tcp_input.c 142 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890389138923893389438953896389738983899390039013902390339043905390639073908390939103911391239133914391539163917391839193920392139223923392439253926392739283929393039313932393339343935393639373938393939403941394239433944394539463947394839493950395139523953395439553956395739583959396039613962396339643965396639673968396939703971397239733974397539763977397839793980398139823983398439853986398739883989399039913992399339943995399639973998399940004001400240034004400540064007400840094010401140124013401440154016401740184019402040214022402340244025402640274028402940304031403240334034403540364037403840394040404140424043404440454046404740484049405040514052405340544055405640574058405940604061406240634064406540664067406840694070407140724073407440754076407740784079408040814082408340844085408640874088408940904091409240934094409540964097409840994100410141024103410441054106410741084109411041114112411341144115411641174118411941204121412241234124412541264127412841294130413141324133413441354136413741384139414041414142414341444145414641474148414941504151415241534154415541564157415841594160416141624163416441654166416741684169417041714172417341744175417641774178417941804181418241834184418541864187418841894190419141924193419441954196419741984199420042014202420342044205420642074208420942104211421242134214421542164217421842194220422142224223422442254226422742284229423042314232423342344235423642374238423942404241424242434244424542464247424842494250425142524253425442554256425742584259426042614262426342644265426642674268426942704271427242734274427542764277427842794280428142824283428442854286428742884289429042914292429342944295429642974298429943004301430243034304430543064307430843094310431143124313431443154316431743184319432043214322432343244325432643274328432943304331433243334334433543364337433843394340434143424343434443454346434743484349435043514352435343544355435643574358435943604361436243634364436543664367436843694370437143724373437443754376437743784379438043814382438343844385438643874388438943904391439243934394439543964397439843994400440144024403440444054406440744084409441044114412441344144415441644174418441944204421442244234424442544264427442844294430443144324433443444354436443744384439444044414442444344444445444644474448444944504451445244534454445544564457445844594460446144624463446444654466446744684469447044714472447344744475447644774478447944804481448244834484448544864487448844894490449144924493449444954496449744984499450045014502450345044505450645074508450945104511451245134514451545164517451845194520452145224523452445254526452745284529453045314532453345344535453645374538453945404541454245434544454545464547454845494550455145524553455445554556455745584559456045614562456345644565456645674568456945704571457245734574457545764577457845794580458145824583458445854586458745884589459045914592459345944595459645974598459946004601460246034604460546064607460846094610461146124613461446154616461746184619462046214622462346244625462646274628462946304631463246334634463546364637463846394640464146424643464446454646464746484649465046514652465346544655465646574658465946604661466246634664466546664667466846694670467146724673467446754676467746784679468046814682468346844685468646874688468946904691469246934694469546964697469846994700470147024703470447054706470747084709471047114712471347144715471647174718471947204721472247234724472547264727472847294730473147324733473447354736473747384739474047414742474347444745474647474748474947504751475247534754475547564757475847594760476147624763476447654766476747684769477047714772477347744775477647774778477947804781478247834784478547864787478847894790479147924793479447954796479747984799480048014802480348044805480648074808480948104811481248134814481548164817481848194820482148224823482448254826482748284829483048314832483348344835483648374838483948404841484248434844484548464847484848494850485148524853485448554856485748584859486048614862486348644865486648674868486948704871487248734874487548764877487848794880488148824883488448854886488748884889489048914892489348944895489648974898489949004901490249034904490549064907490849094910491149124913491449154916491749184919492049214922492349244925492649274928492949304931493249334934493549364937493849394940494149424943494449454946494749484949495049514952495349544955495649574958495949604961496249634964496549664967496849694970497149724973497449754976497749784979498049814982498349844985498649874988498949904991499249934994499549964997499849995000500150025003500450055006500750085009501050115012501350145015501650175018501950205021
  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;
  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 DSACK info */
  101. #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
  102. #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
  103. #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
  104. #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
  105. #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED)
  106. #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
  107. #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
  108. #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
  109. #define IsSackFrto() (sysctl_tcp_frto == 0x2)
  110. #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
  111. #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
  112. /* Adapt the MSS value used to make delayed ack decision to the
  113. * real world.
  114. */
  115. static void tcp_measure_rcv_mss(struct sock *sk,
  116. const struct sk_buff *skb)
  117. {
  118. struct inet_connection_sock *icsk = inet_csk(sk);
  119. const unsigned int lss = icsk->icsk_ack.last_seg_size;
  120. unsigned int len;
  121. icsk->icsk_ack.last_seg_size = 0;
  122. /* skb->len may jitter because of SACKs, even if peer
  123. * sends good full-sized frames.
  124. */
  125. len = skb_shinfo(skb)->gso_size ?: skb->len;
  126. if (len >= icsk->icsk_ack.rcv_mss) {
  127. icsk->icsk_ack.rcv_mss = len;
  128. } else {
  129. /* Otherwise, we make more careful check taking into account,
  130. * that SACKs block is variable.
  131. *
  132. * "len" is invariant segment length, including TCP header.
  133. */
  134. len += skb->data - skb_transport_header(skb);
  135. if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
  136. /* If PSH is not set, packet should be
  137. * full sized, provided peer TCP is not badly broken.
  138. * This observation (if it is correct 8)) allows
  139. * to handle super-low mtu links fairly.
  140. */
  141. (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
  142. !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
  143. /* Subtract also invariant (if peer is RFC compliant),
  144. * tcp header plus fixed timestamp option length.
  145. * Resulting "len" is MSS free of SACK jitter.
  146. */
  147. len -= tcp_sk(sk)->tcp_header_len;
  148. icsk->icsk_ack.last_seg_size = len;
  149. if (len == lss) {
  150. icsk->icsk_ack.rcv_mss = len;
  151. return;
  152. }
  153. }
  154. if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
  155. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
  156. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
  157. }
  158. }
  159. static void tcp_incr_quickack(struct sock *sk)
  160. {
  161. struct inet_connection_sock *icsk = inet_csk(sk);
  162. unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
  163. if (quickacks==0)
  164. quickacks=2;
  165. if (quickacks > icsk->icsk_ack.quick)
  166. icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
  167. }
  168. void tcp_enter_quickack_mode(struct sock *sk)
  169. {
  170. struct inet_connection_sock *icsk = inet_csk(sk);
  171. tcp_incr_quickack(sk);
  172. icsk->icsk_ack.pingpong = 0;
  173. icsk->icsk_ack.ato = TCP_ATO_MIN;
  174. }
  175. /* Send ACKs quickly, if "quick" count is not exhausted
  176. * and the session is not interactive.
  177. */
  178. static inline int tcp_in_quickack_mode(const struct sock *sk)
  179. {
  180. const struct inet_connection_sock *icsk = inet_csk(sk);
  181. return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
  182. }
  183. static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
  184. {
  185. if (tp->ecn_flags&TCP_ECN_OK)
  186. tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
  187. }
  188. static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb)
  189. {
  190. if (tcp_hdr(skb)->cwr)
  191. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  192. }
  193. static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
  194. {
  195. tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
  196. }
  197. static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb)
  198. {
  199. if (tp->ecn_flags&TCP_ECN_OK) {
  200. if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags))
  201. tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
  202. /* Funny extension: if ECT is not set on a segment,
  203. * it is surely retransmit. It is not in ECN RFC,
  204. * but Linux follows this rule. */
  205. else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags)))
  206. tcp_enter_quickack_mode((struct sock *)tp);
  207. }
  208. }
  209. static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th)
  210. {
  211. if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr))
  212. tp->ecn_flags &= ~TCP_ECN_OK;
  213. }
  214. static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th)
  215. {
  216. if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr))
  217. tp->ecn_flags &= ~TCP_ECN_OK;
  218. }
  219. static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th)
  220. {
  221. if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK))
  222. return 1;
  223. return 0;
  224. }
  225. /* Buffer size and advertised window tuning.
  226. *
  227. * 1. Tuning sk->sk_sndbuf, when connection enters established state.
  228. */
  229. static void tcp_fixup_sndbuf(struct sock *sk)
  230. {
  231. int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
  232. sizeof(struct sk_buff);
  233. if (sk->sk_sndbuf < 3 * sndmem)
  234. sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
  235. }
  236. /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
  237. *
  238. * All tcp_full_space() is split to two parts: "network" buffer, allocated
  239. * forward and advertised in receiver window (tp->rcv_wnd) and
  240. * "application buffer", required to isolate scheduling/application
  241. * latencies from network.
  242. * window_clamp is maximal advertised window. It can be less than
  243. * tcp_full_space(), in this case tcp_full_space() - window_clamp
  244. * is reserved for "application" buffer. The less window_clamp is
  245. * the smoother our behaviour from viewpoint of network, but the lower
  246. * throughput and the higher sensitivity of the connection to losses. 8)
  247. *
  248. * rcv_ssthresh is more strict window_clamp used at "slow start"
  249. * phase to predict further behaviour of this connection.
  250. * It is used for two goals:
  251. * - to enforce header prediction at sender, even when application
  252. * requires some significant "application buffer". It is check #1.
  253. * - to prevent pruning of receive queue because of misprediction
  254. * of receiver window. Check #2.
  255. *
  256. * The scheme does not work when sender sends good segments opening
  257. * window and then starts to feed us spaghetti. But it should work
  258. * in common situations. Otherwise, we have to rely on queue collapsing.
  259. */
  260. /* Slow part of check#2. */
  261. static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
  262. {
  263. struct tcp_sock *tp = tcp_sk(sk);
  264. /* Optimize this! */
  265. int truesize = tcp_win_from_space(skb->truesize)/2;
  266. int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
  267. while (tp->rcv_ssthresh <= window) {
  268. if (truesize <= skb->len)
  269. return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
  270. truesize >>= 1;
  271. window >>= 1;
  272. }
  273. return 0;
  274. }
  275. static void tcp_grow_window(struct sock *sk,
  276. struct sk_buff *skb)
  277. {
  278. struct tcp_sock *tp = tcp_sk(sk);
  279. /* Check #1 */
  280. if (tp->rcv_ssthresh < tp->window_clamp &&
  281. (int)tp->rcv_ssthresh < tcp_space(sk) &&
  282. !tcp_memory_pressure) {
  283. int incr;
  284. /* Check #2. Increase window, if skb with such overhead
  285. * will fit to rcvbuf in future.
  286. */
  287. if (tcp_win_from_space(skb->truesize) <= skb->len)
  288. incr = 2*tp->advmss;
  289. else
  290. incr = __tcp_grow_window(sk, skb);
  291. if (incr) {
  292. tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
  293. inet_csk(sk)->icsk_ack.quick |= 1;
  294. }
  295. }
  296. }
  297. /* 3. Tuning rcvbuf, when connection enters established state. */
  298. static void tcp_fixup_rcvbuf(struct sock *sk)
  299. {
  300. struct tcp_sock *tp = tcp_sk(sk);
  301. int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
  302. /* Try to select rcvbuf so that 4 mss-sized segments
  303. * will fit to window and corresponding skbs will fit to our rcvbuf.
  304. * (was 3; 4 is minimum to allow fast retransmit to work.)
  305. */
  306. while (tcp_win_from_space(rcvmem) < tp->advmss)
  307. rcvmem += 128;
  308. if (sk->sk_rcvbuf < 4 * rcvmem)
  309. sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
  310. }
  311. /* 4. Try to fixup all. It is made immediately after connection enters
  312. * established state.
  313. */
  314. static void tcp_init_buffer_space(struct sock *sk)
  315. {
  316. struct tcp_sock *tp = tcp_sk(sk);
  317. int maxwin;
  318. if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
  319. tcp_fixup_rcvbuf(sk);
  320. if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
  321. tcp_fixup_sndbuf(sk);
  322. tp->rcvq_space.space = tp->rcv_wnd;
  323. maxwin = tcp_full_space(sk);
  324. if (tp->window_clamp >= maxwin) {
  325. tp->window_clamp = maxwin;
  326. if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
  327. tp->window_clamp = max(maxwin -
  328. (maxwin >> sysctl_tcp_app_win),
  329. 4 * tp->advmss);
  330. }
  331. /* Force reservation of one segment. */
  332. if (sysctl_tcp_app_win &&
  333. tp->window_clamp > 2 * tp->advmss &&
  334. tp->window_clamp + tp->advmss > maxwin)
  335. tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
  336. tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
  337. tp->snd_cwnd_stamp = tcp_time_stamp;
  338. }
  339. /* 5. Recalculate window clamp after socket hit its memory bounds. */
  340. static void tcp_clamp_window(struct sock *sk)
  341. {
  342. struct tcp_sock *tp = tcp_sk(sk);
  343. struct inet_connection_sock *icsk = inet_csk(sk);
  344. icsk->icsk_ack.quick = 0;
  345. if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
  346. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
  347. !tcp_memory_pressure &&
  348. atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
  349. sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
  350. sysctl_tcp_rmem[2]);
  351. }
  352. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
  353. tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
  354. }
  355. /* Initialize RCV_MSS value.
  356. * RCV_MSS is an our guess about MSS used by the peer.
  357. * We haven't any direct information about the MSS.
  358. * It's better to underestimate the RCV_MSS rather than overestimate.
  359. * Overestimations make us ACKing less frequently than needed.
  360. * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
  361. */
  362. void tcp_initialize_rcv_mss(struct sock *sk)
  363. {
  364. struct tcp_sock *tp = tcp_sk(sk);
  365. unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
  366. hint = min(hint, tp->rcv_wnd/2);
  367. hint = min(hint, TCP_MIN_RCVMSS);
  368. hint = max(hint, TCP_MIN_MSS);
  369. inet_csk(sk)->icsk_ack.rcv_mss = hint;
  370. }
  371. /* Receiver "autotuning" code.
  372. *
  373. * The algorithm for RTT estimation w/o timestamps is based on
  374. * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
  375. * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
  376. *
  377. * More detail on this code can be found at
  378. * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
  379. * though this reference is out of date. A new paper
  380. * is pending.
  381. */
  382. static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
  383. {
  384. u32 new_sample = tp->rcv_rtt_est.rtt;
  385. long m = sample;
  386. if (m == 0)
  387. m = 1;
  388. if (new_sample != 0) {
  389. /* If we sample in larger samples in the non-timestamp
  390. * case, we could grossly overestimate the RTT especially
  391. * with chatty applications or bulk transfer apps which
  392. * are stalled on filesystem I/O.
  393. *
  394. * Also, since we are only going for a minimum in the
  395. * non-timestamp case, we do not smooth things out
  396. * else with timestamps disabled convergence takes too
  397. * long.
  398. */
  399. if (!win_dep) {
  400. m -= (new_sample >> 3);
  401. new_sample += m;
  402. } else if (m < new_sample)
  403. new_sample = m << 3;
  404. } else {
  405. /* No previous measure. */
  406. new_sample = m << 3;
  407. }
  408. if (tp->rcv_rtt_est.rtt != new_sample)
  409. tp->rcv_rtt_est.rtt = new_sample;
  410. }
  411. static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
  412. {
  413. if (tp->rcv_rtt_est.time == 0)
  414. goto new_measure;
  415. if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
  416. return;
  417. tcp_rcv_rtt_update(tp,
  418. jiffies - tp->rcv_rtt_est.time,
  419. 1);
  420. new_measure:
  421. tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
  422. tp->rcv_rtt_est.time = tcp_time_stamp;
  423. }
  424. static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
  425. {
  426. struct tcp_sock *tp = tcp_sk(sk);
  427. if (tp->rx_opt.rcv_tsecr &&
  428. (TCP_SKB_CB(skb)->end_seq -
  429. TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
  430. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
  431. }
  432. /*
  433. * This function should be called every time data is copied to user space.
  434. * It calculates the appropriate TCP receive buffer space.
  435. */
  436. void tcp_rcv_space_adjust(struct sock *sk)
  437. {
  438. struct tcp_sock *tp = tcp_sk(sk);
  439. int time;
  440. int space;
  441. if (tp->rcvq_space.time == 0)
  442. goto new_measure;
  443. time = tcp_time_stamp - tp->rcvq_space.time;
  444. if (time < (tp->rcv_rtt_est.rtt >> 3) ||
  445. tp->rcv_rtt_est.rtt == 0)
  446. return;
  447. space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
  448. space = max(tp->rcvq_space.space, space);
  449. if (tp->rcvq_space.space != space) {
  450. int rcvmem;
  451. tp->rcvq_space.space = space;
  452. if (sysctl_tcp_moderate_rcvbuf &&
  453. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
  454. int new_clamp = space;
  455. /* Receive space grows, normalize in order to
  456. * take into account packet headers and sk_buff
  457. * structure overhead.
