tcp_input.c 166 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138413941404141414241434144414541464147414841494150415141524153415441554156415741584159416041614162416341644165416641674168416941704171417241734174417541764177417841794180418141824183418441854186418741884189419041914192419341944195419641974198419942004201420242034204420542064207420842094210421142124213421442154216421742184219422042214222422342244225422642274228422942304231423242334234423542364237423842394240424142424243424442454246424742484249425042514252425342544255425642574258425942604261426242634264426542664267426842694270427142724273427442754276427742784279428042814282428342844285428642874288428942904291429242934294429542964297429842994300430143024303430443054306430743084309431043114312431343144315431643174318431943204321432243234324432543264327432843294330433143324333433443354336433743384339434043414342434343444345434643474348434943504351435243534354435543564357435843594360436143624363436443654366436743684369437043714372437343744375437643774378437943804381438243834384438543864387438843894390439143924393439443954396439743984399440044014402440344044405440644074408440944104411441244134414441544164417441844194420442144224423442444254426442744284429443044314432443344344435443644374438443944404441444244434444444544464447444844494450445144524453445444554456445744584459446044614462446344644465446644674468446944704471447244734474447544764477447844794480448144824483448444854486448744884489449044914492449344944495449644974498449945004501450245034504450545064507450845094510451145124513451445154516451745184519452045214522452345244525452645274528452945304531453245334534453545364537453845394540454145424543454445454546454745484549455045514552455345544555455645574558455945604561456245634564456545664567456845694570457145724573457445754576457745784579458045814582458345844585458645874588458945904591459245934594459545964597459845994600460146024603460446054606460746084609461046114612461346144615461646174618461946204621462246234624462546264627462846294630463146324633463446354636463746384639464046414642464346444645464646474648464946504651465246534654465546564657465846594660466146624663466446654666466746684669467046714672467346744675467646774678467946804681468246834684468546864687468846894690469146924693469446954696469746984699470047014702470347044705470647074708470947104711471247134714471547164717471847194720472147224723472447254726472747284729473047314732473347344735473647374738473947404741474247434744474547464747474847494750475147524753475447554756475747584759476047614762476347644765476647674768476947704771477247734774477547764777477847794780478147824783478447854786478747884789479047914792479347944795479647974798479948004801480248034804480548064807480848094810481148124813481448154816481748184819482048214822482348244825482648274828482948304831483248334834483548364837483848394840484148424843484448454846484748484849485048514852485348544855485648574858485948604861486248634864486548664867486848694870487148724873487448754876487748784879488048814882488348844885488648874888488948904891489248934894489548964897489848994900490149024903490449054906490749084909491049114912491349144915491649174918491949204921492249234924492549264927492849294930493149324933493449354936493749384939494049414942494349444945494649474948494949504951495249534954495549564957495849594960496149624963496449654966496749684969497049714972497349744975497649774978497949804981498249834984498549864987498849894990499149924993499449954996499749984999500050015002500350045005500650075008500950105011501250135014501550165017501850195020502150225023502450255026502750285029503050315032503350345035503650375038503950405041504250435044504550465047504850495050505150525053505450555056505750585059506050615062506350645065506650675068506950705071507250735074507550765077507850795080508150825083508450855086508750885089509050915092509350945095509650975098509951005101510251035104510551065107510851095110511151125113511451155116511751185119512051215122512351245125512651275128512951305131513251335134513551365137513851395140514151425143514451455146514751485149515051515152515351545155515651575158515951605161516251635164516551665167516851695170517151725173517451755176517751785179518051815182518351845185518651875188518951905191519251935194519551965197519851995200520152025203520452055206520752085209521052115212521352145215521652175218521952205221522252235224522552265227522852295230523152325233523452355236523752385239524052415242524352445245524652475248524952505251525252535254525552565257525852595260526152625263526452655266526752685269527052715272527352745275527652775278527952805281528252835284528552865287528852895290529152925293529452955296529752985299530053015302530353045305530653075308530953105311531253135314531553165317531853195320532153225323532453255326532753285329533053315332533353345335533653375338533953405341534253435344534553465347534853495350535153525353535453555356535753585359536053615362536353645365536653675368536953705371537253735374537553765377537853795380538153825383538453855386538753885389539053915392539353945395539653975398539954005401540254035404540554065407540854095410541154125413541454155416541754185419542054215422542354245425542654275428542954305431543254335434543554365437543854395440544154425443544454455446544754485449545054515452545354545455545654575458545954605461546254635464546554665467546854695470547154725473547454755476547754785479548054815482548354845485548654875488548954905491549254935494549554965497549854995500550155025503550455055506550755085509551055115512551355145515551655175518551955205521552255235524552555265527552855295530553155325533553455355536553755385539554055415542554355445545554655475548554955505551555255535554555555565557555855595560556155625563556455655566556755685569557055715572557355745575557655775578557955805581558255835584558555865587558855895590559155925593559455955596559755985599560056015602560356045605560656075608560956105611561256135614561556165617561856195620562156225623562456255626562756285629563056315632563356345635563656375638563956405641564256435644564556465647564856495650565156525653565456555656565756585659566056615662566356645665566656675668566956705671567256735674567556765677567856795680568156825683568456855686568756885689569056915692569356945695569656975698569957005701570257035704570557065707570857095710571157125713571457155716571757185719572057215722572357245725572657275728572957305731573257335734573557365737573857395740574157425743574457455746574757485749575057515752575357545755575657575758575957605761576257635764576557665767576857695770577157725773577457755776577757785779578057815782578357845785578657875788578957905791579257935794579557965797579857995800580158025803580458055806580758085809581058115812581358145815581658175818581958205821582258235824582558265827582858295830583158325833583458355836583758385839584058415842584358445845584658475848584958505851585258535854585558565857585858595860586158625863586458655866586758685869587058715872587358745875587658775878587958805881588258835884588558865887588858895890589158925893589458955896589758985899590059015902
  1. /*
  2. * INET An implementation of the TCP/IP protocol suite for the LINUX
  3. * operating system. INET is implemented using the BSD Socket
  4. * interface as the means of communication with the user level.
  5. *
  6. * Implementation of the Transmission Control Protocol(TCP).
  7. *
  8. * Authors: Ross Biro
  9. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  11. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  12. * Florian La Roche, <flla@stud.uni-sb.de>
  13. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  15. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  16. * Matthew Dillon, <dillon@apollo.west.oic.com>
  17. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18. * Jorge Cwik, <jorge@laser.satlink.net>
  19. */
  20. /*
  21. * Changes:
  22. * Pedro Roque : Fast Retransmit/Recovery.
  23. * Two receive queues.
  24. * Retransmit queue handled by TCP.
  25. * Better retransmit timer handling.
  26. * New congestion avoidance.
  27. * Header prediction.
  28. * Variable renaming.
  29. *
  30. * Eric : Fast Retransmit.
  31. * Randy Scott : MSS option defines.
  32. * Eric Schenk : Fixes to slow start algorithm.
  33. * Eric Schenk : Yet another double ACK bug.
  34. * Eric Schenk : Delayed ACK bug fixes.
  35. * Eric Schenk : Floyd style fast retrans war avoidance.
  36. * David S. Miller : Don't allow zero congestion window.
  37. * Eric Schenk : Fix retransmitter so that it sends
  38. * next packet on ack of previous packet.
  39. * Andi Kleen : Moved open_request checking here
  40. * and process RSTs for open_requests.
  41. * Andi Kleen : Better prune_queue, and other fixes.
  42. * Andrey Savochkin: Fix RTT measurements in the presence of
  43. * timestamps.
  44. * Andrey Savochkin: Check sequence numbers correctly when
  45. * removing SACKs due to in sequence incoming
  46. * data segments.
  47. * Andi Kleen: Make sure we never ack data there is not
  48. * enough room for. Also make this condition
  49. * a fatal error if it might still happen.
  50. * Andi Kleen: Add tcp_measure_rcv_mss to make
  51. * connections with MSS<min(MTU,ann. MSS)
  52. * work without delayed acks.
  53. * Andi Kleen: Process packets with PSH set in the
  54. * fast path.
  55. * J Hadi Salim: ECN support
  56. * Andrei Gurtov,
  57. * Pasi Sarolahti,
  58. * Panu Kuhlberg: Experimental audit of TCP (re)transmission
  59. * engine. Lots of bugs are found.
  60. * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
  61. */
  62. #define pr_fmt(fmt) "TCP: " fmt
  63. #include <linux/mm.h>
  64. #include <linux/slab.h>
  65. #include <linux/module.h>
  66. #include <linux/sysctl.h>
  67. #include <linux/kernel.h>
  68. #include <net/dst.h>
  69. #include <net/tcp.h>
  70. #include <net/inet_common.h>
  71. #include <linux/ipsec.h>
  72. #include <asm/unaligned.h>
  73. #include <net/netdma.h>
  74. int sysctl_tcp_timestamps __read_mostly = 1;
  75. int sysctl_tcp_window_scaling __read_mostly = 1;
  76. int sysctl_tcp_sack __read_mostly = 1;
  77. int sysctl_tcp_fack __read_mostly = 1;
  78. int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  79. EXPORT_SYMBOL(sysctl_tcp_reordering);
  80. int sysctl_tcp_dsack __read_mostly = 1;
  81. int sysctl_tcp_app_win __read_mostly = 31;
  82. int sysctl_tcp_adv_win_scale __read_mostly = 1;
  83. EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
  84. /* rfc5961 challenge ack rate limiting */
  85. int sysctl_tcp_challenge_ack_limit = 100;
  86. int sysctl_tcp_stdurg __read_mostly;
  87. int sysctl_tcp_rfc1337 __read_mostly;
  88. int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  89. int sysctl_tcp_frto __read_mostly = 2;
  90. int sysctl_tcp_thin_dupack __read_mostly;
  91. int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
  92. int sysctl_tcp_early_retrans __read_mostly = 3;
  93. #define FLAG_DATA 0x01 /* Incoming frame contained data. */
  94. #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
  95. #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
  96. #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
  97. #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
  98. #define FLAG_DATA_SACKED 0x20 /* New SACK. */
  99. #define FLAG_ECE 0x40 /* ECE in this ACK */
  100. #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
  101. #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */
  102. #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
  103. #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
  104. #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */
  105. #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */
  106. #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
  107. #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
  108. #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
  109. #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
  110. #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
  111. #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
  112. /* Adapt the MSS value used to make delayed ack decision to the
  113. * real world.
  114. */
  115. static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
  116. {
  117. struct inet_connection_sock *icsk = inet_csk(sk);
  118. const unsigned int lss = icsk->icsk_ack.last_seg_size;
  119. unsigned int len;
  120. icsk->icsk_ack.last_seg_size = 0;
  121. /* skb->len may jitter because of SACKs, even if peer
  122. * sends good full-sized frames.
  123. */
  124. len = skb_shinfo(skb)->gso_size ? : skb->len;
  125. if (len >= icsk->icsk_ack.rcv_mss) {
  126. icsk->icsk_ack.rcv_mss = len;
  127. } else {
  128. /* Otherwise, we make more careful check taking into account,
  129. * that SACKs block is variable.
  130. *
  131. * "len" is invariant segment length, including TCP header.
  132. */
  133. len += skb->data - skb_transport_header(skb);
  134. if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
  135. /* If PSH is not set, packet should be
  136. * full sized, provided peer TCP is not badly broken.
  137. * This observation (if it is correct 8)) allows
  138. * to handle super-low mtu links fairly.
  139. */
  140. (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
  141. !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
  142. /* Subtract also invariant (if peer is RFC compliant),
  143. * tcp header plus fixed timestamp option length.
  144. * Resulting "len" is MSS free of SACK jitter.
  145. */
  146. len -= tcp_sk(sk)->tcp_header_len;
  147. icsk->icsk_ack.last_seg_size = len;
  148. if (len == lss) {
  149. icsk->icsk_ack.rcv_mss = len;
  150. return;
  151. }
  152. }
  153. if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
  154. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
  155. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
  156. }
  157. }
  158. static void tcp_incr_quickack(struct sock *sk)
  159. {
  160. struct inet_connection_sock *icsk = inet_csk(sk);
  161. unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
  162. if (quickacks == 0)
  163. quickacks = 2;
  164. if (quickacks > icsk->icsk_ack.quick)
  165. icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
  166. }
  167. static void tcp_enter_quickack_mode(struct sock *sk)
  168. {
  169. struct inet_connection_sock *icsk = inet_csk(sk);
  170. tcp_incr_quickack(sk);
  171. icsk->icsk_ack.pingpong = 0;
  172. icsk->icsk_ack.ato = TCP_ATO_MIN;
  173. }
  174. /* Send ACKs quickly, if "quick" count is not exhausted
  175. * and the session is not interactive.
  176. */
  177. static inline bool tcp_in_quickack_mode(const struct sock *sk)
  178. {
  179. const struct inet_connection_sock *icsk = inet_csk(sk);
  180. return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
  181. }
  182. static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
  183. {
  184. if (tp->ecn_flags & TCP_ECN_OK)
  185. tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
  186. }
  187. static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
  188. {
  189. if (tcp_hdr(skb)->cwr)
  190. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  191. }
  192. static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
  193. {
  194. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  195. }
  196. static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
  197. {
  198. if (!(tp->ecn_flags & TCP_ECN_OK))
  199. return;
  200. switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
  201. case INET_ECN_NOT_ECT:
  202. /* Funny extension: if ECT is not set on a segment,
  203. * and we already seen ECT on a previous segment,
  204. * it is probably a retransmit.
  205. */
  206. if (tp->ecn_flags & TCP_ECN_SEEN)
  207. tcp_enter_quickack_mode((struct sock *)tp);
  208. break;
  209. case INET_ECN_CE:
  210. if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
  211. /* Better not delay acks, sender can have a very low cwnd */
  212. tcp_enter_quickack_mode((struct sock *)tp);
  213. tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
  214. }
  215. /* fallinto */
  216. default:
  217. tp->ecn_flags |= TCP_ECN_SEEN;
  218. }
  219. }
  220. static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
  221. {
  222. if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
  223. tp->ecn_flags &= ~TCP_ECN_OK;
  224. }
  225. static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
  226. {
  227. if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
  228. tp->ecn_flags &= ~TCP_ECN_OK;
  229. }
  230. static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
  231. {
  232. if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
  233. return true;
  234. return false;
  235. }
  236. /* Buffer size and advertised window tuning.
  237. *
  238. * 1. Tuning sk->sk_sndbuf, when connection enters established state.
  239. */
  240. static void tcp_sndbuf_expand(struct sock *sk)
  241. {
  242. const struct tcp_sock *tp = tcp_sk(sk);
  243. int sndmem, per_mss;
  244. u32 nr_segs;
  245. /* Worst case is non GSO/TSO : each frame consumes one skb
  246. * and skb->head is kmalloced using power of two area of memory
  247. */
  248. per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
  249. MAX_TCP_HEADER +
  250. SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
  251. per_mss = roundup_pow_of_two(per_mss) +
  252. SKB_DATA_ALIGN(sizeof(struct sk_buff));
  253. nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
  254. nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
  255. /* Fast Recovery (RFC 5681 3.2) :
  256. * Cubic needs 1.7 factor, rounded to 2 to include
  257. * extra cushion (application might react slowly to POLLOUT)
  258. */
  259. sndmem = 2 * nr_segs * per_mss;
  260. if (sk->sk_sndbuf < sndmem)
  261. sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
  262. }
  263. /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
  264. *
  265. * All tcp_full_space() is split to two parts: "network" buffer, allocated
  266. * forward and advertised in receiver window (tp->rcv_wnd) and
  267. * "application buffer", required to isolate scheduling/application
  268. * latencies from network.
  269. * window_clamp is maximal advertised window. It can be less than
  270. * tcp_full_space(), in this case tcp_full_space() - window_clamp
  271. * is reserved for "application" buffer. The less window_clamp is
  272. * the smoother our behaviour from viewpoint of network, but the lower
  273. * throughput and the higher sensitivity of the connection to losses. 8)
  274. *
  275. * rcv_ssthresh is more strict window_clamp used at "slow start"
  276. * phase to predict further behaviour of this connection.
  277. * It is used for two goals:
  278. * - to enforce header prediction at sender, even when application
  279. * requires some significant "application buffer". It is check #1.
  280. * - to prevent pruning of receive queue because of misprediction
  281. * of receiver window. Check #2.
  282. *
  283. * The scheme does not work when sender sends good segments opening
  284. * window and then starts to feed us spaghetti. But it should work
  285. * in common situations. Otherwise, we have to rely on queue collapsing.
  286. */
  287. /* Slow part of check#2. */
  288. static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
  289. {
  290. struct tcp_sock *tp = tcp_sk(sk);
  291. /* Optimize this! */
  292. int truesize = tcp_win_from_space(skb->truesize) >> 1;
  293. int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
  294. while (tp->rcv_ssthresh <= window) {
  295. if (truesize <= skb->len)
  296. return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
  297. truesize >>= 1;
  298. window >>= 1;
  299. }
  300. return 0;
  301. }
  302. static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
  303. {
  304. struct tcp_sock *tp = tcp_sk(sk);
  305. /* Check #1 */
  306. if (tp->rcv_ssthresh < tp->window_clamp &&
  307. (int)tp->rcv_ssthresh < tcp_space(sk) &&
  308. !sk_under_memory_pressure(sk)) {
  309. int incr;
  310. /* Check #2. Increase window, if skb with such overhead
  311. * will fit to rcvbuf in future.
  312. */
  313. if (tcp_win_from_space(skb->truesize) <= skb->len)
  314. incr = 2 * tp->advmss;
  315. else
  316. incr = __tcp_grow_window(sk, skb);
  317. if (incr) {
  318. incr = max_t(int, incr, 2 * skb->len);
  319. tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
  320. tp->window_clamp);
  321. inet_csk(sk)->icsk_ack.quick |= 1;
  322. }
  323. }
  324. }
  325. /* 3. Tuning rcvbuf, when connection enters established state. */
  326. static void tcp_fixup_rcvbuf(struct sock *sk)
  327. {
  328. u32 mss = tcp_sk(sk)->advmss;
  329. int rcvmem;
  330. rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
  331. tcp_default_init_rwnd(mss);
  332. /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
  333. * Allow enough cushion so that sender is not limited by our window
  334. */
  335. if (sysctl_tcp_moderate_rcvbuf)
  336. rcvmem <<= 2;
  337. if (sk->sk_rcvbuf < rcvmem)
  338. sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
  339. }
  340. /* 4. Try to fixup all. It is made immediately after connection enters
  341. * established state.
  342. */
  343. void tcp_init_buffer_space(struct sock *sk)
  344. {
  345. struct tcp_sock *tp = tcp_sk(sk);
  346. int maxwin;
  347. if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
  348. tcp_fixup_rcvbuf(sk);
  349. if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
  350. tcp_sndbuf_expand(sk);
  351. tp->rcvq_space.space = tp->rcv_wnd;
  352. tp->rcvq_space.time = tcp_time_stamp;
  353. tp->rcvq_space.seq = tp->copied_seq;
  354. maxwin = tcp_full_space(sk);
  355. if (tp->window_clamp >= maxwin) {
  356. tp->window_clamp = maxwin;
  357. if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
  358. tp->window_clamp = max(maxwin -
  359. (maxwin >> sysctl_tcp_app_win),
  360. 4 * tp->advmss);
  361. }
  362. /* Force reservation of one segment. */
  363. if (sysctl_tcp_app_win &&
  364. tp->window_clamp > 2 * tp->advmss &&
  365. tp->window_clamp + tp->advmss > maxwin)
  366. tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
  367. tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
  368. tp->snd_cwnd_stamp = tcp_time_stamp;
  369. }
  370. /* 5. Recalculate window clamp after socket hit its memory bounds. */
  371. static void tcp_clamp_window(struct sock *sk)
  372. {
  373. struct tcp_sock *tp = tcp_sk(sk);
  374. struct inet_connection_sock *icsk = inet_csk(sk);
  375. icsk->icsk_ack.quick = 0;
  376. if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
  377. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
  378. !sk_under_memory_pressure(sk) &&
  379. sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
  380. sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
  381. sysctl_tcp_rmem[2]);
  382. }
  383. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
  384. tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
  385. }
  386. /* Initialize RCV_MSS value.
  387. * RCV_MSS is an our guess about MSS used by the peer.
  388. * We haven't any direct information about the MSS.
  389. * It's better to underestimate the RCV_MSS rather than overestimate.
  390. * Overestimations make us ACKing less frequently than needed.
  391. * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
  392. */
  393. void tcp_initialize_rcv_mss(struct sock *sk)
  394. {
  395. const struct tcp_sock *tp = tcp_sk(sk);
  396. unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
  397. hint = min(hint, tp->rcv_wnd / 2);
  398. hint = min(hint, TCP_MSS_DEFAULT);
  399. hint = max(hint, TCP_MIN_MSS);
  400. inet_csk(sk)->icsk_ack.rcv_mss = hint;
  401. }
  402. EXPORT_SYMBOL(tcp_initialize_rcv_mss);
  403. /* Receiver "autotuning" code.
  404. *
  405. * The algorithm for RTT estimation w/o timestamps is based on
  406. * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
  407. * <http://public.lanl.gov/radiant/pubs.html#DRS>
  408. *
  409. * More detail on this code can be found at
  410. * <http://staff.psc.edu/jheffner/>,
  411. * though this reference is out of date. A new paper
  412. * is pending.
  413. */
  414. static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
  415. {
  416. u32 new_sample = tp->rcv_rtt_est.rtt;
  417. long m = sample;
  418. if (m == 0)
  419. m = 1;
  420. if (new_sample != 0) {
  421. /* If we sample in larger samples in the non-timestamp
  422. * case, we could grossly overestimate the RTT especially
  423. * with chatty applications or bulk transfer apps which
  424. * are stalled on filesystem I/O.
  425. *
  426. * Also, since we are only going for a minimum in the
  427. * non-timestamp case, we do not smooth things out
  428. * else with timestamps disabled convergence takes too
  429. * long.
  430. */
  431. if (!win_dep) {
  432. m -= (new_sample >> 3);
  433. new_sample += m;
  434. } else {
  435. m <<= 3;
  436. if (m < new_sample)
  437. new_sample = m;
  438. }
  439. } else {
  440. /* No previous measure. */
  441. new_sample = m << 3;
  442. }
  443. if (tp->rcv_rtt_est.rtt != new_sample)
  444. tp->rcv_rtt_est.rtt = new_sample;
  445. }
  446. static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
  447. {
  448. if (tp->rcv_rtt_est.time == 0)
  449. goto new_measure;
  450. if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
  451. return;
  452. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
  453. new_measure:
  454. tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
  455. tp->rcv_rtt_est.time = tcp_time_stamp;
  456. }
  457. static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
  458. const struct sk_buff *skb)
  459. {
  460. struct tcp_sock *tp = tcp_sk(sk);
  461. if (tp->rx_opt.rcv_tsecr &&
  462. (TCP_SKB_CB(skb)->end_seq -
  463. TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
  464. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
  465. }
  466. /*
  467. * This function should be called every time data is copied to user space.
  468. * It calculates the appropriate TCP receive buffer space.
  469. */
  470. void tcp_rcv_space_adjust(struct sock *sk)
  471. {
  472. struct tcp_sock *tp = tcp_sk(sk);
  473. int time;
  474. int copied;
  475. time = tcp_time_stamp - tp->rcvq_space.time;
  476. if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
  477. return;
  478. /* Number of bytes copied to user in last RTT */
  479. copied = tp->copied_seq - tp->rcvq_space.seq;
  480. if (copied <= tp->rcvq_space.space)
  481. goto new_measure;
  482. /* A bit of theory :
  483. * copied = bytes received in previous RTT, our base window
  484. * To cope with packet losses, we need a 2x factor
  485. * To cope with slow start, and sender growing its cwin by 100 %
  486. * every RTT, we need a 4x factor, because the ACK we are sending
  487. * now is for the next RTT, not the current one :
  488. * <prev RTT . ><current RTT .. ><next RTT .... >
  489. */
  490. if (sysctl_tcp_moderate_rcvbuf &&
  491. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
  492. int rcvwin, rcvmem, rcvbuf;
  493. /* minimal window to cope with packet losses, assuming
  494. * steady state. Add some cushion because of small variations.
  495. */
  496. rcvwin = (copied << 1) + 16 * tp->advmss;
  497. /* If rate increased by 25%,
  498. * assume slow start, rcvwin = 3 * copied
  499. * If rate increased by 50%,
  500. * assume sender can use 2x growth, rcvwin = 4 * copied
  501. */
  502. if (copied >=
  503. tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
  504. if (copied >=
  505. tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
  506. rcvwin <<= 1;
  507. else
  508. rcvwin += (rcvwin >> 1);
  509. }
  510. rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
  511. while (tcp_win_from_space(rcvmem) < tp->advmss)
  512. rcvmem += 128;
  513. rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
  514. if (rcvbuf > sk->sk_rcvbuf) {
  515. sk->sk_rcvbuf = rcvbuf;
  516. /* Make the window clamp follow along. */
  517. tp->window_clamp = rcvwin;
  518. }
  519. }
  520. tp->rcvq_space.space = copied;
  521. new_measure:
  522. tp->rcvq_space.seq = tp->copied_seq;
  523. tp->rcvq_space.time = tcp_time_stamp;
  524. }
  525. /* There is something which you must keep in mind when you analyze the
  526. * behavior of the tp->ato delayed ack timeout interval. When a
  527. * connection starts up, we want to ack as quickly as possible. The
  528. * problem is that "good" TCP's do slow start at the beginning of data
  529. * transmission. The means that until we send the first few ACK's the
  530. * sender will sit on his end and only queue most of his data, because
  531. * he can only send snd_cwnd unacked packets at any given time. For
  532. * each ACK we send, he increments snd_cwnd and transmits more of his
  533. * queue. -DaveM
  534. */
  535. static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
  536. {
  537. struct tcp_sock *tp = tcp_sk(sk);
  538. struct inet_connection_sock *icsk = inet_csk(sk);
  539. u32 now;
  540. inet_csk_schedule_ack(sk);
  541. tcp_measure_rcv_mss(sk, skb);
  542. tcp_rcv_rtt_measure(tp);
  543. now = tcp_time_stamp;
  544. if (!icsk->icsk_ack.ato) {
  545. /* The _first_ data packet received, initialize
  546. * delayed ACK engine.
