tcp_input.c 149 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555355635573558355935603561356235633564356535663567356835693570357135723573357435753576357735783579358035813582358335843585358635873588358935903591359235933594359535963597359835993600360136023603360436053606360736083609361036113612361336143615361636173618361936203621362236233624362536263627362836293630363136323633363436353636363736383639364036413642364336443645364636473648364936503651365236533654365536563657365836593660366136623663366436653666366736683669367036713672367336743675367636773678367936803681368236833684368536863687368836893690369136923693369436953696369736983699370037013702370337043705370637073708370937103711371237133714371537163717371837193720372137223723372437253726372737283729373037313732373337343735373637373738373937403741374237433744374537463747374837493750375137523753375437553756375737583759376037613762376337643765376637673768376937703771377237733774377537763777377837793780378137823783378437853786378737883789379037913792379337943795379637973798379938003801380238033804380538063807380838093810381138123813381438153816381738183819382038213822382338243825382638273828382938303831383238333834383538363837383838393840384138423843384438453846384738483849385038513852385338543855385638573858385938603861386238633864386538663867386838693870387138723873387438753876387738783879388038813882388338843885388638873888388938903891389238933894389538963897389838993900390139023903390439053906390739083909391039113912391339143915391639173918391939203921392239233924392539263927392839293930393139323933393439353936393739383939394039413942394339443945394639473948394939503951395239533954395539563957395839593960396139623963396439653966396739683969397039713972397339743975397639773978397939803981398239833984398539863987398839893990399139923993399439953996399739983999400040014002400340044005400640074008400940104011401240134014401540164017401840194020402140224023402440254026402740284029403040314032403340344035403640374038403940404041404240434044404540464047404840494050405140524053405440554056405740584059406040614062406340644065406640674068406940704071407240734074407540764077407840794080408140824083408440854086408740884089409040914092409340944095409640974098409941004101410241034104410541064107410841094110411141124113411441154116411741184119412041214122412341244125412641274128412941304131413241334134413541364137413841394140414141424143414441454146414741484149415041514152415341544155415641574158415941604161416241634164416541664167416841694170417141724173417441754176417741784179418041814182418341844185418641874188418941904191419241934194419541964197419841994200420142024203420442054206420742084209421042114212421342144215421642174218421942204221422242234224422542264227422842294230423142324233423442354236423742384239424042414242424342444245424642474248424942504251425242534254425542564257425842594260426142624263426442654266426742684269427042714272427342744275427642774278427942804281428242834284428542864287428842894290429142924293429442954296429742984299430043014302430343044305430643074308430943104311431243134314431543164317431843194320432143224323432443254326432743284329433043314332433343344335433643374338433943404341434243434344434543464347434843494350435143524353435443554356435743584359436043614362436343644365436643674368436943704371437243734374437543764377437843794380438143824383438443854386438743884389439043914392439343944395439643974398439944004401440244034404440544064407440844094410441144124413441444154416441744184419442044214422442344244425442644274428442944304431443244334434443544364437443844394440444144424443444444454446444744484449445044514452445344544455445644574458445944604461446244634464446544664467446844694470447144724473447444754476447744784479448044814482448344844485448644874488448944904491449244934494449544964497449844994500450145024503450445054506450745084509451045114512451345144515451645174518451945204521452245234524452545264527452845294530453145324533453445354536453745384539454045414542454345444545454645474548454945504551455245534554455545564557455845594560456145624563456445654566456745684569457045714572457345744575457645774578457945804581458245834584458545864587458845894590459145924593459445954596459745984599460046014602460346044605460646074608460946104611461246134614461546164617461846194620462146224623462446254626462746284629463046314632463346344635463646374638463946404641464246434644464546464647464846494650465146524653465446554656465746584659466046614662466346644665466646674668466946704671467246734674467546764677467846794680468146824683468446854686468746884689469046914692469346944695469646974698469947004701470247034704470547064707470847094710471147124713471447154716471747184719472047214722472347244725472647274728472947304731473247334734473547364737473847394740474147424743474447454746474747484749475047514752475347544755475647574758475947604761476247634764476547664767476847694770477147724773477447754776477747784779478047814782478347844785478647874788478947904791479247934794479547964797479847994800480148024803480448054806480748084809481048114812481348144815481648174818481948204821482248234824482548264827482848294830483148324833483448354836483748384839484048414842484348444845484648474848484948504851485248534854485548564857485848594860486148624863486448654866486748684869487048714872487348744875487648774878487948804881488248834884488548864887488848894890489148924893489448954896489748984899490049014902490349044905490649074908490949104911491249134914491549164917491849194920492149224923492449254926492749284929493049314932493349344935493649374938493949404941494249434944494549464947494849494950495149524953495449554956495749584959496049614962496349644965496649674968496949704971497249734974497549764977497849794980498149824983498449854986498749884989499049914992499349944995499649974998499950005001500250035004500550065007500850095010501150125013501450155016501750185019502050215022502350245025502650275028502950305031503250335034503550365037503850395040504150425043504450455046504750485049505050515052505350545055505650575058505950605061506250635064506550665067506850695070507150725073507450755076507750785079508050815082508350845085508650875088508950905091509250935094509550965097509850995100510151025103510451055106510751085109511051115112511351145115511651175118511951205121512251235124512551265127512851295130513151325133513451355136513751385139514051415142514351445145514651475148514951505151515251535154515551565157515851595160516151625163516451655166516751685169517051715172517351745175517651775178517951805181518251835184518551865187518851895190519151925193519451955196519751985199520052015202520352045205520652075208520952105211
  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. * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
  9. *
  10. * Authors: Ross Biro
  11. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  13. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  14. * Florian La Roche, <flla@stud.uni-sb.de>
  15. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  16. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  17. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  18. * Matthew Dillon, <dillon@apollo.west.oic.com>
  19. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  20. * Jorge Cwik, <jorge@laser.satlink.net>
  21. */
  22. /*
  23. * Changes:
  24. * Pedro Roque : Fast Retransmit/Recovery.
  25. * Two receive queues.
  26. * Retransmit queue handled by TCP.
  27. * Better retransmit timer handling.
  28. * New congestion avoidance.
  29. * Header prediction.
  30. * Variable renaming.
  31. *
  32. * Eric : Fast Retransmit.
  33. * Randy Scott : MSS option defines.
  34. * Eric Schenk : Fixes to slow start algorithm.
  35. * Eric Schenk : Yet another double ACK bug.
  36. * Eric Schenk : Delayed ACK bug fixes.
  37. * Eric Schenk : Floyd style fast retrans war avoidance.
  38. * David S. Miller : Don't allow zero congestion window.
  39. * Eric Schenk : Fix retransmitter so that it sends
  40. * next packet on ack of previous packet.
  41. * Andi Kleen : Moved open_request checking here
  42. * and process RSTs for open_requests.
  43. * Andi Kleen : Better prune_queue, and other fixes.
  44. * Andrey Savochkin: Fix RTT measurements in the presence of
  45. * timestamps.
  46. * Andrey Savochkin: Check sequence numbers correctly when
  47. * removing SACKs due to in sequence incoming
  48. * data segments.
  49. * Andi Kleen: Make sure we never ack data there is not
  50. * enough room for. Also make this condition
  51. * a fatal error if it might still happen.
  52. * Andi Kleen: Add tcp_measure_rcv_mss to make
  53. * connections with MSS<min(MTU,ann. MSS)
  54. * work without delayed acks.
  55. * Andi Kleen: Process packets with PSH set in the
  56. * fast path.
  57. * J Hadi Salim: ECN support
  58. * Andrei Gurtov,
  59. * Pasi Sarolahti,
  60. * Panu Kuhlberg: Experimental audit of TCP (re)transmission
  61. * engine. Lots of bugs are found.
  62. * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
  63. */
  64. #include <linux/mm.h>
  65. #include <linux/module.h>
  66. #include <linux/sysctl.h>
  67. #include <net/tcp.h>
  68. #include <net/inet_common.h>
  69. #include <linux/ipsec.h>
  70. #include <asm/unaligned.h>
  71. #include <net/netdma.h>
  72. int sysctl_tcp_timestamps __read_mostly = 1;
  73. int sysctl_tcp_window_scaling __read_mostly = 1;
  74. int sysctl_tcp_sack __read_mostly = 1;
  75. int sysctl_tcp_fack __read_mostly = 1;
  76. int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  77. int sysctl_tcp_ecn __read_mostly;
  78. int sysctl_tcp_dsack __read_mostly = 1;
  79. int sysctl_tcp_app_win __read_mostly = 31;
  80. int sysctl_tcp_adv_win_scale __read_mostly = 2;
  81. int sysctl_tcp_stdurg __read_mostly;
  82. int sysctl_tcp_rfc1337 __read_mostly;
  83. int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  84. int sysctl_tcp_frto __read_mostly = 2;
  85. int sysctl_tcp_frto_response __read_mostly;
  86. int sysctl_tcp_nometrics_save __read_mostly;
  87. int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
  88. int sysctl_tcp_abc __read_mostly;
  89. #define FLAG_DATA 0x01 /* Incoming frame contained data. */
  90. #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
  91. #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
  92. #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
  93. #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
  94. #define FLAG_DATA_SACKED 0x20 /* New SACK. */
  95. #define FLAG_ECE 0x40 /* ECE in this ACK */
  96. #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
  97. #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
  98. #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
  99. #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
  100. #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */
  101. #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */
  102. #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
  103. #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
  104. #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
  105. #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
  106. #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
  107. #define IsSackFrto() (sysctl_tcp_frto == 0x2)
  108. #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
  109. #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
  110. /* Adapt the MSS value used to make delayed ack decision to the
  111. * real world.
  112. */
  113. static void tcp_measure_rcv_mss(struct sock *sk,
  114. const struct sk_buff *skb)
  115. {
  116. struct inet_connection_sock *icsk = inet_csk(sk);
  117. const unsigned int lss = icsk->icsk_ack.last_seg_size;
  118. unsigned int len;
  119. icsk->icsk_ack.last_seg_size = 0;
  120. /* skb->len may jitter because of SACKs, even if peer
  121. * sends good full-sized frames.
  122. */
  123. len = skb_shinfo(skb)->gso_size ?: skb->len;
  124. if (len >= icsk->icsk_ack.rcv_mss) {
  125. icsk->icsk_ack.rcv_mss = len;
  126. } else {
  127. /* Otherwise, we make more careful check taking into account,
  128. * that SACKs block is variable.
  129. *
  130. * "len" is invariant segment length, including TCP header.
  131. */
  132. len += skb->data - skb_transport_header(skb);
  133. if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
  134. /* If PSH is not set, packet should be
  135. * full sized, provided peer TCP is not badly broken.
  136. * This observation (if it is correct 8)) allows
  137. * to handle super-low mtu links fairly.
  138. */
  139. (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
  140. !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
  141. /* Subtract also invariant (if peer is RFC compliant),
  142. * tcp header plus fixed timestamp option length.
  143. * Resulting "len" is MSS free of SACK jitter.
  144. */
  145. len -= tcp_sk(sk)->tcp_header_len;
  146. icsk->icsk_ack.last_seg_size = len;
  147. if (len == lss) {
  148. icsk->icsk_ack.rcv_mss = len;
  149. return;
  150. }
  151. }
  152. if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
  153. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
  154. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
  155. }
  156. }
  157. static void tcp_incr_quickack(struct sock *sk)
  158. {
  159. struct inet_connection_sock *icsk = inet_csk(sk);
  160. unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
  161. if (quickacks==0)
  162. quickacks=2;
  163. if (quickacks > icsk->icsk_ack.quick)
  164. icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
  165. }
  166. void tcp_enter_quickack_mode(struct sock *sk)
  167. {
  168. struct inet_connection_sock *icsk = inet_csk(sk);
  169. tcp_incr_quickack(sk);
  170. icsk->icsk_ack.pingpong = 0;
  171. icsk->icsk_ack.ato = TCP_ATO_MIN;
  172. }
  173. /* Send ACKs quickly, if "quick" count is not exhausted
  174. * and the session is not interactive.
  175. */
  176. static inline int tcp_in_quickack_mode(const struct sock *sk)
  177. {
  178. const struct inet_connection_sock *icsk = inet_csk(sk);
  179. return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
  180. }
  181. static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
  182. {
  183. if (tp->ecn_flags&TCP_ECN_OK)
  184. tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
  185. }
  186. static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
  187. {
  188. if (tcp_hdr(skb)->cwr)
  189. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  190. }
  191. static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
  192. {
  193. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  194. }
  195. static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
  196. {
  197. if (tp->ecn_flags&TCP_ECN_OK) {
  198. if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
  199. tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
  200. /* Funny extension: if ECT is not set on a segment,
  201. * it is surely retransmit. It is not in ECN RFC,
  202. * but Linux follows this rule. */
  203. else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
  204. tcp_enter_quickack_mode((struct sock *)tp);
  205. }
  206. }
  207. static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
  208. {
  209. if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
  210. tp->ecn_flags &= ~TCP_ECN_OK;
  211. }
  212. static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
  213. {
  214. if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
  215. tp->ecn_flags &= ~TCP_ECN_OK;
  216. }
  217. static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
  218. {
  219. if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
  220. return 1;
  221. return 0;
  222. }
  223. /* Buffer size and advertised window tuning.
  224. *
  225. * 1. Tuning sk->sk_sndbuf, when connection enters established state.
  226. */
  227. static void tcp_fixup_sndbuf(struct sock *sk)
  228. {
  229. int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
  230. sizeof(struct sk_buff);
  231. if (sk->sk_sndbuf < 3 * sndmem)
  232. sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
  233. }
  234. /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
  235. *
  236. * All tcp_full_space() is split to two parts: "network" buffer, allocated
  237. * forward and advertised in receiver window (tp->rcv_wnd) and
  238. * "application buffer", required to isolate scheduling/application
  239. * latencies from network.
  240. * window_clamp is maximal advertised window. It can be less than
  241. * tcp_full_space(), in this case tcp_full_space() - window_clamp
  242. * is reserved for "application" buffer. The less window_clamp is
  243. * the smoother our behaviour from viewpoint of network, but the lower
  244. * throughput and the higher sensitivity of the connection to losses. 8)
  245. *
  246. * rcv_ssthresh is more strict window_clamp used at "slow start"
  247. * phase to predict further behaviour of this connection.
  248. * It is used for two goals:
  249. * - to enforce header prediction at sender, even when application
  250. * requires some significant "application buffer". It is check #1.
  251. * - to prevent pruning of receive queue because of misprediction
  252. * of receiver window. Check #2.
  253. *
  254. * The scheme does not work when sender sends good segments opening
  255. * window and then starts to feed us spaghetti. But it should work
  256. * in common situations. Otherwise, we have to rely on queue collapsing.
  257. */
  258. /* Slow part of check#2. */
  259. static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
  260. {
  261. struct tcp_sock *tp = tcp_sk(sk);
  262. /* Optimize this! */
  263. int truesize = tcp_win_from_space(skb->truesize)/2;
  264. int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
  265. while (tp->rcv_ssthresh <= window) {
  266. if (truesize <= skb->len)
  267. return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
  268. truesize >>= 1;
  269. window >>= 1;
  270. }
  271. return 0;
  272. }
  273. static void tcp_grow_window(struct sock *sk,
  274. struct sk_buff *skb)
  275. {
  276. struct tcp_sock *tp = tcp_sk(sk);
  277. /* Check #1 */
  278. if (tp->rcv_ssthresh < tp->window_clamp &&
  279. (int)tp->rcv_ssthresh < tcp_space(sk) &&
  280. !tcp_memory_pressure) {
  281. int incr;
  282. /* Check #2. Increase window, if skb with such overhead
  283. * will fit to rcvbuf in future.
  284. */
  285. if (tcp_win_from_space(skb->truesize) <= skb->len)
  286. incr = 2*tp->advmss;
  287. else
  288. incr = __tcp_grow_window(sk, skb);
  289. if (incr) {
  290. tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
  291. inet_csk(sk)->icsk_ack.quick |= 1;
  292. }
  293. }
  294. }
  295. /* 3. Tuning rcvbuf, when connection enters established state. */
  296. static void tcp_fixup_rcvbuf(struct sock *sk)
  297. {
  298. struct tcp_sock *tp = tcp_sk(sk);
  299. int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
  300. /* Try to select rcvbuf so that 4 mss-sized segments
  301. * will fit to window and corresponding skbs will fit to our rcvbuf.
  302. * (was 3; 4 is minimum to allow fast retransmit to work.)
  303. */
  304. while (tcp_win_from_space(rcvmem) < tp->advmss)
  305. rcvmem += 128;
  306. if (sk->sk_rcvbuf < 4 * rcvmem)
  307. sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
  308. }
  309. /* 4. Try to fixup all. It is made immediately after connection enters
  310. * established state.
  311. */
  312. static void tcp_init_buffer_space(struct sock *sk)
  313. {
  314. struct tcp_sock *tp = tcp_sk(sk);
  315. int maxwin;
  316. if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
  317. tcp_fixup_rcvbuf(sk);
  318. if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
  319. tcp_fixup_sndbuf(sk);
  320. tp->rcvq_space.space = tp->rcv_wnd;
  321. maxwin = tcp_full_space(sk);
  322. if (tp->window_clamp >= maxwin) {
  323. tp->window_clamp = maxwin;
  324. if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
  325. tp->window_clamp = max(maxwin -
  326. (maxwin >> sysctl_tcp_app_win),
  327. 4 * tp->advmss);
  328. }
  329. /* Force reservation of one segment. */
  330. if (sysctl_tcp_app_win &&
  331. tp->window_clamp > 2 * tp->advmss &&
  332. tp->window_clamp + tp->advmss > maxwin)
  333. tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
  334. tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
  335. tp->snd_cwnd_stamp = tcp_time_stamp;
  336. }
  337. /* 5. Recalculate window clamp after socket hit its memory bounds. */
  338. static void tcp_clamp_window(struct sock *sk)
  339. {
  340. struct tcp_sock *tp = tcp_sk(sk);
  341. struct inet_connection_sock *icsk = inet_csk(sk);
  342. icsk->icsk_ack.quick = 0;
  343. if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
  344. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
  345. !tcp_memory_pressure &&
  346. atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
  347. sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
  348. sysctl_tcp_rmem[2]);
  349. }
  350. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
  351. tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
  352. }
  353. /* Initialize RCV_MSS value.
  354. * RCV_MSS is an our guess about MSS used by the peer.
  355. * We haven't any direct information about the MSS.
  356. * It's better to underestimate the RCV_MSS rather than overestimate.
  357. * Overestimations make us ACKing less frequently than needed.
  358. * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
  359. */
  360. void tcp_initialize_rcv_mss(struct sock *sk)
  361. {
  362. struct tcp_sock *tp = tcp_sk(sk);
  363. unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
  364. hint = min(hint, tp->rcv_wnd/2);
  365. hint = min(hint, TCP_MIN_RCVMSS);
  366. hint = max(hint, TCP_MIN_MSS);
  367. inet_csk(sk)->icsk_ack.rcv_mss = hint;
  368. }
  369. /* Receiver "autotuning" code.
  370. *
  371. * The algorithm for RTT estimation w/o timestamps is based on
  372. * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
  373. * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
  374. *
  375. * More detail on this code can be found at
  376. * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
  377. * though this reference is out of date. A new paper
  378. * is pending.
  379. */
  380. static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
  381. {
  382. u32 new_sample = tp->rcv_rtt_est.rtt;
  383. long m = sample;
  384. if (m == 0)
  385. m = 1;
  386. if (new_sample != 0) {
  387. /* If we sample in larger samples in the non-timestamp
  388. * case, we could grossly overestimate the RTT especially
  389. * with chatty applications or bulk transfer apps which
  390. * are stalled on filesystem I/O.
  391. *
  392. * Also, since we are only going for a minimum in the
  393. * non-timestamp case, we do not smooth things out
  394. * else with timestamps disabled convergence takes too
  395. * long.
  396. */
  397. if (!win_dep) {
  398. m -= (new_sample >> 3);
  399. new_sample += m;
  400. } else if (m < new_sample)
  401. new_sample = m << 3;
  402. } else {
  403. /* No previous measure. */
  404. new_sample = m << 3;
  405. }
  406. if (tp->rcv_rtt_est.rtt != new_sample)
  407. tp->rcv_rtt_est.rtt = new_sample;
  408. }
  409. static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
  410. {
  411. if (tp->rcv_rtt_est.time == 0)
  412. goto new_measure;
  413. if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
  414. return;
  415. tcp_rcv_rtt_update(tp,
  416. jiffies - tp->rcv_rtt_est.time,
  417. 1);
  418. new_measure:
  419. tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
  420. tp->rcv_rtt_est.time = tcp_time_stamp;
  421. }
  422. static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
  423. {
  424. struct tcp_sock *tp = tcp_sk(sk);
  425. if (tp->rx_opt.rcv_tsecr &&
  426. (TCP_SKB_CB(skb)->end_seq -
  427. TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
  428. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
  429. }
  430. /*
  431. * This function should be called every time data is copied to user space.
  432. * It calculates the appropriate TCP receive buffer space.
  433. */
  434. void tcp_rcv_space_adjust(struct sock *sk)
  435. {
  436. struct tcp_sock *tp = tcp_sk(sk);
  437. int time;
  438. int space;
  439. if (tp->rcvq_space.time == 0)
  440. goto new_measure;
  441. time = tcp_time_stamp - tp->rcvq_space.time;
  442. if (time < (tp->rcv_rtt_est.rtt >> 3) ||
  443. tp->rcv_rtt_est.rtt == 0)
  444. return;
  445. space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
  446. space = max(tp->rcvq_space.space, space);
  447. if (tp->rcvq_space.space != space) {
  448. int rcvmem;
  449. tp->rcvq_space.space = space;
  450. if (sysctl_tcp_moderate_rcvbuf &&
  451. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
  452. int new_clamp = space;
  453. /* Receive space grows, normalize in order to
  454. * take into account packet headers and sk_buff
  455. * structure overhead.