  458. */
  459. space /= tp->advmss;
  460. if (!space)
  461. space = 1;
  462. rcvmem = (tp->advmss + MAX_TCP_HEADER +
  463. 16 + sizeof(struct sk_buff));
  464. while (tcp_win_from_space(rcvmem) < tp->advmss)
  465. rcvmem += 128;
  466. space *= rcvmem;
  467. space = min(space, sysctl_tcp_rmem[2]);
  468. if (space > sk->sk_rcvbuf) {
  469. sk->sk_rcvbuf = space;
  470. /* Make the window clamp follow along. */
  471. tp->window_clamp = new_clamp;
  472. }
  473. }
  474. }
  475. new_measure:
  476. tp->rcvq_space.seq = tp->copied_seq;
  477. tp->rcvq_space.time = tcp_time_stamp;
  478. }
  479. /* There is something which you must keep in mind when you analyze the
  480. * behavior of the tp->ato delayed ack timeout interval. When a
  481. * connection starts up, we want to ack as quickly as possible. The
  482. * problem is that "good" TCP's do slow start at the beginning of data
  483. * transmission. The means that until we send the first few ACK's the
  484. * sender will sit on his end and only queue most of his data, because
  485. * he can only send snd_cwnd unacked packets at any given time. For
  486. * each ACK we send, he increments snd_cwnd and transmits more of his
  487. * queue. -DaveM
  488. */
  489. static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
  490. {
  491. struct tcp_sock *tp = tcp_sk(sk);
  492. struct inet_connection_sock *icsk = inet_csk(sk);
  493. u32 now;
  494. inet_csk_schedule_ack(sk);
  495. tcp_measure_rcv_mss(sk, skb);
  496. tcp_rcv_rtt_measure(tp);
  497. now = tcp_time_stamp;
  498. if (!icsk->icsk_ack.ato) {
  499. /* The _first_ data packet received, initialize
  500. * delayed ACK engine.
  501. */
  502. tcp_incr_quickack(sk);
  503. icsk->icsk_ack.ato = TCP_ATO_MIN;
  504. } else {
  505. int m = now - icsk->icsk_ack.lrcvtime;
  506. if (m <= TCP_ATO_MIN/2) {
  507. /* The fastest case is the first. */
  508. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
  509. } else if (m < icsk->icsk_ack.ato) {
  510. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
  511. if (icsk->icsk_ack.ato > icsk->icsk_rto)
  512. icsk->icsk_ack.ato = icsk->icsk_rto;
  513. } else if (m > icsk->icsk_rto) {
  514. /* Too long gap. Apparently sender failed to
  515. * restart window, so that we send ACKs quickly.
  516. */
  517. tcp_incr_quickack(sk);
  518. sk_stream_mem_reclaim(sk);
  519. }
  520. }
  521. icsk->icsk_ack.lrcvtime = now;
  522. TCP_ECN_check_ce(tp, skb);
  523. if (skb->len >= 128)
  524. tcp_grow_window(sk, skb);
  525. }
  526. static u32 tcp_rto_min(struct sock *sk)
  527. {
  528. struct dst_entry *dst = __sk_dst_get(sk);
  529. u32 rto_min = TCP_RTO_MIN;
  530. if (dst && dst_metric_locked(dst, RTAX_RTO_MIN))
  531. rto_min = dst->metrics[RTAX_RTO_MIN-1];
  532. return rto_min;
  533. }
  534. /* Called to compute a smoothed rtt estimate. The data fed to this
  535. * routine either comes from timestamps, or from segments that were
  536. * known _not_ to have been retransmitted [see Karn/Partridge
  537. * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
  538. * piece by Van Jacobson.
  539. * NOTE: the next three routines used to be one big routine.
  540. * To save cycles in the RFC 1323 implementation it was better to break
  541. * it up into three procedures. -- erics
  542. */
  543. static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
  544. {
  545. struct tcp_sock *tp = tcp_sk(sk);
  546. long m = mrtt; /* RTT */
  547. /* The following amusing code comes from Jacobson's
  548. * article in SIGCOMM '88. Note that rtt and mdev
  549. * are scaled versions of rtt and mean deviation.
  550. * This is designed to be as fast as possible
  551. * m stands for "measurement".
  552. *
  553. * On a 1990 paper the rto value is changed to:
  554. * RTO = rtt + 4 * mdev
  555. *
  556. * Funny. This algorithm seems to be very broken.
  557. * These formulae increase RTO, when it should be decreased, increase
  558. * too slowly, when it should be increased quickly, decrease too quickly
  559. * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
  560. * does not matter how to _calculate_ it. Seems, it was trap
  561. * that VJ failed to avoid. 8)
  562. */
  563. if (m == 0)
  564. m = 1;
  565. if (tp->srtt != 0) {
  566. m -= (tp->srtt >> 3); /* m is now error in rtt est */
  567. tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
  568. if (m < 0) {
  569. m = -m; /* m is now abs(error) */
  570. m -= (tp->mdev >> 2); /* similar update on mdev */
  571. /* This is similar to one of Eifel findings.
  572. * Eifel blocks mdev updates when rtt decreases.
  573. * This solution is a bit different: we use finer gain
  574. * for mdev in this case (alpha*beta).
  575. * Like Eifel it also prevents growth of rto,
  576. * but also it limits too fast rto decreases,
  577. * happening in pure Eifel.
  578. */
  579. if (m > 0)
  580. m >>= 3;
  581. } else {
  582. m -= (tp->mdev >> 2); /* similar update on mdev */
  583. }
  584. tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
  585. if (tp->mdev > tp->mdev_max) {
  586. tp->mdev_max = tp->mdev;
  587. if (tp->mdev_max > tp->rttvar)
  588. tp->rttvar = tp->mdev_max;
  589. }
  590. if (after(tp->snd_una, tp->rtt_seq)) {
  591. if (tp->mdev_max < tp->rttvar)
  592. tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
  593. tp->rtt_seq = tp->snd_nxt;
  594. tp->mdev_max = tcp_rto_min(sk);
  595. }
  596. } else {
  597. /* no previous measure. */
  598. tp->srtt = m<<3; /* take the measured time to be rtt */
  599. tp->mdev = m<<1; /* make sure rto = 3*rtt */
  600. tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
  601. tp->rtt_seq = tp->snd_nxt;
  602. }
  603. }
  604. /* Calculate rto without backoff. This is the second half of Van Jacobson's
  605. * routine referred to above.
  606. */
  607. static inline void tcp_set_rto(struct sock *sk)
  608. {
  609. const struct tcp_sock *tp = tcp_sk(sk);
  610. /* Old crap is replaced with new one. 8)
  611. *
  612. * More seriously:
  613. * 1. If rtt variance happened to be less 50msec, it is hallucination.
  614. * It cannot be less due to utterly erratic ACK generation made
  615. * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
  616. * to do with delayed acks, because at cwnd>2 true delack timeout
  617. * is invisible. Actually, Linux-2.4 also generates erratic
  618. * ACKs in some circumstances.
  619. */
  620. inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
  621. /* 2. Fixups made earlier cannot be right.
  622. * If we do not estimate RTO correctly without them,
  623. * all the algo is pure shit and should be replaced
  624. * with correct one. It is exactly, which we pretend to do.
  625. */
  626. }
  627. /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
  628. * guarantees that rto is higher.
  629. */
  630. static inline void tcp_bound_rto(struct sock *sk)
  631. {
  632. if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
  633. inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
  634. }
  635. /* Save metrics learned by this TCP session.
  636. This function is called only, when TCP finishes successfully
  637. i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
  638. */
  639. void tcp_update_metrics(struct sock *sk)
  640. {
  641. struct tcp_sock *tp = tcp_sk(sk);
  642. struct dst_entry *dst = __sk_dst_get(sk);
  643. if (sysctl_tcp_nometrics_save)
  644. return;
  645. dst_confirm(dst);
  646. if (dst && (dst->flags&DST_HOST)) {
  647. const struct inet_connection_sock *icsk = inet_csk(sk);
  648. int m;
  649. if (icsk->icsk_backoff || !tp->srtt) {
  650. /* This session failed to estimate rtt. Why?
  651. * Probably, no packets returned in time.
  652. * Reset our results.
  653. */
  654. if (!(dst_metric_locked(dst, RTAX_RTT)))
  655. dst->metrics[RTAX_RTT-1] = 0;
  656. return;
  657. }
  658. m = dst_metric(dst, RTAX_RTT) - tp->srtt;
  659. /* If newly calculated rtt larger than stored one,
  660. * store new one. Otherwise, use EWMA. Remember,
  661. * rtt overestimation is always better than underestimation.
  662. */
  663. if (!(dst_metric_locked(dst, RTAX_RTT))) {
  664. if (m <= 0)
  665. dst->metrics[RTAX_RTT-1] = tp->srtt;
  666. else
  667. dst->metrics[RTAX_RTT-1] -= (m>>3);
  668. }
  669. if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
  670. if (m < 0)
  671. m = -m;
  672. /* Scale deviation to rttvar fixed point */
  673. m >>= 1;
  674. if (m < tp->mdev)
  675. m = tp->mdev;
  676. if (m >= dst_metric(dst, RTAX_RTTVAR))
  677. dst->metrics[RTAX_RTTVAR-1] = m;
  678. else
  679. dst->metrics[RTAX_RTTVAR-1] -=
  680. (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
  681. }
  682. if (tp->snd_ssthresh >= 0xFFFF) {
  683. /* Slow start still did not finish. */
  684. if (dst_metric(dst, RTAX_SSTHRESH) &&
  685. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  686. (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
  687. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
  688. if (!dst_metric_locked(dst, RTAX_CWND) &&
  689. tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
  690. dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
  691. } else if (tp->snd_cwnd > tp->snd_ssthresh &&
  692. icsk->icsk_ca_state == TCP_CA_Open) {
  693. /* Cong. avoidance phase, cwnd is reliable. */
  694. if (!dst_metric_locked(dst, RTAX_SSTHRESH))
  695. dst->metrics[RTAX_SSTHRESH-1] =
  696. max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
  697. if (!dst_metric_locked(dst, RTAX_CWND))
  698. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
  699. } else {
  700. /* Else slow start did not finish, cwnd is non-sense,
  701. ssthresh may be also invalid.
  702. */
  703. if (!dst_metric_locked(dst, RTAX_CWND))
  704. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
  705. if (dst->metrics[RTAX_SSTHRESH-1] &&
  706. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  707. tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
  708. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
  709. }
  710. if (!dst_metric_locked(dst, RTAX_REORDERING)) {
  711. if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
  712. tp->reordering != sysctl_tcp_reordering)
  713. dst->metrics[RTAX_REORDERING-1] = tp->reordering;
  714. }
  715. }
  716. }
  717. /* Numbers are taken from RFC3390.
  718. *
  719. * John Heffner states:
  720. *
  721. * The RFC specifies a window of no more than 4380 bytes
  722. * unless 2*MSS > 4380. Reading the pseudocode in the RFC
  723. * is a bit misleading because they use a clamp at 4380 bytes
  724. * rather than use a multiplier in the relevant range.
  725. */
  726. __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
  727. {
  728. __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
  729. if (!cwnd) {
  730. if (tp->mss_cache > 1460)
  731. cwnd = 2;
  732. else
  733. cwnd = (tp->mss_cache > 1095) ? 3 : 4;
  734. }
  735. return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
  736. }
  737. /* Set slow start threshold and cwnd not falling to slow start */
  738. void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
  739. {
  740. struct tcp_sock *tp = tcp_sk(sk);
  741. const struct inet_connection_sock *icsk = inet_csk(sk);
  742. tp->prior_ssthresh = 0;
  743. tp->bytes_acked = 0;
  744. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  745. tp->undo_marker = 0;
  746. if (set_ssthresh)
  747. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  748. tp->snd_cwnd = min(tp->snd_cwnd,
  749. tcp_packets_in_flight(tp) + 1U);
  750. tp->snd_cwnd_cnt = 0;
  751. tp->high_seq = tp->snd_nxt;
  752. tp->snd_cwnd_stamp = tcp_time_stamp;
  753. TCP_ECN_queue_cwr(tp);
  754. tcp_set_ca_state(sk, TCP_CA_CWR);
  755. }
  756. }
  757. /* Initialize metrics on socket. */
  758. static void tcp_init_metrics(struct sock *sk)
  759. {
  760. struct tcp_sock *tp = tcp_sk(sk);
  761. struct dst_entry *dst = __sk_dst_get(sk);
  762. if (dst == NULL)
  763. goto reset;
  764. dst_confirm(dst);
  765. if (dst_metric_locked(dst, RTAX_CWND))
  766. tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
  767. if (dst_metric(dst, RTAX_SSTHRESH)) {
  768. tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
  769. if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
  770. tp->snd_ssthresh = tp->snd_cwnd_clamp;
  771. }
  772. if (dst_metric(dst, RTAX_REORDERING) &&
  773. tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
  774. tp->rx_opt.sack_ok &= ~2;
  775. tp->reordering = dst_metric(dst, RTAX_REORDERING);
  776. }
  777. if (dst_metric(dst, RTAX_RTT) == 0)
  778. goto reset;
  779. if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
  780. goto reset;
  781. /* Initial rtt is determined from SYN,SYN-ACK.
  782. * The segment is small and rtt may appear much
  783. * less than real one. Use per-dst memory
  784. * to make it more realistic.
  785. *
  786. * A bit of theory. RTT is time passed after "normal" sized packet
  787. * is sent until it is ACKed. In normal circumstances sending small
  788. * packets force peer to delay ACKs and calculation is correct too.
  789. * The algorithm is adaptive and, provided we follow specs, it
  790. * NEVER underestimate RTT. BUT! If peer tries to make some clever
  791. * tricks sort of "quick acks" for time long enough to decrease RTT
  792. * to low value, and then abruptly stops to do it and starts to delay
  793. * ACKs, wait for troubles.
  794. */
  795. if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
  796. tp->srtt = dst_metric(dst, RTAX_RTT);
  797. tp->rtt_seq = tp->snd_nxt;
  798. }
  799. if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
  800. tp->mdev = dst_metric(dst, RTAX_RTTVAR);
  801. tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
  802. }
  803. tcp_set_rto(sk);
  804. tcp_bound_rto(sk);
  805. if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
  806. goto reset;
  807. tp->snd_cwnd = tcp_init_cwnd(tp, dst);
  808. tp->snd_cwnd_stamp = tcp_time_stamp;
  809. return;
  810. reset:
  811. /* Play conservative. If timestamps are not
  812. * supported, TCP will fail to recalculate correct
  813. * rtt, if initial rto is too small. FORGET ALL AND RESET!
  814. */
  815. if (!tp->rx_opt.saw_tstamp && tp->srtt) {
  816. tp->srtt = 0;
  817. tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
  818. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  819. }
  820. }
  821. static void tcp_update_reordering(struct sock *sk, const int metric,
  822. const int ts)
  823. {
  824. struct tcp_sock *tp = tcp_sk(sk);
  825. if (metric > tp->reordering) {
  826. tp->reordering = min(TCP_MAX_REORDERING, metric);
  827. /* This exciting event is worth to be remembered. 8) */
  828. if (ts)
  829. NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
  830. else if (IsReno(tp))
  831. NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
  832. else if (IsFack(tp))
  833. NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
  834. else
  835. NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
  836. #if FASTRETRANS_DEBUG > 1
  837. printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
  838. tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
  839. tp->reordering,
  840. tp->fackets_out,
  841. tp->sacked_out,
  842. tp->undo_marker ? tp->undo_retrans : 0);
  843. #endif
  844. /* Disable FACK yet. */
  845. tp->rx_opt.sack_ok &= ~2;
  846. }
  847. }
  848. /* This procedure tags the retransmission queue when SACKs arrive.