  547. */
  548. tcp_incr_quickack(sk);
  549. icsk->icsk_ack.ato = TCP_ATO_MIN;
  550. } else {
  551. int m = now - icsk->icsk_ack.lrcvtime;
  552. if (m <= TCP_ATO_MIN / 2) {
  553. /* The fastest case is the first. */
  554. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
  555. } else if (m < icsk->icsk_ack.ato) {
  556. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
  557. if (icsk->icsk_ack.ato > icsk->icsk_rto)
  558. icsk->icsk_ack.ato = icsk->icsk_rto;
  559. } else if (m > icsk->icsk_rto) {
  560. /* Too long gap. Apparently sender failed to
  561. * restart window, so that we send ACKs quickly.
  562. */
  563. tcp_incr_quickack(sk);
  564. sk_mem_reclaim(sk);
  565. }
  566. }
  567. icsk->icsk_ack.lrcvtime = now;
  568. TCP_ECN_check_ce(tp, skb);
  569. if (skb->len >= 128)
  570. tcp_grow_window(sk, skb);
  571. }
  572. /* Called to compute a smoothed rtt estimate. The data fed to this
  573. * routine either comes from timestamps, or from segments that were
  574. * known _not_ to have been retransmitted [see Karn/Partridge
  575. * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
  576. * piece by Van Jacobson.
  577. * NOTE: the next three routines used to be one big routine.
  578. * To save cycles in the RFC 1323 implementation it was better to break
  579. * it up into three procedures. -- erics
  580. */
  581. static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
  582. {
  583. struct tcp_sock *tp = tcp_sk(sk);
  584. long m = mrtt; /* RTT */
  585. /* The following amusing code comes from Jacobson's
  586. * article in SIGCOMM '88. Note that rtt and mdev
  587. * are scaled versions of rtt and mean deviation.
  588. * This is designed to be as fast as possible
  589. * m stands for "measurement".
  590. *
  591. * On a 1990 paper the rto value is changed to:
  592. * RTO = rtt + 4 * mdev
  593. *
  594. * Funny. This algorithm seems to be very broken.
  595. * These formulae increase RTO, when it should be decreased, increase
  596. * too slowly, when it should be increased quickly, decrease too quickly
  597. * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
  598. * does not matter how to _calculate_ it. Seems, it was trap
  599. * that VJ failed to avoid. 8)
  600. */
  601. if (m == 0)
  602. m = 1;
  603. if (tp->srtt != 0) {
  604. m -= (tp->srtt >> 3); /* m is now error in rtt est */
  605. tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
  606. if (m < 0) {
  607. m = -m; /* m is now abs(error) */
  608. m -= (tp->mdev >> 2); /* similar update on mdev */
  609. /* This is similar to one of Eifel findings.
  610. * Eifel blocks mdev updates when rtt decreases.
  611. * This solution is a bit different: we use finer gain
  612. * for mdev in this case (alpha*beta).
  613. * Like Eifel it also prevents growth of rto,
  614. * but also it limits too fast rto decreases,
  615. * happening in pure Eifel.
  616. */
  617. if (m > 0)
  618. m >>= 3;
  619. } else {
  620. m -= (tp->mdev >> 2); /* similar update on mdev */
  621. }
  622. tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
  623. if (tp->mdev > tp->mdev_max) {
  624. tp->mdev_max = tp->mdev;
  625. if (tp->mdev_max > tp->rttvar)
  626. tp->rttvar = tp->mdev_max;
  627. }
  628. if (after(tp->snd_una, tp->rtt_seq)) {
  629. if (tp->mdev_max < tp->rttvar)
  630. tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2;
  631. tp->rtt_seq = tp->snd_nxt;
  632. tp->mdev_max = tcp_rto_min(sk);
  633. }
  634. } else {
  635. /* no previous measure. */
  636. tp->srtt = m << 3; /* take the measured time to be rtt */
  637. tp->mdev = m << 1; /* make sure rto = 3*rtt */
  638. tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
  639. tp->rtt_seq = tp->snd_nxt;
  640. }
  641. }
  642. /* Set the sk_pacing_rate to allow proper sizing of TSO packets.
  643. * Note: TCP stack does not yet implement pacing.
  644. * FQ packet scheduler can be used to implement cheap but effective
  645. * TCP pacing, to smooth the burst on large writes when packets
  646. * in flight is significantly lower than cwnd (or rwin)
  647. */
  648. static void tcp_update_pacing_rate(struct sock *sk)
  649. {
  650. const struct tcp_sock *tp = tcp_sk(sk);
  651. u64 rate;
  652. /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
  653. rate = (u64)tp->mss_cache * 2 * (HZ << 3);
  654. rate *= max(tp->snd_cwnd, tp->packets_out);
  655. /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3),
  656. * be conservative and assume srtt = 1 (125 us instead of 1.25 ms)
  657. * We probably need usec resolution in the future.
  658. * Note: This also takes care of possible srtt=0 case,
  659. * when tcp_rtt_estimator() was not yet called.
  660. */
  661. if (tp->srtt > 8 + 2)
  662. do_div(rate, tp->srtt);
  663. /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
  664. * without any lock. We want to make sure compiler wont store
  665. * intermediate values in this location.
  666. */
  667. ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
  668. sk->sk_max_pacing_rate);
  669. }
  670. /* Calculate rto without backoff. This is the second half of Van Jacobson's
  671. * routine referred to above.
  672. */
  673. void tcp_set_rto(struct sock *sk)
  674. {
  675. const struct tcp_sock *tp = tcp_sk(sk);
  676. /* Old crap is replaced with new one. 8)
  677. *
  678. * More seriously:
  679. * 1. If rtt variance happened to be less 50msec, it is hallucination.
  680. * It cannot be less due to utterly erratic ACK generation made
  681. * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
  682. * to do with delayed acks, because at cwnd>2 true delack timeout
  683. * is invisible. Actually, Linux-2.4 also generates erratic
  684. * ACKs in some circumstances.
  685. */
  686. inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
  687. /* 2. Fixups made earlier cannot be right.
  688. * If we do not estimate RTO correctly without them,
  689. * all the algo is pure shit and should be replaced
  690. * with correct one. It is exactly, which we pretend to do.
  691. */
  692. /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
  693. * guarantees that rto is higher.
  694. */
  695. tcp_bound_rto(sk);
  696. }
  697. __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
  698. {
  699. __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
  700. if (!cwnd)
  701. cwnd = TCP_INIT_CWND;
  702. return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
  703. }
  704. /*
  705. * Packet counting of FACK is based on in-order assumptions, therefore TCP
  706. * disables it when reordering is detected
  707. */
  708. void tcp_disable_fack(struct tcp_sock *tp)
  709. {
  710. /* RFC3517 uses different metric in lost marker => reset on change */
  711. if (tcp_is_fack(tp))
  712. tp->lost_skb_hint = NULL;
  713. tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
  714. }
  715. /* Take a notice that peer is sending D-SACKs */
  716. static void tcp_dsack_seen(struct tcp_sock *tp)
  717. {
  718. tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
  719. }
  720. static void tcp_update_reordering(struct sock *sk, const int metric,
  721. const int ts)
  722. {
  723. struct tcp_sock *tp = tcp_sk(sk);
  724. if (metric > tp->reordering) {
  725. int mib_idx;
  726. tp->reordering = min(TCP_MAX_REORDERING, metric);
  727. /* This exciting event is worth to be remembered. 8) */
  728. if (ts)
  729. mib_idx = LINUX_MIB_TCPTSREORDER;
  730. else if (tcp_is_reno(tp))
  731. mib_idx = LINUX_MIB_TCPRENOREORDER;
  732. else if (tcp_is_fack(tp))
  733. mib_idx = LINUX_MIB_TCPFACKREORDER;
  734. else
  735. mib_idx = LINUX_MIB_TCPSACKREORDER;
  736. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  737. #if FASTRETRANS_DEBUG > 1
  738. pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
  739. tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
  740. tp->reordering,
  741. tp->fackets_out,
  742. tp->sacked_out,
  743. tp->undo_marker ? tp->undo_retrans : 0);
  744. #endif
  745. tcp_disable_fack(tp);
  746. }
  747. if (metric > 0)
  748. tcp_disable_early_retrans(tp);
  749. }
  750. /* This must be called before lost_out is incremented */
  751. static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
  752. {
  753. if ((tp->retransmit_skb_hint == NULL) ||
  754. before(TCP_SKB_CB(skb)->seq,
  755. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  756. tp->retransmit_skb_hint = skb;
  757. if (!tp->lost_out ||
  758. after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
  759. tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
  760. }
  761. static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
  762. {
  763. if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  764. tcp_verify_retransmit_hint(tp, skb);
  765. tp->lost_out += tcp_skb_pcount(skb);
  766. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  767. }
  768. }
  769. static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
  770. struct sk_buff *skb)
  771. {
  772. tcp_verify_retransmit_hint(tp, skb);
  773. if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  774. tp->lost_out += tcp_skb_pcount(skb);
  775. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  776. }
  777. }
  778. /* This procedure tags the retransmission queue when SACKs arrive.
  779. *
  780. * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
  781. * Packets in queue with these bits set are counted in variables
  782. * sacked_out, retrans_out and lost_out, correspondingly.
  783. *
  784. * Valid combinations are:
  785. * Tag InFlight Description
  786. * 0 1 - orig segment is in flight.
  787. * S 0 - nothing flies, orig reached receiver.
  788. * L 0 - nothing flies, orig lost by net.
  789. * R 2 - both orig and retransmit are in flight.
  790. * L|R 1 - orig is lost, retransmit is in flight.
  791. * S|R 1 - orig reached receiver, retrans is still in flight.
  792. * (L|S|R is logically valid, it could occur when L|R is sacked,
  793. * but it is equivalent to plain S and code short-curcuits it to S.
  794. * L|S is logically invalid, it would mean -1 packet in flight 8))
  795. *
  796. * These 6 states form finite state machine, controlled by the following events:
  797. * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
  798. * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
  799. * 3. Loss detection event of two flavors:
  800. * A. Scoreboard estimator decided the packet is lost.
  801. * A'. Reno "three dupacks" marks head of queue lost.
  802. * A''. Its FACK modification, head until snd.fack is lost.
  803. * B. SACK arrives sacking SND.NXT at the moment, when the
  804. * segment was retransmitted.
  805. * 4. D-SACK added new rule: D-SACK changes any tag to S.
  806. *
  807. * It is pleasant to note, that state diagram turns out to be commutative,
  808. * so that we are allowed not to be bothered by order of our actions,
  809. * when multiple events arrive simultaneously. (see the function below).
  810. *
  811. * Reordering detection.
  812. * --------------------
  813. * Reordering metric is maximal distance, which a packet can be displaced
  814. * in packet stream. With SACKs we can estimate it:
  815. *
  816. * 1. SACK fills old hole and the corresponding segment was not
  817. * ever retransmitted -> reordering. Alas, we cannot use it
  818. * when segment was retransmitted.
  819. * 2. The last flaw is solved with D-SACK. D-SACK arrives
  820. * for retransmitted and already SACKed segment -> reordering..
  821. * Both of these heuristics are not used in Loss state, when we cannot
  822. * account for retransmits accurately.
  823. *
  824. * SACK block validation.
  825. * ----------------------
  826. *
  827. * SACK block range validation checks that the received SACK block fits to
  828. * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
  829. * Note that SND.UNA is not included to the range though being valid because
  830. * it means that the receiver is rather inconsistent with itself reporting
  831. * SACK reneging when it should advance SND.UNA. Such SACK block this is
  832. * perfectly valid, however, in light of RFC2018 which explicitly states
  833. * that "SACK block MUST reflect the newest segment. Even if the newest
  834. * segment is going to be discarded ...", not that it looks very clever
  835. * in case of head skb. Due to potentional receiver driven attacks, we
  836. * choose to avoid immediate execution of a walk in write queue due to
  837. * reneging and defer head skb's loss recovery to standard loss recovery
  838. * procedure that will eventually trigger (nothing forbids us doing this).
  839. *
  840. * Implements also blockage to start_seq wrap-around. Problem lies in the
  841. * fact that though start_seq (s) is before end_seq (i.e., not reversed),
  842. * there's no guarantee that it will be before snd_nxt (n). The problem
  843. * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
  844. * wrap (s_w):
  845. *
  846. * <- outs wnd -> <- wrapzone ->
  847. * u e n u_w e_w s n_w
  848. * | | | | | | |
  849. * |<------------+------+----- TCP seqno space --------------+---------->|
  850. * ...-- <2^31 ->| |<--------...
  851. * ...---- >2^31 ------>| |<--------...
  852. *
  853. * Current code wouldn't be vulnerable but it's better still to discard such
  854. * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
  855. * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
  856. * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
  857. * equal to the ideal case (infinite seqno space without wrap caused issues).
  858. *
  859. * With D-SACK the lower bound is extended to cover sequence space below
  860. * SND.UNA down to undo_marker, which is the last point of interest. Yet
  861. * again, D-SACK block must not to go across snd_una (for the same reason as
  862. * for the normal SACK blocks, explained above). But there all simplicity
  863. * ends, TCP might receive valid D-SACKs below that. As long as they reside
  864. * fully below undo_marker they do not affect behavior in anyway and can
  865. * therefore be safely ignored. In rare cases (which are more or less
  866. * theoretical ones), the D-SACK will nicely cross that boundary due to skb
  867. * fragmentation and packet reordering past skb's retransmission. To consider
  868. * them correctly, the acceptable range must be extended even more though
  869. * the exact amount is rather hard to quantify. However, tp->max_window can
  870. * be used as an exaggerated estimate.
  871. */
  872. static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
  873. u32 start_seq, u32 end_seq)
  874. {
  875. /* Too far in future, or reversed (interpretation is ambiguous) */
  876. if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
  877. return false;
  878. /* Nasty start_seq wrap-around check (see comments above) */
  879. if (!before(start_seq, tp->snd_nxt))
  880. return false;
  881. /* In outstanding window? ...This is valid exit for D-SACKs too.
  882. * start_seq == snd_una is non-sensical (see comments above)
  883. */
  884. if (after(start_seq, tp->snd_una))
  885. return true;
  886. if (!is_dsack || !tp->undo_marker)
  887. return false;
  888. /* ...Then it's D-SACK, and must reside below snd_una completely */
  889. if (after(end_seq, tp->snd_una))
  890. return false;
  891. if (!before(start_seq, tp->undo_marker))
  892. return true;
  893. /* Too old */
  894. if (!after(end_seq, tp->undo_marker))
  895. return false;
  896. /* Undo_marker boundary crossing (overestimates a lot). Known already:
  897. * start_seq < undo_marker and end_seq >= undo_marker.
  898. */
  899. return !before(start_seq, end_seq - tp->max_window);
  900. }
  901. /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
  902. * Event "B". Later note: FACK people cheated me again 8), we have to account
  903. * for reordering! Ugly, but should help.
  904. *
  905. * Search retransmitted skbs from write_queue that were sent when snd_nxt was
  906. * less than what is now known to be received by the other end (derived from
  907. * highest SACK block). Also calculate the lowest snd_nxt among the remaining
  908. * retransmitted skbs to avoid some costly processing per ACKs.
  909. */
  910. static void tcp_mark_lost_retrans(struct sock *sk)
  911. {
  912. const struct inet_connection_sock *icsk = inet_csk(sk);
  913. struct tcp_sock *tp = tcp_sk(sk);
  914. struct sk_buff *skb;
  915. int cnt = 0;
  916. u32 new_low_seq = tp->snd_nxt;
  917. u32 received_upto = tcp_highest_sack_seq(tp);
  918. if (!tcp_is_fack(tp) || !tp->retrans_out ||
  919. !after(received_upto, tp->lost_retrans_low) ||
  920. icsk->icsk_ca_state != TCP_CA_Recovery)
  921. return;
  922. tcp_for_write_queue(skb, sk) {
  923. u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
  924. if (skb == tcp_send_head(sk))
  925. break;
  926. if (cnt == tp->retrans_out)
  927. break;
  928. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  929. continue;
  930. if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
  931. continue;
  932. /* TODO: We would like to get rid of tcp_is_fack(tp) only
  933. * constraint here (see above) but figuring out that at
  934. * least tp->reordering SACK blocks reside between ack_seq
  935. * and received_upto is not easy task to do cheaply with
  936. * the available datastructures.
  937. *
  938. * Whether FACK should check here for tp->reordering segs
  939. * in-between one could argue for either way (it would be
  940. * rather simple to implement as we could count fack_count
  941. * during the walk and do tp->fackets_out - fack_count).
  942. */
  943. if (after(received_upto, ack_seq)) {
  944. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  945. tp->retrans_out -= tcp_skb_pcount(skb);
  946. tcp_skb_mark_lost_uncond_verify(tp, skb);
  947. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
  948. } else {
  949. if (before(ack_seq, new_low_seq))
  950. new_low_seq = ack_seq;
  951. cnt += tcp_skb_pcount(skb);
  952. }
  953. }
  954. if (tp->retrans_out)
  955. tp->lost_retrans_low = new_low_seq;
  956. }
  957. static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
  958. struct tcp_sack_block_wire *sp, int num_sacks,
  959. u32 prior_snd_una)
  960. {
  961. struct tcp_sock *tp = tcp_sk(sk);
  962. u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
  963. u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
  964. bool dup_sack = false;
  965. if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
  966. dup_sack = true;
  967. tcp_dsack_seen(tp);
  968. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
  969. } else if (num_sacks > 1) {
  970. u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
  971. u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
  972. if (!after(end_seq_0, end_seq_1) &&
  973. !before(start_seq_0, start_seq_1)) {
  974. dup_sack = true;
  975. tcp_dsack_seen(tp);
  976. NET_INC_STATS_BH(sock_net(sk),
  977. LINUX_MIB_TCPDSACKOFORECV);
  978. }
  979. }
  980. /* D-SACK for already forgotten data... Do dumb counting. */
  981. if (dup_sack && tp->undo_marker && tp->undo_retrans &&
  982. !after(end_seq_0, prior_snd_una) &&
  983. after(end_seq_0, tp->undo_marker))
  984. tp->undo_retrans--;
  985. return dup_sack;
  986. }
  987. struct tcp_sacktag_state {
  988. int reord;
  989. int fack_count;
  990. int flag;
  991. s32 rtt; /* RTT measured by SACKing never-retransmitted data */
  992. };
  993. /* Check if skb is fully within the SACK block. In presence of GSO skbs,
  994. * the incoming SACK may not exactly match but we can find smaller MSS
  995. * aligned portion of it that matches. Therefore we might need to fragment
  996. * which may fail and creates some hassle (caller must handle error case
  997. * returns).
  998. *
  999. * FIXME: this could be merged to shift decision code
  1000. */
  1001. static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
  1002. u32 start_seq, u32 end_seq)
  1003. {
  1004. int err;
  1005. bool in_sack;
  1006. unsigned int pkt_len;
  1007. unsigned int mss;
  1008. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  1009. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  1010. if (tcp_skb_pcount(skb) > 1 && !in_sack &&
  1011. after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
  1012. mss = tcp_skb_mss(skb);
  1013. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
  1014. if (!in_sack) {
  1015. pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
  1016. if (pkt_len < mss)
  1017. pkt_len = mss;
  1018. } else {
  1019. pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
  1020. if (pkt_len < mss)
  1021. return -EINVAL;
  1022. }
  1023. /* Round if necessary so that SACKs cover only full MSSes
  1024. * and/or the remaining small portion (if present)
  1025. */
  1026. if (pkt_len > mss) {
  1027. unsigned int new_len = (pkt_len / mss) * mss;
  1028. if (!in_sack && new_len < pkt_len) {
  1029. new_len += mss;
  1030. if (new_len > skb->len)
  1031. return 0;
  1032. }
  1033. pkt_len = new_len;
  1034. }
  1035. err = tcp_fragment(sk, skb, pkt_len, mss);
  1036. if (err < 0)
  1037. return err;
  1038. }
  1039. return in_sack;
  1040. }
  1041. /* Mark the given newly-SACKed range as such, adjusting counters and hints. */
  1042. static u8 tcp_sacktag_one(struct sock *sk,
  1043. struct tcp_sacktag_state *state, u8 sacked,
  1044. u32 start_seq, u32 end_seq,
  1045. int dup_sack, int pcount, u32 xmit_time)
  1046. {
  1047. struct tcp_sock *tp = tcp_sk(sk);
  1048. int fack_count = state->fack_count;
  1049. /* Account D-SACK for retransmitted packet. */
  1050. if (dup_sack && (sacked & TCPCB_RETRANS)) {
  1051. if (tp->undo_marker && tp->undo_retrans &&
  1052. after(end_seq, tp->undo_marker))
  1053. tp->undo_retrans--;
  1054. if (sacked & TCPCB_SACKED_ACKED)
  1055. state->reord = min(fack_count, state->reord);
  1056. }
  1057. /* Nothing to do; acked frame is about to be dropped (was ACKed). */
  1058. if (!after(end_seq, tp->snd_una))
  1059. return sacked;
  1060. if (!(sacked & TCPCB_SACKED_ACKED)) {
  1061. if (sacked & TCPCB_SACKED_RETRANS) {
  1062. /* If the segment is not tagged as lost,
  1063. * we do not clear RETRANS, believing
  1064. * that retransmission is still in flight.
  1065. */
  1066. if (sacked & TCPCB_LOST) {
  1067. sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1068. tp->lost_out -= pcount;
  1069. tp->retrans_out -= pcount;
  1070. }
  1071. } else {
  1072. if (!(sacked & TCPCB_RETRANS)) {
  1073. /* New sack for not retransmitted frame,
  1074. * which was in hole. It is reordering.
  1075. */
  1076. if (before(start_seq,
  1077. tcp_highest_sack_seq(tp)))
  1078. state->reord = min(fack_count,
  1079. state->reord);
  1080. if (!after(end_seq, tp->high_seq))
  1081. state->flag |= FLAG_ORIG_SACK_ACKED;
  1082. /* Pick the earliest sequence sacked for RTT */
  1083. if (state->rtt < 0)
  1084. state->rtt = tcp_time_stamp - xmit_time;
  1085. }
  1086. if (sacked & TCPCB_LOST) {
  1087. sacked &= ~TCPCB_LOST;
  1088. tp->lost_out -= pcount;
  1089. }
  1090. }
  1091. sacked |= TCPCB_SACKED_ACKED;
  1092. state->flag |= FLAG_DATA_SACKED;
  1093. tp->sacked_out += pcount;
  1094. fack_count += pcount;
  1095. /* Lost marker hint past SACKed? Tweak RFC3517 cnt */
  1096. if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
  1097. before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
  1098. tp->lost_cnt_hint += pcount;
  1099. if (fack_count > tp->fackets_out)
  1100. tp->fackets_out = fack_count;
  1101. }
  1102. /* D-SACK. We can detect redundant retransmission in S|R and plain R
  1103. * frames and clear it. undo_retrans is decreased above, L|R frames
  1104. * are accounted above as well.
  1105. */
  1106. if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
  1107. sacked &= ~TCPCB_SACKED_RETRANS;
  1108. tp->retrans_out -= pcount;
  1109. }
  1110. return sacked;
  1111. }
  1112. /* Shift newly-SACKed bytes from this skb to the immediately previous
  1113. * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
  1114. */
  1115. static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
  1116. struct tcp_sacktag_state *state,
  1117. unsigned int pcount, int shifted, int mss,
  1118. bool dup_sack)
  1119. {
  1120. struct tcp_sock *tp = tcp_sk(sk);
  1121. struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
  1122. u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */
  1123. u32 end_seq = start_seq + shifted; /* end of newly-SACKed */
  1124. BUG_ON(!pcount);
  1125. /* Adjust counters and hints for the newly sacked sequence
  1126. * range but discard the return value since prev is already
  1127. * marked. We must tag the range first because the seq
  1128. * advancement below implicitly advances
  1129. * tcp_highest_sack_seq() when skb is highest_sack.
  1130. */
  1131. tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
  1132. start_seq, end_seq, dup_sack, pcount,
  1133. TCP_SKB_CB(skb)->when);
  1134. if (skb == tp->lost_skb_hint)
  1135. tp->lost_cnt_hint += pcount;
  1136. TCP_SKB_CB(prev)->end_seq += shifted;
  1137. TCP_SKB_CB(skb)->seq += shifted;
  1138. skb_shinfo(prev)->gso_segs += pcount;
  1139. BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
  1140. skb_shinfo(skb)->gso_segs -= pcount;
  1141. /* When we're adding to gso_segs == 1, gso_size will be zero,
  1142. * in theory this shouldn't be necessary but as long as DSACK
  1143. * code can come after this skb later on it's better to keep
  1144. * setting gso_size to something.