  456. */
  457. space /= tp->advmss;
  458. if (!space)
  459. space = 1;
  460. rcvmem = (tp->advmss + MAX_TCP_HEADER +
  461. 16 + sizeof(struct sk_buff));
  462. while (tcp_win_from_space(rcvmem) < tp->advmss)
  463. rcvmem += 128;
  464. space *= rcvmem;
  465. space = min(space, sysctl_tcp_rmem[2]);
  466. if (space > sk->sk_rcvbuf) {
  467. sk->sk_rcvbuf = space;
  468. /* Make the window clamp follow along. */
  469. tp->window_clamp = new_clamp;
  470. }
  471. }
  472. }
  473. new_measure:
  474. tp->rcvq_space.seq = tp->copied_seq;
  475. tp->rcvq_space.time = tcp_time_stamp;
  476. }
  477. /* There is something which you must keep in mind when you analyze the
  478. * behavior of the tp->ato delayed ack timeout interval. When a
  479. * connection starts up, we want to ack as quickly as possible. The
  480. * problem is that "good" TCP's do slow start at the beginning of data
  481. * transmission. The means that until we send the first few ACK's the
  482. * sender will sit on his end and only queue most of his data, because
  483. * he can only send snd_cwnd unacked packets at any given time. For
  484. * each ACK we send, he increments snd_cwnd and transmits more of his
  485. * queue. -DaveM
  486. */
  487. static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
  488. {
  489. struct tcp_sock *tp = tcp_sk(sk);
  490. struct inet_connection_sock *icsk = inet_csk(sk);
  491. u32 now;
  492. inet_csk_schedule_ack(sk);
  493. tcp_measure_rcv_mss(sk, skb);
  494. tcp_rcv_rtt_measure(tp);
  495. now = tcp_time_stamp;
  496. if (!icsk->icsk_ack.ato) {
  497. /* The _first_ data packet received, initialize
  498. * delayed ACK engine.
  499. */
  500. tcp_incr_quickack(sk);
  501. icsk->icsk_ack.ato = TCP_ATO_MIN;
  502. } else {
  503. int m = now - icsk->icsk_ack.lrcvtime;
  504. if (m <= TCP_ATO_MIN/2) {
  505. /* The fastest case is the first. */
  506. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
  507. } else if (m < icsk->icsk_ack.ato) {
  508. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
  509. if (icsk->icsk_ack.ato > icsk->icsk_rto)
  510. icsk->icsk_ack.ato = icsk->icsk_rto;
  511. } else if (m > icsk->icsk_rto) {
  512. /* Too long gap. Apparently sender failed to
  513. * restart window, so that we send ACKs quickly.
  514. */
  515. tcp_incr_quickack(sk);
  516. sk_stream_mem_reclaim(sk);
  517. }
  518. }
  519. icsk->icsk_ack.lrcvtime = now;
  520. TCP_ECN_check_ce(tp, skb);
  521. if (skb->len >= 128)
  522. tcp_grow_window(sk, skb);
  523. }
  524. static u32 tcp_rto_min(struct sock *sk)
  525. {
  526. struct dst_entry *dst = __sk_dst_get(sk);
  527. u32 rto_min = TCP_RTO_MIN;
  528. if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
  529. rto_min = dst->metrics[RTAX_RTO_MIN-1];
  530. return rto_min;
  531. }
  532. /* Called to compute a smoothed rtt estimate. The data fed to this
  533. * routine either comes from timestamps, or from segments that were
  534. * known _not_ to have been retransmitted [see Karn/Partridge
  535. * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
  536. * piece by Van Jacobson.
  537. * NOTE: the next three routines used to be one big routine.
  538. * To save cycles in the RFC 1323 implementation it was better to break
  539. * it up into three procedures. -- erics
  540. */
  541. static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
  542. {
  543. struct tcp_sock *tp = tcp_sk(sk);
  544. long m = mrtt; /* RTT */
  545. /* The following amusing code comes from Jacobson's
  546. * article in SIGCOMM '88. Note that rtt and mdev
  547. * are scaled versions of rtt and mean deviation.
  548. * This is designed to be as fast as possible
  549. * m stands for "measurement".
  550. *
  551. * On a 1990 paper the rto value is changed to:
  552. * RTO = rtt + 4 * mdev
  553. *
  554. * Funny. This algorithm seems to be very broken.
  555. * These formulae increase RTO, when it should be decreased, increase
  556. * too slowly, when it should be increased quickly, decrease too quickly
  557. * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
  558. * does not matter how to _calculate_ it. Seems, it was trap
  559. * that VJ failed to avoid. 8)
  560. */
  561. if (m == 0)
  562. m = 1;
  563. if (tp->srtt != 0) {
  564. m -= (tp->srtt >> 3); /* m is now error in rtt est */
  565. tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
  566. if (m < 0) {
  567. m = -m; /* m is now abs(error) */
  568. m -= (tp->mdev >> 2); /* similar update on mdev */
  569. /* This is similar to one of Eifel findings.
  570. * Eifel blocks mdev updates when rtt decreases.
  571. * This solution is a bit different: we use finer gain
  572. * for mdev in this case (alpha*beta).
  573. * Like Eifel it also prevents growth of rto,
  574. * but also it limits too fast rto decreases,
  575. * happening in pure Eifel.
  576. */
  577. if (m > 0)
  578. m >>= 3;
  579. } else {
  580. m -= (tp->mdev >> 2); /* similar update on mdev */
  581. }
  582. tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
  583. if (tp->mdev > tp->mdev_max) {
  584. tp->mdev_max = tp->mdev;
  585. if (tp->mdev_max > tp->rttvar)
  586. tp->rttvar = tp->mdev_max;
  587. }
  588. if (after(tp->snd_una, tp->rtt_seq)) {
  589. if (tp->mdev_max < tp->rttvar)
  590. tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
  591. tp->rtt_seq = tp->snd_nxt;
  592. tp->mdev_max = tcp_rto_min(sk);
  593. }
  594. } else {
  595. /* no previous measure. */
  596. tp->srtt = m<<3; /* take the measured time to be rtt */
  597. tp->mdev = m<<1; /* make sure rto = 3*rtt */
  598. tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
  599. tp->rtt_seq = tp->snd_nxt;
  600. }
  601. }
  602. /* Calculate rto without backoff. This is the second half of Van Jacobson's
  603. * routine referred to above.
  604. */
  605. static inline void tcp_set_rto(struct sock *sk)
  606. {
  607. const struct tcp_sock *tp = tcp_sk(sk);
  608. /* Old crap is replaced with new one. 8)
  609. *
  610. * More seriously:
  611. * 1. If rtt variance happened to be less 50msec, it is hallucination.
  612. * It cannot be less due to utterly erratic ACK generation made
  613. * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
  614. * to do with delayed acks, because at cwnd>2 true delack timeout
  615. * is invisible. Actually, Linux-2.4 also generates erratic
  616. * ACKs in some circumstances.
  617. */
  618. inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
  619. /* 2. Fixups made earlier cannot be right.
  620. * If we do not estimate RTO correctly without them,
  621. * all the algo is pure shit and should be replaced
  622. * with correct one. It is exactly, which we pretend to do.
  623. */
  624. }
  625. /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
  626. * guarantees that rto is higher.
  627. */
  628. static inline void tcp_bound_rto(struct sock *sk)
  629. {
  630. if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
  631. inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
  632. }
  633. /* Save metrics learned by this TCP session.
  634. This function is called only, when TCP finishes successfully
  635. i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
  636. */
  637. void tcp_update_metrics(struct sock *sk)
  638. {
  639. struct tcp_sock *tp = tcp_sk(sk);
  640. struct dst_entry *dst = __sk_dst_get(sk);
  641. if (sysctl_tcp_nometrics_save)
  642. return;
  643. dst_confirm(dst);
  644. if (dst && (dst->flags&DST_HOST)) {
  645. const struct inet_connection_sock *icsk = inet_csk(sk);
  646. int m;
  647. if (icsk->icsk_backoff || !tp->srtt) {
  648. /* This session failed to estimate rtt. Why?
  649. * Probably, no packets returned in time.
  650. * Reset our results.
  651. */
  652. if (!(dst_metric_locked(dst, RTAX_RTT)))
  653. dst->metrics[RTAX_RTT-1] = 0;
  654. return;
  655. }
  656. m = dst_metric(dst, RTAX_RTT) - tp->srtt;
  657. /* If newly calculated rtt larger than stored one,
  658. * store new one. Otherwise, use EWMA. Remember,
  659. * rtt overestimation is always better than underestimation.
  660. */
  661. if (!(dst_metric_locked(dst, RTAX_RTT))) {
  662. if (m <= 0)
  663. dst->metrics[RTAX_RTT-1] = tp->srtt;
  664. else
  665. dst->metrics[RTAX_RTT-1] -= (m>>3);
  666. }
  667. if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
  668. if (m < 0)
  669. m = -m;
  670. /* Scale deviation to rttvar fixed point */
  671. m >>= 1;
  672. if (m < tp->mdev)
  673. m = tp->mdev;
  674. if (m >= dst_metric(dst, RTAX_RTTVAR))
  675. dst->metrics[RTAX_RTTVAR-1] = m;
  676. else
  677. dst->metrics[RTAX_RTTVAR-1] -=
  678. (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
  679. }
  680. if (tp->snd_ssthresh >= 0xFFFF) {
  681. /* Slow start still did not finish. */
  682. if (dst_metric(dst, RTAX_SSTHRESH) &&
  683. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  684. (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
  685. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
  686. if (!dst_metric_locked(dst, RTAX_CWND) &&
  687. tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
  688. dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
  689. } else if (tp->snd_cwnd > tp->snd_ssthresh &&
  690. icsk->icsk_ca_state == TCP_CA_Open) {
  691. /* Cong. avoidance phase, cwnd is reliable. */
  692. if (!dst_metric_locked(dst, RTAX_SSTHRESH))
  693. dst->metrics[RTAX_SSTHRESH-1] =
  694. max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
  695. if (!dst_metric_locked(dst, RTAX_CWND))
  696. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
  697. } else {
  698. /* Else slow start did not finish, cwnd is non-sense,
  699. ssthresh may be also invalid.
  700. */
  701. if (!dst_metric_locked(dst, RTAX_CWND))
  702. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
  703. if (dst->metrics[RTAX_SSTHRESH-1] &&
  704. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  705. tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
  706. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
  707. }
  708. if (!dst_metric_locked(dst, RTAX_REORDERING)) {
  709. if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
  710. tp->reordering != sysctl_tcp_reordering)
  711. dst->metrics[RTAX_REORDERING-1] = tp->reordering;
  712. }
  713. }
  714. }
  715. /* Numbers are taken from RFC3390.
  716. *
  717. * John Heffner states:
  718. *
  719. * The RFC specifies a window of no more than 4380 bytes
  720. * unless 2*MSS > 4380. Reading the pseudocode in the RFC
  721. * is a bit misleading because they use a clamp at 4380 bytes
  722. * rather than use a multiplier in the relevant range.
  723. */
  724. __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
  725. {
  726. __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
  727. if (!cwnd) {
  728. if (tp->mss_cache > 1460)
  729. cwnd = 2;
  730. else
  731. cwnd = (tp->mss_cache > 1095) ? 3 : 4;
  732. }
  733. return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
  734. }
  735. /* Set slow start threshold and cwnd not falling to slow start */
  736. void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
  737. {
  738. struct tcp_sock *tp = tcp_sk(sk);
  739. const struct inet_connection_sock *icsk = inet_csk(sk);
  740. tp->prior_ssthresh = 0;
  741. tp->bytes_acked = 0;
  742. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  743. tp->undo_marker = 0;
  744. if (set_ssthresh)
  745. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  746. tp->snd_cwnd = min(tp->snd_cwnd,
  747. tcp_packets_in_flight(tp) + 1U);
  748. tp->snd_cwnd_cnt = 0;
  749. tp->high_seq = tp->snd_nxt;
  750. tp->snd_cwnd_stamp = tcp_time_stamp;
  751. TCP_ECN_queue_cwr(tp);
  752. tcp_set_ca_state(sk, TCP_CA_CWR);
  753. }
  754. }
  755. /*
  756. * Packet counting of FACK is based on in-order assumptions, therefore TCP
  757. * disables it when reordering is detected
  758. */
  759. static void tcp_disable_fack(struct tcp_sock *tp)
  760. {
  761. tp->rx_opt.sack_ok &= ~2;
  762. }
  763. /* Take a notice that peer is sending D-SACKs */
  764. static void tcp_dsack_seen(struct tcp_sock *tp)
  765. {
  766. tp->rx_opt.sack_ok |= 4;
  767. }
  768. /* Initialize metrics on socket. */
  769. static void tcp_init_metrics(struct sock *sk)
  770. {
  771. struct tcp_sock *tp = tcp_sk(sk);
  772. struct dst_entry *dst = __sk_dst_get(sk);
  773. if (dst == NULL)
  774. goto reset;
  775. dst_confirm(dst);
  776. if (dst_metric_locked(dst, RTAX_CWND))
  777. tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
  778. if (dst_metric(dst, RTAX_SSTHRESH)) {
  779. tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
  780. if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
  781. tp->snd_ssthresh = tp->snd_cwnd_clamp;
  782. }
  783. if (dst_metric(dst, RTAX_REORDERING) &&
  784. tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
  785. tcp_disable_fack(tp);
  786. tp->reordering = dst_metric(dst, RTAX_REORDERING);
  787. }
  788. if (dst_metric(dst, RTAX_RTT) == 0)
  789. goto reset;
  790. if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
  791. goto reset;
  792. /* Initial rtt is determined from SYN,SYN-ACK.
  793. * The segment is small and rtt may appear much
  794. * less than real one. Use per-dst memory
  795. * to make it more realistic.
  796. *
  797. * A bit of theory. RTT is time passed after "normal" sized packet
  798. * is sent until it is ACKed. In normal circumstances sending small
  799. * packets force peer to delay ACKs and calculation is correct too.
  800. * The algorithm is adaptive and, provided we follow specs, it
  801. * NEVER underestimate RTT. BUT! If peer tries to make some clever
  802. * tricks sort of "quick acks" for time long enough to decrease RTT
  803. * to low value, and then abruptly stops to do it and starts to delay
  804. * ACKs, wait for troubles.
  805. */
  806. if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
  807. tp->srtt = dst_metric(dst, RTAX_RTT);
  808. tp->rtt_seq = tp->snd_nxt;
  809. }
  810. if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
  811. tp->mdev = dst_metric(dst, RTAX_RTTVAR);
  812. tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
  813. }
  814. tcp_set_rto(sk);
  815. tcp_bound_rto(sk);
  816. if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
  817. goto reset;
  818. tp->snd_cwnd = tcp_init_cwnd(tp, dst);
  819. tp->snd_cwnd_stamp = tcp_time_stamp;
  820. return;
  821. reset:
  822. /* Play conservative. If timestamps are not
  823. * supported, TCP will fail to recalculate correct
  824. * rtt, if initial rto is too small. FORGET ALL AND RESET!
  825. */
  826. if (!tp->rx_opt.saw_tstamp && tp->srtt) {
  827. tp->srtt = 0;
  828. tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
  829. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  830. }
  831. }
  832. static void tcp_update_reordering(struct sock *sk, const int metric,
  833. const int ts)
  834. {
  835. struct tcp_sock *tp = tcp_sk(sk);
  836. if (metric > tp->reordering) {
  837. tp->reordering = min(TCP_MAX_REORDERING, metric);
  838. /* This exciting event is worth to be remembered. 8) */
  839. if (ts)
  840. NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
  841. else if (tcp_is_reno(tp))
  842. NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
  843. else if (tcp_is_fack(tp))
  844. NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
  845. else
  846. NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
  847. #if FASTRETRANS_DEBUG > 1
  848. printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
  849. tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
  850. tp->reordering,
  851. tp->fackets_out,
  852. tp->sacked_out,
  853. tp->undo_marker ? tp->undo_retrans : 0);
  854. #endif
  855. tcp_disable_fack(tp);
  856. }
  857. }
  858. /* This procedure tags the retransmission queue when SACKs arrive.
  859. *
  860. * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
  861. * Packets in queue with these bits set are counted in variables
  862. * sacked_out, retrans_out and lost_out, correspondingly.
  863. *
  864. * Valid combinations are:
  865. * Tag InFlight Description
  866. * 0 1 - orig segment is in flight.
  867. * S 0 - nothing flies, orig reached receiver.
  868. * L 0 - nothing flies, orig lost by net.
  869. * R 2 - both orig and retransmit are in flight.
  870. * L|R 1 - orig is lost, retransmit is in flight.
  871. * S|R 1 - orig reached receiver, retrans is still in flight.
  872. * (L|S|R is logically valid, it could occur when L|R is sacked,
  873. * but it is equivalent to plain S and code short-curcuits it to S.
  874. * L|S is logically invalid, it would mean -1 packet in flight 8))
  875. *
  876. * These 6 states form finite state machine, controlled by the following events:
  877. * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
  878. * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
  879. * 3. Loss detection event of one of three flavors:
  880. * A. Scoreboard estimator decided the packet is lost.
  881. * A'. Reno "three dupacks" marks head of queue lost.
  882. * A''. Its FACK modfication, head until snd.fack is lost.
  883. * B. SACK arrives sacking data transmitted after never retransmitted
  884. * hole was sent out.
  885. * C. SACK arrives sacking SND.NXT at the moment, when the
  886. * segment was retransmitted.
  887. * 4. D-SACK added new rule: D-SACK changes any tag to S.
  888. *
  889. * It is pleasant to note, that state diagram turns out to be commutative,
  890. * so that we are allowed not to be bothered by order of our actions,
  891. * when multiple events arrive simultaneously. (see the function below).
  892. *
  893. * Reordering detection.
  894. * --------------------
  895. * Reordering metric is maximal distance, which a packet can be displaced
  896. * in packet stream. With SACKs we can estimate it:
  897. *
  898. * 1. SACK fills old hole and the corresponding segment was not
  899. * ever retransmitted -> reordering. Alas, we cannot use it
  900. * when segment was retransmitted.
  901. * 2. The last flaw is solved with D-SACK. D-SACK arrives
  902. * for retransmitted and already SACKed segment -> reordering..
  903. * Both of these heuristics are not used in Loss state, when we cannot
  904. * account for retransmits accurately.
  905. *
  906. * SACK block validation.
  907. * ----------------------
  908. *
  909. * SACK block range validation checks that the received SACK block fits to
  910. * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
  911. * Note that SND.UNA is not included to the range though being valid because
  912. * it means that the receiver is rather inconsistent with itself reporting
  913. * SACK reneging when it should advance SND.UNA. Such SACK block this is
  914. * perfectly valid, however, in light of RFC2018 which explicitly states
  915. * that "SACK block MUST reflect the newest segment. Even if the newest
  916. * segment is going to be discarded ...", not that it looks very clever
  917. * in case of head skb. Due to potentional receiver driven attacks, we
  918. * choose to avoid immediate execution of a walk in write queue due to
  919. * reneging and defer head skb's loss recovery to standard loss recovery
  920. * procedure that will eventually trigger (nothing forbids us doing this).
  921. *
  922. * Implements also blockage to start_seq wrap-around. Problem lies in the
  923. * fact that though start_seq (s) is before end_seq (i.e., not reversed),
  924. * there's no guarantee that it will be before snd_nxt (n). The problem
  925. * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
  926. * wrap (s_w):
  927. *
  928. * <- outs wnd -> <- wrapzone ->
  929. * u e n u_w e_w s n_w
  930. * | | | | | | |
  931. * |<------------+------+----- TCP seqno space --------------+---------->|
  932. * ...-- <2^31 ->| |<--------...
  933. * ...---- >2^31 ------>| |<--------...
  934. *
  935. * Current code wouldn't be vulnerable but it's better still to discard such
  936. * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
  937. * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
  938. * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
  939. * equal to the ideal case (infinite seqno space without wrap caused issues).
  940. *
  941. * With D-SACK the lower bound is extended to cover sequence space below
  942. * SND.UNA down to undo_marker, which is the last point of interest. Yet
  943. * again, D-SACK block must not to go across snd_una (for the same reason as
  944. * for the normal SACK blocks, explained above). But there all simplicity
  945. * ends, TCP might receive valid D-SACKs below that. As long as they reside
  946. * fully below undo_marker they do not affect behavior in anyway and can
  947. * therefore be safely ignored. In rare cases (which are more or less
  948. * theoretical ones), the D-SACK will nicely cross that boundary due to skb
  949. * fragmentation and packet reordering past skb's retransmission. To consider
  950. * them correctly, the acceptable range must be extended even more though
  951. * the exact amount is rather hard to quantify. However, tp->max_window can
  952. * be used as an exaggerated estimate.
  953. */
  954. static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack,
  955. u32 start_seq, u32 end_seq)
  956. {
  957. /* Too far in future, or reversed (interpretation is ambiguous) */
  958. if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
  959. return 0;
  960. /* Nasty start_seq wrap-around check (see comments above) */
  961. if (!before(start_seq, tp->snd_nxt))
  962. return 0;
  963. /* In outstanding window? ...This is valid exit for D-SACKs too.
  964. * start_seq == snd_una is non-sensical (see comments above)
  965. */
  966. if (after(start_seq, tp->snd_una))
  967. return 1;
  968. if (!is_dsack || !tp->undo_marker)
  969. return 0;
  970. /* ...Then it's D-SACK, and must reside below snd_una completely */
  971. if (!after(end_seq, tp->snd_una))
  972. return 0;
  973. if (!before(start_seq, tp->undo_marker))
  974. return 1;
  975. /* Too old */
  976. if (!after(end_seq, tp->undo_marker))
  977. return 0;
  978. /* Undo_marker boundary crossing (overestimates a lot). Known already:
  979. * start_seq < undo_marker and end_seq >= undo_marker.
  980. */
  981. return !before(start_seq, end_seq - tp->max_window);
  982. }
  983. /* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
  984. * Event "C". Later note: FACK people cheated me again 8), we have to account
  985. * for reordering! Ugly, but should help.
  986. *
  987. * Search retransmitted skbs from write_queue that were sent when snd_nxt was
  988. * less than what is now known to be received by the other end (derived from
  989. * SACK blocks by the caller). Also calculate the lowest snd_nxt among the
  990. * remaining retransmitted skbs to avoid some costly processing per ACKs.
  991. */
  992. static int tcp_mark_lost_retrans(struct sock *sk, u32 received_upto)
  993. {
  994. struct tcp_sock *tp = tcp_sk(sk);
  995. struct sk_buff *skb;
  996. int flag = 0;
  997. int cnt = 0;
  998. u32 new_low_seq = tp->snd_nxt;
  999. tcp_for_write_queue(skb, sk) {
  1000. u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
  1001. if (skb == tcp_send_head(sk))
  1002. break;
  1003. if (cnt == tp->retrans_out)
  1004. break;
  1005. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1006. continue;
  1007. if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
  1008. continue;
  1009. if (after(received_upto, ack_seq) &&
  1010. (tcp_is_fack(tp) ||
  1011. !before(received_upto,
  1012. ack_seq + tp->reordering * tp->mss_cache))) {
  1013. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1014. tp->retrans_out -= tcp_skb_pcount(skb);
  1015. /* clear lost hint */
  1016. tp->retransmit_skb_hint = NULL;
  1017. if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  1018. tp->lost_out += tcp_skb_pcount(skb);
  1019. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1020. flag |= FLAG_DATA_SACKED;
  1021. NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
  1022. }
  1023. } else {
  1024. if (before(ack_seq, new_low_seq))
  1025. new_low_seq = ack_seq;
  1026. cnt += tcp_skb_pcount(skb);
  1027. }
  1028. }
  1029. if (tp->retrans_out)
  1030. tp->lost_retrans_low = new_low_seq;
  1031. return flag;
  1032. }
  1033. static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
  1034. struct tcp_sack_block_wire *sp, int num_sacks,
  1035. u32 prior_snd_una)
  1036. {
  1037. u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
  1038. u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
  1039. int dup_sack = 0;
  1040. if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
  1041. dup_sack = 1;
  1042. tcp_dsack_seen(tp);
  1043. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
  1044. } else if (num_sacks > 1) {
  1045. u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
  1046. u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
  1047. if (!after(end_seq_0, end_seq_1) &&
  1048. !before(start_seq_0, start_seq_1)) {
  1049. dup_sack = 1;
  1050. tcp_dsack_seen(tp);
  1051. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
  1052. }
  1053. }
  1054. /* D-SACK for already forgotten data... Do dumb counting. */
  1055. if (dup_sack &&
  1056. !after(end_seq_0, prior_snd_una) &&
  1057. after(end_seq_0, tp->undo_marker))
  1058. tp->undo_retrans--;
  1059. return dup_sack;
  1060. }
  1061. /* Check if skb is fully within the SACK block. In presence of GSO skbs,
  1062. * the incoming SACK may not exactly match but we can find smaller MSS
  1063. * aligned portion of it that matches. Therefore we might need to fragment
  1064. * which may fail and creates some hassle (caller must handle error case
  1065. * returns).