  849. *
  850. * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
  851. * Packets in queue with these bits set are counted in variables
  852. * sacked_out, retrans_out and lost_out, correspondingly.
  853. *
  854. * Valid combinations are:
  855. * Tag InFlight Description
  856. * 0 1 - orig segment is in flight.
  857. * S 0 - nothing flies, orig reached receiver.
  858. * L 0 - nothing flies, orig lost by net.
  859. * R 2 - both orig and retransmit are in flight.
  860. * L|R 1 - orig is lost, retransmit is in flight.
  861. * S|R 1 - orig reached receiver, retrans is still in flight.
  862. * (L|S|R is logically valid, it could occur when L|R is sacked,
  863. * but it is equivalent to plain S and code short-curcuits it to S.
  864. * L|S is logically invalid, it would mean -1 packet in flight 8))
  865. *
  866. * These 6 states form finite state machine, controlled by the following events:
  867. * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
  868. * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
  869. * 3. Loss detection event of one of three flavors:
  870. * A. Scoreboard estimator decided the packet is lost.
  871. * A'. Reno "three dupacks" marks head of queue lost.
  872. * A''. Its FACK modfication, head until snd.fack is lost.
  873. * B. SACK arrives sacking data transmitted after never retransmitted
  874. * hole was sent out.
  875. * C. SACK arrives sacking SND.NXT at the moment, when the
  876. * segment was retransmitted.
  877. * 4. D-SACK added new rule: D-SACK changes any tag to S.
  878. *
  879. * It is pleasant to note, that state diagram turns out to be commutative,
  880. * so that we are allowed not to be bothered by order of our actions,
  881. * when multiple events arrive simultaneously. (see the function below).
  882. *
  883. * Reordering detection.
  884. * --------------------
  885. * Reordering metric is maximal distance, which a packet can be displaced
  886. * in packet stream. With SACKs we can estimate it:
  887. *
  888. * 1. SACK fills old hole and the corresponding segment was not
  889. * ever retransmitted -> reordering. Alas, we cannot use it
  890. * when segment was retransmitted.
  891. * 2. The last flaw is solved with D-SACK. D-SACK arrives
  892. * for retransmitted and already SACKed segment -> reordering..
  893. * Both of these heuristics are not used in Loss state, when we cannot
  894. * account for retransmits accurately.
  895. */
  896. static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb,
  897. struct tcp_sack_block_wire *sp, int num_sacks,
  898. u32 prior_snd_una)
  899. {
  900. u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq));
  901. u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq));
  902. int dup_sack = 0;
  903. if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
  904. dup_sack = 1;
  905. tp->rx_opt.sack_ok |= 4;
  906. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
  907. } else if (num_sacks > 1) {
  908. u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq));
  909. u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq));
  910. if (!after(end_seq_0, end_seq_1) &&
  911. !before(start_seq_0, start_seq_1)) {
  912. dup_sack = 1;
  913. tp->rx_opt.sack_ok |= 4;
  914. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
  915. }
  916. }
  917. /* D-SACK for already forgotten data... Do dumb counting. */
  918. if (dup_sack &&
  919. !after(end_seq_0, prior_snd_una) &&
  920. after(end_seq_0, tp->undo_marker))
  921. tp->undo_retrans--;
  922. return dup_sack;
  923. }
  924. static int
  925. tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
  926. {
  927. const struct inet_connection_sock *icsk = inet_csk(sk);
  928. struct tcp_sock *tp = tcp_sk(sk);
  929. unsigned char *ptr = (skb_transport_header(ack_skb) +
  930. TCP_SKB_CB(ack_skb)->sacked);
  931. struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
  932. struct sk_buff *cached_skb;
  933. int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
  934. int reord = tp->packets_out;
  935. int prior_fackets;
  936. u32 lost_retrans = 0;
  937. int flag = 0;
  938. int found_dup_sack = 0;
  939. int cached_fack_count;
  940. int i;
  941. int first_sack_index;
  942. if (!tp->sacked_out) {
  943. tp->fackets_out = 0;
  944. tp->highest_sack = tp->snd_una;
  945. }
  946. prior_fackets = tp->fackets_out;
  947. found_dup_sack = tcp_check_dsack(tp, ack_skb, sp,
  948. num_sacks, prior_snd_una);
  949. if (found_dup_sack)
  950. flag |= FLAG_DSACKING_ACK;
  951. /* Eliminate too old ACKs, but take into
  952. * account more or less fresh ones, they can
  953. * contain valid SACK info.
  954. */
  955. if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
  956. return 0;
  957. /* SACK fastpath:
  958. * if the only SACK change is the increase of the end_seq of
  959. * the first block then only apply that SACK block
  960. * and use retrans queue hinting otherwise slowpath */
  961. flag = 1;
  962. for (i = 0; i < num_sacks; i++) {
  963. __be32 start_seq = sp[i].start_seq;
  964. __be32 end_seq = sp[i].end_seq;
  965. if (i == 0) {
  966. if (tp->recv_sack_cache[i].start_seq != start_seq)
  967. flag = 0;
  968. } else {
  969. if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
  970. (tp->recv_sack_cache[i].end_seq != end_seq))
  971. flag = 0;
  972. }
  973. tp->recv_sack_cache[i].start_seq = start_seq;
  974. tp->recv_sack_cache[i].end_seq = end_seq;
  975. }
  976. /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
  977. for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
  978. tp->recv_sack_cache[i].start_seq = 0;
  979. tp->recv_sack_cache[i].end_seq = 0;
  980. }
  981. first_sack_index = 0;
  982. if (flag)
  983. num_sacks = 1;
  984. else {
  985. int j;
  986. tp->fastpath_skb_hint = NULL;
  987. /* order SACK blocks to allow in order walk of the retrans queue */
  988. for (i = num_sacks-1; i > 0; i--) {
  989. for (j = 0; j < i; j++){
  990. if (after(ntohl(sp[j].start_seq),
  991. ntohl(sp[j+1].start_seq))){
  992. struct tcp_sack_block_wire tmp;
  993. tmp = sp[j];
  994. sp[j] = sp[j+1];
  995. sp[j+1] = tmp;
  996. /* Track where the first SACK block goes to */
  997. if (j == first_sack_index)
  998. first_sack_index = j+1;
  999. }
  1000. }
  1001. }
  1002. }
  1003. /* clear flag as used for different purpose in following code */
  1004. flag = 0;
  1005. /* Use SACK fastpath hint if valid */
  1006. cached_skb = tp->fastpath_skb_hint;
  1007. cached_fack_count = tp->fastpath_cnt_hint;
  1008. if (!cached_skb) {
  1009. cached_skb = tcp_write_queue_head(sk);
  1010. cached_fack_count = 0;
  1011. }
  1012. for (i=0; i<num_sacks; i++, sp++) {
  1013. struct sk_buff *skb;
  1014. __u32 start_seq = ntohl(sp->start_seq);
  1015. __u32 end_seq = ntohl(sp->end_seq);
  1016. int fack_count;
  1017. int dup_sack = (found_dup_sack && (i == first_sack_index));
  1018. skb = cached_skb;
  1019. fack_count = cached_fack_count;
  1020. /* Event "B" in the comment above. */
  1021. if (after(end_seq, tp->high_seq))
  1022. flag |= FLAG_DATA_LOST;
  1023. tcp_for_write_queue_from(skb, sk) {
  1024. int in_sack, pcount;
  1025. u8 sacked;
  1026. if (skb == tcp_send_head(sk))
  1027. break;
  1028. cached_skb = skb;
  1029. cached_fack_count = fack_count;
  1030. if (i == first_sack_index) {
  1031. tp->fastpath_skb_hint = skb;
  1032. tp->fastpath_cnt_hint = fack_count;
  1033. }
  1034. /* The retransmission queue is always in order, so
  1035. * we can short-circuit the walk early.
  1036. */
  1037. if (!before(TCP_SKB_CB(skb)->seq, end_seq))
  1038. break;
  1039. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  1040. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  1041. pcount = tcp_skb_pcount(skb);
  1042. if (pcount > 1 && !in_sack &&
  1043. after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
  1044. unsigned int pkt_len;
  1045. in_sack = !after(start_seq,
  1046. TCP_SKB_CB(skb)->seq);
  1047. if (!in_sack)
  1048. pkt_len = (start_seq -
  1049. TCP_SKB_CB(skb)->seq);
  1050. else
  1051. pkt_len = (end_seq -
  1052. TCP_SKB_CB(skb)->seq);
  1053. if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
  1054. break;
  1055. pcount = tcp_skb_pcount(skb);
  1056. }
  1057. fack_count += pcount;
  1058. sacked = TCP_SKB_CB(skb)->sacked;
  1059. /* Account D-SACK for retransmitted packet. */
  1060. if ((dup_sack && in_sack) &&
  1061. (sacked & TCPCB_RETRANS) &&
  1062. after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
  1063. tp->undo_retrans--;
  1064. /* The frame is ACKed. */
  1065. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
  1066. if (sacked&TCPCB_RETRANS) {
  1067. if ((dup_sack && in_sack) &&
  1068. (sacked&TCPCB_SACKED_ACKED))
  1069. reord = min(fack_count, reord);
  1070. } else {
  1071. /* If it was in a hole, we detected reordering. */
  1072. if (fack_count < prior_fackets &&
  1073. !(sacked&TCPCB_SACKED_ACKED))
  1074. reord = min(fack_count, reord);
  1075. }
  1076. /* Nothing to do; acked frame is about to be dropped. */
  1077. continue;
  1078. }
  1079. if ((sacked&TCPCB_SACKED_RETRANS) &&
  1080. after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
  1081. (!lost_retrans || after(end_seq, lost_retrans)))
  1082. lost_retrans = end_seq;
  1083. if (!in_sack)
  1084. continue;
  1085. if (!(sacked&TCPCB_SACKED_ACKED)) {
  1086. if (sacked & TCPCB_SACKED_RETRANS) {
  1087. /* If the segment is not tagged as lost,
  1088. * we do not clear RETRANS, believing
  1089. * that retransmission is still in flight.
  1090. */
  1091. if (sacked & TCPCB_LOST) {
  1092. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1093. tp->lost_out -= tcp_skb_pcount(skb);
  1094. tp->retrans_out -= tcp_skb_pcount(skb);
  1095. /* clear lost hint */
  1096. tp->retransmit_skb_hint = NULL;
  1097. }
  1098. } else {
  1099. /* New sack for not retransmitted frame,
  1100. * which was in hole. It is reordering.
  1101. */
  1102. if (!(sacked & TCPCB_RETRANS) &&
  1103. fack_count < prior_fackets)
  1104. reord = min(fack_count, reord);
  1105. if (sacked & TCPCB_LOST) {
  1106. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1107. tp->lost_out -= tcp_skb_pcount(skb);
  1108. /* clear lost hint */
  1109. tp->retransmit_skb_hint = NULL;
  1110. }
  1111. /* SACK enhanced F-RTO detection.
  1112. * Set flag if and only if non-rexmitted
  1113. * segments below frto_highmark are
  1114. * SACKed (RFC4138; Appendix B).
  1115. * Clearing correct due to in-order walk
  1116. */
  1117. if (after(end_seq, tp->frto_highmark)) {
  1118. flag &= ~FLAG_ONLY_ORIG_SACKED;
  1119. } else {
  1120. if (!(sacked & TCPCB_RETRANS))
  1121. flag |= FLAG_ONLY_ORIG_SACKED;
  1122. }
  1123. }
  1124. TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
  1125. flag |= FLAG_DATA_SACKED;
  1126. tp->sacked_out += tcp_skb_pcount(skb);
  1127. if (fack_count > tp->fackets_out)
  1128. tp->fackets_out = fack_count;
  1129. if (after(TCP_SKB_CB(skb)->seq,
  1130. tp->highest_sack))
  1131. tp->highest_sack = TCP_SKB_CB(skb)->seq;
  1132. } else {
  1133. if (dup_sack && (sacked&TCPCB_RETRANS))
  1134. reord = min(fack_count, reord);
  1135. }
  1136. /* D-SACK. We can detect redundant retransmission
  1137. * in S|R and plain R frames and clear it.
  1138. * undo_retrans is decreased above, L|R frames
  1139. * are accounted above as well.
  1140. */
  1141. if (dup_sack &&
  1142. (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
  1143. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1144. tp->retrans_out -= tcp_skb_pcount(skb);
  1145. tp->retransmit_skb_hint = NULL;
  1146. }
  1147. }
  1148. }
  1149. /* Check for lost retransmit. This superb idea is
  1150. * borrowed from "ratehalving". Event "C".
  1151. * Later note: FACK people cheated me again 8),
  1152. * we have to account for reordering! Ugly,
  1153. * but should help.
  1154. */
  1155. if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
  1156. struct sk_buff *skb;
  1157. tcp_for_write_queue(skb, sk) {
  1158. if (skb == tcp_send_head(sk))
  1159. break;
  1160. if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
  1161. break;
  1162. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1163. continue;
  1164. if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
  1165. after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
  1166. (IsFack(tp) ||
  1167. !before(lost_retrans,
  1168. TCP_SKB_CB(skb)->ack_seq + tp->reordering *
  1169. tp->mss_cache))) {
  1170. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1171. tp->retrans_out -= tcp_skb_pcount(skb);
  1172. /* clear lost hint */
  1173. tp->retransmit_skb_hint = NULL;
  1174. if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  1175. tp->lost_out += tcp_skb_pcount(skb);
  1176. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1177. flag |= FLAG_DATA_SACKED;
  1178. NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
  1179. }
  1180. }
  1181. }
  1182. }
  1183. tp->left_out = tp->sacked_out + tp->lost_out;
  1184. if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
  1185. (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
  1186. tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
  1187. #if FASTRETRANS_DEBUG > 0
  1188. BUG_TRAP((int)tp->sacked_out >= 0);
  1189. BUG_TRAP((int)tp->lost_out >= 0);
  1190. BUG_TRAP((int)tp->retrans_out >= 0);
  1191. BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
  1192. #endif
  1193. return flag;
  1194. }
  1195. /* F-RTO can only be used if TCP has never retransmitted anything other than
  1196. * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
  1197. */
  1198. static void tcp_check_reno_reordering(struct sock *sk, const int addend)
  1199. {
  1200. struct tcp_sock *tp = tcp_sk(sk);
  1201. u32 holes;
  1202. holes = max(tp->lost_out, 1U);
  1203. holes = min(holes, tp->packets_out);
  1204. if ((tp->sacked_out + holes) > tp->packets_out) {
  1205. tp->sacked_out = tp->packets_out - holes;
  1206. tcp_update_reordering(sk, tp->packets_out + addend, 0);
  1207. }
  1208. }
  1209. /* Emulate SACKs for SACKless connection: account for a new dupack. */
  1210. static void tcp_add_reno_sack(struct sock *sk)
  1211. {
  1212. struct tcp_sock *tp = tcp_sk(sk);
  1213. tp->sacked_out++;
  1214. tcp_check_reno_reordering(sk, 0);
  1215. tcp_sync_left_out(tp);
  1216. }
  1217. /* Account for ACK, ACKing some data in Reno Recovery phase. */
  1218. static void tcp_remove_reno_sacks(struct sock *sk, int acked)
  1219. {
  1220. struct tcp_sock *tp = tcp_sk(sk);
  1221. if (acked > 0) {
  1222. /* One ACK acked hole. The rest eat duplicate ACKs. */
  1223. if (acked-1 >= tp->sacked_out)
  1224. tp->sacked_out = 0;
  1225. else
  1226. tp->sacked_out -= acked-1;
  1227. }
  1228. tcp_check_reno_reordering(sk, acked);
  1229. tcp_sync_left_out(tp);
  1230. }
  1231. static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
  1232. {
  1233. tp->sacked_out = 0;
  1234. tp->left_out = tp->lost_out;
  1235. }
  1236. int tcp_use_frto(struct sock *sk)
  1237. {
  1238. const struct tcp_sock *tp = tcp_sk(sk);
  1239. struct sk_buff *skb;
  1240. if (!sysctl_tcp_frto)
  1241. return 0;
  1242. if (IsSackFrto())
  1243. return 1;
  1244. /* Avoid expensive walking of rexmit queue if possible */
  1245. if (tp->retrans_out > 1)
  1246. return 0;
  1247. skb = tcp_write_queue_head(sk);
  1248. skb = tcp_write_queue_next(sk, skb); /* Skips head */
  1249. tcp_for_write_queue_from(skb, sk) {
  1250. if (skb == tcp_send_head(sk))
  1251. break;
  1252. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1253. return 0;
  1254. /* Short-circuit when first non-SACKed skb has been checked */
  1255. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
  1256. break;
  1257. }
  1258. return 1;
  1259. }
  1260. /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
  1261. * recovery a bit and use heuristics in tcp_process_frto() to detect if
  1262. * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
  1263. * keep retrans_out counting accurate (with SACK F-RTO, other than head
  1264. * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
  1265. * bits are handled if the Loss state is really to be entered (in
  1266. * tcp_enter_frto_loss).