  1145. */
  1146. if (!skb_shinfo(prev)->gso_size) {
  1147. skb_shinfo(prev)->gso_size = mss;
  1148. skb_shinfo(prev)->gso_type = sk->sk_gso_type;
  1149. }
  1150. /* CHECKME: To clear or not to clear? Mimics normal skb currently */
  1151. if (skb_shinfo(skb)->gso_segs <= 1) {
  1152. skb_shinfo(skb)->gso_size = 0;
  1153. skb_shinfo(skb)->gso_type = 0;
  1154. }
  1155. /* Difference in this won't matter, both ACKed by the same cumul. ACK */
  1156. TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
  1157. if (skb->len > 0) {
  1158. BUG_ON(!tcp_skb_pcount(skb));
  1159. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
  1160. return false;
  1161. }
  1162. /* Whole SKB was eaten :-) */
  1163. if (skb == tp->retransmit_skb_hint)
  1164. tp->retransmit_skb_hint = prev;
  1165. if (skb == tp->lost_skb_hint) {
  1166. tp->lost_skb_hint = prev;
  1167. tp->lost_cnt_hint -= tcp_skb_pcount(prev);
  1168. }
  1169. TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
  1170. if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
  1171. TCP_SKB_CB(prev)->end_seq++;
  1172. if (skb == tcp_highest_sack(sk))
  1173. tcp_advance_highest_sack(sk, skb);
  1174. tcp_unlink_write_queue(skb, sk);
  1175. sk_wmem_free_skb(sk, skb);
  1176. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
  1177. return true;
  1178. }
  1179. /* I wish gso_size would have a bit more sane initialization than
  1180. * something-or-zero which complicates things
  1181. */
  1182. static int tcp_skb_seglen(const struct sk_buff *skb)
  1183. {
  1184. return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
  1185. }
  1186. /* Shifting pages past head area doesn't work */
  1187. static int skb_can_shift(const struct sk_buff *skb)
  1188. {
  1189. return !skb_headlen(skb) && skb_is_nonlinear(skb);
  1190. }
  1191. /* Try collapsing SACK blocks spanning across multiple skbs to a single
  1192. * skb.
  1193. */
  1194. static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
  1195. struct tcp_sacktag_state *state,
  1196. u32 start_seq, u32 end_seq,
  1197. bool dup_sack)
  1198. {
  1199. struct tcp_sock *tp = tcp_sk(sk);
  1200. struct sk_buff *prev;
  1201. int mss;
  1202. int pcount = 0;
  1203. int len;
  1204. int in_sack;
  1205. if (!sk_can_gso(sk))
  1206. goto fallback;
  1207. /* Normally R but no L won't result in plain S */
  1208. if (!dup_sack &&
  1209. (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
  1210. goto fallback;
  1211. if (!skb_can_shift(skb))
  1212. goto fallback;
  1213. /* This frame is about to be dropped (was ACKed). */
  1214. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1215. goto fallback;
  1216. /* Can only happen with delayed DSACK + discard craziness */
  1217. if (unlikely(skb == tcp_write_queue_head(sk)))
  1218. goto fallback;
  1219. prev = tcp_write_queue_prev(sk, skb);
  1220. if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
  1221. goto fallback;
  1222. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  1223. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  1224. if (in_sack) {
  1225. len = skb->len;
  1226. pcount = tcp_skb_pcount(skb);
  1227. mss = tcp_skb_seglen(skb);
  1228. /* TODO: Fix DSACKs to not fragment already SACKed and we can
  1229. * drop this restriction as unnecessary
  1230. */
  1231. if (mss != tcp_skb_seglen(prev))
  1232. goto fallback;
  1233. } else {
  1234. if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
  1235. goto noop;
  1236. /* CHECKME: This is non-MSS split case only?, this will
  1237. * cause skipped skbs due to advancing loop btw, original
  1238. * has that feature too
  1239. */
  1240. if (tcp_skb_pcount(skb) <= 1)
  1241. goto noop;
  1242. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
  1243. if (!in_sack) {
  1244. /* TODO: head merge to next could be attempted here
  1245. * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
  1246. * though it might not be worth of the additional hassle
  1247. *
  1248. * ...we can probably just fallback to what was done
  1249. * previously. We could try merging non-SACKed ones
  1250. * as well but it probably isn't going to buy off
  1251. * because later SACKs might again split them, and
  1252. * it would make skb timestamp tracking considerably
  1253. * harder problem.
  1254. */
  1255. goto fallback;
  1256. }
  1257. len = end_seq - TCP_SKB_CB(skb)->seq;
  1258. BUG_ON(len < 0);
  1259. BUG_ON(len > skb->len);
  1260. /* MSS boundaries should be honoured or else pcount will
  1261. * severely break even though it makes things bit trickier.
  1262. * Optimize common case to avoid most of the divides
  1263. */
  1264. mss = tcp_skb_mss(skb);
  1265. /* TODO: Fix DSACKs to not fragment already SACKed and we can
  1266. * drop this restriction as unnecessary
  1267. */
  1268. if (mss != tcp_skb_seglen(prev))
  1269. goto fallback;
  1270. if (len == mss) {
  1271. pcount = 1;
  1272. } else if (len < mss) {
  1273. goto noop;
  1274. } else {
  1275. pcount = len / mss;
  1276. len = pcount * mss;
  1277. }
  1278. }
  1279. /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
  1280. if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
  1281. goto fallback;
  1282. if (!skb_shift(prev, skb, len))
  1283. goto fallback;
  1284. if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
  1285. goto out;
  1286. /* Hole filled allows collapsing with the next as well, this is very
  1287. * useful when hole on every nth skb pattern happens
  1288. */
  1289. if (prev == tcp_write_queue_tail(sk))
  1290. goto out;
  1291. skb = tcp_write_queue_next(sk, prev);
  1292. if (!skb_can_shift(skb) ||
  1293. (skb == tcp_send_head(sk)) ||
  1294. ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
  1295. (mss != tcp_skb_seglen(skb)))
  1296. goto out;
  1297. len = skb->len;
  1298. if (skb_shift(prev, skb, len)) {
  1299. pcount += tcp_skb_pcount(skb);
  1300. tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
  1301. }
  1302. out:
  1303. state->fack_count += pcount;
  1304. return prev;
  1305. noop:
  1306. return skb;
  1307. fallback:
  1308. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
  1309. return NULL;
  1310. }
  1311. static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
  1312. struct tcp_sack_block *next_dup,
  1313. struct tcp_sacktag_state *state,
  1314. u32 start_seq, u32 end_seq,
  1315. bool dup_sack_in)
  1316. {
  1317. struct tcp_sock *tp = tcp_sk(sk);
  1318. struct sk_buff *tmp;
  1319. tcp_for_write_queue_from(skb, sk) {
  1320. int in_sack = 0;
  1321. bool dup_sack = dup_sack_in;
  1322. if (skb == tcp_send_head(sk))
  1323. break;
  1324. /* queue is in-order => we can short-circuit the walk early */
  1325. if (!before(TCP_SKB_CB(skb)->seq, end_seq))
  1326. break;
  1327. if ((next_dup != NULL) &&
  1328. before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
  1329. in_sack = tcp_match_skb_to_sack(sk, skb,
  1330. next_dup->start_seq,
  1331. next_dup->end_seq);
  1332. if (in_sack > 0)
  1333. dup_sack = true;
  1334. }
  1335. /* skb reference here is a bit tricky to get right, since
  1336. * shifting can eat and free both this skb and the next,
  1337. * so not even _safe variant of the loop is enough.
  1338. */
  1339. if (in_sack <= 0) {
  1340. tmp = tcp_shift_skb_data(sk, skb, state,
  1341. start_seq, end_seq, dup_sack);
  1342. if (tmp != NULL) {
  1343. if (tmp != skb) {
  1344. skb = tmp;
  1345. continue;
  1346. }
  1347. in_sack = 0;
  1348. } else {
  1349. in_sack = tcp_match_skb_to_sack(sk, skb,
  1350. start_seq,
  1351. end_seq);
  1352. }
  1353. }
  1354. if (unlikely(in_sack < 0))
  1355. break;
  1356. if (in_sack) {
  1357. TCP_SKB_CB(skb)->sacked =
  1358. tcp_sacktag_one(sk,
  1359. state,
  1360. TCP_SKB_CB(skb)->sacked,
  1361. TCP_SKB_CB(skb)->seq,
  1362. TCP_SKB_CB(skb)->end_seq,
  1363. dup_sack,
  1364. tcp_skb_pcount(skb),
  1365. TCP_SKB_CB(skb)->when);
  1366. if (!before(TCP_SKB_CB(skb)->seq,
  1367. tcp_highest_sack_seq(tp)))
  1368. tcp_advance_highest_sack(sk, skb);
  1369. }
  1370. state->fack_count += tcp_skb_pcount(skb);
  1371. }
  1372. return skb;
  1373. }
  1374. /* Avoid all extra work that is being done by sacktag while walking in
  1375. * a normal way
  1376. */
  1377. static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
  1378. struct tcp_sacktag_state *state,
  1379. u32 skip_to_seq)
  1380. {
  1381. tcp_for_write_queue_from(skb, sk) {
  1382. if (skb == tcp_send_head(sk))
  1383. break;
  1384. if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
  1385. break;
  1386. state->fack_count += tcp_skb_pcount(skb);
  1387. }
  1388. return skb;
  1389. }
  1390. static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
  1391. struct sock *sk,
  1392. struct tcp_sack_block *next_dup,
  1393. struct tcp_sacktag_state *state,
  1394. u32 skip_to_seq)
  1395. {
  1396. if (next_dup == NULL)
  1397. return skb;
  1398. if (before(next_dup->start_seq, skip_to_seq)) {
  1399. skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
  1400. skb = tcp_sacktag_walk(skb, sk, NULL, state,
  1401. next_dup->start_seq, next_dup->end_seq,
  1402. 1);
  1403. }
  1404. return skb;
  1405. }
  1406. static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
  1407. {
  1408. return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
  1409. }
  1410. static int
  1411. tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
  1412. u32 prior_snd_una, s32 *sack_rtt)
  1413. {
  1414. struct tcp_sock *tp = tcp_sk(sk);
  1415. const unsigned char *ptr = (skb_transport_header(ack_skb) +
  1416. TCP_SKB_CB(ack_skb)->sacked);
  1417. struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
  1418. struct tcp_sack_block sp[TCP_NUM_SACKS];
  1419. struct tcp_sack_block *cache;
  1420. struct tcp_sacktag_state state;
  1421. struct sk_buff *skb;
  1422. int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
  1423. int used_sacks;
  1424. bool found_dup_sack = false;
  1425. int i, j;
  1426. int first_sack_index;
  1427. state.flag = 0;
  1428. state.reord = tp->packets_out;
  1429. state.rtt = -1;
  1430. if (!tp->sacked_out) {
  1431. if (WARN_ON(tp->fackets_out))
  1432. tp->fackets_out = 0;
  1433. tcp_highest_sack_reset(sk);
  1434. }
  1435. found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
  1436. num_sacks, prior_snd_una);
  1437. if (found_dup_sack)
  1438. state.flag |= FLAG_DSACKING_ACK;
  1439. /* Eliminate too old ACKs, but take into
  1440. * account more or less fresh ones, they can
  1441. * contain valid SACK info.
  1442. */
  1443. if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
  1444. return 0;
  1445. if (!tp->packets_out)
  1446. goto out;
  1447. used_sacks = 0;
  1448. first_sack_index = 0;
  1449. for (i = 0; i < num_sacks; i++) {
  1450. bool dup_sack = !i && found_dup_sack;
  1451. sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
  1452. sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
  1453. if (!tcp_is_sackblock_valid(tp, dup_sack,
  1454. sp[used_sacks].start_seq,
  1455. sp[used_sacks].end_seq)) {
  1456. int mib_idx;
  1457. if (dup_sack) {
  1458. if (!tp->undo_marker)
  1459. mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
  1460. else
  1461. mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
  1462. } else {
  1463. /* Don't count olds caused by ACK reordering */
  1464. if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
  1465. !after(sp[used_sacks].end_seq, tp->snd_una))
  1466. continue;
  1467. mib_idx = LINUX_MIB_TCPSACKDISCARD;
  1468. }
  1469. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  1470. if (i == 0)
  1471. first_sack_index = -1;
  1472. continue;
  1473. }
  1474. /* Ignore very old stuff early */
  1475. if (!after(sp[used_sacks].end_seq, prior_snd_una))
  1476. continue;
  1477. used_sacks++;
  1478. }
  1479. /* order SACK blocks to allow in order walk of the retrans queue */
  1480. for (i = used_sacks - 1; i > 0; i--) {
  1481. for (j = 0; j < i; j++) {
  1482. if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
  1483. swap(sp[j], sp[j + 1]);
  1484. /* Track where the first SACK block goes to */
  1485. if (j == first_sack_index)
  1486. first_sack_index = j + 1;
  1487. }
  1488. }
  1489. }
  1490. skb = tcp_write_queue_head(sk);
  1491. state.fack_count = 0;
  1492. i = 0;
  1493. if (!tp->sacked_out) {
  1494. /* It's already past, so skip checking against it */
  1495. cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
  1496. } else {
  1497. cache = tp->recv_sack_cache;
  1498. /* Skip empty blocks in at head of the cache */
  1499. while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
  1500. !cache->end_seq)
  1501. cache++;
  1502. }
  1503. while (i < used_sacks) {
  1504. u32 start_seq = sp[i].start_seq;
  1505. u32 end_seq = sp[i].end_seq;
  1506. bool dup_sack = (found_dup_sack && (i == first_sack_index));
  1507. struct tcp_sack_block *next_dup = NULL;
  1508. if (found_dup_sack && ((i + 1) == first_sack_index))
  1509. next_dup = &sp[i + 1];
  1510. /* Skip too early cached blocks */
  1511. while (tcp_sack_cache_ok(tp, cache) &&
  1512. !before(start_seq, cache->end_seq))
  1513. cache++;
  1514. /* Can skip some work by looking recv_sack_cache? */
  1515. if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
  1516. after(end_seq, cache->start_seq)) {
  1517. /* Head todo? */
  1518. if (before(start_seq, cache->start_seq)) {
  1519. skb = tcp_sacktag_skip(skb, sk, &state,
  1520. start_seq);
  1521. skb = tcp_sacktag_walk(skb, sk, next_dup,
  1522. &state,
  1523. start_seq,
  1524. cache->start_seq,
  1525. dup_sack);
  1526. }
  1527. /* Rest of the block already fully processed? */
  1528. if (!after(end_seq, cache->end_seq))
  1529. goto advance_sp;
  1530. skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
  1531. &state,
  1532. cache->end_seq);
  1533. /* ...tail remains todo... */
  1534. if (tcp_highest_sack_seq(tp) == cache->end_seq) {
  1535. /* ...but better entrypoint exists! */
  1536. skb = tcp_highest_sack(sk);
  1537. if (skb == NULL)
  1538. break;
  1539. state.fack_count = tp->fackets_out;
  1540. cache++;
  1541. goto walk;
  1542. }
  1543. skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
  1544. /* Check overlap against next cached too (past this one already) */
  1545. cache++;
  1546. continue;
  1547. }
  1548. if (!before(start_seq, tcp_highest_sack_seq(tp))) {
  1549. skb = tcp_highest_sack(sk);
  1550. if (skb == NULL)
  1551. break;
  1552. state.fack_count = tp->fackets_out;
  1553. }
  1554. skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
  1555. walk:
  1556. skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
  1557. start_seq, end_seq, dup_sack);
  1558. advance_sp:
  1559. i++;
  1560. }
  1561. /* Clear the head of the cache sack blocks so we can skip it next time */
  1562. for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
  1563. tp->recv_sack_cache[i].start_seq = 0;
  1564. tp->recv_sack_cache[i].end_seq = 0;
  1565. }
  1566. for (j = 0; j < used_sacks; j++)
  1567. tp->recv_sack_cache[i++] = sp[j];
  1568. tcp_mark_lost_retrans(sk);
  1569. tcp_verify_left_out(tp);
  1570. if ((state.reord < tp->fackets_out) &&
  1571. ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
  1572. tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
  1573. out:
  1574. #if FASTRETRANS_DEBUG > 0
  1575. WARN_ON((int)tp->sacked_out < 0);
  1576. WARN_ON((int)tp->lost_out < 0);
  1577. WARN_ON((int)tp->retrans_out < 0);
  1578. WARN_ON((int)tcp_packets_in_flight(tp) < 0);
  1579. #endif
  1580. *sack_rtt = state.rtt;
  1581. return state.flag;
  1582. }
  1583. /* Limits sacked_out so that sum with lost_out isn't ever larger than
  1584. * packets_out. Returns false if sacked_out adjustement wasn't necessary.
  1585. */
  1586. static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
  1587. {
  1588. u32 holes;
  1589. holes = max(tp->lost_out, 1U);
  1590. holes = min(holes, tp->packets_out);
  1591. if ((tp->sacked_out + holes) > tp->packets_out) {
  1592. tp->sacked_out = tp->packets_out - holes;
  1593. return true;
  1594. }
  1595. return false;
  1596. }
  1597. /* If we receive more dupacks than we expected counting segments
  1598. * in assumption of absent reordering, interpret this as reordering.
  1599. * The only another reason could be bug in receiver TCP.
  1600. */
  1601. static void tcp_check_reno_reordering(struct sock *sk, const int addend)
  1602. {
  1603. struct tcp_sock *tp = tcp_sk(sk);
  1604. if (tcp_limit_reno_sacked(tp))
  1605. tcp_update_reordering(sk, tp->packets_out + addend, 0);
  1606. }
  1607. /* Emulate SACKs for SACKless connection: account for a new dupack. */
  1608. static void tcp_add_reno_sack(struct sock *sk)
  1609. {
  1610. struct tcp_sock *tp = tcp_sk(sk);
  1611. tp->sacked_out++;
  1612. tcp_check_reno_reordering(sk, 0);
  1613. tcp_verify_left_out(tp);
  1614. }
  1615. /* Account for ACK, ACKing some data in Reno Recovery phase. */
  1616. static void tcp_remove_reno_sacks(struct sock *sk, int acked)
  1617. {
  1618. struct tcp_sock *tp = tcp_sk(sk);
  1619. if (acked > 0) {
  1620. /* One ACK acked hole. The rest eat duplicate ACKs. */
  1621. if (acked - 1 >= tp->sacked_out)
  1622. tp->sacked_out = 0;
  1623. else
  1624. tp->sacked_out -= acked - 1;
  1625. }
  1626. tcp_check_reno_reordering(sk, acked);
  1627. tcp_verify_left_out(tp);
  1628. }
  1629. static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
  1630. {
  1631. tp->sacked_out = 0;
  1632. }
  1633. static void tcp_clear_retrans_partial(struct tcp_sock *tp)
  1634. {
  1635. tp->retrans_out = 0;
  1636. tp->lost_out = 0;
  1637. tp->undo_marker = 0;
  1638. tp->undo_retrans = 0;
  1639. }
  1640. void tcp_clear_retrans(struct tcp_sock *tp)
  1641. {
  1642. tcp_clear_retrans_partial(tp);
  1643. tp->fackets_out = 0;
  1644. tp->sacked_out = 0;
  1645. }
  1646. /* Enter Loss state. If "how" is not zero, forget all SACK information
  1647. * and reset tags completely, otherwise preserve SACKs. If receiver
  1648. * dropped its ofo queue, we will know this due to reneging detection.
  1649. */
  1650. void tcp_enter_loss(struct sock *sk, int how)
  1651. {
  1652. const struct inet_connection_sock *icsk = inet_csk(sk);
  1653. struct tcp_sock *tp = tcp_sk(sk);
  1654. struct sk_buff *skb;
  1655. bool new_recovery = false;
  1656. /* Reduce ssthresh if it has not yet been made inside this window. */
  1657. if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
  1658. !after(tp->high_seq, tp->snd_una) ||
  1659. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1660. new_recovery = true;
  1661. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1662. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1663. tcp_ca_event(sk, CA_EVENT_LOSS);
  1664. }
  1665. tp->snd_cwnd = 1;
  1666. tp->snd_cwnd_cnt = 0;
  1667. tp->snd_cwnd_stamp = tcp_time_stamp;
  1668. tcp_clear_retrans_partial(tp);
  1669. if (tcp_is_reno(tp))
  1670. tcp_reset_reno_sack(tp);
  1671. tp->undo_marker = tp->snd_una;
  1672. if (how) {
  1673. tp->sacked_out = 0;
  1674. tp->fackets_out = 0;
  1675. }
  1676. tcp_clear_all_retrans_hints(tp);
  1677. tcp_for_write_queue(skb, sk) {
  1678. if (skb == tcp_send_head(sk))
  1679. break;
  1680. if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
  1681. tp->undo_marker = 0;
  1682. TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
  1683. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
  1684. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
  1685. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1686. tp->lost_out += tcp_skb_pcount(skb);
  1687. tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
  1688. }
  1689. }
  1690. tcp_verify_left_out(tp);
  1691. /* Timeout in disordered state after receiving substantial DUPACKs
  1692. * suggests that the degree of reordering is over-estimated.
  1693. */
  1694. if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
  1695. tp->sacked_out >= sysctl_tcp_reordering)
  1696. tp->reordering = min_t(unsigned int, tp->reordering,
  1697. sysctl_tcp_reordering);
  1698. tcp_set_ca_state(sk, TCP_CA_Loss);
  1699. tp->high_seq = tp->snd_nxt;
  1700. TCP_ECN_queue_cwr(tp);
  1701. /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
  1702. * loss recovery is underway except recurring timeout(s) on
  1703. * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
  1704. */
  1705. tp->frto = sysctl_tcp_frto &&
  1706. (new_recovery || icsk->icsk_retransmits) &&
  1707. !inet_csk(sk)->icsk_mtup.probe_size;
  1708. }
  1709. /* If ACK arrived pointing to a remembered SACK, it means that our
  1710. * remembered SACKs do not reflect real state of receiver i.e.
  1711. * receiver _host_ is heavily congested (or buggy).
  1712. *
  1713. * Do processing similar to RTO timeout.
  1714. */
  1715. static bool tcp_check_sack_reneging(struct sock *sk, int flag)
  1716. {
  1717. if (flag & FLAG_SACK_RENEGING) {
  1718. struct inet_connection_sock *icsk = inet_csk(sk);
  1719. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
  1720. tcp_enter_loss(sk, 1);
  1721. icsk->icsk_retransmits++;
  1722. tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
  1723. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1724. icsk->icsk_rto, TCP_RTO_MAX);
  1725. return true;
  1726. }
  1727. return false;
  1728. }
  1729. static inline int tcp_fackets_out(const struct tcp_sock *tp)
  1730. {
  1731. return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
  1732. }
  1733. /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
  1734. * counter when SACK is enabled (without SACK, sacked_out is used for
  1735. * that purpose).
  1736. *
  1737. * Instead, with FACK TCP uses fackets_out that includes both SACKed
  1738. * segments up to the highest received SACK block so far and holes in
  1739. * between them.
  1740. *
  1741. * With reordering, holes may still be in flight, so RFC3517 recovery
  1742. * uses pure sacked_out (total number of SACKed segments) even though
  1743. * it violates the RFC that uses duplicate ACKs, often these are equal
  1744. * but when e.g. out-of-window ACKs or packet duplication occurs,
  1745. * they differ. Since neither occurs due to loss, TCP should really
  1746. * ignore them.
  1747. */
  1748. static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
  1749. {
  1750. return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
  1751. }
  1752. static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
  1753. {
  1754. struct tcp_sock *tp = tcp_sk(sk);
  1755. unsigned long delay;
  1756. /* Delay early retransmit and entering fast recovery for
  1757. * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
  1758. * available, or RTO is scheduled to fire first.
  1759. */
  1760. if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
  1761. (flag & FLAG_ECE) || !tp->srtt)
  1762. return false;
  1763. delay = max_t(unsigned long, (tp->srtt >> 5), msecs_to_jiffies(2));
  1764. if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
  1765. return false;
  1766. inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
  1767. TCP_RTO_MAX);
  1768. return true;
  1769. }
  1770. /* Linux NewReno/SACK/FACK/ECN state machine.
  1771. * --------------------------------------
  1772. *
  1773. * "Open" Normal state, no dubious events, fast path.
  1774. * "Disorder" In all the respects it is "Open",
  1775. * but requires a bit more attention. It is entered when
  1776. * we see some SACKs or dupacks. It is split of "Open"
  1777. * mainly to move some processing from fast path to slow one.
  1778. * "CWR" CWND was reduced due to some Congestion Notification event.
  1779. * It can be ECN, ICMP source quench, local device congestion.
  1780. * "Recovery" CWND was reduced, we are fast-retransmitting.
  1781. * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
  1782. *
  1783. * tcp_fastretrans_alert() is entered:
  1784. * - each incoming ACK, if state is not "Open"
  1785. * - when arrived ACK is unusual, namely:
  1786. * * SACK
  1787. * * Duplicate ACK.
  1788. * * ECN ECE.
  1789. *
  1790. * Counting packets in flight is pretty simple.
  1791. *
  1792. * in_flight = packets_out - left_out + retrans_out
  1793. *
  1794. * packets_out is SND.NXT-SND.UNA counted in packets.
  1795. *
  1796. * retrans_out is number of retransmitted segments.
  1797. *
  1798. * left_out is number of segments left network, but not ACKed yet.
  1799. *
  1800. * left_out = sacked_out + lost_out
  1801. *
  1802. * sacked_out: Packets, which arrived to receiver out of order
  1803. * and hence not ACKed. With SACKs this number is simply
  1804. * amount of SACKed data. Even without SACKs
  1805. * it is easy to give pretty reliable estimate of this number,
  1806. * counting duplicate ACKs.
  1807. *
  1808. * lost_out: Packets lost by network. TCP has no explicit
  1809. * "loss notification" feedback from network (for now).