  1066. */
  1067. static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
  1068. u32 start_seq, u32 end_seq)
  1069. {
  1070. int in_sack, err;
  1071. unsigned int pkt_len;
  1072. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  1073. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  1074. if (tcp_skb_pcount(skb) > 1 && !in_sack &&
  1075. after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
  1076. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
  1077. if (!in_sack)
  1078. pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
  1079. else
  1080. pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
  1081. err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size);
  1082. if (err < 0)
  1083. return err;
  1084. }
  1085. return in_sack;
  1086. }
  1087. static int
  1088. tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
  1089. {
  1090. const struct inet_connection_sock *icsk = inet_csk(sk);
  1091. struct tcp_sock *tp = tcp_sk(sk);
  1092. unsigned char *ptr = (skb_transport_header(ack_skb) +
  1093. TCP_SKB_CB(ack_skb)->sacked);
  1094. struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
  1095. struct sk_buff *cached_skb;
  1096. int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
  1097. int reord = tp->packets_out;
  1098. int prior_fackets;
  1099. u32 highest_sack_end_seq = tp->lost_retrans_low;
  1100. int flag = 0;
  1101. int found_dup_sack = 0;
  1102. int cached_fack_count;
  1103. int i;
  1104. int first_sack_index;
  1105. int force_one_sack;
  1106. if (!tp->sacked_out) {
  1107. if (WARN_ON(tp->fackets_out))
  1108. tp->fackets_out = 0;
  1109. tp->highest_sack = tp->snd_una;
  1110. }
  1111. prior_fackets = tp->fackets_out;
  1112. found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
  1113. num_sacks, prior_snd_una);
  1114. if (found_dup_sack)
  1115. flag |= FLAG_DSACKING_ACK;
  1116. /* Eliminate too old ACKs, but take into
  1117. * account more or less fresh ones, they can
  1118. * contain valid SACK info.
  1119. */
  1120. if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
  1121. return 0;
  1122. if (!tp->packets_out)
  1123. goto out;
  1124. /* SACK fastpath:
  1125. * if the only SACK change is the increase of the end_seq of
  1126. * the first block then only apply that SACK block
  1127. * and use retrans queue hinting otherwise slowpath */
  1128. force_one_sack = 1;
  1129. for (i = 0; i < num_sacks; i++) {
  1130. __be32 start_seq = sp[i].start_seq;
  1131. __be32 end_seq = sp[i].end_seq;
  1132. if (i == 0) {
  1133. if (tp->recv_sack_cache[i].start_seq != start_seq)
  1134. force_one_sack = 0;
  1135. } else {
  1136. if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
  1137. (tp->recv_sack_cache[i].end_seq != end_seq))
  1138. force_one_sack = 0;
  1139. }
  1140. tp->recv_sack_cache[i].start_seq = start_seq;
  1141. tp->recv_sack_cache[i].end_seq = end_seq;
  1142. }
  1143. /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
  1144. for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
  1145. tp->recv_sack_cache[i].start_seq = 0;
  1146. tp->recv_sack_cache[i].end_seq = 0;
  1147. }
  1148. first_sack_index = 0;
  1149. if (force_one_sack)
  1150. num_sacks = 1;
  1151. else {
  1152. int j;
  1153. tp->fastpath_skb_hint = NULL;
  1154. /* order SACK blocks to allow in order walk of the retrans queue */
  1155. for (i = num_sacks-1; i > 0; i--) {
  1156. for (j = 0; j < i; j++){
  1157. if (after(ntohl(sp[j].start_seq),
  1158. ntohl(sp[j+1].start_seq))){
  1159. struct tcp_sack_block_wire tmp;
  1160. tmp = sp[j];
  1161. sp[j] = sp[j+1];
  1162. sp[j+1] = tmp;
  1163. /* Track where the first SACK block goes to */
  1164. if (j == first_sack_index)
  1165. first_sack_index = j+1;
  1166. }
  1167. }
  1168. }
  1169. }
  1170. /* Use SACK fastpath hint if valid */
  1171. cached_skb = tp->fastpath_skb_hint;
  1172. cached_fack_count = tp->fastpath_cnt_hint;
  1173. if (!cached_skb) {
  1174. cached_skb = tcp_write_queue_head(sk);
  1175. cached_fack_count = 0;
  1176. }
  1177. for (i = 0; i < num_sacks; i++) {
  1178. struct sk_buff *skb;
  1179. __u32 start_seq = ntohl(sp->start_seq);
  1180. __u32 end_seq = ntohl(sp->end_seq);
  1181. int fack_count;
  1182. int dup_sack = (found_dup_sack && (i == first_sack_index));
  1183. int next_dup = (found_dup_sack && (i+1 == first_sack_index));
  1184. sp++;
  1185. if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) {
  1186. if (dup_sack) {
  1187. if (!tp->undo_marker)
  1188. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO);
  1189. else
  1190. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD);
  1191. } else {
  1192. /* Don't count olds caused by ACK reordering */
  1193. if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
  1194. !after(end_seq, tp->snd_una))
  1195. continue;
  1196. NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD);
  1197. }
  1198. continue;
  1199. }
  1200. skb = cached_skb;
  1201. fack_count = cached_fack_count;
  1202. /* Event "B" in the comment above. */
  1203. if (after(end_seq, tp->high_seq))
  1204. flag |= FLAG_DATA_LOST;
  1205. tcp_for_write_queue_from(skb, sk) {
  1206. int in_sack = 0;
  1207. u8 sacked;
  1208. if (skb == tcp_send_head(sk))
  1209. break;
  1210. cached_skb = skb;
  1211. cached_fack_count = fack_count;
  1212. if (i == first_sack_index) {
  1213. tp->fastpath_skb_hint = skb;
  1214. tp->fastpath_cnt_hint = fack_count;
  1215. }
  1216. /* The retransmission queue is always in order, so
  1217. * we can short-circuit the walk early.
  1218. */
  1219. if (!before(TCP_SKB_CB(skb)->seq, end_seq))
  1220. break;
  1221. dup_sack = (found_dup_sack && (i == first_sack_index));
  1222. /* Due to sorting DSACK may reside within this SACK block! */
  1223. if (next_dup) {
  1224. u32 dup_start = ntohl(sp->start_seq);
  1225. u32 dup_end = ntohl(sp->end_seq);
  1226. if (before(TCP_SKB_CB(skb)->seq, dup_end)) {
  1227. in_sack = tcp_match_skb_to_sack(sk, skb, dup_start, dup_end);
  1228. if (in_sack > 0)
  1229. dup_sack = 1;
  1230. }
  1231. }
  1232. /* DSACK info lost if out-of-mem, try SACK still */
  1233. if (in_sack <= 0)
  1234. in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, end_seq);
  1235. if (unlikely(in_sack < 0))
  1236. break;
  1237. if (!in_sack) {
  1238. fack_count += tcp_skb_pcount(skb);
  1239. continue;
  1240. }
  1241. sacked = TCP_SKB_CB(skb)->sacked;
  1242. /* Account D-SACK for retransmitted packet. */
  1243. if (dup_sack && (sacked & TCPCB_RETRANS)) {
  1244. if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
  1245. tp->undo_retrans--;
  1246. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una) &&
  1247. (sacked & TCPCB_SACKED_ACKED))
  1248. reord = min(fack_count, reord);
  1249. }
  1250. /* Nothing to do; acked frame is about to be dropped (was ACKed). */
  1251. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
  1252. fack_count += tcp_skb_pcount(skb);
  1253. continue;
  1254. }
  1255. if (!(sacked&TCPCB_SACKED_ACKED)) {
  1256. if (sacked & TCPCB_SACKED_RETRANS) {
  1257. /* If the segment is not tagged as lost,
  1258. * we do not clear RETRANS, believing
  1259. * that retransmission is still in flight.
  1260. */
  1261. if (sacked & TCPCB_LOST) {
  1262. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1263. tp->lost_out -= tcp_skb_pcount(skb);
  1264. tp->retrans_out -= tcp_skb_pcount(skb);
  1265. /* clear lost hint */
  1266. tp->retransmit_skb_hint = NULL;
  1267. }
  1268. } else {
  1269. if (!(sacked & TCPCB_RETRANS)) {
  1270. /* New sack for not retransmitted frame,
  1271. * which was in hole. It is reordering.
  1272. */
  1273. if (fack_count < prior_fackets)
  1274. reord = min(fack_count, reord);
  1275. /* SACK enhanced F-RTO (RFC4138; Appendix B) */
  1276. if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark))
  1277. flag |= FLAG_ONLY_ORIG_SACKED;
  1278. }
  1279. if (sacked & TCPCB_LOST) {
  1280. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1281. tp->lost_out -= tcp_skb_pcount(skb);
  1282. /* clear lost hint */
  1283. tp->retransmit_skb_hint = NULL;
  1284. }
  1285. }
  1286. TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
  1287. flag |= FLAG_DATA_SACKED;
  1288. tp->sacked_out += tcp_skb_pcount(skb);
  1289. fack_count += tcp_skb_pcount(skb);
  1290. if (fack_count > tp->fackets_out)
  1291. tp->fackets_out = fack_count;
  1292. if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) {
  1293. tp->highest_sack = TCP_SKB_CB(skb)->seq;
  1294. highest_sack_end_seq = TCP_SKB_CB(skb)->end_seq;
  1295. }
  1296. } else {
  1297. if (dup_sack && (sacked&TCPCB_RETRANS))
  1298. reord = min(fack_count, reord);
  1299. fack_count += tcp_skb_pcount(skb);
  1300. }
  1301. /* D-SACK. We can detect redundant retransmission
  1302. * in S|R and plain R frames and clear it.
  1303. * undo_retrans is decreased above, L|R frames
  1304. * are accounted above as well.
  1305. */
  1306. if (dup_sack &&
  1307. (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
  1308. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1309. tp->retrans_out -= tcp_skb_pcount(skb);
  1310. tp->retransmit_skb_hint = NULL;
  1311. }
  1312. }
  1313. /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct
  1314. * due to in-order walk
  1315. */
  1316. if (after(end_seq, tp->frto_highmark))
  1317. flag &= ~FLAG_ONLY_ORIG_SACKED;
  1318. }
  1319. if (tp->retrans_out &&
  1320. after(highest_sack_end_seq, tp->lost_retrans_low) &&
  1321. icsk->icsk_ca_state == TCP_CA_Recovery)
  1322. flag |= tcp_mark_lost_retrans(sk, highest_sack_end_seq);
  1323. tcp_verify_left_out(tp);
  1324. if ((reord < tp->fackets_out) &&
  1325. ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) &&
  1326. (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
  1327. tcp_update_reordering(sk, tp->fackets_out - reord, 0);
  1328. out:
  1329. #if FASTRETRANS_DEBUG > 0
  1330. BUG_TRAP((int)tp->sacked_out >= 0);
  1331. BUG_TRAP((int)tp->lost_out >= 0);
  1332. BUG_TRAP((int)tp->retrans_out >= 0);
  1333. BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
  1334. #endif
  1335. return flag;
  1336. }
  1337. /* If we receive more dupacks than we expected counting segments
  1338. * in assumption of absent reordering, interpret this as reordering.
  1339. * The only another reason could be bug in receiver TCP.
  1340. */
  1341. static void tcp_check_reno_reordering(struct sock *sk, const int addend)
  1342. {
  1343. struct tcp_sock *tp = tcp_sk(sk);
  1344. u32 holes;
  1345. holes = max(tp->lost_out, 1U);
  1346. holes = min(holes, tp->packets_out);
  1347. if ((tp->sacked_out + holes) > tp->packets_out) {
  1348. tp->sacked_out = tp->packets_out - holes;
  1349. tcp_update_reordering(sk, tp->packets_out + addend, 0);
  1350. }
  1351. }
  1352. /* Emulate SACKs for SACKless connection: account for a new dupack. */
  1353. static void tcp_add_reno_sack(struct sock *sk)
  1354. {
  1355. struct tcp_sock *tp = tcp_sk(sk);
  1356. tp->sacked_out++;
  1357. tcp_check_reno_reordering(sk, 0);
  1358. tcp_verify_left_out(tp);
  1359. }
  1360. /* Account for ACK, ACKing some data in Reno Recovery phase. */
  1361. static void tcp_remove_reno_sacks(struct sock *sk, int acked)
  1362. {
  1363. struct tcp_sock *tp = tcp_sk(sk);
  1364. if (acked > 0) {
  1365. /* One ACK acked hole. The rest eat duplicate ACKs. */
  1366. if (acked-1 >= tp->sacked_out)
  1367. tp->sacked_out = 0;
  1368. else
  1369. tp->sacked_out -= acked-1;
  1370. }
  1371. tcp_check_reno_reordering(sk, acked);
  1372. tcp_verify_left_out(tp);
  1373. }
  1374. static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
  1375. {
  1376. tp->sacked_out = 0;
  1377. }
  1378. /* F-RTO can only be used if TCP has never retransmitted anything other than
  1379. * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
  1380. */
  1381. int tcp_use_frto(struct sock *sk)
  1382. {
  1383. const struct tcp_sock *tp = tcp_sk(sk);
  1384. struct sk_buff *skb;
  1385. if (!sysctl_tcp_frto)
  1386. return 0;
  1387. if (IsSackFrto())
  1388. return 1;
  1389. /* Avoid expensive walking of rexmit queue if possible */
  1390. if (tp->retrans_out > 1)
  1391. return 0;
  1392. skb = tcp_write_queue_head(sk);
  1393. skb = tcp_write_queue_next(sk, skb); /* Skips head */
  1394. tcp_for_write_queue_from(skb, sk) {
  1395. if (skb == tcp_send_head(sk))
  1396. break;
  1397. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1398. return 0;
  1399. /* Short-circuit when first non-SACKed skb has been checked */
  1400. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
  1401. break;
  1402. }
  1403. return 1;
  1404. }
  1405. /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
  1406. * recovery a bit and use heuristics in tcp_process_frto() to detect if
  1407. * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
  1408. * keep retrans_out counting accurate (with SACK F-RTO, other than head
  1409. * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
  1410. * bits are handled if the Loss state is really to be entered (in
  1411. * tcp_enter_frto_loss).
  1412. *
  1413. * Do like tcp_enter_loss() would; when RTO expires the second time it
  1414. * does:
  1415. * "Reduce ssthresh if it has not yet been made inside this window."
  1416. */
  1417. void tcp_enter_frto(struct sock *sk)
  1418. {
  1419. const struct inet_connection_sock *icsk = inet_csk(sk);
  1420. struct tcp_sock *tp = tcp_sk(sk);
  1421. struct sk_buff *skb;
  1422. if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
  1423. tp->snd_una == tp->high_seq ||
  1424. ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
  1425. !icsk->icsk_retransmits)) {
  1426. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1427. /* Our state is too optimistic in ssthresh() call because cwnd
  1428. * is not reduced until tcp_enter_frto_loss() when previous F-RTO
  1429. * recovery has not yet completed. Pattern would be this: RTO,
  1430. * Cumulative ACK, RTO (2xRTO for the same segment does not end
  1431. * up here twice).
  1432. * RFC4138 should be more specific on what to do, even though
  1433. * RTO is quite unlikely to occur after the first Cumulative ACK
  1434. * due to back-off and complexity of triggering events ...
  1435. */
  1436. if (tp->frto_counter) {
  1437. u32 stored_cwnd;
  1438. stored_cwnd = tp->snd_cwnd;
  1439. tp->snd_cwnd = 2;
  1440. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1441. tp->snd_cwnd = stored_cwnd;
  1442. } else {
  1443. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1444. }
  1445. /* ... in theory, cong.control module could do "any tricks" in
  1446. * ssthresh(), which means that ca_state, lost bits and lost_out
  1447. * counter would have to be faked before the call occurs. We
  1448. * consider that too expensive, unlikely and hacky, so modules
  1449. * using these in ssthresh() must deal these incompatibility
  1450. * issues if they receives CA_EVENT_FRTO and frto_counter != 0
  1451. */
  1452. tcp_ca_event(sk, CA_EVENT_FRTO);
  1453. }
  1454. tp->undo_marker = tp->snd_una;
  1455. tp->undo_retrans = 0;
  1456. skb = tcp_write_queue_head(sk);
  1457. if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
  1458. tp->undo_marker = 0;
  1459. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  1460. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1461. tp->retrans_out -= tcp_skb_pcount(skb);
  1462. }
  1463. tcp_verify_left_out(tp);
  1464. /* Too bad if TCP was application limited */
  1465. tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1);
  1466. /* Earlier loss recovery underway (see RFC4138; Appendix B).
  1467. * The last condition is necessary at least in tp->frto_counter case.
  1468. */
  1469. if (IsSackFrto() && (tp->frto_counter ||
  1470. ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
  1471. after(tp->high_seq, tp->snd_una)) {
  1472. tp->frto_highmark = tp->high_seq;
  1473. } else {
  1474. tp->frto_highmark = tp->snd_nxt;
  1475. }
  1476. tcp_set_ca_state(sk, TCP_CA_Disorder);
  1477. tp->high_seq = tp->snd_nxt;
  1478. tp->frto_counter = 1;
  1479. }
  1480. /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
  1481. * which indicates that we should follow the traditional RTO recovery,
  1482. * i.e. mark everything lost and do go-back-N retransmission.
  1483. */
  1484. static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
  1485. {
  1486. struct tcp_sock *tp = tcp_sk(sk);
  1487. struct sk_buff *skb;
  1488. tp->lost_out = 0;
  1489. tp->retrans_out = 0;
  1490. if (tcp_is_reno(tp))
  1491. tcp_reset_reno_sack(tp);
  1492. tcp_for_write_queue(skb, sk) {
  1493. if (skb == tcp_send_head(sk))
  1494. break;
  1495. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1496. /*
  1497. * Count the retransmission made on RTO correctly (only when
  1498. * waiting for the first ACK and did not get it)...
  1499. */
  1500. if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
  1501. /* For some reason this R-bit might get cleared? */
  1502. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  1503. tp->retrans_out += tcp_skb_pcount(skb);
  1504. /* ...enter this if branch just for the first segment */
  1505. flag |= FLAG_DATA_ACKED;
  1506. } else {
  1507. if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS)
  1508. tp->undo_marker = 0;
  1509. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1510. }
  1511. /* Don't lost mark skbs that were fwd transmitted after RTO */
  1512. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
  1513. !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
  1514. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1515. tp->lost_out += tcp_skb_pcount(skb);
  1516. }
  1517. }
  1518. tcp_verify_left_out(tp);
  1519. tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
  1520. tp->snd_cwnd_cnt = 0;
  1521. tp->snd_cwnd_stamp = tcp_time_stamp;
  1522. tp->frto_counter = 0;
  1523. tp->bytes_acked = 0;
  1524. tp->reordering = min_t(unsigned int, tp->reordering,
  1525. sysctl_tcp_reordering);
  1526. tcp_set_ca_state(sk, TCP_CA_Loss);
  1527. tp->high_seq = tp->frto_highmark;
  1528. TCP_ECN_queue_cwr(tp);
  1529. tcp_clear_retrans_hints_partial(tp);
  1530. }
  1531. static void tcp_clear_retrans_partial(struct tcp_sock *tp)
  1532. {
  1533. tp->retrans_out = 0;
  1534. tp->lost_out = 0;
  1535. tp->undo_marker = 0;
  1536. tp->undo_retrans = 0;
  1537. }
  1538. void tcp_clear_retrans(struct tcp_sock *tp)
  1539. {
  1540. tcp_clear_retrans_partial(tp);
  1541. tp->fackets_out = 0;
  1542. tp->sacked_out = 0;
  1543. }
  1544. /* Enter Loss state. If "how" is not zero, forget all SACK information
  1545. * and reset tags completely, otherwise preserve SACKs. If receiver
  1546. * dropped its ofo queue, we will know this due to reneging detection.
  1547. */
  1548. void tcp_enter_loss(struct sock *sk, int how)
  1549. {
  1550. const struct inet_connection_sock *icsk = inet_csk(sk);
  1551. struct tcp_sock *tp = tcp_sk(sk);
  1552. struct sk_buff *skb;
  1553. /* Reduce ssthresh if it has not yet been made inside this window. */
  1554. if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
  1555. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1556. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1557. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1558. tcp_ca_event(sk, CA_EVENT_LOSS);
  1559. }
  1560. tp->snd_cwnd = 1;
  1561. tp->snd_cwnd_cnt = 0;
  1562. tp->snd_cwnd_stamp = tcp_time_stamp;
  1563. tp->bytes_acked = 0;
  1564. tcp_clear_retrans_partial(tp);
  1565. if (tcp_is_reno(tp))
  1566. tcp_reset_reno_sack(tp);
  1567. if (!how) {
  1568. /* Push undo marker, if it was plain RTO and nothing
  1569. * was retransmitted. */
  1570. tp->undo_marker = tp->snd_una;
  1571. tcp_clear_retrans_hints_partial(tp);
  1572. } else {
  1573. tp->sacked_out = 0;
  1574. tp->fackets_out = 0;
  1575. tcp_clear_all_retrans_hints(tp);
  1576. }
  1577. tcp_for_write_queue(skb, sk) {
  1578. if (skb == tcp_send_head(sk))
  1579. break;
  1580. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1581. tp->undo_marker = 0;
  1582. TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
  1583. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
  1584. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
  1585. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1586. tp->lost_out += tcp_skb_pcount(skb);
  1587. }
  1588. }
  1589. tcp_verify_left_out(tp);
  1590. tp->reordering = min_t(unsigned int, tp->reordering,
  1591. sysctl_tcp_reordering);
  1592. tcp_set_ca_state(sk, TCP_CA_Loss);
  1593. tp->high_seq = tp->snd_nxt;
  1594. TCP_ECN_queue_cwr(tp);
  1595. /* Abort F-RTO algorithm if one is in progress */
  1596. tp->frto_counter = 0;
  1597. }
  1598. static int tcp_check_sack_reneging(struct sock *sk)
  1599. {
  1600. struct sk_buff *skb;
  1601. /* If ACK arrived pointing to a remembered SACK,
  1602. * it means that our remembered SACKs do not reflect
  1603. * real state of receiver i.e.