  1267. *
  1268. * Do like tcp_enter_loss() would; when RTO expires the second time it
  1269. * does:
  1270. * "Reduce ssthresh if it has not yet been made inside this window."
  1271. */
  1272. void tcp_enter_frto(struct sock *sk)
  1273. {
  1274. const struct inet_connection_sock *icsk = inet_csk(sk);
  1275. struct tcp_sock *tp = tcp_sk(sk);
  1276. struct sk_buff *skb;
  1277. if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
  1278. tp->snd_una == tp->high_seq ||
  1279. ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
  1280. !icsk->icsk_retransmits)) {
  1281. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1282. /* Our state is too optimistic in ssthresh() call because cwnd
  1283. * is not reduced until tcp_enter_frto_loss() when previous FRTO
  1284. * recovery has not yet completed. Pattern would be this: RTO,
  1285. * Cumulative ACK, RTO (2xRTO for the same segment does not end
  1286. * up here twice).
  1287. * RFC4138 should be more specific on what to do, even though
  1288. * RTO is quite unlikely to occur after the first Cumulative ACK
  1289. * due to back-off and complexity of triggering events ...
  1290. */
  1291. if (tp->frto_counter) {
  1292. u32 stored_cwnd;
  1293. stored_cwnd = tp->snd_cwnd;
  1294. tp->snd_cwnd = 2;
  1295. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1296. tp->snd_cwnd = stored_cwnd;
  1297. } else {
  1298. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1299. }
  1300. /* ... in theory, cong.control module could do "any tricks" in
  1301. * ssthresh(), which means that ca_state, lost bits and lost_out
  1302. * counter would have to be faked before the call occurs. We
  1303. * consider that too expensive, unlikely and hacky, so modules
  1304. * using these in ssthresh() must deal these incompatibility
  1305. * issues if they receives CA_EVENT_FRTO and frto_counter != 0
  1306. */
  1307. tcp_ca_event(sk, CA_EVENT_FRTO);
  1308. }
  1309. tp->undo_marker = tp->snd_una;
  1310. tp->undo_retrans = 0;
  1311. skb = tcp_write_queue_head(sk);
  1312. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  1313. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1314. tp->retrans_out -= tcp_skb_pcount(skb);
  1315. }
  1316. tcp_sync_left_out(tp);
  1317. /* Earlier loss recovery underway (see RFC4138; Appendix B).
  1318. * The last condition is necessary at least in tp->frto_counter case.
  1319. */
  1320. if (IsSackFrto() && (tp->frto_counter ||
  1321. ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
  1322. after(tp->high_seq, tp->snd_una)) {
  1323. tp->frto_highmark = tp->high_seq;
  1324. } else {
  1325. tp->frto_highmark = tp->snd_nxt;
  1326. }
  1327. tcp_set_ca_state(sk, TCP_CA_Disorder);
  1328. tp->high_seq = tp->snd_nxt;
  1329. tp->frto_counter = 1;
  1330. }
  1331. /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
  1332. * which indicates that we should follow the traditional RTO recovery,
  1333. * i.e. mark everything lost and do go-back-N retransmission.
  1334. */
  1335. static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
  1336. {
  1337. struct tcp_sock *tp = tcp_sk(sk);
  1338. struct sk_buff *skb;
  1339. tp->lost_out = 0;
  1340. tp->retrans_out = 0;
  1341. if (IsReno(tp))
  1342. tcp_reset_reno_sack(tp);
  1343. tcp_for_write_queue(skb, sk) {
  1344. if (skb == tcp_send_head(sk))
  1345. break;
  1346. /*
  1347. * Count the retransmission made on RTO correctly (only when
  1348. * waiting for the first ACK and did not get it)...
  1349. */
  1350. if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
  1351. /* For some reason this R-bit might get cleared? */
  1352. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
  1353. tp->retrans_out += tcp_skb_pcount(skb);
  1354. /* ...enter this if branch just for the first segment */
  1355. flag |= FLAG_DATA_ACKED;
  1356. } else {
  1357. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1358. }
  1359. /* Don't lost mark skbs that were fwd transmitted after RTO */
  1360. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) &&
  1361. !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) {
  1362. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1363. tp->lost_out += tcp_skb_pcount(skb);
  1364. }
  1365. }
  1366. tcp_sync_left_out(tp);
  1367. tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
  1368. tp->snd_cwnd_cnt = 0;
  1369. tp->snd_cwnd_stamp = tcp_time_stamp;
  1370. tp->undo_marker = 0;
  1371. tp->frto_counter = 0;
  1372. tp->reordering = min_t(unsigned int, tp->reordering,
  1373. sysctl_tcp_reordering);
  1374. tcp_set_ca_state(sk, TCP_CA_Loss);
  1375. tp->high_seq = tp->frto_highmark;
  1376. TCP_ECN_queue_cwr(tp);
  1377. clear_all_retrans_hints(tp);
  1378. }
  1379. void tcp_clear_retrans(struct tcp_sock *tp)
  1380. {
  1381. tp->left_out = 0;
  1382. tp->retrans_out = 0;
  1383. tp->fackets_out = 0;
  1384. tp->sacked_out = 0;
  1385. tp->lost_out = 0;
  1386. tp->undo_marker = 0;
  1387. tp->undo_retrans = 0;
  1388. }
  1389. /* Enter Loss state. If "how" is not zero, forget all SACK information
  1390. * and reset tags completely, otherwise preserve SACKs. If receiver
  1391. * dropped its ofo queue, we will know this due to reneging detection.
  1392. */
  1393. void tcp_enter_loss(struct sock *sk, int how)
  1394. {
  1395. const struct inet_connection_sock *icsk = inet_csk(sk);
  1396. struct tcp_sock *tp = tcp_sk(sk);
  1397. struct sk_buff *skb;
  1398. int cnt = 0;
  1399. /* Reduce ssthresh if it has not yet been made inside this window. */
  1400. if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
  1401. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1402. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1403. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1404. tcp_ca_event(sk, CA_EVENT_LOSS);
  1405. }
  1406. tp->snd_cwnd = 1;
  1407. tp->snd_cwnd_cnt = 0;
  1408. tp->snd_cwnd_stamp = tcp_time_stamp;
  1409. tp->bytes_acked = 0;
  1410. tcp_clear_retrans(tp);
  1411. /* Push undo marker, if it was plain RTO and nothing
  1412. * was retransmitted. */
  1413. if (!how)
  1414. tp->undo_marker = tp->snd_una;
  1415. tcp_for_write_queue(skb, sk) {
  1416. if (skb == tcp_send_head(sk))
  1417. break;
  1418. cnt += tcp_skb_pcount(skb);
  1419. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1420. tp->undo_marker = 0;
  1421. TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
  1422. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
  1423. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
  1424. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1425. tp->lost_out += tcp_skb_pcount(skb);
  1426. } else {
  1427. tp->sacked_out += tcp_skb_pcount(skb);
  1428. tp->fackets_out = cnt;
  1429. }
  1430. }
  1431. tcp_sync_left_out(tp);
  1432. tp->reordering = min_t(unsigned int, tp->reordering,
  1433. sysctl_tcp_reordering);
  1434. tcp_set_ca_state(sk, TCP_CA_Loss);
  1435. tp->high_seq = tp->snd_nxt;
  1436. TCP_ECN_queue_cwr(tp);
  1437. /* Abort FRTO algorithm if one is in progress */
  1438. tp->frto_counter = 0;
  1439. clear_all_retrans_hints(tp);
  1440. }
  1441. static int tcp_check_sack_reneging(struct sock *sk)
  1442. {
  1443. struct sk_buff *skb;
  1444. /* If ACK arrived pointing to a remembered SACK,
  1445. * it means that our remembered SACKs do not reflect
  1446. * real state of receiver i.e.
  1447. * receiver _host_ is heavily congested (or buggy).
  1448. * Do processing similar to RTO timeout.
  1449. */
  1450. if ((skb = tcp_write_queue_head(sk)) != NULL &&
  1451. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
  1452. struct inet_connection_sock *icsk = inet_csk(sk);
  1453. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
  1454. tcp_enter_loss(sk, 1);
  1455. icsk->icsk_retransmits++;
  1456. tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
  1457. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1458. icsk->icsk_rto, TCP_RTO_MAX);
  1459. return 1;
  1460. }
  1461. return 0;
  1462. }
  1463. static inline int tcp_fackets_out(struct tcp_sock *tp)
  1464. {
  1465. return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
  1466. }
  1467. static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
  1468. {
  1469. return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
  1470. }
  1471. static inline int tcp_head_timedout(struct sock *sk)
  1472. {
  1473. struct tcp_sock *tp = tcp_sk(sk);
  1474. return tp->packets_out &&
  1475. tcp_skb_timedout(sk, tcp_write_queue_head(sk));
  1476. }
  1477. /* Linux NewReno/SACK/FACK/ECN state machine.
  1478. * --------------------------------------
  1479. *
  1480. * "Open" Normal state, no dubious events, fast path.
  1481. * "Disorder" In all the respects it is "Open",
  1482. * but requires a bit more attention. It is entered when
  1483. * we see some SACKs or dupacks. It is split of "Open"
  1484. * mainly to move some processing from fast path to slow one.
  1485. * "CWR" CWND was reduced due to some Congestion Notification event.
  1486. * It can be ECN, ICMP source quench, local device congestion.
  1487. * "Recovery" CWND was reduced, we are fast-retransmitting.
  1488. * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
  1489. *
  1490. * tcp_fastretrans_alert() is entered:
  1491. * - each incoming ACK, if state is not "Open"
  1492. * - when arrived ACK is unusual, namely:
  1493. * * SACK
  1494. * * Duplicate ACK.
  1495. * * ECN ECE.
  1496. *
  1497. * Counting packets in flight is pretty simple.
  1498. *
  1499. * in_flight = packets_out - left_out + retrans_out
  1500. *
  1501. * packets_out is SND.NXT-SND.UNA counted in packets.
  1502. *
  1503. * retrans_out is number of retransmitted segments.
  1504. *
  1505. * left_out is number of segments left network, but not ACKed yet.
  1506. *
  1507. * left_out = sacked_out + lost_out
  1508. *
  1509. * sacked_out: Packets, which arrived to receiver out of order
  1510. * and hence not ACKed. With SACKs this number is simply
  1511. * amount of SACKed data. Even without SACKs
  1512. * it is easy to give pretty reliable estimate of this number,
  1513. * counting duplicate ACKs.
  1514. *
  1515. * lost_out: Packets lost by network. TCP has no explicit
  1516. * "loss notification" feedback from network (for now).
  1517. * It means that this number can be only _guessed_.
  1518. * Actually, it is the heuristics to predict lossage that
  1519. * distinguishes different algorithms.
  1520. *
  1521. * F.e. after RTO, when all the queue is considered as lost,
  1522. * lost_out = packets_out and in_flight = retrans_out.
  1523. *
  1524. * Essentially, we have now two algorithms counting
  1525. * lost packets.
  1526. *
  1527. * FACK: It is the simplest heuristics. As soon as we decided
  1528. * that something is lost, we decide that _all_ not SACKed
  1529. * packets until the most forward SACK are lost. I.e.
  1530. * lost_out = fackets_out - sacked_out and left_out = fackets_out.
  1531. * It is absolutely correct estimate, if network does not reorder
  1532. * packets. And it loses any connection to reality when reordering
  1533. * takes place. We use FACK by default until reordering
  1534. * is suspected on the path to this destination.
  1535. *
  1536. * NewReno: when Recovery is entered, we assume that one segment
  1537. * is lost (classic Reno). While we are in Recovery and
  1538. * a partial ACK arrives, we assume that one more packet
  1539. * is lost (NewReno). This heuristics are the same in NewReno
  1540. * and SACK.
  1541. *
  1542. * Imagine, that's all! Forget about all this shamanism about CWND inflation
  1543. * deflation etc. CWND is real congestion window, never inflated, changes
  1544. * only according to classic VJ rules.
  1545. *
  1546. * Really tricky (and requiring careful tuning) part of algorithm
  1547. * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
  1548. * The first determines the moment _when_ we should reduce CWND and,
  1549. * hence, slow down forward transmission. In fact, it determines the moment
  1550. * when we decide that hole is caused by loss, rather than by a reorder.
  1551. *
  1552. * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
  1553. * holes, caused by lost packets.
  1554. *
  1555. * And the most logically complicated part of algorithm is undo
  1556. * heuristics. We detect false retransmits due to both too early
  1557. * fast retransmit (reordering) and underestimated RTO, analyzing
  1558. * timestamps and D-SACKs. When we detect that some segments were
  1559. * retransmitted by mistake and CWND reduction was wrong, we undo
  1560. * window reduction and abort recovery phase. This logic is hidden
  1561. * inside several functions named tcp_try_undo_<something>.
  1562. */
  1563. /* This function decides, when we should leave Disordered state
  1564. * and enter Recovery phase, reducing congestion window.
  1565. *
  1566. * Main question: may we further continue forward transmission
  1567. * with the same cwnd?
  1568. */
  1569. static int tcp_time_to_recover(struct sock *sk)
  1570. {
  1571. struct tcp_sock *tp = tcp_sk(sk);
  1572. __u32 packets_out;
  1573. /* Do not perform any recovery during FRTO algorithm */
  1574. if (tp->frto_counter)
  1575. return 0;
  1576. /* Trick#1: The loss is proven. */
  1577. if (tp->lost_out)
  1578. return 1;
  1579. /* Not-A-Trick#2 : Classic rule... */
  1580. if (tcp_fackets_out(tp) > tp->reordering)
  1581. return 1;
  1582. /* Trick#3 : when we use RFC2988 timer restart, fast
  1583. * retransmit can be triggered by timeout of queue head.
  1584. */
  1585. if (tcp_head_timedout(sk))
  1586. return 1;
  1587. /* Trick#4: It is still not OK... But will it be useful to delay
  1588. * recovery more?
  1589. */
  1590. packets_out = tp->packets_out;
  1591. if (packets_out <= tp->reordering &&
  1592. tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
  1593. !tcp_may_send_now(sk)) {
  1594. /* We have nothing to send. This connection is limited
  1595. * either by receiver window or by application.
  1596. */
  1597. return 1;
  1598. }
  1599. return 0;
  1600. }
  1601. /* RFC: This is from the original, I doubt that this is necessary at all:
  1602. * clear xmit_retrans hint if seq of this skb is beyond hint. How could we
  1603. * retransmitted past LOST markings in the first place? I'm not fully sure
  1604. * about undo and end of connection cases, which can cause R without L?