  1810. * It means that this number can be only _guessed_.
  1811. * Actually, it is the heuristics to predict lossage that
  1812. * distinguishes different algorithms.
  1813. *
  1814. * F.e. after RTO, when all the queue is considered as lost,
  1815. * lost_out = packets_out and in_flight = retrans_out.
  1816. *
  1817. * Essentially, we have now two algorithms counting
  1818. * lost packets.
  1819. *
  1820. * FACK: It is the simplest heuristics. As soon as we decided
  1821. * that something is lost, we decide that _all_ not SACKed
  1822. * packets until the most forward SACK are lost. I.e.
  1823. * lost_out = fackets_out - sacked_out and left_out = fackets_out.
  1824. * It is absolutely correct estimate, if network does not reorder
  1825. * packets. And it loses any connection to reality when reordering
  1826. * takes place. We use FACK by default until reordering
  1827. * is suspected on the path to this destination.
  1828. *
  1829. * NewReno: when Recovery is entered, we assume that one segment
  1830. * is lost (classic Reno). While we are in Recovery and
  1831. * a partial ACK arrives, we assume that one more packet
  1832. * is lost (NewReno). This heuristics are the same in NewReno
  1833. * and SACK.
  1834. *
  1835. * Imagine, that's all! Forget about all this shamanism about CWND inflation
  1836. * deflation etc. CWND is real congestion window, never inflated, changes
  1837. * only according to classic VJ rules.
  1838. *
  1839. * Really tricky (and requiring careful tuning) part of algorithm
  1840. * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
  1841. * The first determines the moment _when_ we should reduce CWND and,
  1842. * hence, slow down forward transmission. In fact, it determines the moment
  1843. * when we decide that hole is caused by loss, rather than by a reorder.
  1844. *
  1845. * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
  1846. * holes, caused by lost packets.
  1847. *
  1848. * And the most logically complicated part of algorithm is undo
  1849. * heuristics. We detect false retransmits due to both too early
  1850. * fast retransmit (reordering) and underestimated RTO, analyzing
  1851. * timestamps and D-SACKs. When we detect that some segments were
  1852. * retransmitted by mistake and CWND reduction was wrong, we undo
  1853. * window reduction and abort recovery phase. This logic is hidden
  1854. * inside several functions named tcp_try_undo_<something>.
  1855. */
  1856. /* This function decides, when we should leave Disordered state
  1857. * and enter Recovery phase, reducing congestion window.
  1858. *
  1859. * Main question: may we further continue forward transmission
  1860. * with the same cwnd?
  1861. */
  1862. static bool tcp_time_to_recover(struct sock *sk, int flag)
  1863. {
  1864. struct tcp_sock *tp = tcp_sk(sk);
  1865. __u32 packets_out;
  1866. /* Trick#1: The loss is proven. */
  1867. if (tp->lost_out)
  1868. return true;
  1869. /* Not-A-Trick#2 : Classic rule... */
  1870. if (tcp_dupack_heuristics(tp) > tp->reordering)
  1871. return true;
  1872. /* Trick#4: It is still not OK... But will it be useful to delay
  1873. * recovery more?
  1874. */
  1875. packets_out = tp->packets_out;
  1876. if (packets_out <= tp->reordering &&
  1877. tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
  1878. !tcp_may_send_now(sk)) {
  1879. /* We have nothing to send. This connection is limited
  1880. * either by receiver window or by application.
  1881. */
  1882. return true;
  1883. }
  1884. /* If a thin stream is detected, retransmit after first
  1885. * received dupack. Employ only if SACK is supported in order
  1886. * to avoid possible corner-case series of spurious retransmissions
  1887. * Use only if there are no unsent data.
  1888. */
  1889. if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
  1890. tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
  1891. tcp_is_sack(tp) && !tcp_send_head(sk))
  1892. return true;
  1893. /* Trick#6: TCP early retransmit, per RFC5827. To avoid spurious
  1894. * retransmissions due to small network reorderings, we implement
  1895. * Mitigation A.3 in the RFC and delay the retransmission for a short
  1896. * interval if appropriate.
  1897. */
  1898. if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
  1899. (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
  1900. !tcp_may_send_now(sk))
  1901. return !tcp_pause_early_retransmit(sk, flag);
  1902. return false;
  1903. }
  1904. /* Detect loss in event "A" above by marking head of queue up as lost.
  1905. * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
  1906. * are considered lost. For RFC3517 SACK, a segment is considered lost if it
  1907. * has at least tp->reordering SACKed seqments above it; "packets" refers to
  1908. * the maximum SACKed segments to pass before reaching this limit.
  1909. */
  1910. static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
  1911. {
  1912. struct tcp_sock *tp = tcp_sk(sk);
  1913. struct sk_buff *skb;
  1914. int cnt, oldcnt;
  1915. int err;
  1916. unsigned int mss;
  1917. /* Use SACK to deduce losses of new sequences sent during recovery */
  1918. const u32 loss_high = tcp_is_sack(tp) ? tp->snd_nxt : tp->high_seq;
  1919. WARN_ON(packets > tp->packets_out);
  1920. if (tp->lost_skb_hint) {
  1921. skb = tp->lost_skb_hint;
  1922. cnt = tp->lost_cnt_hint;
  1923. /* Head already handled? */
  1924. if (mark_head && skb != tcp_write_queue_head(sk))
  1925. return;
  1926. } else {
  1927. skb = tcp_write_queue_head(sk);
  1928. cnt = 0;
  1929. }
  1930. tcp_for_write_queue_from(skb, sk) {
  1931. if (skb == tcp_send_head(sk))
  1932. break;
  1933. /* TODO: do this better */
  1934. /* this is not the most efficient way to do this... */
  1935. tp->lost_skb_hint = skb;
  1936. tp->lost_cnt_hint = cnt;
  1937. if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
  1938. break;
  1939. oldcnt = cnt;
  1940. if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
  1941. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
  1942. cnt += tcp_skb_pcount(skb);
  1943. if (cnt > packets) {
  1944. if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
  1945. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
  1946. (oldcnt >= packets))
  1947. break;
  1948. mss = skb_shinfo(skb)->gso_size;
  1949. err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
  1950. if (err < 0)
  1951. break;
  1952. cnt = packets;
  1953. }
  1954. tcp_skb_mark_lost(tp, skb);
  1955. if (mark_head)
  1956. break;
  1957. }
  1958. tcp_verify_left_out(tp);
  1959. }
  1960. /* Account newly detected lost packet(s) */
  1961. static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
  1962. {
  1963. struct tcp_sock *tp = tcp_sk(sk);
  1964. if (tcp_is_reno(tp)) {
  1965. tcp_mark_head_lost(sk, 1, 1);
  1966. } else if (tcp_is_fack(tp)) {
  1967. int lost = tp->fackets_out - tp->reordering;
  1968. if (lost <= 0)
  1969. lost = 1;
  1970. tcp_mark_head_lost(sk, lost, 0);
  1971. } else {
  1972. int sacked_upto = tp->sacked_out - tp->reordering;
  1973. if (sacked_upto >= 0)
  1974. tcp_mark_head_lost(sk, sacked_upto, 0);
  1975. else if (fast_rexmit)
  1976. tcp_mark_head_lost(sk, 1, 1);
  1977. }
  1978. }
  1979. /* CWND moderation, preventing bursts due to too big ACKs
  1980. * in dubious situations.
  1981. */
  1982. static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
  1983. {
  1984. tp->snd_cwnd = min(tp->snd_cwnd,
  1985. tcp_packets_in_flight(tp) + tcp_max_burst(tp));
  1986. tp->snd_cwnd_stamp = tcp_time_stamp;
  1987. }
  1988. /* Nothing was retransmitted or returned timestamp is less
  1989. * than timestamp of the first retransmission.
  1990. */
  1991. static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
  1992. {
  1993. return !tp->retrans_stamp ||
  1994. (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  1995. before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
  1996. }
  1997. /* Undo procedures. */
  1998. #if FASTRETRANS_DEBUG > 1
  1999. static void DBGUNDO(struct sock *sk, const char *msg)
  2000. {
  2001. struct tcp_sock *tp = tcp_sk(sk);
  2002. struct inet_sock *inet = inet_sk(sk);
  2003. if (sk->sk_family == AF_INET) {
  2004. pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
  2005. msg,
  2006. &inet->inet_daddr, ntohs(inet->inet_dport),
  2007. tp->snd_cwnd, tcp_left_out(tp),
  2008. tp->snd_ssthresh, tp->prior_ssthresh,
  2009. tp->packets_out);
  2010. }
  2011. #if IS_ENABLED(CONFIG_IPV6)
  2012. else if (sk->sk_family == AF_INET6) {
  2013. struct ipv6_pinfo *np = inet6_sk(sk);
  2014. pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
  2015. msg,
  2016. &np->daddr, ntohs(inet->inet_dport),
  2017. tp->snd_cwnd, tcp_left_out(tp),
  2018. tp->snd_ssthresh, tp->prior_ssthresh,
  2019. tp->packets_out);
  2020. }
  2021. #endif
  2022. }
  2023. #else
  2024. #define DBGUNDO(x...) do { } while (0)
  2025. #endif
  2026. static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
  2027. {
  2028. struct tcp_sock *tp = tcp_sk(sk);
  2029. if (unmark_loss) {
  2030. struct sk_buff *skb;
  2031. tcp_for_write_queue(skb, sk) {
  2032. if (skb == tcp_send_head(sk))
  2033. break;
  2034. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  2035. }
  2036. tp->lost_out = 0;
  2037. tcp_clear_all_retrans_hints(tp);
  2038. }
  2039. if (tp->prior_ssthresh) {
  2040. const struct inet_connection_sock *icsk = inet_csk(sk);
  2041. if (icsk->icsk_ca_ops->undo_cwnd)
  2042. tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
  2043. else
  2044. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
  2045. if (tp->prior_ssthresh > tp->snd_ssthresh) {
  2046. tp->snd_ssthresh = tp->prior_ssthresh;
  2047. TCP_ECN_withdraw_cwr(tp);
  2048. }
  2049. } else {
  2050. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
  2051. }
  2052. tp->snd_cwnd_stamp = tcp_time_stamp;
  2053. tp->undo_marker = 0;
  2054. }
  2055. static inline bool tcp_may_undo(const struct tcp_sock *tp)
  2056. {
  2057. return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
  2058. }
  2059. /* People celebrate: "We love our President!" */
  2060. static bool tcp_try_undo_recovery(struct sock *sk)
  2061. {
  2062. struct tcp_sock *tp = tcp_sk(sk);
  2063. if (tcp_may_undo(tp)) {
  2064. int mib_idx;
  2065. /* Happy end! We did not retransmit anything
  2066. * or our original transmission succeeded.
  2067. */
  2068. DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
  2069. tcp_undo_cwnd_reduction(sk, false);
  2070. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
  2071. mib_idx = LINUX_MIB_TCPLOSSUNDO;
  2072. else
  2073. mib_idx = LINUX_MIB_TCPFULLUNDO;
  2074. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  2075. }
  2076. if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
  2077. /* Hold old state until something *above* high_seq
  2078. * is ACKed. For Reno it is MUST to prevent false
  2079. * fast retransmits (RFC2582). SACK TCP is safe. */
  2080. tcp_moderate_cwnd(tp);
  2081. return true;
  2082. }
  2083. tcp_set_ca_state(sk, TCP_CA_Open);
  2084. return false;
  2085. }
  2086. /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
  2087. static bool tcp_try_undo_dsack(struct sock *sk)
  2088. {
  2089. struct tcp_sock *tp = tcp_sk(sk);
  2090. if (tp->undo_marker && !tp->undo_retrans) {
  2091. DBGUNDO(sk, "D-SACK");
  2092. tcp_undo_cwnd_reduction(sk, false);
  2093. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
  2094. return true;
  2095. }
  2096. return false;
  2097. }
  2098. /* We can clear retrans_stamp when there are no retransmissions in the
  2099. * window. It would seem that it is trivially available for us in
  2100. * tp->retrans_out, however, that kind of assumptions doesn't consider
  2101. * what will happen if errors occur when sending retransmission for the
  2102. * second time. ...It could the that such segment has only
  2103. * TCPCB_EVER_RETRANS set at the present time. It seems that checking
  2104. * the head skb is enough except for some reneging corner cases that
  2105. * are not worth the effort.
  2106. *
  2107. * Main reason for all this complexity is the fact that connection dying
  2108. * time now depends on the validity of the retrans_stamp, in particular,
  2109. * that successive retransmissions of a segment must not advance
  2110. * retrans_stamp under any conditions.
  2111. */
  2112. static bool tcp_any_retrans_done(const struct sock *sk)
  2113. {
  2114. const struct tcp_sock *tp = tcp_sk(sk);
  2115. struct sk_buff *skb;
  2116. if (tp->retrans_out)
  2117. return true;
  2118. skb = tcp_write_queue_head(sk);
  2119. if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
  2120. return true;
  2121. return false;
  2122. }
  2123. /* Undo during loss recovery after partial ACK or using F-RTO. */
  2124. static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
  2125. {
  2126. struct tcp_sock *tp = tcp_sk(sk);
  2127. if (frto_undo || tcp_may_undo(tp)) {
  2128. tcp_undo_cwnd_reduction(sk, true);
  2129. DBGUNDO(sk, "partial loss");
  2130. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
  2131. if (frto_undo)
  2132. NET_INC_STATS_BH(sock_net(sk),
  2133. LINUX_MIB_TCPSPURIOUSRTOS);
  2134. inet_csk(sk)->icsk_retransmits = 0;
  2135. if (frto_undo || tcp_is_sack(tp))
  2136. tcp_set_ca_state(sk, TCP_CA_Open);
  2137. return true;
  2138. }
  2139. return false;
  2140. }
  2141. /* The cwnd reduction in CWR and Recovery use the PRR algorithm
  2142. * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
  2143. * It computes the number of packets to send (sndcnt) based on packets newly
  2144. * delivered:
  2145. * 1) If the packets in flight is larger than ssthresh, PRR spreads the
  2146. * cwnd reductions across a full RTT.
  2147. * 2) If packets in flight is lower than ssthresh (such as due to excess
  2148. * losses and/or application stalls), do not perform any further cwnd
  2149. * reductions, but instead slow start up to ssthresh.
  2150. */
  2151. static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
  2152. {
  2153. struct tcp_sock *tp = tcp_sk(sk);
  2154. tp->high_seq = tp->snd_nxt;
  2155. tp->tlp_high_seq = 0;
  2156. tp->snd_cwnd_cnt = 0;
  2157. tp->prior_cwnd = tp->snd_cwnd;
  2158. tp->prr_delivered = 0;
  2159. tp->prr_out = 0;
  2160. if (set_ssthresh)
  2161. tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
  2162. TCP_ECN_queue_cwr(tp);
  2163. }
  2164. static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
  2165. int fast_rexmit)
  2166. {
  2167. struct tcp_sock *tp = tcp_sk(sk);
  2168. int sndcnt = 0;
  2169. int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
  2170. int newly_acked_sacked = prior_unsacked -
  2171. (tp->packets_out - tp->sacked_out);
  2172. tp->prr_delivered += newly_acked_sacked;
  2173. if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
  2174. u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
  2175. tp->prior_cwnd - 1;
  2176. sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
  2177. } else {
  2178. sndcnt = min_t(int, delta,
  2179. max_t(int, tp->prr_delivered - tp->prr_out,
  2180. newly_acked_sacked) + 1);
  2181. }
  2182. sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
  2183. tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
  2184. }
  2185. static inline void tcp_end_cwnd_reduction(struct sock *sk)
  2186. {
  2187. struct tcp_sock *tp = tcp_sk(sk);
  2188. /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
  2189. if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
  2190. (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
  2191. tp->snd_cwnd = tp->snd_ssthresh;
  2192. tp->snd_cwnd_stamp = tcp_time_stamp;
  2193. }
  2194. tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
  2195. }
  2196. /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
  2197. void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
  2198. {
  2199. struct tcp_sock *tp = tcp_sk(sk);
  2200. tp->prior_ssthresh = 0;
  2201. if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
  2202. tp->undo_marker = 0;
  2203. tcp_init_cwnd_reduction(sk, set_ssthresh);
  2204. tcp_set_ca_state(sk, TCP_CA_CWR);
  2205. }
  2206. }
  2207. static void tcp_try_keep_open(struct sock *sk)
  2208. {
  2209. struct tcp_sock *tp = tcp_sk(sk);
  2210. int state = TCP_CA_Open;
  2211. if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
  2212. state = TCP_CA_Disorder;
  2213. if (inet_csk(sk)->icsk_ca_state != state) {
  2214. tcp_set_ca_state(sk, state);
  2215. tp->high_seq = tp->snd_nxt;
  2216. }
  2217. }
  2218. static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
  2219. {
  2220. struct tcp_sock *tp = tcp_sk(sk);
  2221. tcp_verify_left_out(tp);
  2222. if (!tcp_any_retrans_done(sk))
  2223. tp->retrans_stamp = 0;
  2224. if (flag & FLAG_ECE)
  2225. tcp_enter_cwr(sk, 1);
  2226. if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
  2227. tcp_try_keep_open(sk);
  2228. } else {
  2229. tcp_cwnd_reduction(sk, prior_unsacked, 0);
  2230. }
  2231. }
  2232. static void tcp_mtup_probe_failed(struct sock *sk)
  2233. {
  2234. struct inet_connection_sock *icsk = inet_csk(sk);
  2235. icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
  2236. icsk->icsk_mtup.probe_size = 0;
  2237. }
  2238. static void tcp_mtup_probe_success(struct sock *sk)
  2239. {
  2240. struct tcp_sock *tp = tcp_sk(sk);
  2241. struct inet_connection_sock *icsk = inet_csk(sk);
  2242. /* FIXME: breaks with very large cwnd */
  2243. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  2244. tp->snd_cwnd = tp->snd_cwnd *
  2245. tcp_mss_to_mtu(sk, tp->mss_cache) /
  2246. icsk->icsk_mtup.probe_size;
  2247. tp->snd_cwnd_cnt = 0;
  2248. tp->snd_cwnd_stamp = tcp_time_stamp;
  2249. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  2250. icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
  2251. icsk->icsk_mtup.probe_size = 0;
  2252. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  2253. }
  2254. /* Do a simple retransmit without using the backoff mechanisms in
  2255. * tcp_timer. This is used for path mtu discovery.
  2256. * The socket is already locked here.
  2257. */
  2258. void tcp_simple_retransmit(struct sock *sk)
  2259. {
  2260. const struct inet_connection_sock *icsk = inet_csk(sk);
  2261. struct tcp_sock *tp = tcp_sk(sk);
  2262. struct sk_buff *skb;
  2263. unsigned int mss = tcp_current_mss(sk);
  2264. u32 prior_lost = tp->lost_out;
  2265. tcp_for_write_queue(skb, sk) {
  2266. if (skb == tcp_send_head(sk))
  2267. break;
  2268. if (tcp_skb_seglen(skb) > mss &&
  2269. !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
  2270. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  2271. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  2272. tp->retrans_out -= tcp_skb_pcount(skb);
  2273. }
  2274. tcp_skb_mark_lost_uncond_verify(tp, skb);
  2275. }
  2276. }
  2277. tcp_clear_retrans_hints_partial(tp);
  2278. if (prior_lost == tp->lost_out)
  2279. return;
  2280. if (tcp_is_reno(tp))
  2281. tcp_limit_reno_sacked(tp);
  2282. tcp_verify_left_out(tp);
  2283. /* Don't muck with the congestion window here.
  2284. * Reason is that we do not increase amount of _data_
  2285. * in network, but units changed and effective
  2286. * cwnd/ssthresh really reduced now.
  2287. */
  2288. if (icsk->icsk_ca_state != TCP_CA_Loss) {
  2289. tp->high_seq = tp->snd_nxt;
  2290. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  2291. tp->prior_ssthresh = 0;
  2292. tp->undo_marker = 0;
  2293. tcp_set_ca_state(sk, TCP_CA_Loss);
  2294. }
  2295. tcp_xmit_retransmit_queue(sk);
  2296. }
  2297. EXPORT_SYMBOL(tcp_simple_retransmit);
  2298. static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
  2299. {
  2300. struct tcp_sock *tp = tcp_sk(sk);
  2301. int mib_idx;
  2302. if (tcp_is_reno(tp))
  2303. mib_idx = LINUX_MIB_TCPRENORECOVERY;
  2304. else
  2305. mib_idx = LINUX_MIB_TCPSACKRECOVERY;
  2306. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  2307. tp->prior_ssthresh = 0;
  2308. tp->undo_marker = tp->snd_una;
  2309. tp->undo_retrans = tp->retrans_out;
  2310. if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
  2311. if (!ece_ack)
  2312. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  2313. tcp_init_cwnd_reduction(sk, true);
  2314. }
  2315. tcp_set_ca_state(sk, TCP_CA_Recovery);
  2316. }
  2317. /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
  2318. * recovered or spurious. Otherwise retransmits more on partial ACKs.
  2319. */
  2320. static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
  2321. {
  2322. struct inet_connection_sock *icsk = inet_csk(sk);
  2323. struct tcp_sock *tp = tcp_sk(sk);
  2324. bool recovered = !before(tp->snd_una, tp->high_seq);
  2325. if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
  2326. if (flag & FLAG_ORIG_SACK_ACKED) {
  2327. /* Step 3.b. A timeout is spurious if not all data are
  2328. * lost, i.e., never-retransmitted data are (s)acked.
  2329. */
  2330. tcp_try_undo_loss(sk, true);
  2331. return;
  2332. }
  2333. if (after(tp->snd_nxt, tp->high_seq) &&
  2334. (flag & FLAG_DATA_SACKED || is_dupack)) {
  2335. tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
  2336. } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
  2337. tp->high_seq = tp->snd_nxt;
  2338. __tcp_push_pending_frames(sk, tcp_current_mss(sk),
  2339. TCP_NAGLE_OFF);
  2340. if (after(tp->snd_nxt, tp->high_seq))
  2341. return; /* Step 2.b */
  2342. tp->frto = 0;
  2343. }
  2344. }
  2345. if (recovered) {
  2346. /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
  2347. icsk->icsk_retransmits = 0;
  2348. tcp_try_undo_recovery(sk);
  2349. return;
  2350. }
  2351. if (flag & FLAG_DATA_ACKED)
  2352. icsk->icsk_retransmits = 0;
  2353. if (tcp_is_reno(tp)) {
  2354. /* A Reno DUPACK means new data in F-RTO step 2.b above are
  2355. * delivered. Lower inflight to clock out (re)tranmissions.
  2356. */
  2357. if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
  2358. tcp_add_reno_sack(sk);
  2359. else if (flag & FLAG_SND_UNA_ADVANCED)
  2360. tcp_reset_reno_sack(tp);
  2361. }
  2362. if (tcp_try_undo_loss(sk, false))
  2363. return;
  2364. tcp_xmit_retransmit_queue(sk);
  2365. }
  2366. /* Undo during fast recovery after partial ACK. */
  2367. static bool tcp_try_undo_partial(struct sock *sk, const int acked,
  2368. const int prior_unsacked)
  2369. {
  2370. struct tcp_sock *tp = tcp_sk(sk);
  2371. if (tp->undo_marker && tcp_packet_delayed(tp)) {
  2372. /* Plain luck! Hole if filled with delayed
  2373. * packet, rather than with a retransmit.
  2374. */
  2375. tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
  2376. /* We are getting evidence that the reordering degree is higher
  2377. * than we realized. If there are no retransmits out then we
  2378. * can undo. Otherwise we clock out new packets but do not
  2379. * mark more packets lost or retransmit more.
  2380. */
  2381. if (tp->retrans_out) {
  2382. tcp_cwnd_reduction(sk, prior_unsacked, 0);
  2383. return true;
  2384. }
  2385. if (!tcp_any_retrans_done(sk))
  2386. tp->retrans_stamp = 0;
  2387. DBGUNDO(sk, "partial recovery");
  2388. tcp_undo_cwnd_reduction(sk, true);
  2389. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
  2390. tcp_try_keep_open(sk);
  2391. return true;
  2392. }
  2393. return false;
  2394. }
  2395. /* Process an event, which can update packets-in-flight not trivially.
  2396. * Main goal of this function is to calculate new estimate for left_out,
  2397. * taking into account both packets sitting in receiver's buffer and
  2398. * packets lost by network.
  2399. *
  2400. * Besides that it does CWND reduction, when packet loss is detected
  2401. * and changes state of machine.
  2402. *
  2403. * It does _not_ decide what to send, it is made in function
  2404. * tcp_xmit_retransmit_queue().
  2405. */
  2406. static void tcp_fastretrans_alert(struct sock *sk, const int acked,
  2407. const int prior_unsacked,
  2408. bool is_dupack, int flag)
  2409. {
  2410. struct inet_connection_sock *icsk = inet_csk(sk);
  2411. struct tcp_sock *tp = tcp_sk(sk);
  2412. bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
  2413. (tcp_fackets_out(tp) > tp->reordering));
  2414. int fast_rexmit = 0;
  2415. if (WARN_ON(!tp->packets_out && tp->sacked_out))
  2416. tp->sacked_out = 0;
  2417. if (WARN_ON(!tp->sacked_out && tp->fackets_out))
  2418. tp->fackets_out = 0;
  2419. /* Now state machine starts.