  1604. * receiver _host_ is heavily congested (or buggy).
  1605. * Do processing similar to RTO timeout.
  1606. */
  1607. if ((skb = tcp_write_queue_head(sk)) != NULL &&
  1608. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
  1609. struct inet_connection_sock *icsk = inet_csk(sk);
  1610. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
  1611. tcp_enter_loss(sk, 1);
  1612. icsk->icsk_retransmits++;
  1613. tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
  1614. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1615. icsk->icsk_rto, TCP_RTO_MAX);
  1616. return 1;
  1617. }
  1618. return 0;
  1619. }
  1620. static inline int tcp_fackets_out(struct tcp_sock *tp)
  1621. {
  1622. return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out;
  1623. }
  1624. static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
  1625. {
  1626. return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
  1627. }
  1628. static inline int tcp_head_timedout(struct sock *sk)
  1629. {
  1630. struct tcp_sock *tp = tcp_sk(sk);
  1631. return tp->packets_out &&
  1632. tcp_skb_timedout(sk, tcp_write_queue_head(sk));
  1633. }
  1634. /* Linux NewReno/SACK/FACK/ECN state machine.
  1635. * --------------------------------------
  1636. *
  1637. * "Open" Normal state, no dubious events, fast path.
  1638. * "Disorder" In all the respects it is "Open",
  1639. * but requires a bit more attention. It is entered when
  1640. * we see some SACKs or dupacks. It is split of "Open"
  1641. * mainly to move some processing from fast path to slow one.
  1642. * "CWR" CWND was reduced due to some Congestion Notification event.
  1643. * It can be ECN, ICMP source quench, local device congestion.
  1644. * "Recovery" CWND was reduced, we are fast-retransmitting.
  1645. * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
  1646. *
  1647. * tcp_fastretrans_alert() is entered:
  1648. * - each incoming ACK, if state is not "Open"
  1649. * - when arrived ACK is unusual, namely:
  1650. * * SACK
  1651. * * Duplicate ACK.
  1652. * * ECN ECE.
  1653. *
  1654. * Counting packets in flight is pretty simple.
  1655. *
  1656. * in_flight = packets_out - left_out + retrans_out
  1657. *
  1658. * packets_out is SND.NXT-SND.UNA counted in packets.
  1659. *
  1660. * retrans_out is number of retransmitted segments.
  1661. *
  1662. * left_out is number of segments left network, but not ACKed yet.
  1663. *
  1664. * left_out = sacked_out + lost_out
  1665. *
  1666. * sacked_out: Packets, which arrived to receiver out of order
  1667. * and hence not ACKed. With SACKs this number is simply
  1668. * amount of SACKed data. Even without SACKs
  1669. * it is easy to give pretty reliable estimate of this number,
  1670. * counting duplicate ACKs.
  1671. *
  1672. * lost_out: Packets lost by network. TCP has no explicit
  1673. * "loss notification" feedback from network (for now).
  1674. * It means that this number can be only _guessed_.
  1675. * Actually, it is the heuristics to predict lossage that
  1676. * distinguishes different algorithms.
  1677. *
  1678. * F.e. after RTO, when all the queue is considered as lost,
  1679. * lost_out = packets_out and in_flight = retrans_out.
  1680. *
  1681. * Essentially, we have now two algorithms counting
  1682. * lost packets.
  1683. *
  1684. * FACK: It is the simplest heuristics. As soon as we decided
  1685. * that something is lost, we decide that _all_ not SACKed
  1686. * packets until the most forward SACK are lost. I.e.
  1687. * lost_out = fackets_out - sacked_out and left_out = fackets_out.
  1688. * It is absolutely correct estimate, if network does not reorder
  1689. * packets. And it loses any connection to reality when reordering
  1690. * takes place. We use FACK by default until reordering
  1691. * is suspected on the path to this destination.
  1692. *
  1693. * NewReno: when Recovery is entered, we assume that one segment
  1694. * is lost (classic Reno). While we are in Recovery and
  1695. * a partial ACK arrives, we assume that one more packet
  1696. * is lost (NewReno). This heuristics are the same in NewReno
  1697. * and SACK.
  1698. *
  1699. * Imagine, that's all! Forget about all this shamanism about CWND inflation
  1700. * deflation etc. CWND is real congestion window, never inflated, changes
  1701. * only according to classic VJ rules.
  1702. *
  1703. * Really tricky (and requiring careful tuning) part of algorithm
  1704. * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
  1705. * The first determines the moment _when_ we should reduce CWND and,
  1706. * hence, slow down forward transmission. In fact, it determines the moment
  1707. * when we decide that hole is caused by loss, rather than by a reorder.
  1708. *
  1709. * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
  1710. * holes, caused by lost packets.
  1711. *
  1712. * And the most logically complicated part of algorithm is undo
  1713. * heuristics. We detect false retransmits due to both too early
  1714. * fast retransmit (reordering) and underestimated RTO, analyzing
  1715. * timestamps and D-SACKs. When we detect that some segments were
  1716. * retransmitted by mistake and CWND reduction was wrong, we undo
  1717. * window reduction and abort recovery phase. This logic is hidden
  1718. * inside several functions named tcp_try_undo_<something>.
  1719. */
  1720. /* This function decides, when we should leave Disordered state
  1721. * and enter Recovery phase, reducing congestion window.
  1722. *
  1723. * Main question: may we further continue forward transmission
  1724. * with the same cwnd?
  1725. */
  1726. static int tcp_time_to_recover(struct sock *sk)
  1727. {
  1728. struct tcp_sock *tp = tcp_sk(sk);
  1729. __u32 packets_out;
  1730. /* Do not perform any recovery during F-RTO algorithm */
  1731. if (tp->frto_counter)
  1732. return 0;
  1733. /* Trick#1: The loss is proven. */
  1734. if (tp->lost_out)
  1735. return 1;
  1736. /* Not-A-Trick#2 : Classic rule... */
  1737. if (tcp_fackets_out(tp) > tp->reordering)
  1738. return 1;
  1739. /* Trick#3 : when we use RFC2988 timer restart, fast
  1740. * retransmit can be triggered by timeout of queue head.
  1741. */
  1742. if (tcp_head_timedout(sk))
  1743. return 1;
  1744. /* Trick#4: It is still not OK... But will it be useful to delay
  1745. * recovery more?
  1746. */
  1747. packets_out = tp->packets_out;
  1748. if (packets_out <= tp->reordering &&
  1749. tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
  1750. !tcp_may_send_now(sk)) {
  1751. /* We have nothing to send. This connection is limited
  1752. * either by receiver window or by application.
  1753. */
  1754. return 1;
  1755. }
  1756. return 0;
  1757. }
  1758. /* RFC: This is from the original, I doubt that this is necessary at all:
  1759. * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
  1760. * retransmitted past LOST markings in the first place? I'm not fully sure
  1761. * about undo and end of connection cases, which can cause R without L?
  1762. */
  1763. static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
  1764. struct sk_buff *skb)
  1765. {
  1766. if ((tp->retransmit_skb_hint != NULL) &&
  1767. before(TCP_SKB_CB(skb)->seq,
  1768. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  1769. tp->retransmit_skb_hint = NULL;
  1770. }
  1771. /* Mark head of queue up as lost. */
  1772. static void tcp_mark_head_lost(struct sock *sk, int packets)
  1773. {
  1774. struct tcp_sock *tp = tcp_sk(sk);
  1775. struct sk_buff *skb;
  1776. int cnt;
  1777. BUG_TRAP(packets <= tp->packets_out);
  1778. if (tp->lost_skb_hint) {
  1779. skb = tp->lost_skb_hint;
  1780. cnt = tp->lost_cnt_hint;
  1781. } else {
  1782. skb = tcp_write_queue_head(sk);
  1783. cnt = 0;
  1784. }
  1785. tcp_for_write_queue_from(skb, sk) {
  1786. if (skb == tcp_send_head(sk))
  1787. break;
  1788. /* TODO: do this better */
  1789. /* this is not the most efficient way to do this... */
  1790. tp->lost_skb_hint = skb;
  1791. tp->lost_cnt_hint = cnt;
  1792. cnt += tcp_skb_pcount(skb);
  1793. if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, tp->high_seq))
  1794. break;
  1795. if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
  1796. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1797. tp->lost_out += tcp_skb_pcount(skb);
  1798. tcp_verify_retransmit_hint(tp, skb);
  1799. }
  1800. }
  1801. tcp_verify_left_out(tp);
  1802. }
  1803. /* Account newly detected lost packet(s) */
  1804. static void tcp_update_scoreboard(struct sock *sk)
  1805. {
  1806. struct tcp_sock *tp = tcp_sk(sk);
  1807. if (tcp_is_fack(tp)) {
  1808. int lost = tp->fackets_out - tp->reordering;
  1809. if (lost <= 0)
  1810. lost = 1;
  1811. tcp_mark_head_lost(sk, lost);
  1812. } else {
  1813. tcp_mark_head_lost(sk, 1);
  1814. }
  1815. /* New heuristics: it is possible only after we switched
  1816. * to restart timer each time when something is ACKed.
  1817. * Hence, we can detect timed out packets during fast
  1818. * retransmit without falling to slow start.
  1819. */
  1820. if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) {
  1821. struct sk_buff *skb;
  1822. skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
  1823. : tcp_write_queue_head(sk);
  1824. tcp_for_write_queue_from(skb, sk) {
  1825. if (skb == tcp_send_head(sk))
  1826. break;
  1827. if (!tcp_skb_timedout(sk, skb))
  1828. break;
  1829. if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_SACKED_ACKED|TCPCB_LOST))) {
  1830. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1831. tp->lost_out += tcp_skb_pcount(skb);
  1832. tcp_verify_retransmit_hint(tp, skb);
  1833. }
  1834. }
  1835. tp->scoreboard_skb_hint = skb;
  1836. tcp_verify_left_out(tp);
  1837. }
  1838. }
  1839. /* CWND moderation, preventing bursts due to too big ACKs
  1840. * in dubious situations.
  1841. */
  1842. static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
  1843. {
  1844. tp->snd_cwnd = min(tp->snd_cwnd,
  1845. tcp_packets_in_flight(tp)+tcp_max_burst(tp));
  1846. tp->snd_cwnd_stamp = tcp_time_stamp;
  1847. }
  1848. /* Lower bound on congestion window is slow start threshold
  1849. * unless congestion avoidance choice decides to overide it.
  1850. */
  1851. static inline u32 tcp_cwnd_min(const struct sock *sk)
  1852. {
  1853. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1854. return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
  1855. }
  1856. /* Decrease cwnd each second ack. */
  1857. static void tcp_cwnd_down(struct sock *sk, int flag)
  1858. {
  1859. struct tcp_sock *tp = tcp_sk(sk);
  1860. int decr = tp->snd_cwnd_cnt + 1;
  1861. if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
  1862. (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) {
  1863. tp->snd_cwnd_cnt = decr&1;
  1864. decr >>= 1;
  1865. if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
  1866. tp->snd_cwnd -= decr;
  1867. tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
  1868. tp->snd_cwnd_stamp = tcp_time_stamp;
  1869. }
  1870. }
  1871. /* Nothing was retransmitted or returned timestamp is less
  1872. * than timestamp of the first retransmission.
  1873. */
  1874. static inline int tcp_packet_delayed(struct tcp_sock *tp)
  1875. {
  1876. return !tp->retrans_stamp ||
  1877. (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  1878. (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
  1879. }
  1880. /* Undo procedures. */
  1881. #if FASTRETRANS_DEBUG > 1
  1882. static void DBGUNDO(struct sock *sk, const char *msg)
  1883. {
  1884. struct tcp_sock *tp = tcp_sk(sk);
  1885. struct inet_sock *inet = inet_sk(sk);
  1886. printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
  1887. msg,
  1888. NIPQUAD(inet->daddr), ntohs(inet->dport),
  1889. tp->snd_cwnd, tcp_left_out(tp),
  1890. tp->snd_ssthresh, tp->prior_ssthresh,
  1891. tp->packets_out);
  1892. }
  1893. #else
  1894. #define DBGUNDO(x...) do { } while (0)
  1895. #endif
  1896. static void tcp_undo_cwr(struct sock *sk, const int undo)
  1897. {
  1898. struct tcp_sock *tp = tcp_sk(sk);
  1899. if (tp->prior_ssthresh) {
  1900. const struct inet_connection_sock *icsk = inet_csk(sk);
  1901. if (icsk->icsk_ca_ops->undo_cwnd)
  1902. tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
  1903. else
  1904. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
  1905. if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
  1906. tp->snd_ssthresh = tp->prior_ssthresh;
  1907. TCP_ECN_withdraw_cwr(tp);
  1908. }
  1909. } else {
  1910. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
  1911. }
  1912. tcp_moderate_cwnd(tp);
  1913. tp->snd_cwnd_stamp = tcp_time_stamp;
  1914. /* There is something screwy going on with the retrans hints after
  1915. an undo */
  1916. tcp_clear_all_retrans_hints(tp);
  1917. }
  1918. static inline int tcp_may_undo(struct tcp_sock *tp)
  1919. {
  1920. return tp->undo_marker &&
  1921. (!tp->undo_retrans || tcp_packet_delayed(tp));
  1922. }
  1923. /* People celebrate: "We love our President!" */
  1924. static int tcp_try_undo_recovery(struct sock *sk)
  1925. {
  1926. struct tcp_sock *tp = tcp_sk(sk);
  1927. if (tcp_may_undo(tp)) {
  1928. /* Happy end! We did not retransmit anything
  1929. * or our original transmission succeeded.
  1930. */
  1931. DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
  1932. tcp_undo_cwr(sk, 1);
  1933. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
  1934. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1935. else
  1936. NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
  1937. tp->undo_marker = 0;
  1938. }
  1939. if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
  1940. /* Hold old state until something *above* high_seq
  1941. * is ACKed. For Reno it is MUST to prevent false
  1942. * fast retransmits (RFC2582). SACK TCP is safe. */
  1943. tcp_moderate_cwnd(tp);
  1944. return 1;
  1945. }
  1946. tcp_set_ca_state(sk, TCP_CA_Open);
  1947. return 0;
  1948. }
  1949. /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
  1950. static void tcp_try_undo_dsack(struct sock *sk)
  1951. {
  1952. struct tcp_sock *tp = tcp_sk(sk);
  1953. if (tp->undo_marker && !tp->undo_retrans) {
  1954. DBGUNDO(sk, "D-SACK");
  1955. tcp_undo_cwr(sk, 1);
  1956. tp->undo_marker = 0;
  1957. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
  1958. }
  1959. }
  1960. /* Undo during fast recovery after partial ACK. */
  1961. static int tcp_try_undo_partial(struct sock *sk, int acked)
  1962. {
  1963. struct tcp_sock *tp = tcp_sk(sk);
  1964. /* Partial ACK arrived. Force Hoe's retransmit. */
  1965. int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering;
  1966. if (tcp_may_undo(tp)) {
  1967. /* Plain luck! Hole if filled with delayed
  1968. * packet, rather than with a retransmit.
  1969. */
  1970. if (tp->retrans_out == 0)
  1971. tp->retrans_stamp = 0;
  1972. tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
  1973. DBGUNDO(sk, "Hoe");
  1974. tcp_undo_cwr(sk, 0);
  1975. NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
  1976. /* So... Do not make Hoe's retransmit yet.
  1977. * If the first packet was delayed, the rest
  1978. * ones are most probably delayed as well.
  1979. */
  1980. failed = 0;
  1981. }
  1982. return failed;
  1983. }
  1984. /* Undo during loss recovery after partial ACK. */
  1985. static int tcp_try_undo_loss(struct sock *sk)
  1986. {
  1987. struct tcp_sock *tp = tcp_sk(sk);
  1988. if (tcp_may_undo(tp)) {
  1989. struct sk_buff *skb;
  1990. tcp_for_write_queue(skb, sk) {
  1991. if (skb == tcp_send_head(sk))
  1992. break;
  1993. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1994. }
  1995. tcp_clear_all_retrans_hints(tp);
  1996. DBGUNDO(sk, "partial loss");
  1997. tp->lost_out = 0;
  1998. tcp_undo_cwr(sk, 1);
  1999. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  2000. inet_csk(sk)->icsk_retransmits = 0;
  2001. tp->undo_marker = 0;
  2002. if (tcp_is_sack(tp))
  2003. tcp_set_ca_state(sk, TCP_CA_Open);
  2004. return 1;
  2005. }
  2006. return 0;
  2007. }
  2008. static inline void tcp_complete_cwr(struct sock *sk)
  2009. {
  2010. struct tcp_sock *tp = tcp_sk(sk);
  2011. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  2012. tp->snd_cwnd_stamp = tcp_time_stamp;
  2013. tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
  2014. }
  2015. static void tcp_try_to_open(struct sock *sk, int flag)
  2016. {
  2017. struct tcp_sock *tp = tcp_sk(sk);
  2018. tcp_verify_left_out(tp);
  2019. if (tp->retrans_out == 0)
  2020. tp->retrans_stamp = 0;
  2021. if (flag&FLAG_ECE)
  2022. tcp_enter_cwr(sk, 1);
  2023. if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
  2024. int state = TCP_CA_Open;
  2025. if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker)
  2026. state = TCP_CA_Disorder;
  2027. if (inet_csk(sk)->icsk_ca_state != state) {
  2028. tcp_set_ca_state(sk, state);
  2029. tp->high_seq = tp->snd_nxt;
  2030. }
  2031. tcp_moderate_cwnd(tp);
  2032. } else {
  2033. tcp_cwnd_down(sk, flag);
  2034. }
  2035. }
  2036. static void tcp_mtup_probe_failed(struct sock *sk)
  2037. {
  2038. struct inet_connection_sock *icsk = inet_csk(sk);
  2039. icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
  2040. icsk->icsk_mtup.probe_size = 0;
  2041. }
  2042. static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
  2043. {
  2044. struct tcp_sock *tp = tcp_sk(sk);
  2045. struct inet_connection_sock *icsk = inet_csk(sk);
  2046. /* FIXME: breaks with very large cwnd */
  2047. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  2048. tp->snd_cwnd = tp->snd_cwnd *
  2049. tcp_mss_to_mtu(sk, tp->mss_cache) /
  2050. icsk->icsk_mtup.probe_size;
  2051. tp->snd_cwnd_cnt = 0;
  2052. tp->snd_cwnd_stamp = tcp_time_stamp;
  2053. tp->rcv_ssthresh = tcp_current_ssthresh(sk);
  2054. icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
  2055. icsk->icsk_mtup.probe_size = 0;
  2056. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  2057. }
  2058. /* Process an event, which can update packets-in-flight not trivially.
  2059. * Main goal of this function is to calculate new estimate for left_out,
  2060. * taking into account both packets sitting in receiver's buffer and
  2061. * packets lost by network.
  2062. *
  2063. * Besides that it does CWND reduction, when packet loss is detected
  2064. * and changes state of machine.
  2065. *
  2066. * It does _not_ decide what to send, it is made in function
  2067. * tcp_xmit_retransmit_queue().
  2068. */
  2069. static void
  2070. tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag)
  2071. {
  2072. struct inet_connection_sock *icsk = inet_csk(sk);
  2073. struct tcp_sock *tp = tcp_sk(sk);
  2074. int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
  2075. int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
  2076. (tp->fackets_out > tp->reordering));
  2077. /* Some technical things:
  2078. * 1. Reno does not count dupacks (sacked_out) automatically. */
  2079. if (!tp->packets_out)
  2080. tp->sacked_out = 0;
  2081. if (WARN_ON(!tp->sacked_out && tp->fackets_out))
  2082. tp->fackets_out = 0;
  2083. /* Now state machine starts.