  1605. */
  1606. static void tcp_verify_retransmit_hint(struct tcp_sock *tp,
  1607. struct sk_buff *skb)
  1608. {
  1609. if ((tp->retransmit_skb_hint != NULL) &&
  1610. before(TCP_SKB_CB(skb)->seq,
  1611. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  1612. tp->retransmit_skb_hint = NULL;
  1613. }
  1614. /* Mark head of queue up as lost. */
  1615. static void tcp_mark_head_lost(struct sock *sk,
  1616. int packets, u32 high_seq)
  1617. {
  1618. struct tcp_sock *tp = tcp_sk(sk);
  1619. struct sk_buff *skb;
  1620. int cnt;
  1621. BUG_TRAP(packets <= tp->packets_out);
  1622. if (tp->lost_skb_hint) {
  1623. skb = tp->lost_skb_hint;
  1624. cnt = tp->lost_cnt_hint;
  1625. } else {
  1626. skb = tcp_write_queue_head(sk);
  1627. cnt = 0;
  1628. }
  1629. tcp_for_write_queue_from(skb, sk) {
  1630. if (skb == tcp_send_head(sk))
  1631. break;
  1632. /* TODO: do this better */
  1633. /* this is not the most efficient way to do this... */
  1634. tp->lost_skb_hint = skb;
  1635. tp->lost_cnt_hint = cnt;
  1636. cnt += tcp_skb_pcount(skb);
  1637. if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
  1638. break;
  1639. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1640. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1641. tp->lost_out += tcp_skb_pcount(skb);
  1642. tcp_verify_retransmit_hint(tp, skb);
  1643. }
  1644. }
  1645. tcp_sync_left_out(tp);
  1646. }
  1647. /* Account newly detected lost packet(s) */
  1648. static void tcp_update_scoreboard(struct sock *sk)
  1649. {
  1650. struct tcp_sock *tp = tcp_sk(sk);
  1651. if (IsFack(tp)) {
  1652. int lost = tp->fackets_out - tp->reordering;
  1653. if (lost <= 0)
  1654. lost = 1;
  1655. tcp_mark_head_lost(sk, lost, tp->high_seq);
  1656. } else {
  1657. tcp_mark_head_lost(sk, 1, tp->high_seq);
  1658. }
  1659. /* New heuristics: it is possible only after we switched
  1660. * to restart timer each time when something is ACKed.
  1661. * Hence, we can detect timed out packets during fast
  1662. * retransmit without falling to slow start.
  1663. */
  1664. if (!IsReno(tp) && tcp_head_timedout(sk)) {
  1665. struct sk_buff *skb;
  1666. skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
  1667. : tcp_write_queue_head(sk);
  1668. tcp_for_write_queue_from(skb, sk) {
  1669. if (skb == tcp_send_head(sk))
  1670. break;
  1671. if (!tcp_skb_timedout(sk, skb))
  1672. break;
  1673. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1674. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1675. tp->lost_out += tcp_skb_pcount(skb);
  1676. tcp_verify_retransmit_hint(tp, skb);
  1677. }
  1678. }
  1679. tp->scoreboard_skb_hint = skb;
  1680. tcp_sync_left_out(tp);
  1681. }
  1682. }
  1683. /* CWND moderation, preventing bursts due to too big ACKs
  1684. * in dubious situations.
  1685. */
  1686. static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
  1687. {
  1688. tp->snd_cwnd = min(tp->snd_cwnd,
  1689. tcp_packets_in_flight(tp)+tcp_max_burst(tp));
  1690. tp->snd_cwnd_stamp = tcp_time_stamp;
  1691. }
  1692. /* Lower bound on congestion window is slow start threshold
  1693. * unless congestion avoidance choice decides to overide it.
  1694. */
  1695. static inline u32 tcp_cwnd_min(const struct sock *sk)
  1696. {
  1697. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1698. return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
  1699. }
  1700. /* Decrease cwnd each second ack. */
  1701. static void tcp_cwnd_down(struct sock *sk, int flag)
  1702. {
  1703. struct tcp_sock *tp = tcp_sk(sk);
  1704. int decr = tp->snd_cwnd_cnt + 1;
  1705. if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) ||
  1706. (IsReno(tp) && !(flag&FLAG_NOT_DUP))) {
  1707. tp->snd_cwnd_cnt = decr&1;
  1708. decr >>= 1;
  1709. if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
  1710. tp->snd_cwnd -= decr;
  1711. tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
  1712. tp->snd_cwnd_stamp = tcp_time_stamp;
  1713. }
  1714. }
  1715. /* Nothing was retransmitted or returned timestamp is less
  1716. * than timestamp of the first retransmission.
  1717. */
  1718. static inline int tcp_packet_delayed(struct tcp_sock *tp)
  1719. {
  1720. return !tp->retrans_stamp ||
  1721. (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  1722. (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
  1723. }
  1724. /* Undo procedures. */
  1725. #if FASTRETRANS_DEBUG > 1
  1726. static void DBGUNDO(struct sock *sk, const char *msg)
  1727. {
  1728. struct tcp_sock *tp = tcp_sk(sk);
  1729. struct inet_sock *inet = inet_sk(sk);
  1730. printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
  1731. msg,
  1732. NIPQUAD(inet->daddr), ntohs(inet->dport),
  1733. tp->snd_cwnd, tp->left_out,
  1734. tp->snd_ssthresh, tp->prior_ssthresh,
  1735. tp->packets_out);
  1736. }
  1737. #else
  1738. #define DBGUNDO(x...) do { } while (0)
  1739. #endif
  1740. static void tcp_undo_cwr(struct sock *sk, const int undo)
  1741. {
  1742. struct tcp_sock *tp = tcp_sk(sk);
  1743. if (tp->prior_ssthresh) {
  1744. const struct inet_connection_sock *icsk = inet_csk(sk);
  1745. if (icsk->icsk_ca_ops->undo_cwnd)
  1746. tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
  1747. else
  1748. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
  1749. if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
  1750. tp->snd_ssthresh = tp->prior_ssthresh;
  1751. TCP_ECN_withdraw_cwr(tp);
  1752. }
  1753. } else {
  1754. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
  1755. }
  1756. tcp_moderate_cwnd(tp);
  1757. tp->snd_cwnd_stamp = tcp_time_stamp;
  1758. /* There is something screwy going on with the retrans hints after
  1759. an undo */
  1760. clear_all_retrans_hints(tp);
  1761. }
  1762. static inline int tcp_may_undo(struct tcp_sock *tp)
  1763. {
  1764. return tp->undo_marker &&
  1765. (!tp->undo_retrans || tcp_packet_delayed(tp));
  1766. }
  1767. /* People celebrate: "We love our President!" */
  1768. static int tcp_try_undo_recovery(struct sock *sk)
  1769. {
  1770. struct tcp_sock *tp = tcp_sk(sk);
  1771. if (tcp_may_undo(tp)) {
  1772. /* Happy end! We did not retransmit anything
  1773. * or our original transmission succeeded.
  1774. */
  1775. DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
  1776. tcp_undo_cwr(sk, 1);
  1777. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
  1778. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1779. else
  1780. NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
  1781. tp->undo_marker = 0;
  1782. }
  1783. if (tp->snd_una == tp->high_seq && IsReno(tp)) {
  1784. /* Hold old state until something *above* high_seq
  1785. * is ACKed. For Reno it is MUST to prevent false
  1786. * fast retransmits (RFC2582). SACK TCP is safe. */
  1787. tcp_moderate_cwnd(tp);
  1788. return 1;
  1789. }
  1790. tcp_set_ca_state(sk, TCP_CA_Open);
  1791. return 0;
  1792. }
  1793. /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
  1794. static void tcp_try_undo_dsack(struct sock *sk)
  1795. {
  1796. struct tcp_sock *tp = tcp_sk(sk);
  1797. if (tp->undo_marker && !tp->undo_retrans) {
  1798. DBGUNDO(sk, "D-SACK");
  1799. tcp_undo_cwr(sk, 1);
  1800. tp->undo_marker = 0;
  1801. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
  1802. }
  1803. }
  1804. /* Undo during fast recovery after partial ACK. */
  1805. static int tcp_try_undo_partial(struct sock *sk, int acked)
  1806. {
  1807. struct tcp_sock *tp = tcp_sk(sk);
  1808. /* Partial ACK arrived. Force Hoe's retransmit. */
  1809. int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
  1810. if (tcp_may_undo(tp)) {
  1811. /* Plain luck! Hole if filled with delayed
  1812. * packet, rather than with a retransmit.
  1813. */
  1814. if (tp->retrans_out == 0)
  1815. tp->retrans_stamp = 0;
  1816. tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
  1817. DBGUNDO(sk, "Hoe");
  1818. tcp_undo_cwr(sk, 0);
  1819. NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
  1820. /* So... Do not make Hoe's retransmit yet.
  1821. * If the first packet was delayed, the rest
  1822. * ones are most probably delayed as well.
  1823. */
  1824. failed = 0;
  1825. }
  1826. return failed;
  1827. }
  1828. /* Undo during loss recovery after partial ACK. */
  1829. static int tcp_try_undo_loss(struct sock *sk)
  1830. {
  1831. struct tcp_sock *tp = tcp_sk(sk);
  1832. if (tcp_may_undo(tp)) {
  1833. struct sk_buff *skb;
  1834. tcp_for_write_queue(skb, sk) {
  1835. if (skb == tcp_send_head(sk))
  1836. break;
  1837. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1838. }
  1839. clear_all_retrans_hints(tp);
  1840. DBGUNDO(sk, "partial loss");
  1841. tp->lost_out = 0;
  1842. tp->left_out = tp->sacked_out;
  1843. tcp_undo_cwr(sk, 1);
  1844. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1845. inet_csk(sk)->icsk_retransmits = 0;
  1846. tp->undo_marker = 0;
  1847. if (!IsReno(tp))
  1848. tcp_set_ca_state(sk, TCP_CA_Open);
  1849. return 1;
  1850. }
  1851. return 0;
  1852. }
  1853. static inline void tcp_complete_cwr(struct sock *sk)
  1854. {
  1855. struct tcp_sock *tp = tcp_sk(sk);
  1856. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  1857. tp->snd_cwnd_stamp = tcp_time_stamp;
  1858. tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
  1859. }
  1860. static void tcp_try_to_open(struct sock *sk, int flag)
  1861. {
  1862. struct tcp_sock *tp = tcp_sk(sk);
  1863. tcp_sync_left_out(tp);
  1864. if (tp->retrans_out == 0)
  1865. tp->retrans_stamp = 0;
  1866. if (flag&FLAG_ECE)
  1867. tcp_enter_cwr(sk, 1);
  1868. if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
  1869. int state = TCP_CA_Open;
  1870. if (tp->left_out || tp->retrans_out || tp->undo_marker)
  1871. state = TCP_CA_Disorder;
  1872. if (inet_csk(sk)->icsk_ca_state != state) {
  1873. tcp_set_ca_state(sk, state);
  1874. tp->high_seq = tp->snd_nxt;
  1875. }
  1876. tcp_moderate_cwnd(tp);
  1877. } else {
  1878. tcp_cwnd_down(sk, flag);
  1879. }
  1880. }
  1881. static void tcp_mtup_probe_failed(struct sock *sk)
  1882. {
  1883. struct inet_connection_sock *icsk = inet_csk(sk);
  1884. icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
  1885. icsk->icsk_mtup.probe_size = 0;
  1886. }
  1887. static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
  1888. {
  1889. struct tcp_sock *tp = tcp_sk(sk);
  1890. struct inet_connection_sock *icsk = inet_csk(sk);
  1891. /* FIXME: breaks with very large cwnd */
  1892. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1893. tp->snd_cwnd = tp->snd_cwnd *
  1894. tcp_mss_to_mtu(sk, tp->mss_cache) /
  1895. icsk->icsk_mtup.probe_size;
  1896. tp->snd_cwnd_cnt = 0;
  1897. tp->snd_cwnd_stamp = tcp_time_stamp;
  1898. tp->rcv_ssthresh = tcp_current_ssthresh(sk);
  1899. icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
  1900. icsk->icsk_mtup.probe_size = 0;
  1901. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  1902. }
  1903. /* Process an event, which can update packets-in-flight not trivially.
  1904. * Main goal of this function is to calculate new estimate for left_out,
  1905. * taking into account both packets sitting in receiver's buffer and
  1906. * packets lost by network.
  1907. *
  1908. * Besides that it does CWND reduction, when packet loss is detected
  1909. * and changes state of machine.
  1910. *
  1911. * It does _not_ decide what to send, it is made in function
  1912. * tcp_xmit_retransmit_queue().
  1913. */
  1914. static void
  1915. tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag)
  1916. {
  1917. struct inet_connection_sock *icsk = inet_csk(sk);
  1918. struct tcp_sock *tp = tcp_sk(sk);
  1919. int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP));
  1920. int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) &&
  1921. (tp->fackets_out > tp->reordering));
  1922. /* Some technical things:
  1923. * 1. Reno does not count dupacks (sacked_out) automatically. */
  1924. if (!tp->packets_out)
  1925. tp->sacked_out = 0;
  1926. /* 2. SACK counts snd_fack in packets inaccurately. */
  1927. if (tp->sacked_out == 0)
  1928. tp->fackets_out = 0;
  1929. /* Now state machine starts.
  1930. * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
  1931. if (flag&FLAG_ECE)
  1932. tp->prior_ssthresh = 0;
  1933. /* B. In all the states check for reneging SACKs. */
  1934. if (tp->sacked_out && tcp_check_sack_reneging(sk))
  1935. return;
  1936. /* C. Process data loss notification, provided it is valid. */
  1937. if ((flag&FLAG_DATA_LOST) &&
  1938. before(tp->snd_una, tp->high_seq) &&
  1939. icsk->icsk_ca_state != TCP_CA_Open &&
  1940. tp->fackets_out > tp->reordering) {
  1941. tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
  1942. NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
  1943. }
  1944. /* D. Synchronize left_out to current state. */
  1945. tcp_sync_left_out(tp);
  1946. /* E. Check state exit conditions. State can be terminated
  1947. * when high_seq is ACKed. */
  1948. if (icsk->icsk_ca_state == TCP_CA_Open) {
  1949. BUG_TRAP(tp->retrans_out == 0);
  1950. tp->retrans_stamp = 0;
  1951. } else if (!before(tp->snd_una, tp->high_seq)) {
  1952. switch (icsk->icsk_ca_state) {
  1953. case TCP_CA_Loss:
  1954. icsk->icsk_retransmits = 0;
  1955. if (tcp_try_undo_recovery(sk))
  1956. return;
  1957. break;
  1958. case TCP_CA_CWR:
  1959. /* CWR is to be held something *above* high_seq
  1960. * is ACKed for CWR bit to reach receiver. */
  1961. if (tp->snd_una != tp->high_seq) {
  1962. tcp_complete_cwr(sk);
  1963. tcp_set_ca_state(sk, TCP_CA_Open);
  1964. }
  1965. break;
  1966. case TCP_CA_Disorder:
  1967. tcp_try_undo_dsack(sk);
  1968. if (!tp->undo_marker ||
  1969. /* For SACK case do not Open to allow to undo
  1970. * catching for all duplicate ACKs. */
  1971. IsReno(tp) || tp->snd_una != tp->high_seq) {
  1972. tp->undo_marker = 0;
  1973. tcp_set_ca_state(sk, TCP_CA_Open);
  1974. }
  1975. break;
  1976. case TCP_CA_Recovery:
  1977. if (IsReno(tp))
  1978. tcp_reset_reno_sack(tp);
  1979. if (tcp_try_undo_recovery(sk))
  1980. return;
  1981. tcp_complete_cwr(sk);
  1982. break;
  1983. }
  1984. }
  1985. /* F. Process state. */
  1986. switch (icsk->icsk_ca_state) {
  1987. case TCP_CA_Recovery:
  1988. if (!(flag & FLAG_SND_UNA_ADVANCED)) {
  1989. if (IsReno(tp) && is_dupack)
  1990. tcp_add_reno_sack(sk);
  1991. } else {
  1992. int acked = prior_packets - tp->packets_out;
  1993. if (IsReno(tp))
  1994. tcp_remove_reno_sacks(sk, acked);
  1995. do_lost = tcp_try_undo_partial(sk, acked);
  1996. }
  1997. break;
  1998. case TCP_CA_Loss:
  1999. if (flag&FLAG_DATA_ACKED)
  2000. icsk->icsk_retransmits = 0;
  2001. if (!tcp_try_undo_loss(sk)) {
  2002. tcp_moderate_cwnd(tp);
  2003. tcp_xmit_retransmit_queue(sk);
  2004. return;
  2005. }
  2006. if (icsk->icsk_ca_state != TCP_CA_Open)
  2007. return;
  2008. /* Loss is undone; fall through to processing in Open state. */
  2009. default:
  2010. if (IsReno(tp)) {
  2011. if (flag & FLAG_SND_UNA_ADVANCED)
  2012. tcp_reset_reno_sack(tp);
  2013. if (is_dupack)
  2014. tcp_add_reno_sack(sk);
  2015. }
  2016. if (icsk->icsk_ca_state == TCP_CA_Disorder)
  2017. tcp_try_undo_dsack(sk);
  2018. if (!tcp_time_to_recover(sk)) {
  2019. tcp_try_to_open(sk, flag);
  2020. return;
  2021. }
  2022. /* MTU probe failure: don't reduce cwnd */
  2023. if (icsk->icsk_ca_state < TCP_CA_CWR &&
  2024. icsk->icsk_mtup.probe_size &&
  2025. tp->snd_una == tp->mtu_probe.probe_seq_start) {
  2026. tcp_mtup_probe_failed(sk);
  2027. /* Restores the reduction we did in tcp_mtup_probe() */
  2028. tp->snd_cwnd++;
  2029. tcp_simple_retransmit(sk);
  2030. return;
  2031. }
  2032. /* Otherwise enter Recovery state */
  2033. if (IsReno(tp))
  2034. NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
  2035. else
  2036. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
  2037. tp->high_seq = tp->snd_nxt;
  2038. tp->prior_ssthresh = 0;
  2039. tp->undo_marker = tp->snd_una;
  2040. tp->undo_retrans = tp->retrans_out;
  2041. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  2042. if (!(flag&FLAG_ECE))
  2043. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  2044. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  2045. TCP_ECN_queue_cwr(tp);
  2046. }
  2047. tp->bytes_acked = 0;
  2048. tp->snd_cwnd_cnt = 0;
  2049. tcp_set_ca_state(sk, TCP_CA_Recovery);
  2050. }
  2051. if (do_lost || tcp_head_timedout(sk))
  2052. tcp_update_scoreboard(sk);
  2053. tcp_cwnd_down(sk, flag);
  2054. tcp_xmit_retransmit_queue(sk);
  2055. }
  2056. /* Read draft-ietf-tcplw-high-performance before mucking
  2057. * with this code. (Supersedes RFC1323)
  2058. */
  2059. static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
  2060. {
  2061. /* RTTM Rule: A TSecr value received in a segment is used to
  2062. * update the averaged RTT measurement only if the segment
  2063. * acknowledges some new data, i.e., only if it advances the
  2064. * left edge of the send window.