  2420. * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
  2421. if (flag & FLAG_ECE)
  2422. tp->prior_ssthresh = 0;
  2423. /* B. In all the states check for reneging SACKs. */
  2424. if (tcp_check_sack_reneging(sk, flag))
  2425. return;
  2426. /* C. Check consistency of the current state. */
  2427. tcp_verify_left_out(tp);
  2428. /* D. Check state exit conditions. State can be terminated
  2429. * when high_seq is ACKed. */
  2430. if (icsk->icsk_ca_state == TCP_CA_Open) {
  2431. WARN_ON(tp->retrans_out != 0);
  2432. tp->retrans_stamp = 0;
  2433. } else if (!before(tp->snd_una, tp->high_seq)) {
  2434. switch (icsk->icsk_ca_state) {
  2435. case TCP_CA_CWR:
  2436. /* CWR is to be held something *above* high_seq
  2437. * is ACKed for CWR bit to reach receiver. */
  2438. if (tp->snd_una != tp->high_seq) {
  2439. tcp_end_cwnd_reduction(sk);
  2440. tcp_set_ca_state(sk, TCP_CA_Open);
  2441. }
  2442. break;
  2443. case TCP_CA_Recovery:
  2444. if (tcp_is_reno(tp))
  2445. tcp_reset_reno_sack(tp);
  2446. if (tcp_try_undo_recovery(sk))
  2447. return;
  2448. tcp_end_cwnd_reduction(sk);
  2449. break;
  2450. }
  2451. }
  2452. /* E. Process state. */
  2453. switch (icsk->icsk_ca_state) {
  2454. case TCP_CA_Recovery:
  2455. if (!(flag & FLAG_SND_UNA_ADVANCED)) {
  2456. if (tcp_is_reno(tp) && is_dupack)
  2457. tcp_add_reno_sack(sk);
  2458. } else {
  2459. if (tcp_try_undo_partial(sk, acked, prior_unsacked))
  2460. return;
  2461. /* Partial ACK arrived. Force fast retransmit. */
  2462. do_lost = tcp_is_reno(tp) ||
  2463. tcp_fackets_out(tp) > tp->reordering;
  2464. }
  2465. if (tcp_try_undo_dsack(sk)) {
  2466. tcp_try_keep_open(sk);
  2467. return;
  2468. }
  2469. break;
  2470. case TCP_CA_Loss:
  2471. tcp_process_loss(sk, flag, is_dupack);
  2472. if (icsk->icsk_ca_state != TCP_CA_Open)
  2473. return;
  2474. /* Fall through to processing in Open state. */
  2475. default:
  2476. if (tcp_is_reno(tp)) {
  2477. if (flag & FLAG_SND_UNA_ADVANCED)
  2478. tcp_reset_reno_sack(tp);
  2479. if (is_dupack)
  2480. tcp_add_reno_sack(sk);
  2481. }
  2482. if (icsk->icsk_ca_state <= TCP_CA_Disorder)
  2483. tcp_try_undo_dsack(sk);
  2484. if (!tcp_time_to_recover(sk, flag)) {
  2485. tcp_try_to_open(sk, flag, prior_unsacked);
  2486. return;
  2487. }
  2488. /* MTU probe failure: don't reduce cwnd */
  2489. if (icsk->icsk_ca_state < TCP_CA_CWR &&
  2490. icsk->icsk_mtup.probe_size &&
  2491. tp->snd_una == tp->mtu_probe.probe_seq_start) {
  2492. tcp_mtup_probe_failed(sk);
  2493. /* Restores the reduction we did in tcp_mtup_probe() */
  2494. tp->snd_cwnd++;
  2495. tcp_simple_retransmit(sk);
  2496. return;
  2497. }
  2498. /* Otherwise enter Recovery state */
  2499. tcp_enter_recovery(sk, (flag & FLAG_ECE));
  2500. fast_rexmit = 1;
  2501. }
  2502. if (do_lost)
  2503. tcp_update_scoreboard(sk, fast_rexmit);
  2504. tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
  2505. tcp_xmit_retransmit_queue(sk);
  2506. }
  2507. static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
  2508. s32 seq_rtt, s32 sack_rtt)
  2509. {
  2510. const struct tcp_sock *tp = tcp_sk(sk);
  2511. /* Prefer RTT measured from ACK's timing to TS-ECR. This is because
  2512. * broken middle-boxes or peers may corrupt TS-ECR fields. But
  2513. * Karn's algorithm forbids taking RTT if some retransmitted data
  2514. * is acked (RFC6298).
  2515. */
  2516. if (flag & FLAG_RETRANS_DATA_ACKED)
  2517. seq_rtt = -1;
  2518. if (seq_rtt < 0)
  2519. seq_rtt = sack_rtt;
  2520. /* RTTM Rule: A TSecr value received in a segment is used to
  2521. * update the averaged RTT measurement only if the segment
  2522. * acknowledges some new data, i.e., only if it advances the
  2523. * left edge of the send window.
  2524. * See draft-ietf-tcplw-high-performance-00, section 3.3.
  2525. */
  2526. if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  2527. flag & FLAG_ACKED)
  2528. seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
  2529. if (seq_rtt < 0)
  2530. return false;
  2531. tcp_rtt_estimator(sk, seq_rtt);
  2532. tcp_set_rto(sk);
  2533. /* RFC6298: only reset backoff on valid RTT measurement. */
  2534. inet_csk(sk)->icsk_backoff = 0;
  2535. return true;
  2536. }
  2537. /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
  2538. static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
  2539. {
  2540. struct tcp_sock *tp = tcp_sk(sk);
  2541. s32 seq_rtt = -1;
  2542. if (synack_stamp && !tp->total_retrans)
  2543. seq_rtt = tcp_time_stamp - synack_stamp;
  2544. /* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
  2545. * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
  2546. */
  2547. if (!tp->srtt)
  2548. tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
  2549. }
  2550. static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
  2551. {
  2552. const struct inet_connection_sock *icsk = inet_csk(sk);
  2553. icsk->icsk_ca_ops->cong_avoid(sk, ack, acked, in_flight);
  2554. tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
  2555. }
  2556. /* Restart timer after forward progress on connection.
  2557. * RFC2988 recommends to restart timer to now+rto.
  2558. */
  2559. void tcp_rearm_rto(struct sock *sk)
  2560. {
  2561. const struct inet_connection_sock *icsk = inet_csk(sk);
  2562. struct tcp_sock *tp = tcp_sk(sk);
  2563. /* If the retrans timer is currently being used by Fast Open
  2564. * for SYN-ACK retrans purpose, stay put.
  2565. */
  2566. if (tp->fastopen_rsk)
  2567. return;
  2568. if (!tp->packets_out) {
  2569. inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
  2570. } else {
  2571. u32 rto = inet_csk(sk)->icsk_rto;
  2572. /* Offset the time elapsed after installing regular RTO */
  2573. if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
  2574. icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
  2575. struct sk_buff *skb = tcp_write_queue_head(sk);
  2576. const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
  2577. s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
  2578. /* delta may not be positive if the socket is locked
  2579. * when the retrans timer fires and is rescheduled.
  2580. */
  2581. if (delta > 0)
  2582. rto = delta;
  2583. }
  2584. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
  2585. TCP_RTO_MAX);
  2586. }
  2587. }
  2588. /* This function is called when the delayed ER timer fires. TCP enters
  2589. * fast recovery and performs fast-retransmit.
  2590. */
  2591. void tcp_resume_early_retransmit(struct sock *sk)
  2592. {
  2593. struct tcp_sock *tp = tcp_sk(sk);
  2594. tcp_rearm_rto(sk);
  2595. /* Stop if ER is disabled after the delayed ER timer is scheduled */
  2596. if (!tp->do_early_retrans)
  2597. return;
  2598. tcp_enter_recovery(sk, false);
  2599. tcp_update_scoreboard(sk, 1);
  2600. tcp_xmit_retransmit_queue(sk);
  2601. }
  2602. /* If we get here, the whole TSO packet has not been acked. */
  2603. static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
  2604. {
  2605. struct tcp_sock *tp = tcp_sk(sk);
  2606. u32 packets_acked;
  2607. BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
  2608. packets_acked = tcp_skb_pcount(skb);
  2609. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  2610. return 0;
  2611. packets_acked -= tcp_skb_pcount(skb);
  2612. if (packets_acked) {
  2613. BUG_ON(tcp_skb_pcount(skb) == 0);
  2614. BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
  2615. }
  2616. return packets_acked;
  2617. }
  2618. /* Remove acknowledged frames from the retransmission queue. If our packet
  2619. * is before the ack sequence we can discard it as it's confirmed to have
  2620. * arrived at the other end.
  2621. */
  2622. static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
  2623. u32 prior_snd_una, s32 sack_rtt)
  2624. {
  2625. struct tcp_sock *tp = tcp_sk(sk);
  2626. const struct inet_connection_sock *icsk = inet_csk(sk);
  2627. struct sk_buff *skb;
  2628. u32 now = tcp_time_stamp;
  2629. bool fully_acked = true;
  2630. int flag = 0;
  2631. u32 pkts_acked = 0;
  2632. u32 reord = tp->packets_out;
  2633. u32 prior_sacked = tp->sacked_out;
  2634. s32 seq_rtt = -1;
  2635. s32 ca_seq_rtt = -1;
  2636. ktime_t last_ackt = net_invalid_timestamp();
  2637. bool rtt_update;
  2638. while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
  2639. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2640. u32 acked_pcount;
  2641. u8 sacked = scb->sacked;
  2642. /* Determine how many packets and what bytes were acked, tso and else */
  2643. if (after(scb->end_seq, tp->snd_una)) {
  2644. if (tcp_skb_pcount(skb) == 1 ||
  2645. !after(tp->snd_una, scb->seq))
  2646. break;
  2647. acked_pcount = tcp_tso_acked(sk, skb);
  2648. if (!acked_pcount)
  2649. break;
  2650. fully_acked = false;
  2651. } else {
  2652. acked_pcount = tcp_skb_pcount(skb);
  2653. }
  2654. if (sacked & TCPCB_RETRANS) {
  2655. if (sacked & TCPCB_SACKED_RETRANS)
  2656. tp->retrans_out -= acked_pcount;
  2657. flag |= FLAG_RETRANS_DATA_ACKED;
  2658. } else {
  2659. ca_seq_rtt = now - scb->when;
  2660. last_ackt = skb->tstamp;
  2661. if (seq_rtt < 0) {
  2662. seq_rtt = ca_seq_rtt;
  2663. }
  2664. if (!(sacked & TCPCB_SACKED_ACKED))
  2665. reord = min(pkts_acked, reord);
  2666. if (!after(scb->end_seq, tp->high_seq))
  2667. flag |= FLAG_ORIG_SACK_ACKED;
  2668. }
  2669. if (sacked & TCPCB_SACKED_ACKED)
  2670. tp->sacked_out -= acked_pcount;
  2671. if (sacked & TCPCB_LOST)
  2672. tp->lost_out -= acked_pcount;
  2673. tp->packets_out -= acked_pcount;
  2674. pkts_acked += acked_pcount;
  2675. /* Initial outgoing SYN's get put onto the write_queue
  2676. * just like anything else we transmit. It is not
  2677. * true data, and if we misinform our callers that
  2678. * this ACK acks real data, we will erroneously exit
  2679. * connection startup slow start one packet too
  2680. * quickly. This is severely frowned upon behavior.
  2681. */
  2682. if (!(scb->tcp_flags & TCPHDR_SYN)) {
  2683. flag |= FLAG_DATA_ACKED;
  2684. } else {
  2685. flag |= FLAG_SYN_ACKED;
  2686. tp->retrans_stamp = 0;
  2687. }
  2688. if (!fully_acked)
  2689. break;
  2690. tcp_unlink_write_queue(skb, sk);
  2691. sk_wmem_free_skb(sk, skb);
  2692. if (skb == tp->retransmit_skb_hint)
  2693. tp->retransmit_skb_hint = NULL;
  2694. if (skb == tp->lost_skb_hint)
  2695. tp->lost_skb_hint = NULL;
  2696. }
  2697. if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
  2698. tp->snd_up = tp->snd_una;
  2699. if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
  2700. flag |= FLAG_SACK_RENEGING;
  2701. rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt);
  2702. if (flag & FLAG_ACKED) {
  2703. const struct tcp_congestion_ops *ca_ops
  2704. = inet_csk(sk)->icsk_ca_ops;
  2705. tcp_rearm_rto(sk);
  2706. if (unlikely(icsk->icsk_mtup.probe_size &&
  2707. !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
  2708. tcp_mtup_probe_success(sk);
  2709. }
  2710. if (tcp_is_reno(tp)) {
  2711. tcp_remove_reno_sacks(sk, pkts_acked);
  2712. } else {
  2713. int delta;
  2714. /* Non-retransmitted hole got filled? That's reordering */
  2715. if (reord < prior_fackets)
  2716. tcp_update_reordering(sk, tp->fackets_out - reord, 0);
  2717. delta = tcp_is_fack(tp) ? pkts_acked :
  2718. prior_sacked - tp->sacked_out;
  2719. tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
  2720. }
  2721. tp->fackets_out -= min(pkts_acked, tp->fackets_out);
  2722. if (ca_ops->pkts_acked) {
  2723. s32 rtt_us = -1;
  2724. /* Is the ACK triggering packet unambiguous? */
  2725. if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
  2726. /* High resolution needed and available? */
  2727. if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
  2728. !ktime_equal(last_ackt,
  2729. net_invalid_timestamp()))
  2730. rtt_us = ktime_us_delta(ktime_get_real(),
  2731. last_ackt);
  2732. else if (ca_seq_rtt >= 0)
  2733. rtt_us = jiffies_to_usecs(ca_seq_rtt);
  2734. }
  2735. ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
  2736. }
  2737. } else if (skb && rtt_update && sack_rtt >= 0 &&
  2738. sack_rtt > (s32)(now - TCP_SKB_CB(skb)->when)) {
  2739. /* Do not re-arm RTO if the sack RTT is measured from data sent
  2740. * after when the head was last (re)transmitted. Otherwise the
  2741. * timeout may continue to extend in loss recovery.
  2742. */
  2743. tcp_rearm_rto(sk);
  2744. }
  2745. #if FASTRETRANS_DEBUG > 0
  2746. WARN_ON((int)tp->sacked_out < 0);
  2747. WARN_ON((int)tp->lost_out < 0);
  2748. WARN_ON((int)tp->retrans_out < 0);
  2749. if (!tp->packets_out && tcp_is_sack(tp)) {
  2750. icsk = inet_csk(sk);
  2751. if (tp->lost_out) {
  2752. pr_debug("Leak l=%u %d\n",
  2753. tp->lost_out, icsk->icsk_ca_state);
  2754. tp->lost_out = 0;
  2755. }
  2756. if (tp->sacked_out) {
  2757. pr_debug("Leak s=%u %d\n",
  2758. tp->sacked_out, icsk->icsk_ca_state);
  2759. tp->sacked_out = 0;
  2760. }
  2761. if (tp->retrans_out) {
  2762. pr_debug("Leak r=%u %d\n",
  2763. tp->retrans_out, icsk->icsk_ca_state);
  2764. tp->retrans_out = 0;
  2765. }
  2766. }
  2767. #endif
  2768. return flag;
  2769. }
  2770. static void tcp_ack_probe(struct sock *sk)
  2771. {
  2772. const struct tcp_sock *tp = tcp_sk(sk);
  2773. struct inet_connection_sock *icsk = inet_csk(sk);
  2774. /* Was it a usable window open? */
  2775. if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
  2776. icsk->icsk_backoff = 0;
  2777. inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
  2778. /* Socket must be waked up by subsequent tcp_data_snd_check().
  2779. * This function is not for random using!
  2780. */
  2781. } else {
  2782. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2783. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2784. TCP_RTO_MAX);
  2785. }
  2786. }
  2787. static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
  2788. {
  2789. return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
  2790. inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
  2791. }
  2792. /* Decide wheather to run the increase function of congestion control. */
  2793. static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
  2794. {
  2795. if (tcp_in_cwnd_reduction(sk))
  2796. return false;
  2797. /* If reordering is high then always grow cwnd whenever data is
  2798. * delivered regardless of its ordering. Otherwise stay conservative
  2799. * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
  2800. * new SACK or ECE mark may first advance cwnd here and later reduce
  2801. * cwnd in tcp_fastretrans_alert() based on more states.
  2802. */
  2803. if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
  2804. return flag & FLAG_FORWARD_PROGRESS;
  2805. return flag & FLAG_DATA_ACKED;
  2806. }
  2807. /* Check that window update is acceptable.
  2808. * The function assumes that snd_una<=ack<=snd_next.
  2809. */
  2810. static inline bool tcp_may_update_window(const struct tcp_sock *tp,
  2811. const u32 ack, const u32 ack_seq,
  2812. const u32 nwin)
  2813. {
  2814. return after(ack, tp->snd_una) ||
  2815. after(ack_seq, tp->snd_wl1) ||
  2816. (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
  2817. }
  2818. /* Update our send window.
  2819. *
  2820. * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
  2821. * and in FreeBSD. NetBSD's one is even worse.) is wrong.
  2822. */
  2823. static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
  2824. u32 ack_seq)
  2825. {
  2826. struct tcp_sock *tp = tcp_sk(sk);
  2827. int flag = 0;
  2828. u32 nwin = ntohs(tcp_hdr(skb)->window);
  2829. if (likely(!tcp_hdr(skb)->syn))
  2830. nwin <<= tp->rx_opt.snd_wscale;
  2831. if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
  2832. flag |= FLAG_WIN_UPDATE;
  2833. tcp_update_wl(tp, ack_seq);
  2834. if (tp->snd_wnd != nwin) {
  2835. tp->snd_wnd = nwin;
  2836. /* Note, it is the only place, where
  2837. * fast path is recovered for sending TCP.
  2838. */
  2839. tp->pred_flags = 0;
  2840. tcp_fast_path_check(sk);
  2841. if (nwin > tp->max_window) {
  2842. tp->max_window = nwin;
  2843. tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
  2844. }
  2845. }
  2846. }
  2847. tp->snd_una = ack;
  2848. return flag;
  2849. }
  2850. /* RFC 5961 7 [ACK Throttling] */
  2851. static void tcp_send_challenge_ack(struct sock *sk)
  2852. {
  2853. /* unprotected vars, we dont care of overwrites */
  2854. static u32 challenge_timestamp;
  2855. static unsigned int challenge_count;
  2856. u32 now = jiffies / HZ;
  2857. if (now != challenge_timestamp) {
  2858. challenge_timestamp = now;
  2859. challenge_count = 0;
  2860. }
  2861. if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
  2862. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
  2863. tcp_send_ack(sk);
  2864. }
  2865. }
  2866. static void tcp_store_ts_recent(struct tcp_sock *tp)
  2867. {
  2868. tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
  2869. tp->rx_opt.ts_recent_stamp = get_seconds();
  2870. }
  2871. static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
  2872. {
  2873. if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
  2874. /* PAWS bug workaround wrt. ACK frames, the PAWS discard
  2875. * extra check below makes sure this can only happen
  2876. * for pure ACK frames. -DaveM
  2877. *
  2878. * Not only, also it occurs for expired timestamps.
  2879. */
  2880. if (tcp_paws_check(&tp->rx_opt, 0))
  2881. tcp_store_ts_recent(tp);
  2882. }
  2883. }
  2884. /* This routine deals with acks during a TLP episode.
  2885. * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
  2886. */
  2887. static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
  2888. {
  2889. struct tcp_sock *tp = tcp_sk(sk);
  2890. bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
  2891. !(flag & (FLAG_SND_UNA_ADVANCED |
  2892. FLAG_NOT_DUP | FLAG_DATA_SACKED));
  2893. /* Mark the end of TLP episode on receiving TLP dupack or when
  2894. * ack is after tlp_high_seq.
  2895. */
  2896. if (is_tlp_dupack) {
  2897. tp->tlp_high_seq = 0;
  2898. return;
  2899. }
  2900. if (after(ack, tp->tlp_high_seq)) {
  2901. tp->tlp_high_seq = 0;
  2902. /* Don't reduce cwnd if DSACK arrives for TLP retrans. */
  2903. if (!(flag & FLAG_DSACKING_ACK)) {
  2904. tcp_init_cwnd_reduction(sk, true);
  2905. tcp_set_ca_state(sk, TCP_CA_CWR);
  2906. tcp_end_cwnd_reduction(sk);
  2907. tcp_try_keep_open(sk);
  2908. NET_INC_STATS_BH(sock_net(sk),
  2909. LINUX_MIB_TCPLOSSPROBERECOVERY);
  2910. }
  2911. }
  2912. }
  2913. /* This routine deals with incoming acks, but not outgoing ones. */
  2914. static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
  2915. {
  2916. struct inet_connection_sock *icsk = inet_csk(sk);
  2917. struct tcp_sock *tp = tcp_sk(sk);
  2918. u32 prior_snd_una = tp->snd_una;
  2919. u32 ack_seq = TCP_SKB_CB(skb)->seq;
  2920. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2921. bool is_dupack = false;
  2922. u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt;
  2923. u32 prior_fackets;
  2924. int prior_packets = tp->packets_out;
  2925. const int prior_unsacked = tp->packets_out - tp->sacked_out;
  2926. int acked = 0; /* Number of packets newly acked */
  2927. s32 sack_rtt = -1;
  2928. /* If the ack is older than previous acks
  2929. * then we can probably ignore it.
  2930. */
  2931. if (before(ack, prior_snd_una)) {
  2932. /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
  2933. if (before(ack, prior_snd_una - tp->max_window)) {
  2934. tcp_send_challenge_ack(sk);
  2935. return -1;
  2936. }
  2937. goto old_ack;
  2938. }
  2939. /* If the ack includes data we haven't sent yet, discard
  2940. * this segment (RFC793 Section 3.9).
  2941. */
  2942. if (after(ack, tp->snd_nxt))
  2943. goto invalid_ack;
  2944. if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
  2945. icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  2946. tcp_rearm_rto(sk);
  2947. if (after(ack, prior_snd_una))
  2948. flag |= FLAG_SND_UNA_ADVANCED;
  2949. prior_fackets = tp->fackets_out;
  2950. prior_in_flight = tcp_packets_in_flight(tp);
  2951. /* ts_recent update must be made after we are sure that the packet
  2952. * is in window.
  2953. */
  2954. if (flag & FLAG_UPDATE_TS_RECENT)
  2955. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  2956. if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
  2957. /* Window is constant, pure forward advance.
  2958. * No more checks are required.
  2959. * Note, we use the fact that SND.UNA>=SND.WL2.
  2960. */
  2961. tcp_update_wl(tp, ack_seq);
  2962. tp->snd_una = ack;
  2963. flag |= FLAG_WIN_UPDATE;
  2964. tcp_ca_event(sk, CA_EVENT_FAST_ACK);
  2965. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
  2966. } else {
  2967. if (ack_seq != TCP_SKB_CB(skb)->end_seq)
  2968. flag |= FLAG_DATA;
  2969. else
  2970. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
  2971. flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
  2972. if (TCP_SKB_CB(skb)->sacked)
  2973. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
  2974. &sack_rtt);
  2975. if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
  2976. flag |= FLAG_ECE;
  2977. tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
  2978. }
  2979. /* We passed data and got it acked, remove any soft error
  2980. * log. Something worked...
  2981. */
  2982. sk->sk_err_soft = 0;
  2983. icsk->icsk_probes_out = 0;
  2984. tp->rcv_tstamp = tcp_time_stamp;
  2985. if (!prior_packets)
  2986. goto no_queue;
  2987. /* See if we can take anything off of the retransmit queue. */
  2988. acked = tp->packets_out;
  2989. flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt);
  2990. acked -= tp->packets_out;
  2991. /* Advance cwnd if state allows */
  2992. if (tcp_may_raise_cwnd(sk, flag))
  2993. tcp_cong_avoid(sk, ack, acked, prior_in_flight);
  2994. if (tcp_ack_is_dubious(sk, flag)) {
  2995. is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
  2996. tcp_fastretrans_alert(sk, acked, prior_unsacked,
  2997. is_dupack, flag);
  2998. }
  2999. if (tp->tlp_high_seq)
  3000. tcp_process_tlp_ack(sk, ack, flag);
  3001. if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
  3002. struct dst_entry *dst = __sk_dst_get(sk);
  3003. if (dst)
  3004. dst_confirm(dst);
  3005. }
  3006. if (icsk->icsk_pending == ICSK_TIME_RETRANS)
  3007. tcp_schedule_loss_probe(sk);
  3008. if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd)
  3009. tcp_update_pacing_rate(sk);
  3010. return 1;
  3011. no_queue:
  3012. /* If data was DSACKed, see if we can undo a cwnd reduction. */
  3013. if (flag & FLAG_DSACKING_ACK)
  3014. tcp_fastretrans_alert(sk, acked, prior_unsacked,
  3015. is_dupack, flag);
  3016. /* If this ack opens up a zero window, clear backoff. It was
  3017. * being used to time the probes, and is probably far higher than
  3018. * it needs to be for normal retransmission.
  3019. */
  3020. if (tcp_send_head(sk))
  3021. tcp_ack_probe(sk);
  3022. if (tp->tlp_high_seq)
  3023. tcp_process_tlp_ack(sk, ack, flag);
  3024. return 1;
  3025. invalid_ack:
  3026. SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  3027. return -1;
  3028. old_ack:
  3029. /* If data was SACKed, tag it and see if we should send more data.
  3030. * If data was DSACKed, see if we can undo a cwnd reduction.
  3031. */
  3032. if (TCP_SKB_CB(skb)->sacked) {
  3033. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
  3034. &sack_rtt);
  3035. tcp_fastretrans_alert(sk, acked, prior_unsacked,
  3036. is_dupack, flag);
  3037. }
  3038. SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  3039. return 0;
  3040. }
  3041. /* Look for tcp options. Normally only called on SYN and SYNACK packets.