  2084. * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
  2085. if (flag&FLAG_ECE)
  2086. tp->prior_ssthresh = 0;
  2087. /* B. In all the states check for reneging SACKs. */
  2088. if (tp->sacked_out && tcp_check_sack_reneging(sk))
  2089. return;
  2090. /* C. Process data loss notification, provided it is valid. */
  2091. if ((flag&FLAG_DATA_LOST) &&
  2092. before(tp->snd_una, tp->high_seq) &&
  2093. icsk->icsk_ca_state != TCP_CA_Open &&
  2094. tp->fackets_out > tp->reordering) {
  2095. tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering);
  2096. NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
  2097. }
  2098. /* D. Check consistency of the current state. */
  2099. tcp_verify_left_out(tp);
  2100. /* E. Check state exit conditions. State can be terminated
  2101. * when high_seq is ACKed. */
  2102. if (icsk->icsk_ca_state == TCP_CA_Open) {
  2103. BUG_TRAP(tp->retrans_out == 0);
  2104. tp->retrans_stamp = 0;
  2105. } else if (!before(tp->snd_una, tp->high_seq)) {
  2106. switch (icsk->icsk_ca_state) {
  2107. case TCP_CA_Loss:
  2108. icsk->icsk_retransmits = 0;
  2109. if (tcp_try_undo_recovery(sk))
  2110. return;
  2111. break;
  2112. case TCP_CA_CWR:
  2113. /* CWR is to be held something *above* high_seq
  2114. * is ACKed for CWR bit to reach receiver. */
  2115. if (tp->snd_una != tp->high_seq) {
  2116. tcp_complete_cwr(sk);
  2117. tcp_set_ca_state(sk, TCP_CA_Open);
  2118. }
  2119. break;
  2120. case TCP_CA_Disorder:
  2121. tcp_try_undo_dsack(sk);
  2122. if (!tp->undo_marker ||
  2123. /* For SACK case do not Open to allow to undo
  2124. * catching for all duplicate ACKs. */
  2125. tcp_is_reno(tp) || tp->snd_una != tp->high_seq) {
  2126. tp->undo_marker = 0;
  2127. tcp_set_ca_state(sk, TCP_CA_Open);
  2128. }
  2129. break;
  2130. case TCP_CA_Recovery:
  2131. if (tcp_is_reno(tp))
  2132. tcp_reset_reno_sack(tp);
  2133. if (tcp_try_undo_recovery(sk))
  2134. return;
  2135. tcp_complete_cwr(sk);
  2136. break;
  2137. }
  2138. }
  2139. /* F. Process state. */
  2140. switch (icsk->icsk_ca_state) {
  2141. case TCP_CA_Recovery:
  2142. if (!(flag & FLAG_SND_UNA_ADVANCED)) {
  2143. if (tcp_is_reno(tp) && is_dupack)
  2144. tcp_add_reno_sack(sk);
  2145. } else
  2146. do_lost = tcp_try_undo_partial(sk, pkts_acked);
  2147. break;
  2148. case TCP_CA_Loss:
  2149. if (flag&FLAG_DATA_ACKED)
  2150. icsk->icsk_retransmits = 0;
  2151. if (!tcp_try_undo_loss(sk)) {
  2152. tcp_moderate_cwnd(tp);
  2153. tcp_xmit_retransmit_queue(sk);
  2154. return;
  2155. }
  2156. if (icsk->icsk_ca_state != TCP_CA_Open)
  2157. return;
  2158. /* Loss is undone; fall through to processing in Open state. */
  2159. default:
  2160. if (tcp_is_reno(tp)) {
  2161. if (flag & FLAG_SND_UNA_ADVANCED)
  2162. tcp_reset_reno_sack(tp);
  2163. if (is_dupack)
  2164. tcp_add_reno_sack(sk);
  2165. }
  2166. if (icsk->icsk_ca_state == TCP_CA_Disorder)
  2167. tcp_try_undo_dsack(sk);
  2168. if (!tcp_time_to_recover(sk)) {
  2169. tcp_try_to_open(sk, flag);
  2170. return;
  2171. }
  2172. /* MTU probe failure: don't reduce cwnd */
  2173. if (icsk->icsk_ca_state < TCP_CA_CWR &&
  2174. icsk->icsk_mtup.probe_size &&
  2175. tp->snd_una == tp->mtu_probe.probe_seq_start) {
  2176. tcp_mtup_probe_failed(sk);
  2177. /* Restores the reduction we did in tcp_mtup_probe() */
  2178. tp->snd_cwnd++;
  2179. tcp_simple_retransmit(sk);
  2180. return;
  2181. }
  2182. /* Otherwise enter Recovery state */
  2183. if (tcp_is_reno(tp))
  2184. NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
  2185. else
  2186. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
  2187. tp->high_seq = tp->snd_nxt;
  2188. tp->prior_ssthresh = 0;
  2189. tp->undo_marker = tp->snd_una;
  2190. tp->undo_retrans = tp->retrans_out;
  2191. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  2192. if (!(flag&FLAG_ECE))
  2193. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  2194. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  2195. TCP_ECN_queue_cwr(tp);
  2196. }
  2197. tp->bytes_acked = 0;
  2198. tp->snd_cwnd_cnt = 0;
  2199. tcp_set_ca_state(sk, TCP_CA_Recovery);
  2200. }
  2201. if (do_lost || tcp_head_timedout(sk))
  2202. tcp_update_scoreboard(sk);
  2203. tcp_cwnd_down(sk, flag);
  2204. tcp_xmit_retransmit_queue(sk);
  2205. }
  2206. /* Read draft-ietf-tcplw-high-performance before mucking
  2207. * with this code. (Supersedes RFC1323)
  2208. */
  2209. static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
  2210. {
  2211. /* RTTM Rule: A TSecr value received in a segment is used to
  2212. * update the averaged RTT measurement only if the segment
  2213. * acknowledges some new data, i.e., only if it advances the
  2214. * left edge of the send window.
  2215. *
  2216. * See draft-ietf-tcplw-high-performance-00, section 3.3.
  2217. * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
  2218. *
  2219. * Changed: reset backoff as soon as we see the first valid sample.
  2220. * If we do not, we get strongly overestimated rto. With timestamps
  2221. * samples are accepted even from very old segments: f.e., when rtt=1
  2222. * increases to 8, we retransmit 5 times and after 8 seconds delayed
  2223. * answer arrives rto becomes 120 seconds! If at least one of segments
  2224. * in window is lost... Voila. --ANK (010210)
  2225. */
  2226. struct tcp_sock *tp = tcp_sk(sk);
  2227. const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
  2228. tcp_rtt_estimator(sk, seq_rtt);
  2229. tcp_set_rto(sk);
  2230. inet_csk(sk)->icsk_backoff = 0;
  2231. tcp_bound_rto(sk);
  2232. }
  2233. static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
  2234. {
  2235. /* We don't have a timestamp. Can only use
  2236. * packets that are not retransmitted to determine
  2237. * rtt estimates. Also, we must not reset the
  2238. * backoff for rto until we get a non-retransmitted
  2239. * packet. This allows us to deal with a situation
  2240. * where the network delay has increased suddenly.
  2241. * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
  2242. */
  2243. if (flag & FLAG_RETRANS_DATA_ACKED)
  2244. return;
  2245. tcp_rtt_estimator(sk, seq_rtt);
  2246. tcp_set_rto(sk);
  2247. inet_csk(sk)->icsk_backoff = 0;
  2248. tcp_bound_rto(sk);
  2249. }
  2250. static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
  2251. const s32 seq_rtt)
  2252. {
  2253. const struct tcp_sock *tp = tcp_sk(sk);
  2254. /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
  2255. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  2256. tcp_ack_saw_tstamp(sk, flag);
  2257. else if (seq_rtt >= 0)
  2258. tcp_ack_no_tstamp(sk, seq_rtt, flag);
  2259. }
  2260. static void tcp_cong_avoid(struct sock *sk, u32 ack,
  2261. u32 in_flight, int good)
  2262. {
  2263. const struct inet_connection_sock *icsk = inet_csk(sk);
  2264. icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
  2265. tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
  2266. }
  2267. /* Restart timer after forward progress on connection.
  2268. * RFC2988 recommends to restart timer to now+rto.
  2269. */
  2270. static void tcp_rearm_rto(struct sock *sk)
  2271. {
  2272. struct tcp_sock *tp = tcp_sk(sk);
  2273. if (!tp->packets_out) {
  2274. inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
  2275. } else {
  2276. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2277. }
  2278. }
  2279. /* If we get here, the whole TSO packet has not been acked. */
  2280. static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
  2281. {
  2282. struct tcp_sock *tp = tcp_sk(sk);
  2283. u32 packets_acked;
  2284. BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
  2285. packets_acked = tcp_skb_pcount(skb);
  2286. if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
  2287. return 0;
  2288. packets_acked -= tcp_skb_pcount(skb);
  2289. if (packets_acked) {
  2290. BUG_ON(tcp_skb_pcount(skb) == 0);
  2291. BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
  2292. }
  2293. return packets_acked;
  2294. }
  2295. /* Remove acknowledged frames from the retransmission queue. If our packet
  2296. * is before the ack sequence we can discard it as it's confirmed to have
  2297. * arrived at the other end.
  2298. */
  2299. static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p,
  2300. int prior_fackets)
  2301. {
  2302. struct tcp_sock *tp = tcp_sk(sk);
  2303. const struct inet_connection_sock *icsk = inet_csk(sk);
  2304. struct sk_buff *skb;
  2305. u32 now = tcp_time_stamp;
  2306. int fully_acked = 1;
  2307. int flag = 0;
  2308. int prior_packets = tp->packets_out;
  2309. u32 cnt = 0;
  2310. u32 reord = tp->packets_out;
  2311. s32 seq_rtt = -1;
  2312. s32 ca_seq_rtt = -1;
  2313. ktime_t last_ackt = net_invalid_timestamp();
  2314. while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
  2315. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2316. u32 end_seq;
  2317. u32 packets_acked;
  2318. u8 sacked = scb->sacked;
  2319. /* Determine how many packets and what bytes were acked, tso and else */
  2320. if (after(scb->end_seq, tp->snd_una)) {
  2321. if (tcp_skb_pcount(skb) == 1 ||
  2322. !after(tp->snd_una, scb->seq))
  2323. break;
  2324. packets_acked = tcp_tso_acked(sk, skb);
  2325. if (!packets_acked)
  2326. break;
  2327. fully_acked = 0;
  2328. end_seq = tp->snd_una;
  2329. } else {
  2330. packets_acked = tcp_skb_pcount(skb);
  2331. end_seq = scb->end_seq;
  2332. }
  2333. /* MTU probing checks */
  2334. if (fully_acked && icsk->icsk_mtup.probe_size &&
  2335. !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) {
  2336. tcp_mtup_probe_success(sk, skb);
  2337. }
  2338. if (sacked) {
  2339. if (sacked & TCPCB_RETRANS) {
  2340. if (sacked & TCPCB_SACKED_RETRANS)
  2341. tp->retrans_out -= packets_acked;
  2342. flag |= FLAG_RETRANS_DATA_ACKED;
  2343. ca_seq_rtt = -1;
  2344. seq_rtt = -1;
  2345. if ((flag & FLAG_DATA_ACKED) ||
  2346. (packets_acked > 1))
  2347. flag |= FLAG_NONHEAD_RETRANS_ACKED;
  2348. } else {
  2349. ca_seq_rtt = now - scb->when;
  2350. last_ackt = skb->tstamp;
  2351. if (seq_rtt < 0) {
  2352. seq_rtt = ca_seq_rtt;
  2353. }
  2354. if (!(sacked & TCPCB_SACKED_ACKED))
  2355. reord = min(cnt, reord);
  2356. }
  2357. if (sacked & TCPCB_SACKED_ACKED)
  2358. tp->sacked_out -= packets_acked;
  2359. if (sacked & TCPCB_LOST)
  2360. tp->lost_out -= packets_acked;
  2361. if ((sacked & TCPCB_URG) && tp->urg_mode &&
  2362. !before(end_seq, tp->snd_up))
  2363. tp->urg_mode = 0;
  2364. } else {
  2365. ca_seq_rtt = now - scb->when;
  2366. last_ackt = skb->tstamp;
  2367. if (seq_rtt < 0) {
  2368. seq_rtt = ca_seq_rtt;
  2369. }
  2370. reord = min(cnt, reord);
  2371. }
  2372. tp->packets_out -= packets_acked;
  2373. cnt += packets_acked;
  2374. /* Initial outgoing SYN's get put onto the write_queue
  2375. * just like anything else we transmit. It is not
  2376. * true data, and if we misinform our callers that
  2377. * this ACK acks real data, we will erroneously exit
  2378. * connection startup slow start one packet too
  2379. * quickly. This is severely frowned upon behavior.
  2380. */
  2381. if (!(scb->flags & TCPCB_FLAG_SYN)) {
  2382. flag |= FLAG_DATA_ACKED;
  2383. } else {
  2384. flag |= FLAG_SYN_ACKED;
  2385. tp->retrans_stamp = 0;
  2386. }
  2387. if (!fully_acked)
  2388. break;
  2389. tcp_unlink_write_queue(skb, sk);
  2390. sk_stream_free_skb(sk, skb);
  2391. tcp_clear_all_retrans_hints(tp);
  2392. }
  2393. if (flag & FLAG_ACKED) {
  2394. u32 pkts_acked = prior_packets - tp->packets_out;
  2395. const struct tcp_congestion_ops *ca_ops
  2396. = inet_csk(sk)->icsk_ca_ops;
  2397. tcp_ack_update_rtt(sk, flag, seq_rtt);
  2398. tcp_rearm_rto(sk);
  2399. if (tcp_is_reno(tp)) {
  2400. tcp_remove_reno_sacks(sk, pkts_acked);
  2401. } else {
  2402. /* Non-retransmitted hole got filled? That's reordering */
  2403. if (reord < prior_fackets)
  2404. tcp_update_reordering(sk, tp->fackets_out - reord, 0);
  2405. }
  2406. tp->fackets_out -= min(pkts_acked, tp->fackets_out);
  2407. /* hint's skb might be NULL but we don't need to care */
  2408. tp->fastpath_cnt_hint -= min_t(u32, pkts_acked,
  2409. tp->fastpath_cnt_hint);
  2410. if (ca_ops->pkts_acked) {
  2411. s32 rtt_us = -1;
  2412. /* Is the ACK triggering packet unambiguous? */
  2413. if (!(flag & FLAG_RETRANS_DATA_ACKED)) {
  2414. /* High resolution needed and available? */
  2415. if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
  2416. !ktime_equal(last_ackt,
  2417. net_invalid_timestamp()))
  2418. rtt_us = ktime_us_delta(ktime_get_real(),
  2419. last_ackt);
  2420. else if (ca_seq_rtt > 0)
  2421. rtt_us = jiffies_to_usecs(ca_seq_rtt);
  2422. }
  2423. ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
  2424. }
  2425. }
  2426. #if FASTRETRANS_DEBUG > 0
  2427. BUG_TRAP((int)tp->sacked_out >= 0);
  2428. BUG_TRAP((int)tp->lost_out >= 0);
  2429. BUG_TRAP((int)tp->retrans_out >= 0);
  2430. if (!tp->packets_out && tcp_is_sack(tp)) {
  2431. icsk = inet_csk(sk);
  2432. if (tp->lost_out) {
  2433. printk(KERN_DEBUG "Leak l=%u %d\n",
  2434. tp->lost_out, icsk->icsk_ca_state);
  2435. tp->lost_out = 0;
  2436. }
  2437. if (tp->sacked_out) {
  2438. printk(KERN_DEBUG "Leak s=%u %d\n",
  2439. tp->sacked_out, icsk->icsk_ca_state);
  2440. tp->sacked_out = 0;
  2441. }
  2442. if (tp->retrans_out) {
  2443. printk(KERN_DEBUG "Leak r=%u %d\n",
  2444. tp->retrans_out, icsk->icsk_ca_state);
  2445. tp->retrans_out = 0;
  2446. }
  2447. }
  2448. #endif
  2449. *seq_rtt_p = seq_rtt;
  2450. return flag;
  2451. }
  2452. static void tcp_ack_probe(struct sock *sk)
  2453. {
  2454. const struct tcp_sock *tp = tcp_sk(sk);
  2455. struct inet_connection_sock *icsk = inet_csk(sk);
  2456. /* Was it a usable window open? */
  2457. if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
  2458. tp->snd_una + tp->snd_wnd)) {
  2459. icsk->icsk_backoff = 0;
  2460. inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
  2461. /* Socket must be waked up by subsequent tcp_data_snd_check().
  2462. * This function is not for random using!
  2463. */
  2464. } else {
  2465. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2466. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2467. TCP_RTO_MAX);
  2468. }
  2469. }
  2470. static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
  2471. {
  2472. return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
  2473. inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
  2474. }
  2475. static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
  2476. {
  2477. const struct tcp_sock *tp = tcp_sk(sk);
  2478. return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
  2479. !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
  2480. }
  2481. /* Check that window update is acceptable.
  2482. * The function assumes that snd_una<=ack<=snd_next.
  2483. */
  2484. static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
  2485. const u32 ack_seq, const u32 nwin)
  2486. {
  2487. return (after(ack, tp->snd_una) ||
  2488. after(ack_seq, tp->snd_wl1) ||
  2489. (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
  2490. }
  2491. /* Update our send window.
  2492. *
  2493. * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
  2494. * and in FreeBSD. NetBSD's one is even worse.) is wrong.
  2495. */
  2496. static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
  2497. u32 ack_seq)
  2498. {
  2499. struct tcp_sock *tp = tcp_sk(sk);
  2500. int flag = 0;
  2501. u32 nwin = ntohs(tcp_hdr(skb)->window);
  2502. if (likely(!tcp_hdr(skb)->syn))
  2503. nwin <<= tp->rx_opt.snd_wscale;
  2504. if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
  2505. flag |= FLAG_WIN_UPDATE;
  2506. tcp_update_wl(tp, ack, ack_seq);
  2507. if (tp->snd_wnd != nwin) {
  2508. tp->snd_wnd = nwin;
  2509. /* Note, it is the only place, where
  2510. * fast path is recovered for sending TCP.
  2511. */
  2512. tp->pred_flags = 0;
  2513. tcp_fast_path_check(sk);
  2514. if (nwin > tp->max_window) {
  2515. tp->max_window = nwin;
  2516. tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
  2517. }
  2518. }
  2519. }
  2520. tp->snd_una = ack;
  2521. return flag;
  2522. }
  2523. /* A very conservative spurious RTO response algorithm: reduce cwnd and
  2524. * continue in congestion avoidance.
  2525. */
  2526. static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
  2527. {
  2528. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  2529. tp->snd_cwnd_cnt = 0;
  2530. tp->bytes_acked = 0;
  2531. TCP_ECN_queue_cwr(tp);
  2532. tcp_moderate_cwnd(tp);
  2533. }
  2534. /* A conservative spurious RTO response algorithm: reduce cwnd using
  2535. * rate halving and continue in congestion avoidance.
  2536. */
  2537. static void tcp_ratehalving_spur_to_response(struct sock *sk)
  2538. {
  2539. tcp_enter_cwr(sk, 0);
  2540. }
  2541. static void tcp_undo_spur_to_response(struct sock *sk, int flag)
  2542. {
  2543. if (flag&FLAG_ECE)
  2544. tcp_ratehalving_spur_to_response(sk);
  2545. else
  2546. tcp_undo_cwr(sk, 1);
  2547. }
  2548. /* F-RTO spurious RTO detection algorithm (RFC4138)
  2549. *
  2550. * F-RTO affects during two new ACKs following RTO (well, almost, see inline
  2551. * comments). State (ACK number) is kept in frto_counter. When ACK advances
  2552. * window (but not to or beyond highest sequence sent before RTO):
  2553. * On First ACK, send two new segments out.
  2554. * On Second ACK, RTO was likely spurious. Do spurious response (response
  2555. * algorithm is not part of the F-RTO detection algorithm
  2556. * given in RFC4138 but can be selected separately).
  2557. * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
  2558. * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
  2559. * of Nagle, this is done using frto_counter states 2 and 3, when a new data
  2560. * segment of any size sent during F-RTO, state 2 is upgraded to 3.
  2561. *
  2562. * Rationale: if the RTO was spurious, new ACKs should arrive from the
  2563. * original window even after we transmit two new data segments.
  2564. *
  2565. * SACK version:
  2566. * on first step, wait until first cumulative ACK arrives, then move to
  2567. * the second step. In second step, the next ACK decides.
  2568. *
  2569. * F-RTO is implemented (mainly) in four functions:
  2570. * - tcp_use_frto() is used to determine if TCP is can use F-RTO
  2571. * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
  2572. * called when tcp_use_frto() showed green light
  2573. * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
  2574. * - tcp_enter_frto_loss() is called if there is not enough evidence
  2575. * to prove that the RTO is indeed spurious. It transfers the control
  2576. * from F-RTO to the conventional RTO recovery
  2577. */
  2578. static int tcp_process_frto(struct sock *sk, int flag)
  2579. {
  2580. struct tcp_sock *tp = tcp_sk(sk);
  2581. tcp_verify_left_out(tp);
  2582. /* Duplicate the behavior from Loss state (fastretrans_alert) */
  2583. if (flag&FLAG_DATA_ACKED)
  2584. inet_csk(sk)->icsk_retransmits = 0;
  2585. if ((flag & FLAG_NONHEAD_RETRANS_ACKED) ||
  2586. ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED)))
  2587. tp->undo_marker = 0;
  2588. if (!before(tp->snd_una, tp->frto_highmark)) {
  2589. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
  2590. return 1;
  2591. }
  2592. if (!IsSackFrto() || tcp_is_reno(tp)) {
  2593. /* RFC4138 shortcoming in step 2; should also have case c):
  2594. * ACK isn't duplicate nor advances window, e.g., opposite dir
  2595. * data, winupdate
  2596. */
  2597. if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
  2598. return 1;
  2599. if (!(flag&FLAG_DATA_ACKED)) {
  2600. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
  2601. flag);
  2602. return 1;
  2603. }
  2604. } else {
  2605. if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
  2606. /* Prevent sending of new data. */
  2607. tp->snd_cwnd = min(tp->snd_cwnd,
  2608. tcp_packets_in_flight(tp));
  2609. return 1;
  2610. }
  2611. if ((tp->frto_counter >= 2) &&
  2612. (!(flag&FLAG_FORWARD_PROGRESS) ||
  2613. ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
  2614. /* RFC4138 shortcoming (see comment above) */
  2615. if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
  2616. return 1;
  2617. tcp_enter_frto_loss(sk, 3, flag);
  2618. return 1;
  2619. }
  2620. }
  2621. if (tp->frto_counter == 1) {
  2622. /* tcp_may_send_now needs to see updated state */
  2623. tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
  2624. tp->frto_counter = 2;
  2625. if (!tcp_may_send_now(sk))
  2626. tcp_enter_frto_loss(sk, 2, flag);
  2627. return 1;
  2628. } else {
  2629. switch (sysctl_tcp_frto_response) {
  2630. case 2:
  2631. tcp_undo_spur_to_response(sk, flag);
  2632. break;
  2633. case 1:
  2634. tcp_conservative_spur_to_response(tp);
  2635. break;
  2636. default:
  2637. tcp_ratehalving_spur_to_response(sk);
  2638. break;
  2639. }
  2640. tp->frto_counter = 0;
  2641. tp->undo_marker = 0;
  2642. NET_INC_STATS_BH(LINUX_MIB_TCPSPURIOUSRTOS);
  2643. }
  2644. return 0;
  2645. }
  2646. /* This routine deals with incoming acks, but not outgoing ones. */
  2647. static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
  2648. {
  2649. struct inet_connection_sock *icsk = inet_csk(sk);
  2650. struct tcp_sock *tp = tcp_sk(sk);
  2651. u32 prior_snd_una = tp->snd_una;
  2652. u32 ack_seq = TCP_SKB_CB(skb)->seq;
  2653. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2654. u32 prior_in_flight;
  2655. u32 prior_fackets;
  2656. s32 seq_rtt;
  2657. int prior_packets;
  2658. int frto_cwnd = 0;
  2659. /* If the ack is newer than sent or older than previous acks
  2660. * then we can probably ignore it.
  2661. */
  2662. if (after(ack, tp->snd_nxt))
  2663. goto uninteresting_ack;
  2664. if (before(ack, prior_snd_una))
  2665. goto old_ack;
  2666. if (after(ack, prior_snd_una))
  2667. flag |= FLAG_SND_UNA_ADVANCED;
  2668. if (sysctl_tcp_abc) {
  2669. if (icsk->icsk_ca_state < TCP_CA_CWR)
  2670. tp->bytes_acked += ack - prior_snd_una;
  2671. else if (icsk->icsk_ca_state == TCP_CA_Loss)
  2672. /* we assume just one segment left network */
  2673. tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
  2674. }
  2675. prior_fackets = tp->fackets_out;
  2676. prior_in_flight = tcp_packets_in_flight(tp);
  2677. if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
  2678. /* Window is constant, pure forward advance.
  2679. * No more checks are required.
  2680. * Note, we use the fact that SND.UNA>=SND.WL2.