  2065. *
  2066. * See draft-ietf-tcplw-high-performance-00, section 3.3.
  2067. * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
  2068. *
  2069. * Changed: reset backoff as soon as we see the first valid sample.
  2070. * If we do not, we get strongly overestimated rto. With timestamps
  2071. * samples are accepted even from very old segments: f.e., when rtt=1
  2072. * increases to 8, we retransmit 5 times and after 8 seconds delayed
  2073. * answer arrives rto becomes 120 seconds! If at least one of segments
  2074. * in window is lost... Voila. --ANK (010210)
  2075. */
  2076. struct tcp_sock *tp = tcp_sk(sk);
  2077. const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
  2078. tcp_rtt_estimator(sk, seq_rtt);
  2079. tcp_set_rto(sk);
  2080. inet_csk(sk)->icsk_backoff = 0;
  2081. tcp_bound_rto(sk);
  2082. }
  2083. static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
  2084. {
  2085. /* We don't have a timestamp. Can only use
  2086. * packets that are not retransmitted to determine
  2087. * rtt estimates. Also, we must not reset the
  2088. * backoff for rto until we get a non-retransmitted
  2089. * packet. This allows us to deal with a situation
  2090. * where the network delay has increased suddenly.
  2091. * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
  2092. */
  2093. if (flag & FLAG_RETRANS_DATA_ACKED)
  2094. return;
  2095. tcp_rtt_estimator(sk, seq_rtt);
  2096. tcp_set_rto(sk);
  2097. inet_csk(sk)->icsk_backoff = 0;
  2098. tcp_bound_rto(sk);
  2099. }
  2100. static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
  2101. const s32 seq_rtt)
  2102. {
  2103. const struct tcp_sock *tp = tcp_sk(sk);
  2104. /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
  2105. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  2106. tcp_ack_saw_tstamp(sk, flag);
  2107. else if (seq_rtt >= 0)
  2108. tcp_ack_no_tstamp(sk, seq_rtt, flag);
  2109. }
  2110. static void tcp_cong_avoid(struct sock *sk, u32 ack,
  2111. u32 in_flight, int good)
  2112. {
  2113. const struct inet_connection_sock *icsk = inet_csk(sk);
  2114. icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good);
  2115. tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
  2116. }
  2117. /* Restart timer after forward progress on connection.
  2118. * RFC2988 recommends to restart timer to now+rto.
  2119. */
  2120. static void tcp_ack_packets_out(struct sock *sk)
  2121. {
  2122. struct tcp_sock *tp = tcp_sk(sk);
  2123. if (!tp->packets_out) {
  2124. inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
  2125. } else {
  2126. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2127. }
  2128. }
  2129. static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
  2130. __u32 now, __s32 *seq_rtt)
  2131. {
  2132. struct tcp_sock *tp = tcp_sk(sk);
  2133. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2134. __u32 seq = tp->snd_una;
  2135. __u32 packets_acked;
  2136. int acked = 0;
  2137. /* If we get here, the whole TSO packet has not been
  2138. * acked.
  2139. */
  2140. BUG_ON(!after(scb->end_seq, seq));
  2141. packets_acked = tcp_skb_pcount(skb);
  2142. if (tcp_trim_head(sk, skb, seq - scb->seq))
  2143. return 0;
  2144. packets_acked -= tcp_skb_pcount(skb);
  2145. if (packets_acked) {
  2146. __u8 sacked = scb->sacked;
  2147. acked |= FLAG_DATA_ACKED;
  2148. if (sacked) {
  2149. if (sacked & TCPCB_RETRANS) {
  2150. if (sacked & TCPCB_SACKED_RETRANS)
  2151. tp->retrans_out -= packets_acked;
  2152. acked |= FLAG_RETRANS_DATA_ACKED;
  2153. *seq_rtt = -1;
  2154. } else if (*seq_rtt < 0)
  2155. *seq_rtt = now - scb->when;
  2156. if (sacked & TCPCB_SACKED_ACKED)
  2157. tp->sacked_out -= packets_acked;
  2158. if (sacked & TCPCB_LOST)
  2159. tp->lost_out -= packets_acked;
  2160. if (sacked & TCPCB_URG) {
  2161. if (tp->urg_mode &&
  2162. !before(seq, tp->snd_up))
  2163. tp->urg_mode = 0;
  2164. }
  2165. } else if (*seq_rtt < 0)
  2166. *seq_rtt = now - scb->when;
  2167. if (tp->fackets_out) {
  2168. __u32 dval = min(tp->fackets_out, packets_acked);
  2169. tp->fackets_out -= dval;
  2170. }
  2171. /* hint's skb might be NULL but we don't need to care */
  2172. tp->fastpath_cnt_hint -= min_t(u32, packets_acked,
  2173. tp->fastpath_cnt_hint);
  2174. tp->packets_out -= packets_acked;
  2175. BUG_ON(tcp_skb_pcount(skb) == 0);
  2176. BUG_ON(!before(scb->seq, scb->end_seq));
  2177. }
  2178. return acked;
  2179. }
  2180. /* Remove acknowledged frames from the retransmission queue. */
  2181. static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
  2182. {
  2183. struct tcp_sock *tp = tcp_sk(sk);
  2184. const struct inet_connection_sock *icsk = inet_csk(sk);
  2185. struct sk_buff *skb;
  2186. __u32 now = tcp_time_stamp;
  2187. int acked = 0;
  2188. int prior_packets = tp->packets_out;
  2189. __s32 seq_rtt = -1;
  2190. ktime_t last_ackt = net_invalid_timestamp();
  2191. while ((skb = tcp_write_queue_head(sk)) &&
  2192. skb != tcp_send_head(sk)) {
  2193. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2194. __u8 sacked = scb->sacked;
  2195. /* If our packet is before the ack sequence we can
  2196. * discard it as it's confirmed to have arrived at
  2197. * the other end.
  2198. */
  2199. if (after(scb->end_seq, tp->snd_una)) {
  2200. if (tcp_skb_pcount(skb) > 1 &&
  2201. after(tp->snd_una, scb->seq))
  2202. acked |= tcp_tso_acked(sk, skb,
  2203. now, &seq_rtt);
  2204. break;
  2205. }
  2206. /* Initial outgoing SYN's get put onto the write_queue
  2207. * just like anything else we transmit. It is not
  2208. * true data, and if we misinform our callers that
  2209. * this ACK acks real data, we will erroneously exit
  2210. * connection startup slow start one packet too
  2211. * quickly. This is severely frowned upon behavior.
  2212. */
  2213. if (!(scb->flags & TCPCB_FLAG_SYN)) {
  2214. acked |= FLAG_DATA_ACKED;
  2215. } else {
  2216. acked |= FLAG_SYN_ACKED;
  2217. tp->retrans_stamp = 0;
  2218. }
  2219. /* MTU probing checks */
  2220. if (icsk->icsk_mtup.probe_size) {
  2221. if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
  2222. tcp_mtup_probe_success(sk, skb);
  2223. }
  2224. }
  2225. if (sacked) {
  2226. if (sacked & TCPCB_RETRANS) {
  2227. if (sacked & TCPCB_SACKED_RETRANS)
  2228. tp->retrans_out -= tcp_skb_pcount(skb);
  2229. acked |= FLAG_RETRANS_DATA_ACKED;
  2230. seq_rtt = -1;
  2231. } else if (seq_rtt < 0) {
  2232. seq_rtt = now - scb->when;
  2233. last_ackt = skb->tstamp;
  2234. }
  2235. if (sacked & TCPCB_SACKED_ACKED)
  2236. tp->sacked_out -= tcp_skb_pcount(skb);
  2237. if (sacked & TCPCB_LOST)
  2238. tp->lost_out -= tcp_skb_pcount(skb);
  2239. if (sacked & TCPCB_URG) {
  2240. if (tp->urg_mode &&
  2241. !before(scb->end_seq, tp->snd_up))
  2242. tp->urg_mode = 0;
  2243. }
  2244. } else if (seq_rtt < 0) {
  2245. seq_rtt = now - scb->when;
  2246. last_ackt = skb->tstamp;
  2247. }
  2248. tcp_dec_pcount_approx(&tp->fackets_out, skb);
  2249. tcp_packets_out_dec(tp, skb);
  2250. tcp_unlink_write_queue(skb, sk);
  2251. sk_stream_free_skb(sk, skb);
  2252. clear_all_retrans_hints(tp);
  2253. }
  2254. if (acked&FLAG_ACKED) {
  2255. u32 pkts_acked = prior_packets - tp->packets_out;
  2256. const struct tcp_congestion_ops *ca_ops
  2257. = inet_csk(sk)->icsk_ca_ops;
  2258. tcp_ack_update_rtt(sk, acked, seq_rtt);
  2259. tcp_ack_packets_out(sk);
  2260. if (ca_ops->pkts_acked) {
  2261. s32 rtt_us = -1;
  2262. /* Is the ACK triggering packet unambiguous? */
  2263. if (!(acked & FLAG_RETRANS_DATA_ACKED)) {
  2264. /* High resolution needed and available? */
  2265. if (ca_ops->flags & TCP_CONG_RTT_STAMP &&
  2266. !ktime_equal(last_ackt,
  2267. net_invalid_timestamp()))
  2268. rtt_us = ktime_us_delta(ktime_get_real(),
  2269. last_ackt);
  2270. else if (seq_rtt > 0)
  2271. rtt_us = jiffies_to_usecs(seq_rtt);
  2272. }
  2273. ca_ops->pkts_acked(sk, pkts_acked, rtt_us);
  2274. }
  2275. }
  2276. #if FASTRETRANS_DEBUG > 0
  2277. BUG_TRAP((int)tp->sacked_out >= 0);
  2278. BUG_TRAP((int)tp->lost_out >= 0);
  2279. BUG_TRAP((int)tp->retrans_out >= 0);
  2280. if (!tp->packets_out && tp->rx_opt.sack_ok) {
  2281. const struct inet_connection_sock *icsk = inet_csk(sk);
  2282. if (tp->lost_out) {
  2283. printk(KERN_DEBUG "Leak l=%u %d\n",
  2284. tp->lost_out, icsk->icsk_ca_state);
  2285. tp->lost_out = 0;
  2286. }
  2287. if (tp->sacked_out) {
  2288. printk(KERN_DEBUG "Leak s=%u %d\n",
  2289. tp->sacked_out, icsk->icsk_ca_state);
  2290. tp->sacked_out = 0;
  2291. }
  2292. if (tp->retrans_out) {
  2293. printk(KERN_DEBUG "Leak r=%u %d\n",
  2294. tp->retrans_out, icsk->icsk_ca_state);
  2295. tp->retrans_out = 0;
  2296. }
  2297. }
  2298. #endif
  2299. *seq_rtt_p = seq_rtt;
  2300. return acked;
  2301. }
  2302. static void tcp_ack_probe(struct sock *sk)
  2303. {
  2304. const struct tcp_sock *tp = tcp_sk(sk);
  2305. struct inet_connection_sock *icsk = inet_csk(sk);
  2306. /* Was it a usable window open? */
  2307. if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
  2308. tp->snd_una + tp->snd_wnd)) {
  2309. icsk->icsk_backoff = 0;
  2310. inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
  2311. /* Socket must be waked up by subsequent tcp_data_snd_check().
  2312. * This function is not for random using!
  2313. */
  2314. } else {
  2315. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2316. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2317. TCP_RTO_MAX);
  2318. }
  2319. }
  2320. static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
  2321. {
  2322. return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
  2323. inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
  2324. }
  2325. static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
  2326. {
  2327. const struct tcp_sock *tp = tcp_sk(sk);
  2328. return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
  2329. !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
  2330. }
  2331. /* Check that window update is acceptable.
  2332. * The function assumes that snd_una<=ack<=snd_next.
  2333. */
  2334. static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
  2335. const u32 ack_seq, const u32 nwin)
  2336. {
  2337. return (after(ack, tp->snd_una) ||
  2338. after(ack_seq, tp->snd_wl1) ||
  2339. (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
  2340. }
  2341. /* Update our send window.
  2342. *
  2343. * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
  2344. * and in FreeBSD. NetBSD's one is even worse.) is wrong.
  2345. */
  2346. static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
  2347. u32 ack_seq)
  2348. {
  2349. struct tcp_sock *tp = tcp_sk(sk);
  2350. int flag = 0;
  2351. u32 nwin = ntohs(tcp_hdr(skb)->window);
  2352. if (likely(!tcp_hdr(skb)->syn))
  2353. nwin <<= tp->rx_opt.snd_wscale;
  2354. if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
  2355. flag |= FLAG_WIN_UPDATE;
  2356. tcp_update_wl(tp, ack, ack_seq);
  2357. if (tp->snd_wnd != nwin) {
  2358. tp->snd_wnd = nwin;
  2359. /* Note, it is the only place, where
  2360. * fast path is recovered for sending TCP.
  2361. */
  2362. tp->pred_flags = 0;
  2363. tcp_fast_path_check(sk);
  2364. if (nwin > tp->max_window) {
  2365. tp->max_window = nwin;
  2366. tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
  2367. }
  2368. }
  2369. }
  2370. tp->snd_una = ack;
  2371. return flag;
  2372. }
  2373. /* A very conservative spurious RTO response algorithm: reduce cwnd and
  2374. * continue in congestion avoidance.
  2375. */
  2376. static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
  2377. {
  2378. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  2379. tp->snd_cwnd_cnt = 0;
  2380. TCP_ECN_queue_cwr(tp);
  2381. tcp_moderate_cwnd(tp);
  2382. }
  2383. /* A conservative spurious RTO response algorithm: reduce cwnd using
  2384. * rate halving and continue in congestion avoidance.
  2385. */
  2386. static void tcp_ratehalving_spur_to_response(struct sock *sk)
  2387. {
  2388. tcp_enter_cwr(sk, 0);
  2389. }
  2390. static void tcp_undo_spur_to_response(struct sock *sk, int flag)
  2391. {
  2392. if (flag&FLAG_ECE)
  2393. tcp_ratehalving_spur_to_response(sk);
  2394. else
  2395. tcp_undo_cwr(sk, 1);
  2396. }
  2397. /* F-RTO spurious RTO detection algorithm (RFC4138)
  2398. *
  2399. * F-RTO affects during two new ACKs following RTO (well, almost, see inline
  2400. * comments). State (ACK number) is kept in frto_counter. When ACK advances
  2401. * window (but not to or beyond highest sequence sent before RTO):
  2402. * On First ACK, send two new segments out.
