu;
  5068. int idx;
  5069. idx = srcu_read_lock(&pmus_srcu);
  5070. list_for_each_entry_rcu(pmu, &pmus, entry) {
  5071. ctx = &per_cpu_ptr(pmu->pmu_cpu_context, cpu)->ctx;
  5072. mutex_lock(&ctx->mutex);
  5073. smp_call_function_single(cpu, __perf_event_exit_context, ctx, 1);
  5074. mutex_unlock(&ctx->mutex);
  5075. }
  5076. srcu_read_unlock(&pmus_srcu, idx);
  5077. }
  5078. static void perf_event_exit_cpu(int cpu)
  5079. {
  5080. struct swevent_htable *swhash = &per_cpu(swevent_htable, cpu);
  5081. mutex_lock(&swhash->hlist_mutex);
  5082. swevent_hlist_release(swhash);
  5083. mutex_unlock(&swhash->hlist_mutex);
  5084. perf_event_exit_cpu_context(cpu);
  5085. }
  5086. #else
  5087. static inline void perf_event_exit_cpu(int cpu) { }
  5088. #endif
  5089. static int __cpuinit
  5090. perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
  5091. {
  5092. unsigned int cpu = (long)hcpu;
  5093. switch (action & ~CPU_TASKS_FROZEN) {
  5094. case CPU_UP_PREPARE:
  5095. case CPU_DOWN_FAILED:
  5096. perf_event_init_cpu(cpu);
  5097. break;
  5098. case CPU_UP_CANCELED:
  5099. case CPU_DOWN_PREPARE:
  5100. perf_event_exit_cpu(cpu);
  5101. break;
  5102. default:
  5103. break;
  5104. }
  5105. return NOTIFY_OK;
  5106. }
  5107. void __init perf_event_init(void)
  5108. {
  5109. perf_event_init_all_cpus();
  5110. init_srcu_struct(&pmus_srcu);
  5111. perf_pmu_register(&perf_swevent);
  5112. perf_pmu_register(&perf_cpu_clock);
  5113. perf_pmu_register(&perf_task_clock);
  5114. perf_tp_register();
  5115. perf_cpu_notifier(perf_cpu_notify);
  5116. }