  3042. * But, this can also be called on packets in the established flow when
  3043. * the fast version below fails.
  3044. */
  3045. void tcp_parse_options(const struct sk_buff *skb,
  3046. struct tcp_options_received *opt_rx, int estab,
  3047. struct tcp_fastopen_cookie *foc)
  3048. {
  3049. const unsigned char *ptr;
  3050. const struct tcphdr *th = tcp_hdr(skb);
  3051. int length = (th->doff * 4) - sizeof(struct tcphdr);
  3052. ptr = (const unsigned char *)(th + 1);
  3053. opt_rx->saw_tstamp = 0;
  3054. while (length > 0) {
  3055. int opcode = *ptr++;
  3056. int opsize;
  3057. switch (opcode) {
  3058. case TCPOPT_EOL:
  3059. return;
  3060. case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
  3061. length--;
  3062. continue;
  3063. default:
  3064. opsize = *ptr++;
  3065. if (opsize < 2) /* "silly options" */
  3066. return;
  3067. if (opsize > length)
  3068. return; /* don't parse partial options */
  3069. switch (opcode) {
  3070. case TCPOPT_MSS:
  3071. if (opsize == TCPOLEN_MSS && th->syn && !estab) {
  3072. u16 in_mss = get_unaligned_be16(ptr);
  3073. if (in_mss) {
  3074. if (opt_rx->user_mss &&
  3075. opt_rx->user_mss < in_mss)
  3076. in_mss = opt_rx->user_mss;
  3077. opt_rx->mss_clamp = in_mss;
  3078. }
  3079. }
  3080. break;
  3081. case TCPOPT_WINDOW:
  3082. if (opsize == TCPOLEN_WINDOW && th->syn &&
  3083. !estab && sysctl_tcp_window_scaling) {
  3084. __u8 snd_wscale = *(__u8 *)ptr;
  3085. opt_rx->wscale_ok = 1;
  3086. if (snd_wscale > 14) {
  3087. net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
  3088. __func__,
  3089. snd_wscale);
  3090. snd_wscale = 14;
  3091. }
  3092. opt_rx->snd_wscale = snd_wscale;
  3093. }
  3094. break;
  3095. case TCPOPT_TIMESTAMP:
  3096. if ((opsize == TCPOLEN_TIMESTAMP) &&
  3097. ((estab && opt_rx->tstamp_ok) ||
  3098. (!estab && sysctl_tcp_timestamps))) {
  3099. opt_rx->saw_tstamp = 1;
  3100. opt_rx->rcv_tsval = get_unaligned_be32(ptr);
  3101. opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
  3102. }
  3103. break;
  3104. case TCPOPT_SACK_PERM:
  3105. if (opsize == TCPOLEN_SACK_PERM && th->syn &&
  3106. !estab && sysctl_tcp_sack) {
  3107. opt_rx->sack_ok = TCP_SACK_SEEN;
  3108. tcp_sack_reset(opt_rx);
  3109. }
  3110. break;
  3111. case TCPOPT_SACK:
  3112. if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
  3113. !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
  3114. opt_rx->sack_ok) {
  3115. TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
  3116. }
  3117. break;
  3118. #ifdef CONFIG_TCP_MD5SIG
  3119. case TCPOPT_MD5SIG:
  3120. /*
  3121. * The MD5 Hash has already been
  3122. * checked (see tcp_v{4,6}_do_rcv()).
  3123. */
  3124. break;
  3125. #endif
  3126. case TCPOPT_EXP:
  3127. /* Fast Open option shares code 254 using a
  3128. * 16 bits magic number. It's valid only in
  3129. * SYN or SYN-ACK with an even size.
  3130. */
  3131. if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
  3132. get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
  3133. foc == NULL || !th->syn || (opsize & 1))
  3134. break;
  3135. foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
  3136. if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
  3137. foc->len <= TCP_FASTOPEN_COOKIE_MAX)
  3138. memcpy(foc->val, ptr + 2, foc->len);
  3139. else if (foc->len != 0)
  3140. foc->len = -1;
  3141. break;
  3142. }
  3143. ptr += opsize-2;
  3144. length -= opsize;
  3145. }
  3146. }
  3147. }
  3148. EXPORT_SYMBOL(tcp_parse_options);
  3149. static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
  3150. {
  3151. const __be32 *ptr = (const __be32 *)(th + 1);
  3152. if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  3153. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
  3154. tp->rx_opt.saw_tstamp = 1;
  3155. ++ptr;
  3156. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  3157. ++ptr;
  3158. if (*ptr)
  3159. tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
  3160. else
  3161. tp->rx_opt.rcv_tsecr = 0;
  3162. return true;
  3163. }
  3164. return false;
  3165. }
  3166. /* Fast parse options. This hopes to only see timestamps.
  3167. * If it is wrong it falls back on tcp_parse_options().
  3168. */
  3169. static bool tcp_fast_parse_options(const struct sk_buff *skb,
  3170. const struct tcphdr *th, struct tcp_sock *tp)
  3171. {
  3172. /* In the spirit of fast parsing, compare doff directly to constant
  3173. * values. Because equality is used, short doff can be ignored here.
  3174. */
  3175. if (th->doff == (sizeof(*th) / 4)) {
  3176. tp->rx_opt.saw_tstamp = 0;
  3177. return false;
  3178. } else if (tp->rx_opt.tstamp_ok &&
  3179. th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
  3180. if (tcp_parse_aligned_timestamp(tp, th))
  3181. return true;
  3182. }
  3183. tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
  3184. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  3185. tp->rx_opt.rcv_tsecr -= tp->tsoffset;
  3186. return true;
  3187. }
  3188. #ifdef CONFIG_TCP_MD5SIG
  3189. /*
  3190. * Parse MD5 Signature option
  3191. */
  3192. const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
  3193. {
  3194. int length = (th->doff << 2) - sizeof(*th);
  3195. const u8 *ptr = (const u8 *)(th + 1);
  3196. /* If the TCP option is too short, we can short cut */
  3197. if (length < TCPOLEN_MD5SIG)
  3198. return NULL;
  3199. while (length > 0) {
  3200. int opcode = *ptr++;
  3201. int opsize;
  3202. switch(opcode) {
  3203. case TCPOPT_EOL:
  3204. return NULL;
  3205. case TCPOPT_NOP:
  3206. length--;
  3207. continue;
  3208. default:
  3209. opsize = *ptr++;
  3210. if (opsize < 2 || opsize > length)
  3211. return NULL;
  3212. if (opcode == TCPOPT_MD5SIG)
  3213. return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
  3214. }
  3215. ptr += opsize - 2;
  3216. length -= opsize;
  3217. }
  3218. return NULL;
  3219. }
  3220. EXPORT_SYMBOL(tcp_parse_md5sig_option);
  3221. #endif
  3222. /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
  3223. *
  3224. * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
  3225. * it can pass through stack. So, the following predicate verifies that
  3226. * this segment is not used for anything but congestion avoidance or
  3227. * fast retransmit. Moreover, we even are able to eliminate most of such
  3228. * second order effects, if we apply some small "replay" window (~RTO)
  3229. * to timestamp space.
  3230. *
  3231. * All these measures still do not guarantee that we reject wrapped ACKs
  3232. * on networks with high bandwidth, when sequence space is recycled fastly,
  3233. * but it guarantees that such events will be very rare and do not affect
  3234. * connection seriously. This doesn't look nice, but alas, PAWS is really
  3235. * buggy extension.
  3236. *
  3237. * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
  3238. * states that events when retransmit arrives after original data are rare.
  3239. * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
  3240. * the biggest problem on large power networks even with minor reordering.
  3241. * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
  3242. * up to bandwidth of 18Gigabit/sec. 8) ]
  3243. */
  3244. static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
  3245. {
  3246. const struct tcp_sock *tp = tcp_sk(sk);
  3247. const struct tcphdr *th = tcp_hdr(skb);
  3248. u32 seq = TCP_SKB_CB(skb)->seq;
  3249. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  3250. return (/* 1. Pure ACK with correct sequence number. */
  3251. (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
  3252. /* 2. ... and duplicate ACK. */
  3253. ack == tp->snd_una &&
  3254. /* 3. ... and does not update window. */
  3255. !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
  3256. /* 4. ... and sits in replay window. */
  3257. (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
  3258. }
  3259. static inline bool tcp_paws_discard(const struct sock *sk,
  3260. const struct sk_buff *skb)
  3261. {
  3262. const struct tcp_sock *tp = tcp_sk(sk);
  3263. return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
  3264. !tcp_disordered_ack(sk, skb);
  3265. }
  3266. /* Check segment sequence number for validity.
  3267. *
  3268. * Segment controls are considered valid, if the segment
  3269. * fits to the window after truncation to the window. Acceptability
  3270. * of data (and SYN, FIN, of course) is checked separately.
  3271. * See tcp_data_queue(), for example.
  3272. *
  3273. * Also, controls (RST is main one) are accepted using RCV.WUP instead
  3274. * of RCV.NXT. Peer still did not advance his SND.UNA when we
  3275. * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
  3276. * (borrowed from freebsd)
  3277. */
  3278. static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
  3279. {
  3280. return !before(end_seq, tp->rcv_wup) &&
  3281. !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
  3282. }
  3283. /* When we get a reset we do this. */
  3284. void tcp_reset(struct sock *sk)
  3285. {
  3286. /* We want the right error as BSD sees it (and indeed as we do). */
  3287. switch (sk->sk_state) {
  3288. case TCP_SYN_SENT:
  3289. sk->sk_err = ECONNREFUSED;
  3290. break;
  3291. case TCP_CLOSE_WAIT:
  3292. sk->sk_err = EPIPE;
  3293. break;
  3294. case TCP_CLOSE:
  3295. return;
  3296. default:
  3297. sk->sk_err = ECONNRESET;
  3298. }
  3299. /* This barrier is coupled with smp_rmb() in tcp_poll() */
  3300. smp_wmb();
  3301. if (!sock_flag(sk, SOCK_DEAD))
  3302. sk->sk_error_report(sk);
  3303. tcp_done(sk);
  3304. }
  3305. /*
  3306. * Process the FIN bit. This now behaves as it is supposed to work
  3307. * and the FIN takes effect when it is validly part of sequence
  3308. * space. Not before when we get holes.
  3309. *
  3310. * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
  3311. * (and thence onto LAST-ACK and finally, CLOSE, we never enter
  3312. * TIME-WAIT)
  3313. *
  3314. * If we are in FINWAIT-1, a received FIN indicates simultaneous
  3315. * close and we go into CLOSING (and later onto TIME-WAIT)
  3316. *
  3317. * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
  3318. */
  3319. static void tcp_fin(struct sock *sk)
  3320. {
  3321. struct tcp_sock *tp = tcp_sk(sk);
  3322. const struct dst_entry *dst;
  3323. inet_csk_schedule_ack(sk);
  3324. sk->sk_shutdown |= RCV_SHUTDOWN;
  3325. sock_set_flag(sk, SOCK_DONE);
  3326. switch (sk->sk_state) {
  3327. case TCP_SYN_RECV:
  3328. case TCP_ESTABLISHED:
  3329. /* Move to CLOSE_WAIT */
  3330. tcp_set_state(sk, TCP_CLOSE_WAIT);
  3331. dst = __sk_dst_get(sk);
  3332. if (!dst || !dst_metric(dst, RTAX_QUICKACK))
  3333. inet_csk(sk)->icsk_ack.pingpong = 1;
  3334. break;
  3335. case TCP_CLOSE_WAIT:
  3336. case TCP_CLOSING:
  3337. /* Received a retransmission of the FIN, do
  3338. * nothing.
  3339. */
  3340. break;
  3341. case TCP_LAST_ACK:
  3342. /* RFC793: Remain in the LAST-ACK state. */
  3343. break;
  3344. case TCP_FIN_WAIT1:
  3345. /* This case occurs when a simultaneous close
  3346. * happens, we must ack the received FIN and
  3347. * enter the CLOSING state.
  3348. */
  3349. tcp_send_ack(sk);
  3350. tcp_set_state(sk, TCP_CLOSING);
  3351. break;
  3352. case TCP_FIN_WAIT2:
  3353. /* Received a FIN -- send ACK and enter TIME_WAIT. */
  3354. tcp_send_ack(sk);
  3355. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  3356. break;
  3357. default:
  3358. /* Only TCP_LISTEN and TCP_CLOSE are left, in these
  3359. * cases we should never reach this piece of code.
  3360. */
  3361. pr_err("%s: Impossible, sk->sk_state=%d\n",
  3362. __func__, sk->sk_state);
  3363. break;
  3364. }
  3365. /* It _is_ possible, that we have something out-of-order _after_ FIN.
  3366. * Probably, we should reset in this case. For now drop them.
  3367. */
  3368. __skb_queue_purge(&tp->out_of_order_queue);
  3369. if (tcp_is_sack(tp))
  3370. tcp_sack_reset(&tp->rx_opt);
  3371. sk_mem_reclaim(sk);
  3372. if (!sock_flag(sk, SOCK_DEAD)) {
  3373. sk->sk_state_change(sk);
  3374. /* Do not send POLL_HUP for half duplex close. */
  3375. if (sk->sk_shutdown == SHUTDOWN_MASK ||
  3376. sk->sk_state == TCP_CLOSE)
  3377. sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
  3378. else
  3379. sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
  3380. }
  3381. }
  3382. static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
  3383. u32 end_seq)
  3384. {
  3385. if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
  3386. if (before(seq, sp->start_seq))
  3387. sp->start_seq = seq;
  3388. if (after(end_seq, sp->end_seq))
  3389. sp->end_seq = end_seq;
  3390. return true;
  3391. }
  3392. return false;
  3393. }
  3394. static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
  3395. {
  3396. struct tcp_sock *tp = tcp_sk(sk);
  3397. if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
  3398. int mib_idx;
  3399. if (before(seq, tp->rcv_nxt))
  3400. mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
  3401. else
  3402. mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
  3403. NET_INC_STATS_BH(sock_net(sk), mib_idx);
  3404. tp->rx_opt.dsack = 1;
  3405. tp->duplicate_sack[0].start_seq = seq;
  3406. tp->duplicate_sack[0].end_seq = end_seq;
  3407. }
  3408. }
  3409. static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
  3410. {
  3411. struct tcp_sock *tp = tcp_sk(sk);
  3412. if (!tp->rx_opt.dsack)
  3413. tcp_dsack_set(sk, seq, end_seq);
  3414. else
  3415. tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
  3416. }
  3417. static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
  3418. {
  3419. struct tcp_sock *tp = tcp_sk(sk);
  3420. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  3421. before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3422. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
  3423. tcp_enter_quickack_mode(sk);
  3424. if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
  3425. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  3426. if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
  3427. end_seq = tp->rcv_nxt;
  3428. tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
  3429. }
  3430. }
  3431. tcp_send_ack(sk);
  3432. }
  3433. /* These routines update the SACK block as out-of-order packets arrive or
  3434. * in-order packets close up the sequence space.
  3435. */
  3436. static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
  3437. {
  3438. int this_sack;
  3439. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3440. struct tcp_sack_block *swalk = sp + 1;
  3441. /* See if the recent change to the first SACK eats into
  3442. * or hits the sequence space of other SACK blocks, if so coalesce.
  3443. */
  3444. for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
  3445. if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
  3446. int i;
  3447. /* Zap SWALK, by moving every further SACK up by one slot.
  3448. * Decrease num_sacks.
  3449. */
  3450. tp->rx_opt.num_sacks--;
  3451. for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
  3452. sp[i] = sp[i + 1];
  3453. continue;
  3454. }
  3455. this_sack++, swalk++;
  3456. }
  3457. }
  3458. static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
  3459. {
  3460. struct tcp_sock *tp = tcp_sk(sk);
  3461. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3462. int cur_sacks = tp->rx_opt.num_sacks;
  3463. int this_sack;
  3464. if (!cur_sacks)
  3465. goto new_sack;
  3466. for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
  3467. if (tcp_sack_extend(sp, seq, end_seq)) {
  3468. /* Rotate this_sack to the first one. */
  3469. for (; this_sack > 0; this_sack--, sp--)
  3470. swap(*sp, *(sp - 1));
  3471. if (cur_sacks > 1)
  3472. tcp_sack_maybe_coalesce(tp);
  3473. return;
  3474. }
  3475. }
  3476. /* Could not find an adjacent existing SACK, build a new one,
  3477. * put it at the front, and shift everyone else down. We
  3478. * always know there is at least one SACK present already here.
  3479. *
  3480. * If the sack array is full, forget about the last one.
  3481. */
  3482. if (this_sack >= TCP_NUM_SACKS) {
  3483. this_sack--;
  3484. tp->rx_opt.num_sacks--;
  3485. sp--;
  3486. }
  3487. for (; this_sack > 0; this_sack--, sp--)
  3488. *sp = *(sp - 1);
  3489. new_sack:
  3490. /* Build the new head SACK, and we're done. */
  3491. sp->start_seq = seq;
  3492. sp->end_seq = end_seq;
  3493. tp->rx_opt.num_sacks++;
  3494. }
  3495. /* RCV.NXT advances, some SACKs should be eaten. */
  3496. static void tcp_sack_remove(struct tcp_sock *tp)
  3497. {
  3498. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3499. int num_sacks = tp->rx_opt.num_sacks;
  3500. int this_sack;
  3501. /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
  3502. if (skb_queue_empty(&tp->out_of_order_queue)) {
  3503. tp->rx_opt.num_sacks = 0;
  3504. return;
  3505. }
  3506. for (this_sack = 0; this_sack < num_sacks;) {
  3507. /* Check if the start of the sack is covered by RCV.NXT. */
  3508. if (!before(tp->rcv_nxt, sp->start_seq)) {
  3509. int i;
  3510. /* RCV.NXT must cover all the block! */
  3511. WARN_ON(before(tp->rcv_nxt, sp->end_seq));
  3512. /* Zap this SACK, by moving forward any other SACKS. */
  3513. for (i=this_sack+1; i < num_sacks; i++)
  3514. tp->selective_acks[i-1] = tp->selective_acks[i];
  3515. num_sacks--;
  3516. continue;
  3517. }
  3518. this_sack++;
  3519. sp++;
  3520. }
  3521. tp->rx_opt.num_sacks = num_sacks;
  3522. }
  3523. /* This one checks to see if we can put data from the
  3524. * out_of_order queue into the receive_queue.
  3525. */
  3526. static void tcp_ofo_queue(struct sock *sk)
  3527. {
  3528. struct tcp_sock *tp = tcp_sk(sk);
  3529. __u32 dsack_high = tp->rcv_nxt;
  3530. struct sk_buff *skb;
  3531. while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
  3532. if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  3533. break;
  3534. if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
  3535. __u32 dsack = dsack_high;
  3536. if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
  3537. dsack_high = TCP_SKB_CB(skb)->end_seq;
  3538. tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
  3539. }
  3540. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3541. SOCK_DEBUG(sk, "ofo packet was already received\n");
  3542. __skb_unlink(skb, &tp->out_of_order_queue);
  3543. __kfree_skb(skb);
  3544. continue;
  3545. }
  3546. SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
  3547. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3548. TCP_SKB_CB(skb)->end_seq);
  3549. __skb_unlink(skb, &tp->out_of_order_queue);
  3550. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3551. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3552. if (tcp_hdr(skb)->fin)
  3553. tcp_fin(sk);
  3554. }
  3555. }
  3556. static bool tcp_prune_ofo_queue(struct sock *sk);
  3557. static int tcp_prune_queue(struct sock *sk);
  3558. static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
  3559. unsigned int size)
  3560. {
  3561. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3562. !sk_rmem_schedule(sk, skb, size)) {
  3563. if (tcp_prune_queue(sk) < 0)
  3564. return -1;
  3565. if (!sk_rmem_schedule(sk, skb, size)) {
  3566. if (!tcp_prune_ofo_queue(sk))
  3567. return -1;
  3568. if (!sk_rmem_schedule(sk, skb, size))
  3569. return -1;
  3570. }
  3571. }
  3572. return 0;
  3573. }
  3574. /**
  3575. * tcp_try_coalesce - try to merge skb to prior one
  3576. * @sk: socket
  3577. * @to: prior buffer
  3578. * @from: buffer to add in queue
  3579. * @fragstolen: pointer to boolean
  3580. *
  3581. * Before queueing skb @from after @to, try to merge them
  3582. * to reduce overall memory use and queue lengths, if cost is small.
  3583. * Packets in ofo or receive queues can stay a long time.
  3584. * Better try to coalesce them right now to avoid future collapses.
  3585. * Returns true if caller should free @from instead of queueing it
  3586. */
  3587. static bool tcp_try_coalesce(struct sock *sk,
  3588. struct sk_buff *to,
  3589. struct sk_buff *from,
  3590. bool *fragstolen)
  3591. {
  3592. int delta;
  3593. *fragstolen = false;
  3594. if (tcp_hdr(from)->fin)
  3595. return false;
  3596. /* Its possible this segment overlaps with prior segment in queue */
  3597. if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
  3598. return false;
  3599. if (!skb_try_coalesce(to, from, fragstolen, &delta))
  3600. return false;
  3601. atomic_add(delta, &sk->sk_rmem_alloc);
  3602. sk_mem_charge(sk, delta);
  3603. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
  3604. TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
  3605. TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
  3606. return true;
  3607. }
  3608. static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
  3609. {
  3610. struct tcp_sock *tp = tcp_sk(sk);
  3611. struct sk_buff *skb1;
  3612. u32 seq, end_seq;
  3613. TCP_ECN_check_ce(tp, skb);
  3614. if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
  3615. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
  3616. __kfree_skb(skb);
  3617. return;
  3618. }
  3619. /* Disable header prediction. */
  3620. tp->pred_flags = 0;
  3621. inet_csk_schedule_ack(sk);
  3622. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
  3623. SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
  3624. tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3625. skb1 = skb_peek_tail(&tp->out_of_order_queue);
  3626. if (!skb1) {
  3627. /* Initial out of order segment, build 1 SACK. */
  3628. if (tcp_is_sack(tp)) {
  3629. tp->rx_opt.num_sacks = 1;
  3630. tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
  3631. tp->selective_acks[0].end_seq =
  3632. TCP_SKB_CB(skb)->end_seq;
  3633. }
  3634. __skb_queue_head(&tp->out_of_order_queue, skb);
  3635. goto end;
  3636. }
  3637. seq = TCP_SKB_CB(skb)->seq;
  3638. end_seq = TCP_SKB_CB(skb)->end_seq;
  3639. if (seq == TCP_SKB_CB(skb1)->end_seq) {
  3640. bool fragstolen;
  3641. if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
  3642. __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
  3643. } else {
  3644. tcp_grow_window(sk, skb);
  3645. kfree_skb_partial(skb, fragstolen);
  3646. skb = NULL;
  3647. }
  3648. if (!tp->rx_opt.num_sacks ||
  3649. tp->selective_acks[0].end_seq != seq)
  3650. goto add_sack;
  3651. /* Common case: data arrive in order after hole. */
  3652. tp->selective_acks[0].end_seq = end_seq;
  3653. goto end;
  3654. }
  3655. /* Find place to insert this segment. */
  3656. while (1) {
  3657. if (!after(TCP_SKB_CB(skb1)->seq, seq))
  3658. break;
  3659. if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
  3660. skb1 = NULL;
  3661. break;
  3662. }
  3663. skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
  3664. }
  3665. /* Do skb overlap to previous one? */
  3666. if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
  3667. if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3668. /* All the bits are present. Drop. */
  3669. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
  3670. __kfree_skb(skb);
  3671. skb = NULL;
  3672. tcp_dsack_set(sk, seq, end_seq);
  3673. goto add_sack;
  3674. }
  3675. if (after(seq, TCP_SKB_CB(skb1)->seq)) {
  3676. /* Partial overlap. */
  3677. tcp_dsack_set(sk, seq,
  3678. TCP_SKB_CB(skb1)->end_seq);
  3679. } else {
  3680. if (skb_queue_is_first(&tp->out_of_order_queue,
  3681. skb1))
  3682. skb1 = NULL;
  3683. else
  3684. skb1 = skb_queue_prev(
  3685. &tp->out_of_order_queue,
  3686. skb1);
  3687. }
  3688. }
  3689. if (!skb1)
  3690. __skb_queue_head(&tp->out_of_order_queue, skb);
  3691. else
  3692. __skb_queue_after(&tp->out_of_order_queue, skb1, skb);
  3693. /* And clean segments covered by new one as whole. */
  3694. while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
  3695. skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
  3696. if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
  3697. break;
  3698. if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3699. tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
  3700. end_seq);
  3701. break;
  3702. }
  3703. __skb_unlink(skb1, &tp->out_of_order_queue);
  3704. tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
  3705. TCP_SKB_CB(skb1)->end_seq);
  3706. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
  3707. __kfree_skb(skb1);
  3708. }
  3709. add_sack:
  3710. if (tcp_is_sack(tp))
  3711. tcp_sack_new_ofo_skb(sk, seq, end_seq);
  3712. end:
  3713. if (skb) {
  3714. tcp_grow_window(sk, skb);
  3715. skb_set_owner_r(skb, sk);
  3716. }
  3717. }
  3718. static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
  3719. bool *fragstolen)
  3720. {
  3721. int eaten;
  3722. struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
  3723. __skb_pull(skb, hdrlen);
  3724. eaten = (tail &&
  3725. tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
  3726. tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3727. if (!eaten) {
  3728. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3729. skb_set_owner_r(skb, sk);
  3730. }
  3731. return eaten;
  3732. }
  3733. int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
  3734. {
  3735. struct sk_buff *skb = NULL;
  3736. struct tcphdr *th;
  3737. bool fragstolen;
  3738. if (size == 0)
  3739. return 0;
  3740. skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
  3741. if (!skb)
  3742. goto err;
  3743. if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
  3744. goto err_free;
  3745. th = (struct tcphdr *)skb_put(skb, sizeof(*th));
  3746. skb_reset_transport_header(skb);
  3747. memset(th, 0, sizeof(*th));
  3748. if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
  3749. goto err_free;
  3750. TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
  3751. TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
  3752. TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
  3753. if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
  3754. WARN_ON_ONCE(fragstolen); /* should not happen */
  3755. __kfree_skb(skb);
  3756. }
  3757. return size;
  3758. err_free:
  3759. kfree_skb(skb);
  3760. err:
  3761. return -ENOMEM;
  3762. }
  3763. static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
  3764. {
  3765. const struct tcphdr *th = tcp_hdr(skb);
  3766. struct tcp_sock *tp = tcp_sk(sk);
  3767. int eaten = -1;
  3768. bool fragstolen = false;
  3769. if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
  3770. goto drop;
  3771. skb_dst_drop(skb);
  3772. __skb_pull(skb, th->doff * 4);
  3773. TCP_ECN_accept_cwr(tp, skb);
  3774. tp->rx_opt.dsack = 0;
  3775. /* Queue data for delivery to the user.