  2403. * On Second ACK, RTO was likely spurious. Do spurious response (response
  2404. * algorithm is not part of the F-RTO detection algorithm
  2405. * given in RFC4138 but can be selected separately).
  2406. * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
  2407. * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
  2408. * of Nagle, this is done using frto_counter states 2 and 3, when a new data
  2409. * segment of any size sent during F-RTO, state 2 is upgraded to 3.
  2410. *
  2411. * Rationale: if the RTO was spurious, new ACKs should arrive from the
  2412. * original window even after we transmit two new data segments.
  2413. *
  2414. * SACK version:
  2415. * on first step, wait until first cumulative ACK arrives, then move to
  2416. * the second step. In second step, the next ACK decides.
  2417. *
  2418. * F-RTO is implemented (mainly) in four functions:
  2419. * - tcp_use_frto() is used to determine if TCP is can use F-RTO
  2420. * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
  2421. * called when tcp_use_frto() showed green light
  2422. * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
  2423. * - tcp_enter_frto_loss() is called if there is not enough evidence
  2424. * to prove that the RTO is indeed spurious. It transfers the control
  2425. * from F-RTO to the conventional RTO recovery
  2426. */
  2427. static int tcp_process_frto(struct sock *sk, int flag)
  2428. {
  2429. struct tcp_sock *tp = tcp_sk(sk);
  2430. tcp_sync_left_out(tp);
  2431. /* Duplicate the behavior from Loss state (fastretrans_alert) */
  2432. if (flag&FLAG_DATA_ACKED)
  2433. inet_csk(sk)->icsk_retransmits = 0;
  2434. if (!before(tp->snd_una, tp->frto_highmark)) {
  2435. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
  2436. return 1;
  2437. }
  2438. if (!IsSackFrto() || IsReno(tp)) {
  2439. /* RFC4138 shortcoming in step 2; should also have case c):
  2440. * ACK isn't duplicate nor advances window, e.g., opposite dir
  2441. * data, winupdate
  2442. */
  2443. if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP))
  2444. return 1;
  2445. if (!(flag&FLAG_DATA_ACKED)) {
  2446. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
  2447. flag);
  2448. return 1;
  2449. }
  2450. } else {
  2451. if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
  2452. /* Prevent sending of new data. */
  2453. tp->snd_cwnd = min(tp->snd_cwnd,
  2454. tcp_packets_in_flight(tp));
  2455. return 1;
  2456. }
  2457. if ((tp->frto_counter >= 2) &&
  2458. (!(flag&FLAG_FORWARD_PROGRESS) ||
  2459. ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
  2460. /* RFC4138 shortcoming (see comment above) */
  2461. if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
  2462. return 1;
  2463. tcp_enter_frto_loss(sk, 3, flag);
  2464. return 1;
  2465. }
  2466. }
  2467. if (tp->frto_counter == 1) {
  2468. /* Sending of the next skb must be allowed or no FRTO */
  2469. if (!tcp_send_head(sk) ||
  2470. after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
  2471. tp->snd_una + tp->snd_wnd)) {
  2472. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
  2473. flag);
  2474. return 1;
  2475. }
  2476. tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
  2477. tp->frto_counter = 2;
  2478. return 1;
  2479. } else {
  2480. switch (sysctl_tcp_frto_response) {
  2481. case 2:
  2482. tcp_undo_spur_to_response(sk, flag);
  2483. break;
  2484. case 1:
  2485. tcp_conservative_spur_to_response(tp);
  2486. break;
  2487. default:
  2488. tcp_ratehalving_spur_to_response(sk);
  2489. break;
  2490. }
  2491. tp->frto_counter = 0;
  2492. }
  2493. return 0;
  2494. }
  2495. /* This routine deals with incoming acks, but not outgoing ones. */
  2496. static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
  2497. {
  2498. struct inet_connection_sock *icsk = inet_csk(sk);
  2499. struct tcp_sock *tp = tcp_sk(sk);
  2500. u32 prior_snd_una = tp->snd_una;
  2501. u32 ack_seq = TCP_SKB_CB(skb)->seq;
  2502. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2503. u32 prior_in_flight;
  2504. s32 seq_rtt;
  2505. int prior_packets;
  2506. int frto_cwnd = 0;
  2507. /* If the ack is newer than sent or older than previous acks
  2508. * then we can probably ignore it.
  2509. */
  2510. if (after(ack, tp->snd_nxt))
  2511. goto uninteresting_ack;
  2512. if (before(ack, prior_snd_una))
  2513. goto old_ack;
  2514. if (after(ack, prior_snd_una))
  2515. flag |= FLAG_SND_UNA_ADVANCED;
  2516. if (sysctl_tcp_abc) {
  2517. if (icsk->icsk_ca_state < TCP_CA_CWR)
  2518. tp->bytes_acked += ack - prior_snd_una;
  2519. else if (icsk->icsk_ca_state == TCP_CA_Loss)
  2520. /* we assume just one segment left network */
  2521. tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
  2522. }
  2523. if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
  2524. /* Window is constant, pure forward advance.
  2525. * No more checks are required.
  2526. * Note, we use the fact that SND.UNA>=SND.WL2.
  2527. */
  2528. tcp_update_wl(tp, ack, ack_seq);
  2529. tp->snd_una = ack;
  2530. flag |= FLAG_WIN_UPDATE;
  2531. tcp_ca_event(sk, CA_EVENT_FAST_ACK);
  2532. NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
  2533. } else {
  2534. if (ack_seq != TCP_SKB_CB(skb)->end_seq)
  2535. flag |= FLAG_DATA;
  2536. else
  2537. NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
  2538. flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
  2539. if (TCP_SKB_CB(skb)->sacked)
  2540. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2541. if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
  2542. flag |= FLAG_ECE;
  2543. tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
  2544. }
  2545. /* We passed data and got it acked, remove any soft error
  2546. * log. Something worked...
  2547. */
  2548. sk->sk_err_soft = 0;
  2549. tp->rcv_tstamp = tcp_time_stamp;
  2550. prior_packets = tp->packets_out;
  2551. if (!prior_packets)
  2552. goto no_queue;
  2553. prior_in_flight = tcp_packets_in_flight(tp);
  2554. /* See if we can take anything off of the retransmit queue. */
  2555. flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
  2556. if (tp->frto_counter)
  2557. frto_cwnd = tcp_process_frto(sk, flag);
  2558. if (tcp_ack_is_dubious(sk, flag)) {
  2559. /* Advance CWND, if state allows this. */
  2560. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
  2561. tcp_may_raise_cwnd(sk, flag))
  2562. tcp_cong_avoid(sk, ack, prior_in_flight, 0);
  2563. tcp_fastretrans_alert(sk, prior_packets, flag);
  2564. } else {
  2565. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
  2566. tcp_cong_avoid(sk, ack, prior_in_flight, 1);
  2567. }
  2568. if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
  2569. dst_confirm(sk->sk_dst_cache);
  2570. return 1;
  2571. no_queue:
  2572. icsk->icsk_probes_out = 0;
  2573. /* If this ack opens up a zero window, clear backoff. It was
  2574. * being used to time the probes, and is probably far higher than
  2575. * it needs to be for normal retransmission.
  2576. */
  2577. if (tcp_send_head(sk))
  2578. tcp_ack_probe(sk);
  2579. return 1;
  2580. old_ack:
  2581. if (TCP_SKB_CB(skb)->sacked)
  2582. tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2583. uninteresting_ack:
  2584. SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  2585. return 0;
  2586. }
  2587. /* Look for tcp options. Normally only called on SYN and SYNACK packets.
  2588. * But, this can also be called on packets in the established flow when
  2589. * the fast version below fails.
  2590. */
  2591. void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
  2592. {
  2593. unsigned char *ptr;
  2594. struct tcphdr *th = tcp_hdr(skb);
  2595. int length=(th->doff*4)-sizeof(struct tcphdr);
  2596. ptr = (unsigned char *)(th + 1);
  2597. opt_rx->saw_tstamp = 0;
  2598. while (length > 0) {
  2599. int opcode=*ptr++;
  2600. int opsize;
  2601. switch (opcode) {
  2602. case TCPOPT_EOL:
  2603. return;
  2604. case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
  2605. length--;
  2606. continue;
  2607. default:
  2608. opsize=*ptr++;
  2609. if (opsize < 2) /* "silly options" */
  2610. return;
  2611. if (opsize > length)
  2612. return; /* don't parse partial options */
  2613. switch (opcode) {
  2614. case TCPOPT_MSS:
  2615. if (opsize==TCPOLEN_MSS && th->syn && !estab) {
  2616. u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
  2617. if (in_mss) {
  2618. if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
  2619. in_mss = opt_rx->user_mss;
  2620. opt_rx->mss_clamp = in_mss;
  2621. }
  2622. }
  2623. break;
  2624. case TCPOPT_WINDOW:
  2625. if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
  2626. if (sysctl_tcp_window_scaling) {
  2627. __u8 snd_wscale = *(__u8 *) ptr;
  2628. opt_rx->wscale_ok = 1;
  2629. if (snd_wscale > 14) {
  2630. if (net_ratelimit())
  2631. printk(KERN_INFO "tcp_parse_options: Illegal window "
  2632. "scaling value %d >14 received.\n",
  2633. snd_wscale);
  2634. snd_wscale = 14;
  2635. }
  2636. opt_rx->snd_wscale = snd_wscale;
  2637. }
  2638. break;
  2639. case TCPOPT_TIMESTAMP:
  2640. if (opsize==TCPOLEN_TIMESTAMP) {
  2641. if ((estab && opt_rx->tstamp_ok) ||
  2642. (!estab && sysctl_tcp_timestamps)) {
  2643. opt_rx->saw_tstamp = 1;
  2644. opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
  2645. opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
  2646. }
  2647. }
  2648. break;
  2649. case TCPOPT_SACK_PERM:
  2650. if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
  2651. if (sysctl_tcp_sack) {
  2652. opt_rx->sack_ok = 1;
  2653. tcp_sack_reset(opt_rx);
  2654. }
  2655. }
  2656. break;
  2657. case TCPOPT_SACK:
  2658. if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
  2659. !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
  2660. opt_rx->sack_ok) {
  2661. TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
  2662. }
  2663. break;
  2664. #ifdef CONFIG_TCP_MD5SIG
  2665. case TCPOPT_MD5SIG:
  2666. /*
  2667. * The MD5 Hash has already been
  2668. * checked (see tcp_v{4,6}_do_rcv()).
  2669. */
  2670. break;
  2671. #endif
  2672. }
  2673. ptr+=opsize-2;
  2674. length-=opsize;
  2675. }
  2676. }
  2677. }
  2678. /* Fast parse options. This hopes to only see timestamps.
  2679. * If it is wrong it falls back on tcp_parse_options().
  2680. */
  2681. static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
  2682. struct tcp_sock *tp)
  2683. {
  2684. if (th->doff == sizeof(struct tcphdr)>>2) {
  2685. tp->rx_opt.saw_tstamp = 0;
  2686. return 0;
  2687. } else if (tp->rx_opt.tstamp_ok &&
  2688. th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
  2689. __be32 *ptr = (__be32 *)(th + 1);
  2690. if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  2691. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
  2692. tp->rx_opt.saw_tstamp = 1;
  2693. ++ptr;
  2694. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  2695. ++ptr;
  2696. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  2697. return 1;
  2698. }
  2699. }
  2700. tcp_parse_options(skb, &tp->rx_opt, 1);
  2701. return 1;
  2702. }
  2703. static inline void tcp_store_ts_recent(struct tcp_sock *tp)
  2704. {
  2705. tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
  2706. tp->rx_opt.ts_recent_stamp = get_seconds();
  2707. }
  2708. static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
  2709. {
  2710. if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
  2711. /* PAWS bug workaround wrt. ACK frames, the PAWS discard
  2712. * extra check below makes sure this can only happen
  2713. * for pure ACK frames. -DaveM
  2714. *
  2715. * Not only, also it occurs for expired timestamps.
  2716. */
  2717. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
  2718. get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
  2719. tcp_store_ts_recent(tp);
  2720. }
  2721. }
  2722. /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
  2723. *
  2724. * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
  2725. * it can pass through stack. So, the following predicate verifies that
  2726. * this segment is not used for anything but congestion avoidance or
  2727. * fast retransmit. Moreover, we even are able to eliminate most of such
  2728. * second order effects, if we apply some small "replay" window (~RTO)
  2729. * to timestamp space.
  2730. *
  2731. * All these measures still do not guarantee that we reject wrapped ACKs
  2732. * on networks with high bandwidth, when sequence space is recycled fastly,
  2733. * but it guarantees that such events will be very rare and do not affect
  2734. * connection seriously. This doesn't look nice, but alas, PAWS is really
  2735. * buggy extension.
  2736. *
  2737. * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
  2738. * states that events when retransmit arrives after original data are rare.
  2739. * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
  2740. * the biggest problem on large power networks even with minor reordering.
  2741. * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
  2742. * up to bandwidth of 18Gigabit/sec. 8) ]
  2743. */
  2744. static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
  2745. {
  2746. struct tcp_sock *tp = tcp_sk(sk);
  2747. struct tcphdr *th = tcp_hdr(skb);
  2748. u32 seq = TCP_SKB_CB(skb)->seq;
  2749. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2750. return (/* 1. Pure ACK with correct sequence number. */
  2751. (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
  2752. /* 2. ... and duplicate ACK. */
  2753. ack == tp->snd_una &&
  2754. /* 3. ... and does not update window. */
  2755. !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
  2756. /* 4. ... and sits in replay window. */
  2757. (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
  2758. }
  2759. static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
  2760. {
  2761. const struct tcp_sock *tp = tcp_sk(sk);
  2762. return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
  2763. get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
  2764. !tcp_disordered_ack(sk, skb));
  2765. }
  2766. /* Check segment sequence number for validity.
  2767. *
  2768. * Segment controls are considered valid, if the segment
  2769. * fits to the window after truncation to the window. Acceptability
  2770. * of data (and SYN, FIN, of course) is checked separately.
  2771. * See tcp_data_queue(), for example.
  2772. *
  2773. * Also, controls (RST is main one) are accepted using RCV.WUP instead
  2774. * of RCV.NXT. Peer still did not advance his SND.UNA when we
  2775. * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
  2776. * (borrowed from freebsd)
  2777. */
  2778. static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2779. {
  2780. return !before(end_seq, tp->rcv_wup) &&
  2781. !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
  2782. }
  2783. /* When we get a reset we do this. */
  2784. static void tcp_reset(struct sock *sk)
  2785. {
  2786. /* We want the right error as BSD sees it (and indeed as we do). */
  2787. switch (sk->sk_state) {
  2788. case TCP_SYN_SENT:
  2789. sk->sk_err = ECONNREFUSED;
  2790. break;
  2791. case TCP_CLOSE_WAIT:
  2792. sk->sk_err = EPIPE;
  2793. break;
  2794. case TCP_CLOSE:
  2795. return;
  2796. default:
  2797. sk->sk_err = ECONNRESET;
  2798. }
  2799. if (!sock_flag(sk, SOCK_DEAD))
  2800. sk->sk_error_report(sk);
  2801. tcp_done(sk);
  2802. }
  2803. /*
  2804. * Process the FIN bit. This now behaves as it is supposed to work
  2805. * and the FIN takes effect when it is validly part of sequence
  2806. * space. Not before when we get holes.