  2681. */
  2682. tcp_update_wl(tp, ack, ack_seq);
  2683. tp->snd_una = ack;
  2684. flag |= FLAG_WIN_UPDATE;
  2685. tcp_ca_event(sk, CA_EVENT_FAST_ACK);
  2686. NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
  2687. } else {
  2688. if (ack_seq != TCP_SKB_CB(skb)->end_seq)
  2689. flag |= FLAG_DATA;
  2690. else
  2691. NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
  2692. flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
  2693. if (TCP_SKB_CB(skb)->sacked)
  2694. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2695. if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
  2696. flag |= FLAG_ECE;
  2697. tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
  2698. }
  2699. /* We passed data and got it acked, remove any soft error
  2700. * log. Something worked...
  2701. */
  2702. sk->sk_err_soft = 0;
  2703. tp->rcv_tstamp = tcp_time_stamp;
  2704. prior_packets = tp->packets_out;
  2705. if (!prior_packets)
  2706. goto no_queue;
  2707. /* See if we can take anything off of the retransmit queue. */
  2708. flag |= tcp_clean_rtx_queue(sk, &seq_rtt, prior_fackets);
  2709. if (tp->frto_counter)
  2710. frto_cwnd = tcp_process_frto(sk, flag);
  2711. /* Guarantee sacktag reordering detection against wrap-arounds */
  2712. if (before(tp->frto_highmark, tp->snd_una))
  2713. tp->frto_highmark = 0;
  2714. if (tcp_ack_is_dubious(sk, flag)) {
  2715. /* Advance CWND, if state allows this. */
  2716. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
  2717. tcp_may_raise_cwnd(sk, flag))
  2718. tcp_cong_avoid(sk, ack, prior_in_flight, 0);
  2719. tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag);
  2720. } else {
  2721. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
  2722. tcp_cong_avoid(sk, ack, prior_in_flight, 1);
  2723. }
  2724. if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
  2725. dst_confirm(sk->sk_dst_cache);
  2726. return 1;
  2727. no_queue:
  2728. icsk->icsk_probes_out = 0;
  2729. /* If this ack opens up a zero window, clear backoff. It was
  2730. * being used to time the probes, and is probably far higher than
  2731. * it needs to be for normal retransmission.
  2732. */
  2733. if (tcp_send_head(sk))
  2734. tcp_ack_probe(sk);
  2735. return 1;
  2736. old_ack:
  2737. if (TCP_SKB_CB(skb)->sacked)
  2738. tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2739. uninteresting_ack:
  2740. SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  2741. return 0;
  2742. }
  2743. /* Look for tcp options. Normally only called on SYN and SYNACK packets.
  2744. * But, this can also be called on packets in the established flow when
  2745. * the fast version below fails.
  2746. */
  2747. void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
  2748. {
  2749. unsigned char *ptr;
  2750. struct tcphdr *th = tcp_hdr(skb);
  2751. int length=(th->doff*4)-sizeof(struct tcphdr);
  2752. ptr = (unsigned char *)(th + 1);
  2753. opt_rx->saw_tstamp = 0;
  2754. while (length > 0) {
  2755. int opcode=*ptr++;
  2756. int opsize;
  2757. switch (opcode) {
  2758. case TCPOPT_EOL:
  2759. return;
  2760. case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
  2761. length--;
  2762. continue;
  2763. default:
  2764. opsize=*ptr++;
  2765. if (opsize < 2) /* "silly options" */
  2766. return;
  2767. if (opsize > length)
  2768. return; /* don't parse partial options */
  2769. switch (opcode) {
  2770. case TCPOPT_MSS:
  2771. if (opsize==TCPOLEN_MSS && th->syn && !estab) {
  2772. u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
  2773. if (in_mss) {
  2774. if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
  2775. in_mss = opt_rx->user_mss;
  2776. opt_rx->mss_clamp = in_mss;
  2777. }
  2778. }
  2779. break;
  2780. case TCPOPT_WINDOW:
  2781. if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
  2782. if (sysctl_tcp_window_scaling) {
  2783. __u8 snd_wscale = *(__u8 *) ptr;
  2784. opt_rx->wscale_ok = 1;
  2785. if (snd_wscale > 14) {
  2786. if (net_ratelimit())
  2787. printk(KERN_INFO "tcp_parse_options: Illegal window "
  2788. "scaling value %d >14 received.\n",
  2789. snd_wscale);
  2790. snd_wscale = 14;
  2791. }
  2792. opt_rx->snd_wscale = snd_wscale;
  2793. }
  2794. break;
  2795. case TCPOPT_TIMESTAMP:
  2796. if (opsize==TCPOLEN_TIMESTAMP) {
  2797. if ((estab && opt_rx->tstamp_ok) ||
  2798. (!estab && sysctl_tcp_timestamps)) {
  2799. opt_rx->saw_tstamp = 1;
  2800. opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
  2801. opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
  2802. }
  2803. }
  2804. break;
  2805. case TCPOPT_SACK_PERM:
  2806. if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
  2807. if (sysctl_tcp_sack) {
  2808. opt_rx->sack_ok = 1;
  2809. tcp_sack_reset(opt_rx);
  2810. }
  2811. }
  2812. break;
  2813. case TCPOPT_SACK:
  2814. if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
  2815. !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
  2816. opt_rx->sack_ok) {
  2817. TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
  2818. }
  2819. break;
  2820. #ifdef CONFIG_TCP_MD5SIG
  2821. case TCPOPT_MD5SIG:
  2822. /*
  2823. * The MD5 Hash has already been
  2824. * checked (see tcp_v{4,6}_do_rcv()).
  2825. */
  2826. break;
  2827. #endif
  2828. }
  2829. ptr+=opsize-2;
  2830. length-=opsize;
  2831. }
  2832. }
  2833. }
  2834. /* Fast parse options. This hopes to only see timestamps.
  2835. * If it is wrong it falls back on tcp_parse_options().
  2836. */
  2837. static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
  2838. struct tcp_sock *tp)
  2839. {
  2840. if (th->doff == sizeof(struct tcphdr)>>2) {
  2841. tp->rx_opt.saw_tstamp = 0;
  2842. return 0;
  2843. } else if (tp->rx_opt.tstamp_ok &&
  2844. th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
  2845. __be32 *ptr = (__be32 *)(th + 1);
  2846. if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  2847. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
  2848. tp->rx_opt.saw_tstamp = 1;
  2849. ++ptr;
  2850. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  2851. ++ptr;
  2852. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  2853. return 1;
  2854. }
  2855. }
  2856. tcp_parse_options(skb, &tp->rx_opt, 1);
  2857. return 1;
  2858. }
  2859. static inline void tcp_store_ts_recent(struct tcp_sock *tp)
  2860. {
  2861. tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
  2862. tp->rx_opt.ts_recent_stamp = get_seconds();
  2863. }
  2864. static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
  2865. {
  2866. if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
  2867. /* PAWS bug workaround wrt. ACK frames, the PAWS discard
  2868. * extra check below makes sure this can only happen
  2869. * for pure ACK frames. -DaveM
  2870. *
  2871. * Not only, also it occurs for expired timestamps.
  2872. */
  2873. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
  2874. get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
  2875. tcp_store_ts_recent(tp);
  2876. }
  2877. }
  2878. /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
  2879. *
  2880. * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
  2881. * it can pass through stack. So, the following predicate verifies that
  2882. * this segment is not used for anything but congestion avoidance or
  2883. * fast retransmit. Moreover, we even are able to eliminate most of such
  2884. * second order effects, if we apply some small "replay" window (~RTO)
  2885. * to timestamp space.
  2886. *
  2887. * All these measures still do not guarantee that we reject wrapped ACKs
  2888. * on networks with high bandwidth, when sequence space is recycled fastly,
  2889. * but it guarantees that such events will be very rare and do not affect
  2890. * connection seriously. This doesn't look nice, but alas, PAWS is really
  2891. * buggy extension.
  2892. *
  2893. * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
  2894. * states that events when retransmit arrives after original data are rare.
  2895. * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
  2896. * the biggest problem on large power networks even with minor reordering.
  2897. * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
  2898. * up to bandwidth of 18Gigabit/sec. 8) ]
  2899. */
  2900. static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
  2901. {
  2902. struct tcp_sock *tp = tcp_sk(sk);
  2903. struct tcphdr *th = tcp_hdr(skb);
  2904. u32 seq = TCP_SKB_CB(skb)->seq;
  2905. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2906. return (/* 1. Pure ACK with correct sequence number. */
  2907. (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
  2908. /* 2. ... and duplicate ACK. */
  2909. ack == tp->snd_una &&
  2910. /* 3. ... and does not update window. */
  2911. !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
  2912. /* 4. ... and sits in replay window. */
  2913. (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
  2914. }
  2915. static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
  2916. {
  2917. const struct tcp_sock *tp = tcp_sk(sk);
  2918. return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
  2919. get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
  2920. !tcp_disordered_ack(sk, skb));
  2921. }
  2922. /* Check segment sequence number for validity.
  2923. *
  2924. * Segment controls are considered valid, if the segment
  2925. * fits to the window after truncation to the window. Acceptability
  2926. * of data (and SYN, FIN, of course) is checked separately.
  2927. * See tcp_data_queue(), for example.
  2928. *
  2929. * Also, controls (RST is main one) are accepted using RCV.WUP instead
  2930. * of RCV.NXT. Peer still did not advance his SND.UNA when we
  2931. * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
  2932. * (borrowed from freebsd)
  2933. */
  2934. static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2935. {
  2936. return !before(end_seq, tp->rcv_wup) &&
  2937. !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
  2938. }
  2939. /* When we get a reset we do this. */
  2940. static void tcp_reset(struct sock *sk)
  2941. {
  2942. /* We want the right error as BSD sees it (and indeed as we do). */
  2943. switch (sk->sk_state) {
  2944. case TCP_SYN_SENT:
  2945. sk->sk_err = ECONNREFUSED;
  2946. break;
  2947. case TCP_CLOSE_WAIT:
  2948. sk->sk_err = EPIPE;
  2949. break;
  2950. case TCP_CLOSE:
  2951. return;
  2952. default:
  2953. sk->sk_err = ECONNRESET;
  2954. }
  2955. if (!sock_flag(sk, SOCK_DEAD))
  2956. sk->sk_error_report(sk);
  2957. tcp_done(sk);
  2958. }
  2959. /*
  2960. * Process the FIN bit. This now behaves as it is supposed to work
  2961. * and the FIN takes effect when it is validly part of sequence
  2962. * space. Not before when we get holes.
  2963. *
  2964. * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
  2965. * (and thence onto LAST-ACK and finally, CLOSE, we never enter
  2966. * TIME-WAIT)
  2967. *
  2968. * If we are in FINWAIT-1, a received FIN indicates simultaneous
  2969. * close and we go into CLOSING (and later onto TIME-WAIT)
  2970. *
  2971. * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
  2972. */
  2973. static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
  2974. {
  2975. struct tcp_sock *tp = tcp_sk(sk);
  2976. inet_csk_schedule_ack(sk);
  2977. sk->sk_shutdown |= RCV_SHUTDOWN;
  2978. sock_set_flag(sk, SOCK_DONE);
  2979. switch (sk->sk_state) {
  2980. case TCP_SYN_RECV:
  2981. case TCP_ESTABLISHED:
  2982. /* Move to CLOSE_WAIT */
  2983. tcp_set_state(sk, TCP_CLOSE_WAIT);
  2984. inet_csk(sk)->icsk_ack.pingpong = 1;
  2985. break;
  2986. case TCP_CLOSE_WAIT:
  2987. case TCP_CLOSING:
  2988. /* Received a retransmission of the FIN, do
  2989. * nothing.
  2990. */
  2991. break;
  2992. case TCP_LAST_ACK:
  2993. /* RFC793: Remain in the LAST-ACK state. */
  2994. break;
  2995. case TCP_FIN_WAIT1:
  2996. /* This case occurs when a simultaneous close
  2997. * happens, we must ack the received FIN and
  2998. * enter the CLOSING state.
  2999. */
  3000. tcp_send_ack(sk);
  3001. tcp_set_state(sk, TCP_CLOSING);
  3002. break;
  3003. case TCP_FIN_WAIT2:
  3004. /* Received a FIN -- send ACK and enter TIME_WAIT. */
  3005. tcp_send_ack(sk);
  3006. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  3007. break;
  3008. default:
  3009. /* Only TCP_LISTEN and TCP_CLOSE are left, in these
  3010. * cases we should never reach this piece of code.
  3011. */
  3012. printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
  3013. __FUNCTION__, sk->sk_state);
  3014. break;
  3015. }
  3016. /* It _is_ possible, that we have something out-of-order _after_ FIN.
  3017. * Probably, we should reset in this case. For now drop them.
  3018. */
  3019. __skb_queue_purge(&tp->out_of_order_queue);
  3020. if (tcp_is_sack(tp))
  3021. tcp_sack_reset(&tp->rx_opt);
  3022. sk_stream_mem_reclaim(sk);
  3023. if (!sock_flag(sk, SOCK_DEAD)) {
  3024. sk->sk_state_change(sk);
  3025. /* Do not send POLL_HUP for half duplex close. */
  3026. if (sk->sk_shutdown == SHUTDOWN_MASK ||
  3027. sk->sk_state == TCP_CLOSE)
  3028. sk_wake_async(sk, 1, POLL_HUP);
  3029. else
  3030. sk_wake_async(sk, 1, POLL_IN);
  3031. }
  3032. }
  3033. static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
  3034. {
  3035. if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
  3036. if (before(seq, sp->start_seq))
  3037. sp->start_seq = seq;
  3038. if (after(end_seq, sp->end_seq))
  3039. sp->end_seq = end_seq;
  3040. return 1;
  3041. }
  3042. return 0;
  3043. }
  3044. static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
  3045. {
  3046. if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
  3047. if (before(seq, tp->rcv_nxt))
  3048. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
  3049. else
  3050. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
  3051. tp->rx_opt.dsack = 1;
  3052. tp->duplicate_sack[0].start_seq = seq;
  3053. tp->duplicate_sack[0].end_seq = end_seq;
  3054. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
  3055. }
  3056. }
  3057. static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
  3058. {
  3059. if (!tp->rx_opt.dsack)
  3060. tcp_dsack_set(tp, seq, end_seq);
  3061. else
  3062. tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
  3063. }
  3064. static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
  3065. {
  3066. struct tcp_sock *tp = tcp_sk(sk);
  3067. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  3068. before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3069. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  3070. tcp_enter_quickack_mode(sk);
  3071. if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
  3072. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  3073. if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
  3074. end_seq = tp->rcv_nxt;
  3075. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
  3076. }
  3077. }
  3078. tcp_send_ack(sk);
  3079. }
  3080. /* These routines update the SACK block as out-of-order packets arrive or
  3081. * in-order packets close up the sequence space.
  3082. */
  3083. static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
  3084. {
  3085. int this_sack;
  3086. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3087. struct tcp_sack_block *swalk = sp+1;
  3088. /* See if the recent change to the first SACK eats into
  3089. * or hits the sequence space of other SACK blocks, if so coalesce.
  3090. */
  3091. for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
  3092. if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
  3093. int i;
  3094. /* Zap SWALK, by moving every further SACK up by one slot.
  3095. * Decrease num_sacks.
  3096. */
  3097. tp->rx_opt.num_sacks--;
  3098. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  3099. for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
  3100. sp[i] = sp[i+1];
  3101. continue;
  3102. }
  3103. this_sack++, swalk++;
  3104. }
  3105. }
  3106. static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
  3107. {
  3108. __u32 tmp;
  3109. tmp = sack1->start_seq;
  3110. sack1->start_seq = sack2->start_seq;
  3111. sack2->start_seq = tmp;
  3112. tmp = sack1->end_seq;
  3113. sack1->end_seq = sack2->end_seq;
  3114. sack2->end_seq = tmp;
  3115. }
  3116. static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
  3117. {
  3118. struct tcp_sock *tp = tcp_sk(sk);
  3119. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3120. int cur_sacks = tp->rx_opt.num_sacks;
  3121. int this_sack;
  3122. if (!cur_sacks)
  3123. goto new_sack;
  3124. for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
  3125. if (tcp_sack_extend(sp, seq, end_seq)) {
  3126. /* Rotate this_sack to the first one. */
  3127. for (; this_sack>0; this_sack--, sp--)
  3128. tcp_sack_swap(sp, sp-1);
  3129. if (cur_sacks > 1)
  3130. tcp_sack_maybe_coalesce(tp);
  3131. return;
  3132. }
  3133. }
  3134. /* Could not find an adjacent existing SACK, build a new one,
  3135. * put it at the front, and shift everyone else down. We
  3136. * always know there is at least one SACK present already here.
  3137. *
  3138. * If the sack array is full, forget about the last one.
  3139. */
  3140. if (this_sack >= 4) {
  3141. this_sack--;
  3142. tp->rx_opt.num_sacks--;
  3143. sp--;
  3144. }
  3145. for (; this_sack > 0; this_sack--, sp--)
  3146. *sp = *(sp-1);
  3147. new_sack:
  3148. /* Build the new head SACK, and we're done. */
  3149. sp->start_seq = seq;
  3150. sp->end_seq = end_seq;
  3151. tp->rx_opt.num_sacks++;
  3152. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  3153. }
  3154. /* RCV.NXT advances, some SACKs should be eaten. */
  3155. static void tcp_sack_remove(struct tcp_sock *tp)
  3156. {
  3157. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3158. int num_sacks = tp->rx_opt.num_sacks;
  3159. int this_sack;
  3160. /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
  3161. if (skb_queue_empty(&tp->out_of_order_queue)) {
  3162. tp->rx_opt.num_sacks = 0;
  3163. tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
  3164. return;
  3165. }
  3166. for (this_sack = 0; this_sack < num_sacks; ) {
  3167. /* Check if the start of the sack is covered by RCV.NXT. */
  3168. if (!before(tp->rcv_nxt, sp->start_seq)) {
  3169. int i;
  3170. /* RCV.NXT must cover all the block! */
  3171. BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
  3172. /* Zap this SACK, by moving forward any other SACKS. */
  3173. for (i=this_sack+1; i < num_sacks; i++)
  3174. tp->selective_acks[i-1] = tp->selective_acks[i];
  3175. num_sacks--;
  3176. continue;
  3177. }
  3178. this_sack++;
  3179. sp++;
  3180. }
  3181. if (num_sacks != tp->rx_opt.num_sacks) {
  3182. tp->rx_opt.num_sacks = num_sacks;
  3183. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  3184. }
  3185. }
  3186. /* This one checks to see if we can put data from the
  3187. * out_of_order queue into the receive_queue.
  3188. */
  3189. static void tcp_ofo_queue(struct sock *sk)
  3190. {
  3191. struct tcp_sock *tp = tcp_sk(sk);
  3192. __u32 dsack_high = tp->rcv_nxt;
  3193. struct sk_buff *skb;
  3194. while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
  3195. if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  3196. break;
  3197. if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
  3198. __u32 dsack = dsack_high;
  3199. if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
  3200. dsack_high = TCP_SKB_CB(skb)->end_seq;
  3201. tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
  3202. }
  3203. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3204. SOCK_DEBUG(sk, "ofo packet was already received \n");
  3205. __skb_unlink(skb, &tp->out_of_order_queue);
  3206. __kfree_skb(skb);
  3207. continue;
  3208. }
  3209. SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
  3210. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3211. TCP_SKB_CB(skb)->end_seq);
  3212. __skb_unlink(skb, &tp->out_of_order_queue);
  3213. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3214. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3215. if (tcp_hdr(skb)->fin)
  3216. tcp_fin(skb, sk, tcp_hdr(skb));
  3217. }
  3218. }
  3219. static int tcp_prune_queue(struct sock *sk);
  3220. static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
  3221. {
  3222. struct tcphdr *th = tcp_hdr(skb);
  3223. struct tcp_sock *tp = tcp_sk(sk);
  3224. int eaten = -1;
  3225. if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
  3226. goto drop;
  3227. __skb_pull(skb, th->doff*4);
  3228. TCP_ECN_accept_cwr(tp, skb);
  3229. if (tp->rx_opt.dsack) {
  3230. tp->rx_opt.dsack = 0;
  3231. tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
  3232. 4 - tp->rx_opt.tstamp_ok);
  3233. }
  3234. /* Queue data for delivery to the user.
  3235. * Packets in sequence go to the receive queue.
  3236. * Out of sequence packets to the out_of_order_queue.
  3237. */
  3238. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3239. if (tcp_receive_window(tp) == 0)
  3240. goto out_of_window;
  3241. /* Ok. In sequence. In window. */
  3242. if (tp->ucopy.task == current &&
  3243. tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
  3244. sock_owned_by_user(sk) && !tp->urg_data) {
  3245. int chunk = min_t(unsigned int, skb->len,
  3246. tp->ucopy.len);
  3247. __set_current_state(TASK_RUNNING);
  3248. local_bh_enable();
  3249. if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
  3250. tp->ucopy.len -= chunk;
  3251. tp->copied_seq += chunk;
  3252. eaten = (chunk == skb->len && !th->fin);
  3253. tcp_rcv_space_adjust(sk);
  3254. }
  3255. local_bh_disable();
  3256. }
  3257. if (eaten <= 0) {
  3258. queue_and_out:
  3259. if (eaten < 0 &&
  3260. (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3261. !sk_stream_rmem_schedule(sk, skb))) {
  3262. if (tcp_prune_queue(sk) < 0 ||
  3263. !sk_stream_rmem_schedule(sk, skb))
  3264. goto drop;
  3265. }
  3266. sk_stream_set_owner_r(skb, sk);
  3267. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3268. }
  3269. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3270. if (skb->len)
  3271. tcp_event_data_recv(sk, skb);
  3272. if (th->fin)
  3273. tcp_fin(skb, sk, th);
  3274. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3275. tcp_ofo_queue(sk);
  3276. /* RFC2581. 4.2. SHOULD send immediate ACK, when
  3277. * gap in queue is filled.
  3278. */
  3279. if (skb_queue_empty(&tp->out_of_order_queue))
  3280. inet_csk(sk)->icsk_ack.pingpong = 0;
  3281. }
  3282. if (tp->rx_opt.num_sacks)
  3283. tcp_sack_remove(tp);
  3284. tcp_fast_path_check(sk);
  3285. if (eaten > 0)
  3286. __kfree_skb(skb);
  3287. else if (!sock_flag(sk, SOCK_DEAD))
  3288. sk->sk_data_ready(sk, 0);
  3289. return;
  3290. }
  3291. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3292. /* A retransmit, 2nd most common case. Force an immediate ack. */
  3293. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  3294. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3295. out_of_window:
  3296. tcp_enter_quickack_mode(sk);
  3297. inet_csk_schedule_ack(sk);
  3298. drop:
  3299. __kfree_skb(skb);
  3300. return;
  3301. }
  3302. /* Out of window. F.e. zero window probe. */
  3303. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
  3304. goto out_of_window;
  3305. tcp_enter_quickack_mode(sk);
  3306. if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3307. /* Partial packet, seq < rcv_next < end_seq */
  3308. SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
  3309. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3310. TCP_SKB_CB(skb)->end_seq);
  3311. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
  3312. /* If window is closed, drop tail of packet. But after
  3313. * remembering D-SACK for its head made in previous line.