  3776. * Packets in sequence go to the receive queue.
  3777. * Out of sequence packets to the out_of_order_queue.
  3778. */
  3779. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3780. if (tcp_receive_window(tp) == 0)
  3781. goto out_of_window;
  3782. /* Ok. In sequence. In window. */
  3783. if (tp->ucopy.task == current &&
  3784. tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
  3785. sock_owned_by_user(sk) && !tp->urg_data) {
  3786. int chunk = min_t(unsigned int, skb->len,
  3787. tp->ucopy.len);
  3788. __set_current_state(TASK_RUNNING);
  3789. local_bh_enable();
  3790. if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
  3791. tp->ucopy.len -= chunk;
  3792. tp->copied_seq += chunk;
  3793. eaten = (chunk == skb->len);
  3794. tcp_rcv_space_adjust(sk);
  3795. }
  3796. local_bh_disable();
  3797. }
  3798. if (eaten <= 0) {
  3799. queue_and_out:
  3800. if (eaten < 0 &&
  3801. tcp_try_rmem_schedule(sk, skb, skb->truesize))
  3802. goto drop;
  3803. eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
  3804. }
  3805. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3806. if (skb->len)
  3807. tcp_event_data_recv(sk, skb);
  3808. if (th->fin)
  3809. tcp_fin(sk);
  3810. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3811. tcp_ofo_queue(sk);
  3812. /* RFC2581. 4.2. SHOULD send immediate ACK, when
  3813. * gap in queue is filled.
  3814. */
  3815. if (skb_queue_empty(&tp->out_of_order_queue))
  3816. inet_csk(sk)->icsk_ack.pingpong = 0;
  3817. }
  3818. if (tp->rx_opt.num_sacks)
  3819. tcp_sack_remove(tp);
  3820. tcp_fast_path_check(sk);
  3821. if (eaten > 0)
  3822. kfree_skb_partial(skb, fragstolen);
  3823. if (!sock_flag(sk, SOCK_DEAD))
  3824. sk->sk_data_ready(sk, 0);
  3825. return;
  3826. }
  3827. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3828. /* A retransmit, 2nd most common case. Force an immediate ack. */
  3829. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
  3830. tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3831. out_of_window:
  3832. tcp_enter_quickack_mode(sk);
  3833. inet_csk_schedule_ack(sk);
  3834. drop:
  3835. __kfree_skb(skb);
  3836. return;
  3837. }
  3838. /* Out of window. F.e. zero window probe. */
  3839. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
  3840. goto out_of_window;
  3841. tcp_enter_quickack_mode(sk);
  3842. if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3843. /* Partial packet, seq < rcv_next < end_seq */
  3844. SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
  3845. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3846. TCP_SKB_CB(skb)->end_seq);
  3847. tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
  3848. /* If window is closed, drop tail of packet. But after
  3849. * remembering D-SACK for its head made in previous line.
  3850. */
  3851. if (!tcp_receive_window(tp))
  3852. goto out_of_window;
  3853. goto queue_and_out;
  3854. }
  3855. tcp_data_queue_ofo(sk, skb);
  3856. }
  3857. static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
  3858. struct sk_buff_head *list)
  3859. {
  3860. struct sk_buff *next = NULL;
  3861. if (!skb_queue_is_last(list, skb))
  3862. next = skb_queue_next(list, skb);
  3863. __skb_unlink(skb, list);
  3864. __kfree_skb(skb);
  3865. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
  3866. return next;
  3867. }
  3868. /* Collapse contiguous sequence of skbs head..tail with
  3869. * sequence numbers start..end.
  3870. *
  3871. * If tail is NULL, this means until the end of the list.
  3872. *
  3873. * Segments with FIN/SYN are not collapsed (only because this
  3874. * simplifies code)
  3875. */
  3876. static void
  3877. tcp_collapse(struct sock *sk, struct sk_buff_head *list,
  3878. struct sk_buff *head, struct sk_buff *tail,
  3879. u32 start, u32 end)
  3880. {
  3881. struct sk_buff *skb, *n;
  3882. bool end_of_skbs;
  3883. /* First, check that queue is collapsible and find
  3884. * the point where collapsing can be useful. */
  3885. skb = head;
  3886. restart:
  3887. end_of_skbs = true;
  3888. skb_queue_walk_from_safe(list, skb, n) {
  3889. if (skb == tail)
  3890. break;
  3891. /* No new bits? It is possible on ofo queue. */
  3892. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3893. skb = tcp_collapse_one(sk, skb, list);
  3894. if (!skb)
  3895. break;
  3896. goto restart;
  3897. }
  3898. /* The first skb to collapse is:
  3899. * - not SYN/FIN and
  3900. * - bloated or contains data before "start" or
  3901. * overlaps to the next one.
  3902. */
  3903. if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
  3904. (tcp_win_from_space(skb->truesize) > skb->len ||
  3905. before(TCP_SKB_CB(skb)->seq, start))) {
  3906. end_of_skbs = false;
  3907. break;
  3908. }
  3909. if (!skb_queue_is_last(list, skb)) {
  3910. struct sk_buff *next = skb_queue_next(list, skb);
  3911. if (next != tail &&
  3912. TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
  3913. end_of_skbs = false;
  3914. break;
  3915. }
  3916. }
  3917. /* Decided to skip this, advance start seq. */
  3918. start = TCP_SKB_CB(skb)->end_seq;
  3919. }
  3920. if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
  3921. return;
  3922. while (before(start, end)) {
  3923. struct sk_buff *nskb;
  3924. unsigned int header = skb_headroom(skb);
  3925. int copy = SKB_MAX_ORDER(header, 0);
  3926. /* Too big header? This can happen with IPv6. */
  3927. if (copy < 0)
  3928. return;
  3929. if (end - start < copy)
  3930. copy = end - start;
  3931. nskb = alloc_skb(copy + header, GFP_ATOMIC);
  3932. if (!nskb)
  3933. return;
  3934. skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
  3935. skb_set_network_header(nskb, (skb_network_header(skb) -
  3936. skb->head));
  3937. skb_set_transport_header(nskb, (skb_transport_header(skb) -
  3938. skb->head));
  3939. skb_reserve(nskb, header);
  3940. memcpy(nskb->head, skb->head, header);
  3941. memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
  3942. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
  3943. __skb_queue_before(list, skb, nskb);
  3944. skb_set_owner_r(nskb, sk);
  3945. /* Copy data, releasing collapsed skbs. */
  3946. while (copy > 0) {
  3947. int offset = start - TCP_SKB_CB(skb)->seq;
  3948. int size = TCP_SKB_CB(skb)->end_seq - start;
  3949. BUG_ON(offset < 0);
  3950. if (size > 0) {
  3951. size = min(copy, size);
  3952. if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
  3953. BUG();
  3954. TCP_SKB_CB(nskb)->end_seq += size;
  3955. copy -= size;
  3956. start += size;
  3957. }
  3958. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3959. skb = tcp_collapse_one(sk, skb, list);
  3960. if (!skb ||
  3961. skb == tail ||
  3962. tcp_hdr(skb)->syn ||
  3963. tcp_hdr(skb)->fin)
  3964. return;
  3965. }
  3966. }
  3967. }
  3968. }
  3969. /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
  3970. * and tcp_collapse() them until all the queue is collapsed.
  3971. */
  3972. static void tcp_collapse_ofo_queue(struct sock *sk)
  3973. {
  3974. struct tcp_sock *tp = tcp_sk(sk);
  3975. struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
  3976. struct sk_buff *head;
  3977. u32 start, end;
  3978. if (skb == NULL)
  3979. return;
  3980. start = TCP_SKB_CB(skb)->seq;
  3981. end = TCP_SKB_CB(skb)->end_seq;
  3982. head = skb;
  3983. for (;;) {
  3984. struct sk_buff *next = NULL;
  3985. if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
  3986. next = skb_queue_next(&tp->out_of_order_queue, skb);
  3987. skb = next;
  3988. /* Segment is terminated when we see gap or when
  3989. * we are at the end of all the queue. */
  3990. if (!skb ||
  3991. after(TCP_SKB_CB(skb)->seq, end) ||
  3992. before(TCP_SKB_CB(skb)->end_seq, start)) {
  3993. tcp_collapse(sk, &tp->out_of_order_queue,
  3994. head, skb, start, end);
  3995. head = skb;
  3996. if (!skb)
  3997. break;
  3998. /* Start new segment */
  3999. start = TCP_SKB_CB(skb)->seq;
  4000. end = TCP_SKB_CB(skb)->end_seq;
  4001. } else {
  4002. if (before(TCP_SKB_CB(skb)->seq, start))
  4003. start = TCP_SKB_CB(skb)->seq;
  4004. if (after(TCP_SKB_CB(skb)->end_seq, end))
  4005. end = TCP_SKB_CB(skb)->end_seq;
  4006. }
  4007. }
  4008. }
  4009. /*
  4010. * Purge the out-of-order queue.
  4011. * Return true if queue was pruned.
  4012. */
  4013. static bool tcp_prune_ofo_queue(struct sock *sk)
  4014. {
  4015. struct tcp_sock *tp = tcp_sk(sk);
  4016. bool res = false;
  4017. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  4018. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
  4019. __skb_queue_purge(&tp->out_of_order_queue);
  4020. /* Reset SACK state. A conforming SACK implementation will
  4021. * do the same at a timeout based retransmit. When a connection
  4022. * is in a sad state like this, we care only about integrity
  4023. * of the connection not performance.
  4024. */
  4025. if (tp->rx_opt.sack_ok)
  4026. tcp_sack_reset(&tp->rx_opt);
  4027. sk_mem_reclaim(sk);
  4028. res = true;
  4029. }
  4030. return res;
  4031. }
  4032. /* Reduce allocated memory if we can, trying to get
  4033. * the socket within its memory limits again.
  4034. *
  4035. * Return less than zero if we should start dropping frames
  4036. * until the socket owning process reads some of the data
  4037. * to stabilize the situation.
  4038. */
  4039. static int tcp_prune_queue(struct sock *sk)
  4040. {
  4041. struct tcp_sock *tp = tcp_sk(sk);
  4042. SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
  4043. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
  4044. if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
  4045. tcp_clamp_window(sk);
  4046. else if (sk_under_memory_pressure(sk))
  4047. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
  4048. tcp_collapse_ofo_queue(sk);
  4049. if (!skb_queue_empty(&sk->sk_receive_queue))
  4050. tcp_collapse(sk, &sk->sk_receive_queue,
  4051. skb_peek(&sk->sk_receive_queue),
  4052. NULL,
  4053. tp->copied_seq, tp->rcv_nxt);
  4054. sk_mem_reclaim(sk);
  4055. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  4056. return 0;
  4057. /* Collapsing did not help, destructive actions follow.
  4058. * This must not ever occur. */
  4059. tcp_prune_ofo_queue(sk);
  4060. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  4061. return 0;
  4062. /* If we are really being abused, tell the caller to silently
  4063. * drop receive data on the floor. It will get retransmitted
  4064. * and hopefully then we'll have sufficient space.
  4065. */
  4066. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
  4067. /* Massive buffer overcommit. */
  4068. tp->pred_flags = 0;
  4069. return -1;
  4070. }
  4071. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  4072. * As additional protections, we do not touch cwnd in retransmission phases,
  4073. * and if application hit its sndbuf limit recently.
  4074. */
  4075. void tcp_cwnd_application_limited(struct sock *sk)
  4076. {
  4077. struct tcp_sock *tp = tcp_sk(sk);
  4078. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  4079. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  4080. /* Limited by application or receiver window. */
  4081. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  4082. u32 win_used = max(tp->snd_cwnd_used, init_win);
  4083. if (win_used < tp->snd_cwnd) {
  4084. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  4085. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  4086. }
  4087. tp->snd_cwnd_used = 0;
  4088. }
  4089. tp->snd_cwnd_stamp = tcp_time_stamp;
  4090. }
  4091. static bool tcp_should_expand_sndbuf(const struct sock *sk)
  4092. {
  4093. const struct tcp_sock *tp = tcp_sk(sk);
  4094. /* If the user specified a specific send buffer setting, do
  4095. * not modify it.
  4096. */
  4097. if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
  4098. return false;
  4099. /* If we are under global TCP memory pressure, do not expand. */
  4100. if (sk_under_memory_pressure(sk))
  4101. return false;
  4102. /* If we are under soft global TCP memory pressure, do not expand. */
  4103. if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
  4104. return false;
  4105. /* If we filled the congestion window, do not expand. */
  4106. if (tp->packets_out >= tp->snd_cwnd)
  4107. return false;
  4108. return true;
  4109. }
  4110. /* When incoming ACK allowed to free some skb from write_queue,
  4111. * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
  4112. * on the exit from tcp input handler.
  4113. *
  4114. * PROBLEM: sndbuf expansion does not work well with largesend.
  4115. */
  4116. static void tcp_new_space(struct sock *sk)
  4117. {
  4118. struct tcp_sock *tp = tcp_sk(sk);
  4119. if (tcp_should_expand_sndbuf(sk)) {
  4120. tcp_sndbuf_expand(sk);
  4121. tp->snd_cwnd_stamp = tcp_time_stamp;
  4122. }
  4123. sk->sk_write_space(sk);
  4124. }
  4125. static void tcp_check_space(struct sock *sk)
  4126. {
  4127. if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
  4128. sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
  4129. if (sk->sk_socket &&
  4130. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
  4131. tcp_new_space(sk);
  4132. }
  4133. }
  4134. static inline void tcp_data_snd_check(struct sock *sk)
  4135. {
  4136. tcp_push_pending_frames(sk);
  4137. tcp_check_space(sk);
  4138. }
  4139. /*
  4140. * Check if sending an ack is needed.
  4141. */
  4142. static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
  4143. {
  4144. struct tcp_sock *tp = tcp_sk(sk);
  4145. /* More than one full frame received... */
  4146. if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
  4147. /* ... and right edge of window advances far enough.
  4148. * (tcp_recvmsg() will send ACK otherwise). Or...
  4149. */
  4150. __tcp_select_window(sk) >= tp->rcv_wnd) ||
  4151. /* We ACK each frame or... */
  4152. tcp_in_quickack_mode(sk) ||
  4153. /* We have out of order data. */
  4154. (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
  4155. /* Then ack it now */
  4156. tcp_send_ack(sk);
  4157. } else {
  4158. /* Else, send delayed ack. */
  4159. tcp_send_delayed_ack(sk);
  4160. }
  4161. }
  4162. static inline void tcp_ack_snd_check(struct sock *sk)
  4163. {
  4164. if (!inet_csk_ack_scheduled(sk)) {
  4165. /* We sent a data segment already. */
  4166. return;
  4167. }
  4168. __tcp_ack_snd_check(sk, 1);
  4169. }
  4170. /*
  4171. * This routine is only called when we have urgent data
  4172. * signaled. Its the 'slow' part of tcp_urg. It could be
  4173. * moved inline now as tcp_urg is only called from one
  4174. * place. We handle URGent data wrong. We have to - as
  4175. * BSD still doesn't use the correction from RFC961.
  4176. * For 1003.1g we should support a new option TCP_STDURG to permit
  4177. * either form (or just set the sysctl tcp_stdurg).
  4178. */
  4179. static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
  4180. {
  4181. struct tcp_sock *tp = tcp_sk(sk);
  4182. u32 ptr = ntohs(th->urg_ptr);
  4183. if (ptr && !sysctl_tcp_stdurg)
  4184. ptr--;
  4185. ptr += ntohl(th->seq);
  4186. /* Ignore urgent data that we've already seen and read. */
  4187. if (after(tp->copied_seq, ptr))
  4188. return;
  4189. /* Do not replay urg ptr.
  4190. *
  4191. * NOTE: interesting situation not covered by specs.
  4192. * Misbehaving sender may send urg ptr, pointing to segment,
  4193. * which we already have in ofo queue. We are not able to fetch
  4194. * such data and will stay in TCP_URG_NOTYET until will be eaten
  4195. * by recvmsg(). Seems, we are not obliged to handle such wicked
  4196. * situations. But it is worth to think about possibility of some
  4197. * DoSes using some hypothetical application level deadlock.
  4198. */
  4199. if (before(ptr, tp->rcv_nxt))
  4200. return;
  4201. /* Do we already have a newer (or duplicate) urgent pointer? */
  4202. if (tp->urg_data && !after(ptr, tp->urg_seq))
  4203. return;
  4204. /* Tell the world about our new urgent pointer. */
  4205. sk_send_sigurg(sk);
  4206. /* We may be adding urgent data when the last byte read was
  4207. * urgent. To do this requires some care. We cannot just ignore
  4208. * tp->copied_seq since we would read the last urgent byte again
  4209. * as data, nor can we alter copied_seq until this data arrives
  4210. * or we break the semantics of SIOCATMARK (and thus sockatmark())
  4211. *
  4212. * NOTE. Double Dutch. Rendering to plain English: author of comment
  4213. * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
  4214. * and expect that both A and B disappear from stream. This is _wrong_.
  4215. * Though this happens in BSD with high probability, this is occasional.
  4216. * Any application relying on this is buggy. Note also, that fix "works"
  4217. * only in this artificial test. Insert some normal data between A and B and we will
  4218. * decline of BSD again. Verdict: it is better to remove to trap
  4219. * buggy users.
  4220. */
  4221. if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
  4222. !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
  4223. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  4224. tp->copied_seq++;
  4225. if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
  4226. __skb_unlink(skb, &sk->sk_receive_queue);
  4227. __kfree_skb(skb);
  4228. }
  4229. }
  4230. tp->urg_data = TCP_URG_NOTYET;
  4231. tp->urg_seq = ptr;
  4232. /* Disable header prediction. */
  4233. tp->pred_flags = 0;
  4234. }
  4235. /* This is the 'fast' part of urgent handling. */
  4236. static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
  4237. {
  4238. struct tcp_sock *tp = tcp_sk(sk);
  4239. /* Check if we get a new urgent pointer - normally not. */
  4240. if (th->urg)
  4241. tcp_check_urg(sk, th);
  4242. /* Do we wait for any urgent data? - normally not... */
  4243. if (tp->urg_data == TCP_URG_NOTYET) {
  4244. u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
  4245. th->syn;
  4246. /* Is the urgent pointer pointing into this packet? */
  4247. if (ptr < skb->len) {
  4248. u8 tmp;
  4249. if (skb_copy_bits(skb, ptr, &tmp, 1))
  4250. BUG();
  4251. tp->urg_data = TCP_URG_VALID | tmp;
  4252. if (!sock_flag(sk, SOCK_DEAD))
  4253. sk->sk_data_ready(sk, 0);
  4254. }
  4255. }
  4256. }
  4257. static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
  4258. {
  4259. struct tcp_sock *tp = tcp_sk(sk);
  4260. int chunk = skb->len - hlen;
  4261. int err;
  4262. local_bh_enable();
  4263. if (skb_csum_unnecessary(skb))
  4264. err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
  4265. else
  4266. err = skb_copy_and_csum_datagram_iovec(skb, hlen,
  4267. tp->ucopy.iov);
  4268. if (!err) {
  4269. tp->ucopy.len -= chunk;
  4270. tp->copied_seq += chunk;
  4271. tcp_rcv_space_adjust(sk);
  4272. }
  4273. local_bh_disable();
  4274. return err;
  4275. }
  4276. static __sum16 __tcp_checksum_complete_user(struct sock *sk,
  4277. struct sk_buff *skb)
  4278. {
  4279. __sum16 result;
  4280. if (sock_owned_by_user(sk)) {
  4281. local_bh_enable();
  4282. result = __tcp_checksum_complete(skb);
  4283. local_bh_disable();
  4284. } else {
  4285. result = __tcp_checksum_complete(skb);
  4286. }
  4287. return result;
  4288. }
  4289. static inline bool tcp_checksum_complete_user(struct sock *sk,
  4290. struct sk_buff *skb)
  4291. {
  4292. return !skb_csum_unnecessary(skb) &&
  4293. __tcp_checksum_complete_user(sk, skb);
  4294. }
  4295. #ifdef CONFIG_NET_DMA
  4296. static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
  4297. int hlen)
  4298. {
  4299. struct tcp_sock *tp = tcp_sk(sk);
  4300. int chunk = skb->len - hlen;
  4301. int dma_cookie;
  4302. bool copied_early = false;
  4303. if (tp->ucopy.wakeup)
  4304. return false;
  4305. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  4306. tp->ucopy.dma_chan = net_dma_find_channel();
  4307. if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
  4308. dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
  4309. skb, hlen,
  4310. tp->ucopy.iov, chunk,
  4311. tp->ucopy.pinned_list);
  4312. if (dma_cookie < 0)
  4313. goto out;
  4314. tp->ucopy.dma_cookie = dma_cookie;
  4315. copied_early = true;
  4316. tp->ucopy.len -= chunk;
  4317. tp->copied_seq += chunk;
  4318. tcp_rcv_space_adjust(sk);
  4319. if ((tp->ucopy.len == 0) ||
  4320. (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
  4321. (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
  4322. tp->ucopy.wakeup = 1;
  4323. sk->sk_data_ready(sk, 0);
  4324. }
  4325. } else if (chunk > 0) {
  4326. tp->ucopy.wakeup = 1;
  4327. sk->sk_data_ready(sk, 0);
  4328. }
  4329. out:
  4330. return copied_early;
  4331. }
  4332. #endif /* CONFIG_NET_DMA */
  4333. /* Does PAWS and seqno based validation of an incoming segment, flags will
  4334. * play significant role here.
  4335. */
  4336. static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
  4337. const struct tcphdr *th, int syn_inerr)
  4338. {
  4339. struct tcp_sock *tp = tcp_sk(sk);
  4340. /* RFC1323: H1. Apply PAWS check first. */
  4341. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  4342. tcp_paws_discard(sk, skb)) {
  4343. if (!th->rst) {
  4344. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
  4345. tcp_send_dupack(sk, skb);
  4346. goto discard;
  4347. }
  4348. /* Reset is accepted even if it did not pass PAWS. */
  4349. }
  4350. /* Step 1: check sequence number */
  4351. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  4352. /* RFC793, page 37: "In all states except SYN-SENT, all reset
  4353. * (RST) segments are validated by checking their SEQ-fields."
  4354. * And page 69: "If an incoming segment is not acceptable,
  4355. * an acknowledgment should be sent in reply (unless the RST
  4356. * bit is set, if so drop the segment and return)".
  4357. */
  4358. if (!th->rst) {
  4359. if (th->syn)
  4360. goto syn_challenge;
  4361. tcp_send_dupack(sk, skb);
  4362. }
  4363. goto discard;
  4364. }
  4365. /* Step 2: check RST bit */
  4366. if (th->rst) {
  4367. /* RFC 5961 3.2 :
  4368. * If sequence number exactly matches RCV.NXT, then
  4369. * RESET the connection
  4370. * else
  4371. * Send a challenge ACK
  4372. */
  4373. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
  4374. tcp_reset(sk);
  4375. else
  4376. tcp_send_challenge_ack(sk);
  4377. goto discard;
  4378. }
  4379. /* step 3: check security and precedence [ignored] */
  4380. /* step 4: Check for a SYN
  4381. * RFC 5691 4.2 : Send a challenge ack
  4382. */
  4383. if (th->syn) {
  4384. syn_challenge:
  4385. if (syn_inerr)
  4386. TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
  4387. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
  4388. tcp_send_challenge_ack(sk);
  4389. goto discard;
  4390. }
  4391. return true;
  4392. discard:
  4393. __kfree_skb(skb);
  4394. return false;
  4395. }
  4396. /*
  4397. * TCP receive function for the ESTABLISHED state.
  4398. *
  4399. * It is split into a fast path and a slow path. The fast path is
  4400. * disabled when:
  4401. * - A zero window was announced from us - zero window probing
  4402. * is only handled properly in the slow path.
  4403. * - Out of order segments arrived.
  4404. * - Urgent data is expected.
  4405. * - There is no buffer space left
  4406. * - Unexpected TCP flags/window values/header lengths are received
  4407. * (detected by checking the TCP header against pred_flags)
  4408. * - Data is sent in both directions. Fast path only supports pure senders
  4409. * or pure receivers (this means either the sequence number or the ack
  4410. * value must stay constant)
  4411. * - Unexpected TCP option.
  4412. *
  4413. * When these conditions are not satisfied it drops into a standard
  4414. * receive procedure patterned after RFC793 to handle all cases.
  4415. * The first three cases are guaranteed by proper pred_flags setting,
  4416. * the rest is checked inline. Fast processing is turned on in
  4417. * tcp_data_queue when everything is OK.
  4418. */
  4419. void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  4420. const struct tcphdr *th, unsigned int len)
  4421. {
  4422. struct tcp_sock *tp = tcp_sk(sk);
  4423. if (unlikely(sk->sk_rx_dst == NULL))
  4424. inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
  4425. /*
  4426. * Header prediction.
  4427. * The code loosely follows the one in the famous
  4428. * "30 instruction TCP receive" Van Jacobson mail.
  4429. *
  4430. * Van's trick is to deposit buffers into socket queue
  4431. * on a device interrupt, to call tcp_recv function
  4432. * on the receive process context and checksum and copy
  4433. * the buffer to user space. smart...
  4434. *
  4435. * Our current scheme is not silly either but we take the
  4436. * extra cost of the net_bh soft interrupt processing...
  4437. * We do checksum and copy also but from device to kernel.
  4438. */
  4439. tp->rx_opt.saw_tstamp = 0;
  4440. /* pred_flags is 0xS?10 << 16 + snd_wnd
  4441. * if header_prediction is to be made
  4442. * 'S' will always be tp->tcp_header_len >> 2
  4443. * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
  4444. * turn it off (when there are holes in the receive
  4445. * space for instance)
  4446. * PSH flag is ignored.
  4447. */
  4448. if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
  4449. TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
  4450. !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
  4451. int tcp_header_len = tp->tcp_header_len;
  4452. /* Timestamp header prediction: tcp_header_len
  4453. * is automatically equal to th->doff*4 due to pred_flags
  4454. * match.
  4455. */
  4456. /* Check timestamp */
  4457. if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
  4458. /* No? Slow path! */
  4459. if (!tcp_parse_aligned_timestamp(tp, th))
  4460. goto slow_path;
  4461. /* If PAWS failed, check it more carefully in slow path */
  4462. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
  4463. goto slow_path;
  4464. /* DO NOT update ts_recent here, if checksum fails
  4465. * and timestamp was corrupted part, it will result
  4466. * in a hung connection since we will drop all
  4467. * future packets due to the PAWS test.
  4468. */
  4469. }
  4470. if (len <= tcp_header_len) {
  4471. /* Bulk data transfer: sender */
  4472. if (len == tcp_header_len) {
  4473. /* Predicted packet is in window by definition.
  4474. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  4475. * Hence, check seq<=rcv_wup reduces to:
  4476. */
  4477. if (tcp_header_len ==
  4478. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  4479. tp->rcv_nxt == tp->rcv_wup)
  4480. tcp_store_ts_recent(tp);
  4481. /* We know that such packets are checksummed
  4482. * on entry.
  4483. */
  4484. tcp_ack(sk, skb, 0);
  4485. __kfree_skb(skb);
  4486. tcp_data_snd_check(sk);
  4487. return;
  4488. } else { /* Header too small */
  4489. TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
  4490. goto discard;
  4491. }
  4492. } else {
  4493. int eaten = 0;
  4494. int copied_early = 0;
  4495. bool fragstolen = false;
  4496. if (tp->copied_seq == tp->rcv_nxt &&
  4497. len - tcp_header_len <= tp->ucopy.len) {
  4498. #ifdef CONFIG_NET_DMA
  4499. if (tp->ucopy.task == current &&
  4500. sock_owned_by_user(sk) &&
  4501. tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
  4502. copied_early = 1;
  4503. eaten = 1;
  4504. }
  4505. #endif
  4506. if (tp->ucopy.task == current &&
  4507. sock_owned_by_user(sk) && !copied_early) {
  4508. __set_current_state(TASK_RUNNING);
  4509. if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
  4510. eaten = 1;
  4511. }
  4512. if (eaten) {
  4513. /* Predicted packet is in window by definition.
  4514. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  4515. * Hence, check seq<=rcv_wup reduces to:
  4516. */
  4517. if (tcp_header_len ==
  4518. (sizeof(struct tcphdr) +
  4519. TCPOLEN_TSTAMP_ALIGNED) &&
  4520. tp->rcv_nxt == tp->rcv_wup)
  4521. tcp_store_ts_recent(tp);
  4522. tcp_rcv_rtt_measure_ts(sk, skb);
  4523. __skb_pull(skb, tcp_header_len);
  4524. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  4525. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
  4526. }
  4527. if (copied_early)
  4528. tcp_cleanup_rbuf(sk, skb->len);
  4529. }
  4530. if (!eaten) {
  4531. if (tcp_checksum_complete_user(sk, skb))
  4532. goto csum_error;
  4533. if ((int)skb->truesize > sk->sk_forward_alloc)
  4534. goto step5;
  4535. /* Predicted packet is in window by definition.
  4536. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  4537. * Hence, check seq<=rcv_wup reduces to:
  4538. */
  4539. if (tcp_header_len ==
  4540. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  4541. tp->rcv_nxt == tp->rcv_wup)
  4542. tcp_store_ts_recent(tp);
  4543. tcp_rcv_rtt_measure_ts(sk, skb);
  4544. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
  4545. /* Bulk data transfer: receiver */
  4546. eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
  4547. &fragstolen);
  4548. }
  4549. tcp_event_data_recv(sk, skb);
  4550. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
  4551. /* Well, only one small jumplet in fast path... */
  4552. tcp_ack(sk, skb, FLAG_DATA);
  4553. tcp_data_snd_check(sk);
  4554. if (!inet_csk_ack_scheduled(sk))
  4555. goto no_ack;
  4556. }
  4557. if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
  4558. __tcp_ack_snd_check(sk, 0);
  4559. no_ack:
  4560. #ifdef CONFIG_NET_DMA
  4561. if (copied_early)
  4562. __skb_queue_tail(&sk->sk_async_wait_queue, skb);
  4563. else
  4564. #endif
  4565. if (eaten)
  4566. kfree_skb_partial(skb, fragstolen);
  4567. sk->sk_data_ready(sk, 0);
  4568. return;
  4569. }
  4570. }
  4571. slow_path:
  4572. if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
  4573. goto csum_error;
  4574. if (!th->ack && !th->rst)
  4575. goto discard;
  4576. /*
  4577. * Standard slow path.
  4578. */
  4579. if (!tcp_validate_incoming(sk, skb, th, 1))
  4580. return;
  4581. step5:
  4582. if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
  4583. goto discard;
  4584. tcp_rcv_rtt_measure_ts(sk, skb);
  4585. /* Process urgent data. */
  4586. tcp_urg(sk, skb, th);
  4587. /* step 7: process the segment text */
  4588. tcp_data_queue(sk, skb);
  4589. tcp_data_snd_check(sk);
  4590. tcp_ack_snd_check(sk);
  4591. return;
  4592. csum_error:
  4593. TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
  4594. TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
  4595. discard:
  4596. __kfree_skb(skb);
  4597. }
  4598. EXPORT_SYMBOL(tcp_rcv_established);
  4599. void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
  4600. {
  4601. struct tcp_sock *tp = tcp_sk(sk);
  4602. struct inet_connection_sock *icsk = inet_csk(sk);
  4603. tcp_set_state(sk, TCP_ESTABLISHED);
  4604. if (skb != NULL) {
  4605. icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
  4606. security_inet_conn_established(sk, skb);
  4607. }
  4608. /* Make sure socket is routed, for correct metrics. */
  4609. icsk->icsk_af_ops->rebuild_header(sk);
  4610. tcp_init_metrics(sk);
  4611. tcp_init_congestion_control(sk);
  4612. /* Prevent spurious tcp_cwnd_restart() on first data
  4613. * packet.
  4614. */
  4615. tp->lsndtime = tcp_time_stamp;
  4616. tcp_init_buffer_space(sk);
  4617. if (sock_flag(sk, SOCK_KEEPOPEN))
  4618. inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
  4619. if (!tp->rx_opt.snd_wscale)
  4620. __tcp_fast_path_on(tp, tp->snd_wnd);
  4621. else
  4622. tp->pred_flags = 0;
  4623. if (!sock_flag(sk, SOCK_DEAD)) {
  4624. sk->sk_state_change(sk);
  4625. sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
  4626. }
  4627. }
  4628. static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
  4629. struct tcp_fastopen_cookie *cookie)
  4630. {
  4631. struct tcp_sock *tp = tcp_sk(sk);
  4632. struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
  4633. u16 mss = tp->rx_opt.mss_clamp;
  4634. bool syn_drop;
  4635. if (mss == tp->rx_opt.user_mss) {
  4636. struct tcp_options_received opt;
  4637. /* Get original SYNACK MSS value if user MSS sets mss_clamp */
  4638. tcp_clear_options(&opt);
  4639. opt.user_mss = opt.mss_clamp = 0;
  4640. tcp_parse_options(synack, &opt, 0, NULL);
  4641. mss = opt.mss_clamp;
  4642. }
  4643. if (!tp->syn_fastopen) /* Ignore an unsolicited cookie */
  4644. cookie->len = -1;
  4645. /* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
  4646. * the remote receives only the retransmitted (regular) SYNs: either
  4647. * the original SYN-data or the corresponding SYN-ACK is lost.
  4648. */
  4649. syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
  4650. tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
  4651. if (data) { /* Retransmit unacked data in SYN */
  4652. tcp_for_write_queue_from(data, sk) {
  4653. if (data == tcp_send_head(sk) ||
  4654. __tcp_retransmit_skb(sk, data))
  4655. break;
  4656. }
  4657. tcp_rearm_rto(sk);
  4658. return true;
  4659. }
  4660. tp->syn_data_acked = tp->syn_data;
  4661. return false;
  4662. }
  4663. static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
  4664. const struct tcphdr *th, unsigned int len)
  4665. {
  4666. struct inet_connection_sock *icsk = inet_csk(sk);
  4667. struct tcp_sock *tp = tcp_sk(sk);
  4668. struct tcp_fastopen_cookie foc = { .len = -1 };
  4669. int saved_clamp = tp->rx_opt.mss_clamp;
  4670. tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
  4671. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  4672. tp->rx_opt.rcv_tsecr -= tp->tsoffset;
  4673. if (th->ack) {
  4674. /* rfc793:
  4675. * "If the state is SYN-SENT then
  4676. * first check the ACK bit
  4677. * If the ACK bit is set
  4678. * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
  4679. * a reset (unless the RST bit is set, if so drop
  4680. * the segment and return)"
  4681. */
  4682. if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
  4683. after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
  4684. goto reset_and_undo;
  4685. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  4686. !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
  4687. tcp_time_stamp)) {
  4688. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
  4689. goto reset_and_undo;
  4690. }
  4691. /* Now ACK is acceptable.
  4692. *
  4693. * "If the RST bit is set
  4694. * If the ACK was acceptable then signal the user "error:
  4695. * connection reset", drop the segment, enter CLOSED state,
  4696. * delete TCB, and return."
  4697. */
  4698. if (th->rst) {
  4699. tcp_reset(sk);
  4700. goto discard;
  4701. }
  4702. /* rfc793:
  4703. * "fifth, if neither of the SYN or RST bits is set then
  4704. * drop the segment and return."
  4705. *
  4706. * See note below!
  4707. * --ANK(990513)
  4708. */
  4709. if (!th->syn)
  4710. goto discard_and_undo;
  4711. /* rfc793:
  4712. * "If the SYN bit is on ...
  4713. * are acceptable then ...
  4714. * (our SYN has been ACKed), change the connection
  4715. * state to ESTABLISHED..."
  4716. */
  4717. TCP_ECN_rcv_synack(tp, th);
  4718. tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
  4719. tcp_ack(sk, skb, FLAG_SLOWPATH);
  4720. /* Ok.. it's good. Set up sequence numbers and
  4721. * move to established.
  4722. */
  4723. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  4724. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  4725. /* RFC1323: The window in SYN & SYN/ACK segments is
  4726. * never scaled.
  4727. */
  4728. tp->snd_wnd = ntohs(th->window);
  4729. if (!tp->rx_opt.wscale_ok) {
  4730. tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
  4731. tp->window_clamp = min(tp->window_clamp, 65535U);
  4732. }
  4733. if (tp->rx_opt.saw_tstamp) {
  4734. tp->rx_opt.tstamp_ok = 1;
  4735. tp->tcp_header_len =
  4736. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  4737. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4738. tcp_store_ts_recent(tp);
  4739. } else {
  4740. tp->tcp_header_len = sizeof(struct tcphdr);
  4741. }
  4742. if (tcp_is_sack(tp) && sysctl_tcp_fack)
  4743. tcp_enable_fack(tp);
  4744. tcp_mtup_init(sk);
  4745. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4746. tcp_initialize_rcv_mss(sk);
  4747. /* Remember, tcp_poll() does not lock socket!
  4748. * Change state from SYN-SENT only after copied_seq
  4749. * is initialized. */
  4750. tp->copied_seq = tp->rcv_nxt;
  4751. smp_mb();
  4752. tcp_finish_connect(sk, skb);
  4753. if ((tp->syn_fastopen || tp->syn_data) &&
  4754. tcp_rcv_fastopen_synack(sk, skb, &foc))
  4755. return -1;
  4756. if (sk->sk_write_pending ||
  4757. icsk->icsk_accept_queue.rskq_defer_accept ||
  4758. icsk->icsk_ack.pingpong) {
  4759. /* Save one ACK. Data will be ready after
  4760. * several ticks, if write_pending is set.
  4761. *
  4762. * It may be deleted, but with this feature tcpdumps
  4763. * look so _wonderfully_ clever, that I was not able
  4764. * to stand against the temptation 8) --ANK
  4765. */
  4766. inet_csk_schedule_ack(sk);
  4767. icsk->icsk_ack.lrcvtime = tcp_time_stamp;
  4768. tcp_enter_quickack_mode(sk);
  4769. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  4770. TCP_DELACK_MAX, TCP_RTO_MAX);
  4771. discard:
  4772. __kfree_skb(skb);
  4773. return 0;
  4774. } else {
  4775. tcp_send_ack(sk);
  4776. }
  4777. return -1;
  4778. }
  4779. /* No ACK in the segment */
  4780. if (th->rst) {
  4781. /* rfc793:
  4782. * "If the RST bit is set
  4783. *
  4784. * Otherwise (no ACK) drop the segment and return."
  4785. */
  4786. goto discard_and_undo;
  4787. }
  4788. /* PAWS check. */
  4789. if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
  4790. tcp_paws_reject(&tp->rx_opt, 0))
  4791. goto discard_and_undo;
  4792. if (th->syn) {
  4793. /* We see SYN without ACK. It is attempt of
  4794. * simultaneous connect with crossed SYNs.
  4795. * Particularly, it can be connect to self.
  4796. */
  4797. tcp_set_state(sk, TCP_SYN_RECV);
  4798. if (tp->rx_opt.saw_tstamp) {
  4799. tp->rx_opt.tstamp_ok = 1;
  4800. tcp_store_ts_recent(tp);
  4801. tp->tcp_header_len =
  4802. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  4803. } else {
  4804. tp->tcp_header_len = sizeof(struct tcphdr);
  4805. }
  4806. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  4807. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  4808. /* RFC1323: The window in SYN & SYN/ACK segments is
  4809. * never scaled.
  4810. */
  4811. tp->snd_wnd = ntohs(th->window);
  4812. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  4813. tp->max_window = tp->snd_wnd;
  4814. TCP_ECN_rcv_syn(tp, th);
  4815. tcp_mtup_init(sk);
  4816. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4817. tcp_initialize_rcv_mss(sk);
  4818. tcp_send_synack(sk);
  4819. #if 0
  4820. /* Note, we could accept data and URG from this segment.
  4821. * There are no obstacles to make this (except that we must
  4822. * either change tcp_recvmsg() to prevent it from returning data
  4823. * before 3WHS completes per RFC793, or employ TCP Fast Open).
  4824. *
  4825. * However, if we ignore data in ACKless segments sometimes,
  4826. * we have no reasons to accept it sometimes.
  4827. * Also, seems the code doing it in step6 of tcp_rcv_state_process
  4828. * is not flawless. So, discard packet for sanity.
  4829. * Uncomment this return to process the data.
  4830. */
  4831. return -1;
  4832. #else
  4833. goto discard;
  4834. #endif
  4835. }
  4836. /* "fifth, if neither of the SYN or RST bits is set then
  4837. * drop the segment and return."
  4838. */
  4839. discard_and_undo:
  4840. tcp_clear_options(&tp->rx_opt);
  4841. tp->rx_opt.mss_clamp = saved_clamp;
  4842. goto discard;
  4843. reset_and_undo:
  4844. tcp_clear_options(&tp->rx_opt);
  4845. tp->rx_opt.mss_clamp = saved_clamp;
  4846. return 1;
  4847. }
  4848. /*
  4849. * This function implements the receiving procedure of RFC 793 for
  4850. * all states except ESTABLISHED and TIME_WAIT.
  4851. * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
  4852. * address independent.
  4853. */
  4854. int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
  4855. const struct tcphdr *th, unsigned int len)
  4856. {
  4857. struct tcp_sock *tp = tcp_sk(sk);
  4858. struct inet_connection_sock *icsk = inet_csk(sk);
  4859. struct request_sock *req;
  4860. int queued = 0;
  4861. bool acceptable;
  4862. u32 synack_stamp;
  4863. tp->rx_opt.saw_tstamp = 0;
  4864. switch (sk->sk_state) {
  4865. case TCP_CLOSE:
  4866. goto discard;
  4867. case TCP_LISTEN:
  4868. if (th->ack)
  4869. return 1;
  4870. if (th->rst)
  4871. goto discard;
  4872. if (th->syn) {
  4873. if (th->fin)
  4874. goto discard;
  4875. if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
  4876. return 1;
  4877. /* Now we have several options: In theory there is
  4878. * nothing else in the frame. KA9Q has an option to
  4879. * send data with the syn, BSD accepts data with the
  4880. * syn up to the [to be] advertised window and
  4881. * Solaris 2.1 gives you a protocol error. For now
  4882. * we just ignore it, that fits the spec precisely
  4883. * and avoids incompatibilities. It would be nice in
  4884. * future to drop through and process the data.
  4885. *
  4886. * Now that TTCP is starting to be used we ought to
  4887. * queue this data.
  4888. * But, this leaves one open to an easy denial of
  4889. * service attack, and SYN cookies can't defend
  4890. * against this problem. So, we drop the data
  4891. * in the interest of security over speed unless
  4892. * it's still in use.
  4893. */
  4894. kfree_skb(skb);
  4895. return 0;
  4896. }
  4897. goto discard;
  4898. case TCP_SYN_SENT:
  4899. queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
  4900. if (queued >= 0)
  4901. return queued;
  4902. /* Do step6 onward by hand. */
  4903. tcp_urg(sk, skb, th);
  4904. __kfree_skb(skb);
  4905. tcp_data_snd_check(sk);
  4906. return 0;
  4907. }
  4908. req = tp->fastopen_rsk;
  4909. if (req != NULL) {
  4910. WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
  4911. sk->sk_state != TCP_FIN_WAIT1);
  4912. if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
  4913. goto discard;
  4914. }
  4915. if (!th->ack && !th->rst)
  4916. goto discard;
  4917. if (!tcp_validate_incoming(sk, skb, th, 0))
  4918. return 0;
  4919. /* step 5: check the ACK field */
  4920. acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
  4921. FLAG_UPDATE_TS_RECENT) > 0;
  4922. switch (sk->sk_state) {
  4923. case TCP_SYN_RECV:
  4924. if (!acceptable)
  4925. return 1;
  4926. /* Once we leave TCP_SYN_RECV, we no longer need req
  4927. * so release it.
  4928. */
  4929. if (req) {
  4930. synack_stamp = tcp_rsk(req)->snt_synack;
  4931. tp->total_retrans = req->num_retrans;
  4932. reqsk_fastopen_remove(sk, req, false);
  4933. } else {
  4934. synack_stamp = tp->lsndtime;
  4935. /* Make sure socket is routed, for correct metrics. */
  4936. icsk->icsk_af_ops->rebuild_header(sk);
  4937. tcp_init_congestion_control(sk);
  4938. tcp_mtup_init(sk);
  4939. tp->copied_seq = tp->rcv_nxt;
  4940. tcp_init_buffer_space(sk);
  4941. }
  4942. smp_mb();
  4943. tcp_set_state(sk, TCP_ESTABLISHED);
  4944. sk->sk_state_change(sk);
  4945. /* Note, that this wakeup is only for marginal crossed SYN case.
  4946. * Passively open sockets are not waked up, because
  4947. * sk->sk_sleep == NULL and sk->sk_socket == NULL.
  4948. */
  4949. if (sk->sk_socket)
  4950. sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
  4951. tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
  4952. tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
  4953. tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
  4954. tcp_synack_rtt_meas(sk, synack_stamp);
  4955. if (tp->rx_opt.tstamp_ok)
  4956. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4957. if (req) {
  4958. /* Re-arm the timer because data may have been sent out.
  4959. * This is similar to the regular data transmission case
  4960. * when new data has just been ack'ed.
  4961. *
  4962. * (TFO) - we could try to be more aggressive and
  4963. * retransmitting any data sooner based on when they
  4964. * are sent out.
  4965. */
  4966. tcp_rearm_rto(sk);
  4967. } else
  4968. tcp_init_metrics(sk);
  4969. tcp_update_pacing_rate(sk);
  4970. /* Prevent spurious tcp_cwnd_restart() on first data packet */
  4971. tp->lsndtime = tcp_time_stamp;
  4972. tcp_initialize_rcv_mss(sk);
  4973. tcp_fast_path_on(tp);
  4974. break;
  4975. case TCP_FIN_WAIT1: {
  4976. struct dst_entry *dst;
  4977. int tmo;
  4978. /* If we enter the TCP_FIN_WAIT1 state and we are a
  4979. * Fast Open socket and this is the first acceptable
  4980. * ACK we have received, this would have acknowledged
  4981. * our SYNACK so stop the SYNACK timer.
  4982. */
  4983. if (req != NULL) {
  4984. /* Return RST if ack_seq is invalid.
  4985. * Note that RFC793 only says to generate a
  4986. * DUPACK for it but for TCP Fast Open it seems
  4987. * better to treat this case like TCP_SYN_RECV
  4988. * above.
  4989. */
  4990. if (!acceptable)
  4991. return 1;
  4992. /* We no longer need the request sock. */
  4993. reqsk_fastopen_remove(sk, req, false);
  4994. tcp_rearm_rto(sk);
  4995. }
  4996. if (tp->snd_una != tp->write_seq)
  4997. break;
  4998. tcp_set_state(sk, TCP_FIN_WAIT2);
  4999. sk->sk_shutdown |= SEND_SHUTDOWN;
  5000. dst = __sk_dst_get(sk);
  5001. if (dst)
  5002. dst_confirm(dst);
  5003. if (!sock_flag(sk, SOCK_DEAD)) {
  5004. /* Wake up lingering close() */
  5005. sk->sk_state_change(sk);
  5006. break;
  5007. }
  5008. if (tp->linger2 < 0 ||
  5009. (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  5010. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
  5011. tcp_done(sk);
  5012. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
  5013. return 1;
  5014. }
  5015. tmo = tcp_fin_time(sk);
  5016. if (tmo > TCP_TIMEWAIT_LEN) {
  5017. inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
  5018. } else if (th->fin || sock_owned_by_user(sk)) {
  5019. /* Bad case. We could lose such FIN otherwise.
  5020. * It is not a big problem, but it looks confusing
  5021. * and not so rare event. We still can lose it now,
  5022. * if it spins in bh_lock_sock(), but it is really
  5023. * marginal case.
  5024. */
  5025. inet_csk_reset_keepalive_timer(sk, tmo);
  5026. } else {
  5027. tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
  5028. goto discard;
  5029. }
  5030. break;
  5031. }
  5032. case TCP_CLOSING:
  5033. if (tp->snd_una == tp->write_seq) {
  5034. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  5035. goto discard;
  5036. }
  5037. break;
  5038. case TCP_LAST_ACK:
  5039. if (tp->snd_una == tp->write_seq) {
  5040. tcp_update_metrics(sk);
  5041. tcp_done(sk);
  5042. goto discard;
  5043. }
  5044. break;
  5045. }
  5046. /* step 6: check the URG bit */
  5047. tcp_urg(sk, skb, th);
  5048. /* step 7: process the segment text */
  5049. switch (sk->sk_state) {
  5050. case TCP_CLOSE_WAIT:
  5051. case TCP_CLOSING:
  5052. case TCP_LAST_ACK:
  5053. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  5054. break;
  5055. case TCP_FIN_WAIT1:
  5056. case TCP_FIN_WAIT2:
  5057. /* RFC 793 says to queue data in these states,
  5058. * RFC 1122 says we MUST send a reset.
  5059. * BSD 4.4 also does reset.
  5060. */
  5061. if (sk->sk_shutdown & RCV_SHUTDOWN) {
  5062. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  5063. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
  5064. NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
  5065. tcp_reset(sk);
  5066. return 1;
  5067. }
  5068. }
  5069. /* Fall through */
  5070. case TCP_ESTABLISHED:
  5071. tcp_data_queue(sk, skb);
  5072. queued = 1;
  5073. break;
  5074. }
  5075. /* tcp_data could move socket to TIME-WAIT */
  5076. if (sk->sk_state != TCP_CLOSE) {
  5077. tcp_data_snd_check(sk);
  5078. tcp_ack_snd_check(sk);
  5079. }
  5080. if (!queued) {
  5081. discard:
  5082. __kfree_skb(skb);
  5083. }
  5084. return 0;
  5085. }
  5086. EXPORT_SYMBOL(tcp_rcv_state_process);