  3314. */
  3315. if (!tcp_receive_window(tp))
  3316. goto out_of_window;
  3317. goto queue_and_out;
  3318. }
  3319. TCP_ECN_check_ce(tp, skb);
  3320. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3321. !sk_stream_rmem_schedule(sk, skb)) {
  3322. if (tcp_prune_queue(sk) < 0 ||
  3323. !sk_stream_rmem_schedule(sk, skb))
  3324. goto drop;
  3325. }
  3326. /* Disable header prediction. */
  3327. tp->pred_flags = 0;
  3328. inet_csk_schedule_ack(sk);
  3329. SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
  3330. tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3331. sk_stream_set_owner_r(skb, sk);
  3332. if (!skb_peek(&tp->out_of_order_queue)) {
  3333. /* Initial out of order segment, build 1 SACK. */
  3334. if (tcp_is_sack(tp)) {
  3335. tp->rx_opt.num_sacks = 1;
  3336. tp->rx_opt.dsack = 0;
  3337. tp->rx_opt.eff_sacks = 1;
  3338. tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
  3339. tp->selective_acks[0].end_seq =
  3340. TCP_SKB_CB(skb)->end_seq;
  3341. }
  3342. __skb_queue_head(&tp->out_of_order_queue,skb);
  3343. } else {
  3344. struct sk_buff *skb1 = tp->out_of_order_queue.prev;
  3345. u32 seq = TCP_SKB_CB(skb)->seq;
  3346. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  3347. if (seq == TCP_SKB_CB(skb1)->end_seq) {
  3348. __skb_append(skb1, skb, &tp->out_of_order_queue);
  3349. if (!tp->rx_opt.num_sacks ||
  3350. tp->selective_acks[0].end_seq != seq)
  3351. goto add_sack;
  3352. /* Common case: data arrive in order after hole. */
  3353. tp->selective_acks[0].end_seq = end_seq;
  3354. return;
  3355. }
  3356. /* Find place to insert this segment. */
  3357. do {
  3358. if (!after(TCP_SKB_CB(skb1)->seq, seq))
  3359. break;
  3360. } while ((skb1 = skb1->prev) !=
  3361. (struct sk_buff*)&tp->out_of_order_queue);
  3362. /* Do skb overlap to previous one? */
  3363. if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
  3364. before(seq, TCP_SKB_CB(skb1)->end_seq)) {
  3365. if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3366. /* All the bits are present. Drop. */
  3367. __kfree_skb(skb);
  3368. tcp_dsack_set(tp, seq, end_seq);
  3369. goto add_sack;
  3370. }
  3371. if (after(seq, TCP_SKB_CB(skb1)->seq)) {
  3372. /* Partial overlap. */
  3373. tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
  3374. } else {
  3375. skb1 = skb1->prev;
  3376. }
  3377. }
  3378. __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
  3379. /* And clean segments covered by new one as whole. */
  3380. while ((skb1 = skb->next) !=
  3381. (struct sk_buff*)&tp->out_of_order_queue &&
  3382. after(end_seq, TCP_SKB_CB(skb1)->seq)) {
  3383. if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3384. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
  3385. break;
  3386. }
  3387. __skb_unlink(skb1, &tp->out_of_order_queue);
  3388. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
  3389. __kfree_skb(skb1);
  3390. }
  3391. add_sack:
  3392. if (tcp_is_sack(tp))
  3393. tcp_sack_new_ofo_skb(sk, seq, end_seq);
  3394. }
  3395. }
  3396. /* Collapse contiguous sequence of skbs head..tail with
  3397. * sequence numbers start..end.
  3398. * Segments with FIN/SYN are not collapsed (only because this
  3399. * simplifies code)
  3400. */
  3401. static void
  3402. tcp_collapse(struct sock *sk, struct sk_buff_head *list,
  3403. struct sk_buff *head, struct sk_buff *tail,
  3404. u32 start, u32 end)
  3405. {
  3406. struct sk_buff *skb;
  3407. /* First, check that queue is collapsible and find
  3408. * the point where collapsing can be useful. */
  3409. for (skb = head; skb != tail; ) {
  3410. /* No new bits? It is possible on ofo queue. */
  3411. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3412. struct sk_buff *next = skb->next;
  3413. __skb_unlink(skb, list);
  3414. __kfree_skb(skb);
  3415. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3416. skb = next;
  3417. continue;
  3418. }
  3419. /* The first skb to collapse is:
  3420. * - not SYN/FIN and
  3421. * - bloated or contains data before "start" or
  3422. * overlaps to the next one.
  3423. */
  3424. if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
  3425. (tcp_win_from_space(skb->truesize) > skb->len ||
  3426. before(TCP_SKB_CB(skb)->seq, start) ||
  3427. (skb->next != tail &&
  3428. TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
  3429. break;
  3430. /* Decided to skip this, advance start seq. */
  3431. start = TCP_SKB_CB(skb)->end_seq;
  3432. skb = skb->next;
  3433. }
  3434. if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
  3435. return;
  3436. while (before(start, end)) {
  3437. struct sk_buff *nskb;
  3438. unsigned int header = skb_headroom(skb);
  3439. int copy = SKB_MAX_ORDER(header, 0);
  3440. /* Too big header? This can happen with IPv6. */
  3441. if (copy < 0)
  3442. return;
  3443. if (end-start < copy)
  3444. copy = end-start;
  3445. nskb = alloc_skb(copy+header, GFP_ATOMIC);
  3446. if (!nskb)
  3447. return;
  3448. skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
  3449. skb_set_network_header(nskb, (skb_network_header(skb) -
  3450. skb->head));
  3451. skb_set_transport_header(nskb, (skb_transport_header(skb) -
  3452. skb->head));
  3453. skb_reserve(nskb, header);
  3454. memcpy(nskb->head, skb->head, header);
  3455. memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
  3456. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
  3457. __skb_insert(nskb, skb->prev, skb, list);
  3458. sk_stream_set_owner_r(nskb, sk);
  3459. /* Copy data, releasing collapsed skbs. */
  3460. while (copy > 0) {
  3461. int offset = start - TCP_SKB_CB(skb)->seq;
  3462. int size = TCP_SKB_CB(skb)->end_seq - start;
  3463. BUG_ON(offset < 0);
  3464. if (size > 0) {
  3465. size = min(copy, size);
  3466. if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
  3467. BUG();
  3468. TCP_SKB_CB(nskb)->end_seq += size;
  3469. copy -= size;
  3470. start += size;
  3471. }
  3472. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3473. struct sk_buff *next = skb->next;
  3474. __skb_unlink(skb, list);
  3475. __kfree_skb(skb);
  3476. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3477. skb = next;
  3478. if (skb == tail ||
  3479. tcp_hdr(skb)->syn ||
  3480. tcp_hdr(skb)->fin)
  3481. return;
  3482. }
  3483. }
  3484. }
  3485. }
  3486. /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
  3487. * and tcp_collapse() them until all the queue is collapsed.
  3488. */
  3489. static void tcp_collapse_ofo_queue(struct sock *sk)
  3490. {
  3491. struct tcp_sock *tp = tcp_sk(sk);
  3492. struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
  3493. struct sk_buff *head;
  3494. u32 start, end;
  3495. if (skb == NULL)
  3496. return;
  3497. start = TCP_SKB_CB(skb)->seq;
  3498. end = TCP_SKB_CB(skb)->end_seq;
  3499. head = skb;
  3500. for (;;) {
  3501. skb = skb->next;
  3502. /* Segment is terminated when we see gap or when
  3503. * we are at the end of all the queue. */
  3504. if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
  3505. after(TCP_SKB_CB(skb)->seq, end) ||
  3506. before(TCP_SKB_CB(skb)->end_seq, start)) {
  3507. tcp_collapse(sk, &tp->out_of_order_queue,
  3508. head, skb, start, end);
  3509. head = skb;
  3510. if (skb == (struct sk_buff *)&tp->out_of_order_queue)
  3511. break;
  3512. /* Start new segment */
  3513. start = TCP_SKB_CB(skb)->seq;
  3514. end = TCP_SKB_CB(skb)->end_seq;
  3515. } else {
  3516. if (before(TCP_SKB_CB(skb)->seq, start))
  3517. start = TCP_SKB_CB(skb)->seq;
  3518. if (after(TCP_SKB_CB(skb)->end_seq, end))
  3519. end = TCP_SKB_CB(skb)->end_seq;
  3520. }
  3521. }
  3522. }
  3523. /* Reduce allocated memory if we can, trying to get
  3524. * the socket within its memory limits again.
  3525. *
  3526. * Return less than zero if we should start dropping frames
  3527. * until the socket owning process reads some of the data
  3528. * to stabilize the situation.
  3529. */
  3530. static int tcp_prune_queue(struct sock *sk)
  3531. {
  3532. struct tcp_sock *tp = tcp_sk(sk);
  3533. SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
  3534. NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
  3535. if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
  3536. tcp_clamp_window(sk);
  3537. else if (tcp_memory_pressure)
  3538. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
  3539. tcp_collapse_ofo_queue(sk);
  3540. tcp_collapse(sk, &sk->sk_receive_queue,
  3541. sk->sk_receive_queue.next,
  3542. (struct sk_buff*)&sk->sk_receive_queue,
  3543. tp->copied_seq, tp->rcv_nxt);
  3544. sk_stream_mem_reclaim(sk);
  3545. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3546. return 0;
  3547. /* Collapsing did not help, destructive actions follow.
  3548. * This must not ever occur. */
  3549. /* First, purge the out_of_order queue. */
  3550. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3551. NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
  3552. __skb_queue_purge(&tp->out_of_order_queue);
  3553. /* Reset SACK state. A conforming SACK implementation will
  3554. * do the same at a timeout based retransmit. When a connection
  3555. * is in a sad state like this, we care only about integrity
  3556. * of the connection not performance.
  3557. */
  3558. if (tcp_is_sack(tp))
  3559. tcp_sack_reset(&tp->rx_opt);
  3560. sk_stream_mem_reclaim(sk);
  3561. }
  3562. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3563. return 0;
  3564. /* If we are really being abused, tell the caller to silently
  3565. * drop receive data on the floor. It will get retransmitted
  3566. * and hopefully then we'll have sufficient space.
  3567. */
  3568. NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
  3569. /* Massive buffer overcommit. */
  3570. tp->pred_flags = 0;
  3571. return -1;
  3572. }
  3573. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  3574. * As additional protections, we do not touch cwnd in retransmission phases,
  3575. * and if application hit its sndbuf limit recently.
  3576. */
  3577. void tcp_cwnd_application_limited(struct sock *sk)
  3578. {
  3579. struct tcp_sock *tp = tcp_sk(sk);
  3580. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  3581. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  3582. /* Limited by application or receiver window. */
  3583. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  3584. u32 win_used = max(tp->snd_cwnd_used, init_win);
  3585. if (win_used < tp->snd_cwnd) {
  3586. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  3587. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  3588. }
  3589. tp->snd_cwnd_used = 0;
  3590. }
  3591. tp->snd_cwnd_stamp = tcp_time_stamp;
  3592. }
  3593. static int tcp_should_expand_sndbuf(struct sock *sk)
  3594. {
  3595. struct tcp_sock *tp = tcp_sk(sk);
  3596. /* If the user specified a specific send buffer setting, do
  3597. * not modify it.
  3598. */
  3599. if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
  3600. return 0;
  3601. /* If we are under global TCP memory pressure, do not expand. */
  3602. if (tcp_memory_pressure)
  3603. return 0;
  3604. /* If we are under soft global TCP memory pressure, do not expand. */
  3605. if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
  3606. return 0;
  3607. /* If we filled the congestion window, do not expand. */
  3608. if (tp->packets_out >= tp->snd_cwnd)
  3609. return 0;
  3610. return 1;
  3611. }
  3612. /* When incoming ACK allowed to free some skb from write_queue,
  3613. * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
  3614. * on the exit from tcp input handler.
  3615. *
  3616. * PROBLEM: sndbuf expansion does not work well with largesend.
  3617. */
  3618. static void tcp_new_space(struct sock *sk)
  3619. {
  3620. struct tcp_sock *tp = tcp_sk(sk);
  3621. if (tcp_should_expand_sndbuf(sk)) {
  3622. int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
  3623. MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
  3624. demanded = max_t(unsigned int, tp->snd_cwnd,
  3625. tp->reordering + 1);
  3626. sndmem *= 2*demanded;
  3627. if (sndmem > sk->sk_sndbuf)
  3628. sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
  3629. tp->snd_cwnd_stamp = tcp_time_stamp;
  3630. }
  3631. sk->sk_write_space(sk);
  3632. }
  3633. static void tcp_check_space(struct sock *sk)
  3634. {
  3635. if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
  3636. sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
  3637. if (sk->sk_socket &&
  3638. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
  3639. tcp_new_space(sk);
  3640. }
  3641. }
  3642. static inline void tcp_data_snd_check(struct sock *sk)
  3643. {
  3644. tcp_push_pending_frames(sk);
  3645. tcp_check_space(sk);
  3646. }
  3647. /*
  3648. * Check if sending an ack is needed.
  3649. */
  3650. static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
  3651. {
  3652. struct tcp_sock *tp = tcp_sk(sk);
  3653. /* More than one full frame received... */
  3654. if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
  3655. /* ... and right edge of window advances far enough.
  3656. * (tcp_recvmsg() will send ACK otherwise). Or...
  3657. */
  3658. && __tcp_select_window(sk) >= tp->rcv_wnd) ||
  3659. /* We ACK each frame or... */
  3660. tcp_in_quickack_mode(sk) ||
  3661. /* We have out of order data. */
  3662. (ofo_possible &&
  3663. skb_peek(&tp->out_of_order_queue))) {
  3664. /* Then ack it now */
  3665. tcp_send_ack(sk);
  3666. } else {
  3667. /* Else, send delayed ack. */
  3668. tcp_send_delayed_ack(sk);
  3669. }
  3670. }
  3671. static inline void tcp_ack_snd_check(struct sock *sk)
  3672. {
  3673. if (!inet_csk_ack_scheduled(sk)) {
  3674. /* We sent a data segment already. */
  3675. return;
  3676. }
  3677. __tcp_ack_snd_check(sk, 1);
  3678. }
  3679. /*
  3680. * This routine is only called when we have urgent data
  3681. * signaled. Its the 'slow' part of tcp_urg. It could be
  3682. * moved inline now as tcp_urg is only called from one
  3683. * place. We handle URGent data wrong. We have to - as
  3684. * BSD still doesn't use the correction from RFC961.
  3685. * For 1003.1g we should support a new option TCP_STDURG to permit
  3686. * either form (or just set the sysctl tcp_stdurg).
  3687. */
  3688. static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
  3689. {
  3690. struct tcp_sock *tp = tcp_sk(sk);
  3691. u32 ptr = ntohs(th->urg_ptr);
  3692. if (ptr && !sysctl_tcp_stdurg)
  3693. ptr--;
  3694. ptr += ntohl(th->seq);
  3695. /* Ignore urgent data that we've already seen and read. */
  3696. if (after(tp->copied_seq, ptr))
  3697. return;
  3698. /* Do not replay urg ptr.
  3699. *
  3700. * NOTE: interesting situation not covered by specs.
  3701. * Misbehaving sender may send urg ptr, pointing to segment,
  3702. * which we already have in ofo queue. We are not able to fetch
  3703. * such data and will stay in TCP_URG_NOTYET until will be eaten
  3704. * by recvmsg(). Seems, we are not obliged to handle such wicked
  3705. * situations. But it is worth to think about possibility of some
  3706. * DoSes using some hypothetical application level deadlock.
  3707. */
  3708. if (before(ptr, tp->rcv_nxt))
  3709. return;
  3710. /* Do we already have a newer (or duplicate) urgent pointer? */
  3711. if (tp->urg_data && !after(ptr, tp->urg_seq))
  3712. return;
  3713. /* Tell the world about our new urgent pointer. */
  3714. sk_send_sigurg(sk);
  3715. /* We may be adding urgent data when the last byte read was
  3716. * urgent. To do this requires some care. We cannot just ignore
  3717. * tp->copied_seq since we would read the last urgent byte again
  3718. * as data, nor can we alter copied_seq until this data arrives
  3719. * or we break the semantics of SIOCATMARK (and thus sockatmark())
  3720. *
  3721. * NOTE. Double Dutch. Rendering to plain English: author of comment
  3722. * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
  3723. * and expect that both A and B disappear from stream. This is _wrong_.
  3724. * Though this happens in BSD with high probability, this is occasional.
  3725. * Any application relying on this is buggy. Note also, that fix "works"
  3726. * only in this artificial test. Insert some normal data between A and B and we will
  3727. * decline of BSD again. Verdict: it is better to remove to trap
  3728. * buggy users.
  3729. */
  3730. if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
  3731. !sock_flag(sk, SOCK_URGINLINE) &&
  3732. tp->copied_seq != tp->rcv_nxt) {
  3733. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  3734. tp->copied_seq++;
  3735. if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
  3736. __skb_unlink(skb, &sk->sk_receive_queue);
  3737. __kfree_skb(skb);
  3738. }
  3739. }
  3740. tp->urg_data = TCP_URG_NOTYET;
  3741. tp->urg_seq = ptr;
  3742. /* Disable header prediction. */
  3743. tp->pred_flags = 0;
  3744. }
  3745. /* This is the 'fast' part of urgent handling. */
  3746. static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
  3747. {
  3748. struct tcp_sock *tp = tcp_sk(sk);
  3749. /* Check if we get a new urgent pointer - normally not. */
  3750. if (th->urg)
  3751. tcp_check_urg(sk,th);
  3752. /* Do we wait for any urgent data? - normally not... */
  3753. if (tp->urg_data == TCP_URG_NOTYET) {
  3754. u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
  3755. th->syn;
  3756. /* Is the urgent pointer pointing into this packet? */
  3757. if (ptr < skb->len) {
  3758. u8 tmp;
  3759. if (skb_copy_bits(skb, ptr, &tmp, 1))
  3760. BUG();
  3761. tp->urg_data = TCP_URG_VALID | tmp;
  3762. if (!sock_flag(sk, SOCK_DEAD))
  3763. sk->sk_data_ready(sk, 0);
  3764. }
  3765. }
  3766. }
  3767. static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
  3768. {
  3769. struct tcp_sock *tp = tcp_sk(sk);
  3770. int chunk = skb->len - hlen;
  3771. int err;
  3772. local_bh_enable();
  3773. if (skb_csum_unnecessary(skb))
  3774. err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
  3775. else
  3776. err = skb_copy_and_csum_datagram_iovec(skb, hlen,
  3777. tp->ucopy.iov);
  3778. if (!err) {
  3779. tp->ucopy.len -= chunk;
  3780. tp->copied_seq += chunk;
  3781. tcp_rcv_space_adjust(sk);
  3782. }
  3783. local_bh_disable();
  3784. return err;
  3785. }
  3786. static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3787. {
  3788. __sum16 result;
  3789. if (sock_owned_by_user(sk)) {
  3790. local_bh_enable();
  3791. result = __tcp_checksum_complete(skb);
  3792. local_bh_disable();
  3793. } else {
  3794. result = __tcp_checksum_complete(skb);
  3795. }
  3796. return result;
  3797. }
  3798. static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3799. {
  3800. return !skb_csum_unnecessary(skb) &&
  3801. __tcp_checksum_complete_user(sk, skb);
  3802. }
  3803. #ifdef CONFIG_NET_DMA
  3804. static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
  3805. {
  3806. struct tcp_sock *tp = tcp_sk(sk);
  3807. int chunk = skb->len - hlen;
  3808. int dma_cookie;
  3809. int copied_early = 0;
  3810. if (tp->ucopy.wakeup)
  3811. return 0;
  3812. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  3813. tp->ucopy.dma_chan = get_softnet_dma();
  3814. if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
  3815. dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
  3816. skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
  3817. if (dma_cookie < 0)
  3818. goto out;
  3819. tp->ucopy.dma_cookie = dma_cookie;
  3820. copied_early = 1;
  3821. tp->ucopy.len -= chunk;
  3822. tp->copied_seq += chunk;
  3823. tcp_rcv_space_adjust(sk);
  3824. if ((tp->ucopy.len == 0) ||
  3825. (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
  3826. (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
  3827. tp->ucopy.wakeup = 1;
  3828. sk->sk_data_ready(sk, 0);
  3829. }
  3830. } else if (chunk > 0) {
  3831. tp->ucopy.wakeup = 1;
  3832. sk->sk_data_ready(sk, 0);
  3833. }
  3834. out:
  3835. return copied_early;
  3836. }
  3837. #endif /* CONFIG_NET_DMA */
  3838. /*
  3839. * TCP receive function for the ESTABLISHED state.
  3840. *
  3841. * It is split into a fast path and a slow path. The fast path is
  3842. * disabled when:
  3843. * - A zero window was announced from us - zero window probing
  3844. * is only handled properly in the slow path.
  3845. * - Out of order segments arrived.
  3846. * - Urgent data is expected.
  3847. * - There is no buffer space left
  3848. * - Unexpected TCP flags/window values/header lengths are received
  3849. * (detected by checking the TCP header against pred_flags)
  3850. * - Data is sent in both directions. Fast path only supports pure senders
  3851. * or pure receivers (this means either the sequence number or the ack
  3852. * value must stay constant)
  3853. * - Unexpected TCP option.
  3854. *
  3855. * When these conditions are not satisfied it drops into a standard
  3856. * receive procedure patterned after RFC793 to handle all cases.
  3857. * The first three cases are guaranteed by proper pred_flags setting,
  3858. * the rest is checked inline. Fast processing is turned on in
  3859. * tcp_data_queue when everything is OK.
  3860. */
  3861. int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  3862. struct tcphdr *th, unsigned len)
  3863. {
  3864. struct tcp_sock *tp = tcp_sk(sk);
  3865. /*
  3866. * Header prediction.
  3867. * The code loosely follows the one in the famous
  3868. * "30 instruction TCP receive" Van Jacobson mail.
  3869. *
  3870. * Van's trick is to deposit buffers into socket queue
  3871. * on a device interrupt, to call tcp_recv function
  3872. * on the receive process context and checksum and copy
  3873. * the buffer to user space. smart...
  3874. *
  3875. * Our current scheme is not silly either but we take the
  3876. * extra cost of the net_bh soft interrupt processing...
  3877. * We do checksum and copy also but from device to kernel.
  3878. */
  3879. tp->rx_opt.saw_tstamp = 0;
  3880. /* pred_flags is 0xS?10 << 16 + snd_wnd
  3881. * if header_prediction is to be made
  3882. * 'S' will always be tp->tcp_header_len >> 2
  3883. * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
  3884. * turn it off (when there are holes in the receive
  3885. * space for instance)
  3886. * PSH flag is ignored.
  3887. */
  3888. if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
  3889. TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3890. int tcp_header_len = tp->tcp_header_len;
  3891. /* Timestamp header prediction: tcp_header_len
  3892. * is automatically equal to th->doff*4 due to pred_flags
  3893. * match.
  3894. */
  3895. /* Check timestamp */
  3896. if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
  3897. __be32 *ptr = (__be32 *)(th + 1);
  3898. /* No? Slow path! */
  3899. if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  3900. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
  3901. goto slow_path;
  3902. tp->rx_opt.saw_tstamp = 1;
  3903. ++ptr;
  3904. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  3905. ++ptr;
  3906. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  3907. /* If PAWS failed, check it more carefully in slow path */
  3908. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
  3909. goto slow_path;
  3910. /* DO NOT update ts_recent here, if checksum fails
  3911. * and timestamp was corrupted part, it will result
  3912. * in a hung connection since we will drop all
  3913. * future packets due to the PAWS test.
  3914. */
  3915. }
  3916. if (len <= tcp_header_len) {
  3917. /* Bulk data transfer: sender */
  3918. if (len == tcp_header_len) {
  3919. /* Predicted packet is in window by definition.
  3920. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3921. * Hence, check seq<=rcv_wup reduces to:
  3922. */
  3923. if (tcp_header_len ==
  3924. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3925. tp->rcv_nxt == tp->rcv_wup)
  3926. tcp_store_ts_recent(tp);
  3927. /* We know that such packets are checksummed
  3928. * on entry.
  3929. */
  3930. tcp_ack(sk, skb, 0);
  3931. __kfree_skb(skb);
  3932. tcp_data_snd_check(sk);
  3933. return 0;
  3934. } else { /* Header too small */
  3935. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3936. goto discard;
  3937. }
  3938. } else {
  3939. int eaten = 0;
  3940. int copied_early = 0;
  3941. if (tp->copied_seq == tp->rcv_nxt &&
  3942. len - tcp_header_len <= tp->ucopy.len) {
  3943. #ifdef CONFIG_NET_DMA
  3944. if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
  3945. copied_early = 1;
  3946. eaten = 1;
  3947. }
  3948. #endif
  3949. if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
  3950. __set_current_state(TASK_RUNNING);
  3951. if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
  3952. eaten = 1;
  3953. }
  3954. if (eaten) {
  3955. /* Predicted packet is in window by definition.
  3956. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3957. * Hence, check seq<=rcv_wup reduces to:
  3958. */
  3959. if (tcp_header_len ==
  3960. (sizeof(struct tcphdr) +
  3961. TCPOLEN_TSTAMP_ALIGNED) &&
  3962. tp->rcv_nxt == tp->rcv_wup)
  3963. tcp_store_ts_recent(tp);
  3964. tcp_rcv_rtt_measure_ts(sk, skb);
  3965. __skb_pull(skb, tcp_header_len);
  3966. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3967. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
  3968. }
  3969. if (copied_early)
  3970. tcp_cleanup_rbuf(sk, skb->len);
  3971. }
  3972. if (!eaten) {
  3973. if (tcp_checksum_complete_user(sk, skb))
  3974. goto csum_error;
  3975. /* Predicted packet is in window by definition.
  3976. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3977. * Hence, check seq<=rcv_wup reduces to:
  3978. */
  3979. if (tcp_header_len ==
  3980. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3981. tp->rcv_nxt == tp->rcv_wup)
  3982. tcp_store_ts_recent(tp);
  3983. tcp_rcv_rtt_measure_ts(sk, skb);
  3984. if ((int)skb->truesize > sk->sk_forward_alloc)
  3985. goto step5;
  3986. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
  3987. /* Bulk data transfer: receiver */
  3988. __skb_pull(skb,tcp_header_len);
  3989. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3990. sk_stream_set_owner_r(skb, sk);
  3991. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3992. }
  3993. tcp_event_data_recv(sk, skb);
  3994. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
  3995. /* Well, only one small jumplet in fast path... */
  3996. tcp_ack(sk, skb, FLAG_DATA);
  3997. tcp_data_snd_check(sk);
  3998. if (!inet_csk_ack_scheduled(sk))
  3999. goto no_ack;
  4000. }
  4001. __tcp_ack_snd_check(sk, 0);
  4002. no_ack:
  4003. #ifdef CONFIG_NET_DMA
  4004. if (copied_early)
  4005. __skb_queue_tail(&sk->sk_async_wait_queue, skb);
  4006. else
  4007. #endif
  4008. if (eaten)
  4009. __kfree_skb(skb);
  4010. else
  4011. sk->sk_data_ready(sk, 0);
  4012. return 0;
  4013. }
  4014. }
  4015. slow_path:
  4016. if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
  4017. goto csum_error;
  4018. /*
  4019. * RFC1323: H1. Apply PAWS check first.
  4020. */
  4021. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  4022. tcp_paws_discard(sk, skb)) {
  4023. if (!th->rst) {
  4024. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  4025. tcp_send_dupack(sk, skb);
  4026. goto discard;
  4027. }
  4028. /* Resets are accepted even if PAWS failed.
  4029. ts_recent update must be made after we are sure
  4030. that the packet is in window.
  4031. */
  4032. }
  4033. /*
  4034. * Standard slow path.
  4035. */
  4036. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  4037. /* RFC793, page 37: "In all states except SYN-SENT, all reset
  4038. * (RST) segments are validated by checking their SEQ-fields."
  4039. * And page 69: "If an incoming segment is not acceptable,
  4040. * an acknowledgment should be sent in reply (unless the RST bit
  4041. * is set, if so drop the segment and return)".
  4042. */
  4043. if (!th->rst)
  4044. tcp_send_dupack(sk, skb);
  4045. goto discard;
  4046. }
  4047. if (th->rst) {
  4048. tcp_reset(sk);
  4049. goto discard;
  4050. }
  4051. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  4052. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  4053. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  4054. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  4055. tcp_reset(sk);
  4056. return 1;
  4057. }
  4058. step5:
  4059. if (th->ack)
  4060. tcp_ack(sk, skb, FLAG_SLOWPATH);
  4061. tcp_rcv_rtt_measure_ts(sk, skb);
  4062. /* Process urgent data. */
  4063. tcp_urg(sk, skb, th);
  4064. /* step 7: process the segment text */
  4065. tcp_data_queue(sk, skb);
  4066. tcp_data_snd_check(sk);
  4067. tcp_ack_snd_check(sk);
  4068. return 0;
  4069. csum_error:
  4070. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  4071. discard:
  4072. __kfree_skb(skb);
  4073. return 0;
  4074. }
  4075. static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
  4076. struct tcphdr *th, unsigned len)
  4077. {
  4078. struct tcp_sock *tp = tcp_sk(sk);
  4079. struct inet_connection_sock *icsk = inet_csk(sk);
  4080. int saved_clamp = tp->rx_opt.mss_clamp;
  4081. tcp_parse_options(skb, &tp->rx_opt, 0);
  4082. if (th->ack) {
  4083. /* rfc793:
  4084. * "If the state is SYN-SENT then
  4085. * first check the ACK bit
  4086. * If the ACK bit is set
  4087. * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
  4088. * a reset (unless the RST bit is set, if so drop
  4089. * the segment and return)"
  4090. *
  4091. * We do not send data with SYN, so that RFC-correct
  4092. * test reduces to:
  4093. */
  4094. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
  4095. goto reset_and_undo;
  4096. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  4097. !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
  4098. tcp_time_stamp)) {
  4099. NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
  4100. goto reset_and_undo;
  4101. }
  4102. /* Now ACK is acceptable.
  4103. *
  4104. * "If the RST bit is set
  4105. * If the ACK was acceptable then signal the user "error:
  4106. * connection reset", drop the segment, enter CLOSED state,
  4107. * delete TCB, and return."
  4108. */
  4109. if (th->rst) {
  4110. tcp_reset(sk);
  4111. goto discard;
  4112. }
  4113. /* rfc793:
  4114. * "fifth, if neither of the SYN or RST bits is set then
  4115. * drop the segment and return."
  4116. *
  4117. * See note below!
  4118. * --ANK(990513)
  4119. */
  4120. if (!th->syn)
  4121. goto discard_and_undo;
  4122. /* rfc793:
  4123. * "If the SYN bit is on ...
  4124. * are acceptable then ...
  4125. * (our SYN has been ACKed), change the connection
  4126. * state to ESTABLISHED..."
  4127. */
  4128. TCP_ECN_rcv_synack(tp, th);
  4129. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  4130. tcp_ack(sk, skb, FLAG_SLOWPATH);
  4131. /* Ok.. it's good. Set up sequence numbers and
  4132. * move to established.
  4133. */
  4134. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  4135. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  4136. /* RFC1323: The window in SYN & SYN/ACK segments is
  4137. * never scaled.
  4138. */
  4139. tp->snd_wnd = ntohs(th->window);
  4140. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
  4141. if (!tp->rx_opt.wscale_ok) {
  4142. tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
  4143. tp->window_clamp = min(tp->window_clamp, 65535U);
  4144. }
  4145. if (tp->rx_opt.saw_tstamp) {
  4146. tp->rx_opt.tstamp_ok = 1;
  4147. tp->tcp_header_len =
  4148. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  4149. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4150. tcp_store_ts_recent(tp);
  4151. } else {
  4152. tp->tcp_header_len = sizeof(struct tcphdr);
  4153. }
  4154. if (tcp_is_sack(tp) && sysctl_tcp_fack)
  4155. tcp_enable_fack(tp);
  4156. tcp_mtup_init(sk);
  4157. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4158. tcp_initialize_rcv_mss(sk);
  4159. /* Remember, tcp_poll() does not lock socket!
  4160. * Change state from SYN-SENT only after copied_seq
  4161. * is initialized. */
  4162. tp->copied_seq = tp->rcv_nxt;
  4163. smp_mb();
  4164. tcp_set_state(sk, TCP_ESTABLISHED);
  4165. security_inet_conn_established(sk, skb);
  4166. /* Make sure socket is routed, for correct metrics. */
  4167. icsk->icsk_af_ops->rebuild_header(sk);
  4168. tcp_init_metrics(sk);
  4169. tcp_init_congestion_control(sk);
  4170. /* Prevent spurious tcp_cwnd_restart() on first data
  4171. * packet.
  4172. */
  4173. tp->lsndtime = tcp_time_stamp;
  4174. tcp_init_buffer_space(sk);
  4175. if (sock_flag(sk, SOCK_KEEPOPEN))
  4176. inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
  4177. if (!tp->rx_opt.snd_wscale)
  4178. __tcp_fast_path_on(tp, tp->snd_wnd);
  4179. else
  4180. tp->pred_flags = 0;
  4181. if (!sock_flag(sk, SOCK_DEAD)) {
  4182. sk->sk_state_change(sk);
  4183. sk_wake_async(sk, 0, POLL_OUT);
  4184. }
  4185. if (sk->sk_write_pending ||
  4186. icsk->icsk_accept_queue.rskq_defer_accept ||
  4187. icsk->icsk_ack.pingpong) {
  4188. /* Save one ACK. Data will be ready after
  4189. * several ticks, if write_pending is set.
  4190. *
  4191. * It may be deleted, but with this feature tcpdumps
  4192. * look so _wonderfully_ clever, that I was not able
  4193. * to stand against the temptation 8) --ANK
  4194. */
  4195. inet_csk_schedule_ack(sk);
  4196. icsk->icsk_ack.lrcvtime = tcp_time_stamp;
  4197. icsk->icsk_ack.ato = TCP_ATO_MIN;
  4198. tcp_incr_quickack(sk);
  4199. tcp_enter_quickack_mode(sk);
  4200. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  4201. TCP_DELACK_MAX, TCP_RTO_MAX);
  4202. discard:
  4203. __kfree_skb(skb);
  4204. return 0;
  4205. } else {
  4206. tcp_send_ack(sk);
  4207. }
  4208. return -1;
  4209. }
  4210. /* No ACK in the segment */
  4211. if (th->rst) {
  4212. /* rfc793:
  4213. * "If the RST bit is set
  4214. *
  4215. * Otherwise (no ACK) drop the segment and return."
  4216. */
  4217. goto discard_and_undo;
  4218. }
  4219. /* PAWS check. */
  4220. if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
  4221. goto discard_and_undo;
  4222. if (th->syn) {
  4223. /* We see SYN without ACK. It is attempt of
  4224. * simultaneous connect with crossed SYNs.
  4225. * Particularly, it can be connect to self.
  4226. */
  4227. tcp_set_state(sk, TCP_SYN_RECV);
  4228. if (tp->rx_opt.saw_tstamp) {
  4229. tp->rx_opt.tstamp_ok = 1;
  4230. tcp_store_ts_recent(tp);
  4231. tp->tcp_header_len =
  4232. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  4233. } else {
  4234. tp->tcp_header_len = sizeof(struct tcphdr);
  4235. }
  4236. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  4237. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  4238. /* RFC1323: The window in SYN & SYN/ACK segments is
  4239. * never scaled.
  4240. */
  4241. tp->snd_wnd = ntohs(th->window);
  4242. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  4243. tp->max_window = tp->snd_wnd;
  4244. TCP_ECN_rcv_syn(tp, th);
  4245. tcp_mtup_init(sk);
  4246. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4247. tcp_initialize_rcv_mss(sk);
  4248. tcp_send_synack(sk);
  4249. #if 0
  4250. /* Note, we could accept data and URG from this segment.
  4251. * There are no obstacles to make this.
  4252. *
  4253. * However, if we ignore data in ACKless segments sometimes,
  4254. * we have no reasons to accept it sometimes.
  4255. * Also, seems the code doing it in step6 of tcp_rcv_state_process
  4256. * is not flawless. So, discard packet for sanity.
  4257. * Uncomment this return to process the data.
  4258. */
  4259. return -1;
  4260. #else
  4261. goto discard;
  4262. #endif
  4263. }
  4264. /* "fifth, if neither of the SYN or RST bits is set then
  4265. * drop the segment and return."
  4266. */
  4267. discard_and_undo:
  4268. tcp_clear_options(&tp->rx_opt);
  4269. tp->rx_opt.mss_clamp = saved_clamp;
  4270. goto discard;
  4271. reset_and_undo:
  4272. tcp_clear_options(&tp->rx_opt);
  4273. tp->rx_opt.mss_clamp = saved_clamp;
  4274. return 1;
  4275. }
  4276. /*
  4277. * This function implements the receiving procedure of RFC 793 for
  4278. * all states except ESTABLISHED and TIME_WAIT.
  4279. * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
  4280. * address independent.
  4281. */
  4282. int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
  4283. struct tcphdr *th, unsigned len)
  4284. {
  4285. struct tcp_sock *tp = tcp_sk(sk);
  4286. struct inet_connection_sock *icsk = inet_csk(sk);
  4287. int queued = 0;
  4288. tp->rx_opt.saw_tstamp = 0;
  4289. switch (sk->sk_state) {
  4290. case TCP_CLOSE:
  4291. goto discard;
  4292. case TCP_LISTEN:
  4293. if (th->ack)
  4294. return 1;
  4295. if (th->rst)
  4296. goto discard;
  4297. if (th->syn) {
  4298. if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
  4299. return 1;
  4300. /* Now we have several options: In theory there is
  4301. * nothing else in the frame. KA9Q has an option to
  4302. * send data with the syn, BSD accepts data with the
  4303. * syn up to the [to be] advertised window and
  4304. * Solaris 2.1 gives you a protocol error. For now
  4305. * we just ignore it, that fits the spec precisely
  4306. * and avoids incompatibilities. It would be nice in
  4307. * future to drop through and process the data.
  4308. *
  4309. * Now that TTCP is starting to be used we ought to
  4310. * queue this data.
  4311. * But, this leaves one open to an easy denial of
  4312. * service attack, and SYN cookies can't defend
  4313. * against this problem. So, we drop the data
  4314. * in the interest of security over speed unless
  4315. * it's still in use.
  4316. */
  4317. kfree_skb(skb);
  4318. return 0;
  4319. }
  4320. goto discard;
  4321. case TCP_SYN_SENT:
  4322. queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
  4323. if (queued >= 0)
  4324. return queued;
  4325. /* Do step6 onward by hand. */
  4326. tcp_urg(sk, skb, th);
  4327. __kfree_skb(skb);
  4328. tcp_data_snd_check(sk);
  4329. return 0;
  4330. }
  4331. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  4332. tcp_paws_discard(sk, skb)) {
  4333. if (!th->rst) {
  4334. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  4335. tcp_send_dupack(sk, skb);
  4336. goto discard;
  4337. }
  4338. /* Reset is accepted even if it did not pass PAWS. */
  4339. }
  4340. /* step 1: check sequence number */
  4341. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  4342. if (!th->rst)
  4343. tcp_send_dupack(sk, skb);
  4344. goto discard;
  4345. }
  4346. /* step 2: check RST bit */
  4347. if (th->rst) {
  4348. tcp_reset(sk);
  4349. goto discard;
  4350. }
  4351. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  4352. /* step 3: check security and precedence [ignored] */
  4353. /* step 4:
  4354. *
  4355. * Check for a SYN in window.
  4356. */
  4357. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  4358. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  4359. tcp_reset(sk);
  4360. return 1;
  4361. }
  4362. /* step 5: check the ACK field */
  4363. if (th->ack) {
  4364. int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
  4365. switch (sk->sk_state) {
  4366. case TCP_SYN_RECV:
  4367. if (acceptable) {
  4368. tp->copied_seq = tp->rcv_nxt;
  4369. smp_mb();
  4370. tcp_set_state(sk, TCP_ESTABLISHED);
  4371. sk->sk_state_change(sk);
  4372. /* Note, that this wakeup is only for marginal
  4373. * crossed SYN case. Passively open sockets
  4374. * are not waked up, because sk->sk_sleep ==
  4375. * NULL and sk->sk_socket == NULL.
  4376. */
  4377. if (sk->sk_socket) {
  4378. sk_wake_async(sk,0,POLL_OUT);
  4379. }
  4380. tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
  4381. tp->snd_wnd = ntohs(th->window) <<
  4382. tp->rx_opt.snd_wscale;
  4383. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
  4384. TCP_SKB_CB(skb)->seq);
  4385. /* tcp_ack considers this ACK as duplicate
  4386. * and does not calculate rtt.
  4387. * Fix it at least with timestamps.
  4388. */
  4389. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  4390. !tp->srtt)
  4391. tcp_ack_saw_tstamp(sk, 0);
  4392. if (tp->rx_opt.tstamp_ok)
  4393. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4394. /* Make sure socket is routed, for
  4395. * correct metrics.
  4396. */
  4397. icsk->icsk_af_ops->rebuild_header(sk);
  4398. tcp_init_metrics(sk);
  4399. tcp_init_congestion_control(sk);
  4400. /* Prevent spurious tcp_cwnd_restart() on
  4401. * first data packet.
  4402. */
  4403. tp->lsndtime = tcp_time_stamp;
  4404. tcp_mtup_init(sk);
  4405. tcp_initialize_rcv_mss(sk);
  4406. tcp_init_buffer_space(sk);
  4407. tcp_fast_path_on(tp);
  4408. } else {
  4409. return 1;
  4410. }
  4411. break;
  4412. case TCP_FIN_WAIT1:
  4413. if (tp->snd_una == tp->write_seq) {
  4414. tcp_set_state(sk, TCP_FIN_WAIT2);
  4415. sk->sk_shutdown |= SEND_SHUTDOWN;
  4416. dst_confirm(sk->sk_dst_cache);
  4417. if (!sock_flag(sk, SOCK_DEAD))
  4418. /* Wake up lingering close() */
  4419. sk->sk_state_change(sk);
  4420. else {
  4421. int tmo;
  4422. if (tp->linger2 < 0 ||
  4423. (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4424. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
  4425. tcp_done(sk);
  4426. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4427. return 1;
  4428. }
  4429. tmo = tcp_fin_time(sk);
  4430. if (tmo > TCP_TIMEWAIT_LEN) {
  4431. inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
  4432. } else if (th->fin || sock_owned_by_user(sk)) {
  4433. /* Bad case. We could lose such FIN otherwise.
  4434. * It is not a big problem, but it looks confusing
  4435. * and not so rare event. We still can lose it now,
  4436. * if it spins in bh_lock_sock(), but it is really
  4437. * marginal case.
  4438. */
  4439. inet_csk_reset_keepalive_timer(sk, tmo);
  4440. } else {
  4441. tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
  4442. goto discard;
  4443. }
  4444. }
  4445. }
  4446. break;
  4447. case TCP_CLOSING:
  4448. if (tp->snd_una == tp->write_seq) {
  4449. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  4450. goto discard;
  4451. }
  4452. break;
  4453. case TCP_LAST_ACK:
  4454. if (tp->snd_una == tp->write_seq) {
  4455. tcp_update_metrics(sk);
  4456. tcp_done(sk);
  4457. goto discard;
  4458. }
  4459. break;
  4460. }
  4461. } else
  4462. goto discard;
  4463. /* step 6: check the URG bit */
  4464. tcp_urg(sk, skb, th);
  4465. /* step 7: process the segment text */
  4466. switch (sk->sk_state) {
  4467. case TCP_CLOSE_WAIT:
  4468. case TCP_CLOSING:
  4469. case TCP_LAST_ACK:
  4470. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  4471. break;
  4472. case TCP_FIN_WAIT1:
  4473. case TCP_FIN_WAIT2:
  4474. /* RFC 793 says to queue data in these states,
  4475. * RFC 1122 says we MUST send a reset.
  4476. * BSD 4.4 also does reset.
  4477. */
  4478. if (sk->sk_shutdown & RCV_SHUTDOWN) {
  4479. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4480. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
  4481. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4482. tcp_reset(sk);
  4483. return 1;
  4484. }
  4485. }
  4486. /* Fall through */
  4487. case TCP_ESTABLISHED:
  4488. tcp_data_queue(sk, skb);
  4489. queued = 1;
  4490. break;
  4491. }
  4492. /* tcp_data could move socket to TIME-WAIT */
  4493. if (sk->sk_state != TCP_CLOSE) {
  4494. tcp_data_snd_check(sk);
  4495. tcp_ack_snd_check(sk);
  4496. }
  4497. if (!queued) {
  4498. discard:
  4499. __kfree_skb(skb);
  4500. }
  4501. return 0;
  4502. }
  4503. EXPORT_SYMBOL(sysctl_tcp_ecn);
  4504. EXPORT_SYMBOL(sysctl_tcp_reordering);
  4505. EXPORT_SYMBOL(tcp_parse_options);
  4506. EXPORT_SYMBOL(tcp_rcv_established);
  4507. EXPORT_SYMBOL(tcp_rcv_state_process);
  4508. EXPORT_SYMBOL(tcp_initialize_rcv_mss);