  2807. *
  2808. * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
  2809. * (and thence onto LAST-ACK and finally, CLOSE, we never enter
  2810. * TIME-WAIT)
  2811. *
  2812. * If we are in FINWAIT-1, a received FIN indicates simultaneous
  2813. * close and we go into CLOSING (and later onto TIME-WAIT)
  2814. *
  2815. * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
  2816. */
  2817. static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
  2818. {
  2819. struct tcp_sock *tp = tcp_sk(sk);
  2820. inet_csk_schedule_ack(sk);
  2821. sk->sk_shutdown |= RCV_SHUTDOWN;
  2822. sock_set_flag(sk, SOCK_DONE);
  2823. switch (sk->sk_state) {
  2824. case TCP_SYN_RECV:
  2825. case TCP_ESTABLISHED:
  2826. /* Move to CLOSE_WAIT */
  2827. tcp_set_state(sk, TCP_CLOSE_WAIT);
  2828. inet_csk(sk)->icsk_ack.pingpong = 1;
  2829. break;
  2830. case TCP_CLOSE_WAIT:
  2831. case TCP_CLOSING:
  2832. /* Received a retransmission of the FIN, do
  2833. * nothing.
  2834. */
  2835. break;
  2836. case TCP_LAST_ACK:
  2837. /* RFC793: Remain in the LAST-ACK state. */
  2838. break;
  2839. case TCP_FIN_WAIT1:
  2840. /* This case occurs when a simultaneous close
  2841. * happens, we must ack the received FIN and
  2842. * enter the CLOSING state.
  2843. */
  2844. tcp_send_ack(sk);
  2845. tcp_set_state(sk, TCP_CLOSING);
  2846. break;
  2847. case TCP_FIN_WAIT2:
  2848. /* Received a FIN -- send ACK and enter TIME_WAIT. */
  2849. tcp_send_ack(sk);
  2850. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  2851. break;
  2852. default:
  2853. /* Only TCP_LISTEN and TCP_CLOSE are left, in these
  2854. * cases we should never reach this piece of code.
  2855. */
  2856. printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
  2857. __FUNCTION__, sk->sk_state);
  2858. break;
  2859. }
  2860. /* It _is_ possible, that we have something out-of-order _after_ FIN.
  2861. * Probably, we should reset in this case. For now drop them.
  2862. */
  2863. __skb_queue_purge(&tp->out_of_order_queue);
  2864. if (tp->rx_opt.sack_ok)
  2865. tcp_sack_reset(&tp->rx_opt);
  2866. sk_stream_mem_reclaim(sk);
  2867. if (!sock_flag(sk, SOCK_DEAD)) {
  2868. sk->sk_state_change(sk);
  2869. /* Do not send POLL_HUP for half duplex close. */
  2870. if (sk->sk_shutdown == SHUTDOWN_MASK ||
  2871. sk->sk_state == TCP_CLOSE)
  2872. sk_wake_async(sk, 1, POLL_HUP);
  2873. else
  2874. sk_wake_async(sk, 1, POLL_IN);
  2875. }
  2876. }
  2877. static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
  2878. {
  2879. if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
  2880. if (before(seq, sp->start_seq))
  2881. sp->start_seq = seq;
  2882. if (after(end_seq, sp->end_seq))
  2883. sp->end_seq = end_seq;
  2884. return 1;
  2885. }
  2886. return 0;
  2887. }
  2888. static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2889. {
  2890. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2891. if (before(seq, tp->rcv_nxt))
  2892. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
  2893. else
  2894. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
  2895. tp->rx_opt.dsack = 1;
  2896. tp->duplicate_sack[0].start_seq = seq;
  2897. tp->duplicate_sack[0].end_seq = end_seq;
  2898. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
  2899. }
  2900. }
  2901. static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2902. {
  2903. if (!tp->rx_opt.dsack)
  2904. tcp_dsack_set(tp, seq, end_seq);
  2905. else
  2906. tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
  2907. }
  2908. static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
  2909. {
  2910. struct tcp_sock *tp = tcp_sk(sk);
  2911. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  2912. before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  2913. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  2914. tcp_enter_quickack_mode(sk);
  2915. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2916. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  2917. if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
  2918. end_seq = tp->rcv_nxt;
  2919. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
  2920. }
  2921. }
  2922. tcp_send_ack(sk);
  2923. }
  2924. /* These routines update the SACK block as out-of-order packets arrive or
  2925. * in-order packets close up the sequence space.
  2926. */
  2927. static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
  2928. {
  2929. int this_sack;
  2930. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2931. struct tcp_sack_block *swalk = sp+1;
  2932. /* See if the recent change to the first SACK eats into
  2933. * or hits the sequence space of other SACK blocks, if so coalesce.
  2934. */
  2935. for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
  2936. if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
  2937. int i;
  2938. /* Zap SWALK, by moving every further SACK up by one slot.
  2939. * Decrease num_sacks.
  2940. */
  2941. tp->rx_opt.num_sacks--;
  2942. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2943. for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
  2944. sp[i] = sp[i+1];
  2945. continue;
  2946. }
  2947. this_sack++, swalk++;
  2948. }
  2949. }
  2950. static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
  2951. {
  2952. __u32 tmp;
  2953. tmp = sack1->start_seq;
  2954. sack1->start_seq = sack2->start_seq;
  2955. sack2->start_seq = tmp;
  2956. tmp = sack1->end_seq;
  2957. sack1->end_seq = sack2->end_seq;
  2958. sack2->end_seq = tmp;
  2959. }
  2960. static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
  2961. {
  2962. struct tcp_sock *tp = tcp_sk(sk);
  2963. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2964. int cur_sacks = tp->rx_opt.num_sacks;
  2965. int this_sack;
  2966. if (!cur_sacks)
  2967. goto new_sack;
  2968. for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
  2969. if (tcp_sack_extend(sp, seq, end_seq)) {
  2970. /* Rotate this_sack to the first one. */
  2971. for (; this_sack>0; this_sack--, sp--)
  2972. tcp_sack_swap(sp, sp-1);
  2973. if (cur_sacks > 1)
  2974. tcp_sack_maybe_coalesce(tp);
  2975. return;
  2976. }
  2977. }
  2978. /* Could not find an adjacent existing SACK, build a new one,
  2979. * put it at the front, and shift everyone else down. We
  2980. * always know there is at least one SACK present already here.
  2981. *
  2982. * If the sack array is full, forget about the last one.
  2983. */
  2984. if (this_sack >= 4) {
  2985. this_sack--;
  2986. tp->rx_opt.num_sacks--;
  2987. sp--;
  2988. }
  2989. for (; this_sack > 0; this_sack--, sp--)
  2990. *sp = *(sp-1);
  2991. new_sack:
  2992. /* Build the new head SACK, and we're done. */
  2993. sp->start_seq = seq;
  2994. sp->end_seq = end_seq;
  2995. tp->rx_opt.num_sacks++;
  2996. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2997. }
  2998. /* RCV.NXT advances, some SACKs should be eaten. */
  2999. static void tcp_sack_remove(struct tcp_sock *tp)
  3000. {
  3001. struct tcp_sack_block *sp = &tp->selective_acks[0];
  3002. int num_sacks = tp->rx_opt.num_sacks;
  3003. int this_sack;
  3004. /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
  3005. if (skb_queue_empty(&tp->out_of_order_queue)) {
  3006. tp->rx_opt.num_sacks = 0;
  3007. tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
  3008. return;
  3009. }
  3010. for (this_sack = 0; this_sack < num_sacks; ) {
  3011. /* Check if the start of the sack is covered by RCV.NXT. */
  3012. if (!before(tp->rcv_nxt, sp->start_seq)) {
  3013. int i;
  3014. /* RCV.NXT must cover all the block! */
  3015. BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
  3016. /* Zap this SACK, by moving forward any other SACKS. */
  3017. for (i=this_sack+1; i < num_sacks; i++)
  3018. tp->selective_acks[i-1] = tp->selective_acks[i];
  3019. num_sacks--;
  3020. continue;
  3021. }
  3022. this_sack++;
  3023. sp++;
  3024. }
  3025. if (num_sacks != tp->rx_opt.num_sacks) {
  3026. tp->rx_opt.num_sacks = num_sacks;
  3027. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  3028. }
  3029. }
  3030. /* This one checks to see if we can put data from the
  3031. * out_of_order queue into the receive_queue.
  3032. */
  3033. static void tcp_ofo_queue(struct sock *sk)
  3034. {
  3035. struct tcp_sock *tp = tcp_sk(sk);
  3036. __u32 dsack_high = tp->rcv_nxt;
  3037. struct sk_buff *skb;
  3038. while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
  3039. if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  3040. break;
  3041. if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
  3042. __u32 dsack = dsack_high;
  3043. if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
  3044. dsack_high = TCP_SKB_CB(skb)->end_seq;
  3045. tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
  3046. }
  3047. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3048. SOCK_DEBUG(sk, "ofo packet was already received \n");
  3049. __skb_unlink(skb, &tp->out_of_order_queue);
  3050. __kfree_skb(skb);
  3051. continue;
  3052. }
  3053. SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
  3054. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3055. TCP_SKB_CB(skb)->end_seq);
  3056. __skb_unlink(skb, &tp->out_of_order_queue);
  3057. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3058. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3059. if (tcp_hdr(skb)->fin)
  3060. tcp_fin(skb, sk, tcp_hdr(skb));
  3061. }
  3062. }
  3063. static int tcp_prune_queue(struct sock *sk);
  3064. static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
  3065. {
  3066. struct tcphdr *th = tcp_hdr(skb);
  3067. struct tcp_sock *tp = tcp_sk(sk);
  3068. int eaten = -1;
  3069. if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
  3070. goto drop;
  3071. __skb_pull(skb, th->doff*4);
  3072. TCP_ECN_accept_cwr(tp, skb);
  3073. if (tp->rx_opt.dsack) {
  3074. tp->rx_opt.dsack = 0;
  3075. tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
  3076. 4 - tp->rx_opt.tstamp_ok);
  3077. }
  3078. /* Queue data for delivery to the user.
  3079. * Packets in sequence go to the receive queue.
  3080. * Out of sequence packets to the out_of_order_queue.
  3081. */
  3082. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3083. if (tcp_receive_window(tp) == 0)
  3084. goto out_of_window;
  3085. /* Ok. In sequence. In window. */
  3086. if (tp->ucopy.task == current &&
  3087. tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
  3088. sock_owned_by_user(sk) && !tp->urg_data) {
  3089. int chunk = min_t(unsigned int, skb->len,
  3090. tp->ucopy.len);
  3091. __set_current_state(TASK_RUNNING);
  3092. local_bh_enable();
  3093. if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
  3094. tp->ucopy.len -= chunk;
  3095. tp->copied_seq += chunk;
  3096. eaten = (chunk == skb->len && !th->fin);
  3097. tcp_rcv_space_adjust(sk);
  3098. }
  3099. local_bh_disable();
  3100. }
  3101. if (eaten <= 0) {
  3102. queue_and_out:
  3103. if (eaten < 0 &&
  3104. (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3105. !sk_stream_rmem_schedule(sk, skb))) {
  3106. if (tcp_prune_queue(sk) < 0 ||
  3107. !sk_stream_rmem_schedule(sk, skb))
  3108. goto drop;
  3109. }
  3110. sk_stream_set_owner_r(skb, sk);
  3111. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3112. }
  3113. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3114. if (skb->len)
  3115. tcp_event_data_recv(sk, skb);
  3116. if (th->fin)
  3117. tcp_fin(skb, sk, th);
  3118. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3119. tcp_ofo_queue(sk);
  3120. /* RFC2581. 4.2. SHOULD send immediate ACK, when
  3121. * gap in queue is filled.
  3122. */
  3123. if (skb_queue_empty(&tp->out_of_order_queue))
  3124. inet_csk(sk)->icsk_ack.pingpong = 0;
  3125. }
  3126. if (tp->rx_opt.num_sacks)
  3127. tcp_sack_remove(tp);
  3128. tcp_fast_path_check(sk);
  3129. if (eaten > 0)
  3130. __kfree_skb(skb);
  3131. else if (!sock_flag(sk, SOCK_DEAD))
  3132. sk->sk_data_ready(sk, 0);
  3133. return;
  3134. }
  3135. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3136. /* A retransmit, 2nd most common case. Force an immediate ack. */
  3137. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  3138. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3139. out_of_window:
  3140. tcp_enter_quickack_mode(sk);
  3141. inet_csk_schedule_ack(sk);
  3142. drop:
  3143. __kfree_skb(skb);
  3144. return;
  3145. }
  3146. /* Out of window. F.e. zero window probe. */
  3147. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
  3148. goto out_of_window;
  3149. tcp_enter_quickack_mode(sk);
  3150. if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3151. /* Partial packet, seq < rcv_next < end_seq */
  3152. SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
  3153. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3154. TCP_SKB_CB(skb)->end_seq);
  3155. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
  3156. /* If window is closed, drop tail of packet. But after
  3157. * remembering D-SACK for its head made in previous line.
  3158. */
  3159. if (!tcp_receive_window(tp))
  3160. goto out_of_window;
  3161. goto queue_and_out;
  3162. }
  3163. TCP_ECN_check_ce(tp, skb);
  3164. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3165. !sk_stream_rmem_schedule(sk, skb)) {
  3166. if (tcp_prune_queue(sk) < 0 ||
  3167. !sk_stream_rmem_schedule(sk, skb))
  3168. goto drop;
  3169. }
  3170. /* Disable header prediction. */
  3171. tp->pred_flags = 0;
  3172. inet_csk_schedule_ack(sk);
  3173. SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
  3174. tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3175. sk_stream_set_owner_r(skb, sk);
  3176. if (!skb_peek(&tp->out_of_order_queue)) {
  3177. /* Initial out of order segment, build 1 SACK. */
  3178. if (tp->rx_opt.sack_ok) {
  3179. tp->rx_opt.num_sacks = 1;
  3180. tp->rx_opt.dsack = 0;
  3181. tp->rx_opt.eff_sacks = 1;
  3182. tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
  3183. tp->selective_acks[0].end_seq =
  3184. TCP_SKB_CB(skb)->end_seq;
  3185. }
  3186. __skb_queue_head(&tp->out_of_order_queue,skb);
  3187. } else {
  3188. struct sk_buff *skb1 = tp->out_of_order_queue.prev;
  3189. u32 seq = TCP_SKB_CB(skb)->seq;
  3190. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  3191. if (seq == TCP_SKB_CB(skb1)->end_seq) {
  3192. __skb_append(skb1, skb, &tp->out_of_order_queue);
  3193. if (!tp->rx_opt.num_sacks ||
  3194. tp->selective_acks[0].end_seq != seq)
  3195. goto add_sack;
  3196. /* Common case: data arrive in order after hole. */
  3197. tp->selective_acks[0].end_seq = end_seq;
  3198. return;
  3199. }
  3200. /* Find place to insert this segment. */
  3201. do {
  3202. if (!after(TCP_SKB_CB(skb1)->seq, seq))
  3203. break;
  3204. } while ((skb1 = skb1->prev) !=
  3205. (struct sk_buff*)&tp->out_of_order_queue);
  3206. /* Do skb overlap to previous one? */
  3207. if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
  3208. before(seq, TCP_SKB_CB(skb1)->end_seq)) {
  3209. if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3210. /* All the bits are present. Drop. */
  3211. __kfree_skb(skb);
  3212. tcp_dsack_set(tp, seq, end_seq);
  3213. goto add_sack;
  3214. }
  3215. if (after(seq, TCP_SKB_CB(skb1)->seq)) {
  3216. /* Partial overlap. */
  3217. tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
  3218. } else {
  3219. skb1 = skb1->prev;
  3220. }
  3221. }
  3222. __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
  3223. /* And clean segments covered by new one as whole. */
  3224. while ((skb1 = skb->next) !=
  3225. (struct sk_buff*)&tp->out_of_order_queue &&
  3226. after(end_seq, TCP_SKB_CB(skb1)->seq)) {
  3227. if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3228. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
  3229. break;
  3230. }
  3231. __skb_unlink(skb1, &tp->out_of_order_queue);
  3232. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
  3233. __kfree_skb(skb1);
  3234. }
  3235. add_sack:
  3236. if (tp->rx_opt.sack_ok)
  3237. tcp_sack_new_ofo_skb(sk, seq, end_seq);
  3238. }
  3239. }
  3240. /* Collapse contiguous sequence of skbs head..tail with
  3241. * sequence numbers start..end.
  3242. * Segments with FIN/SYN are not collapsed (only because this
  3243. * simplifies code)
  3244. */
  3245. static void
  3246. tcp_collapse(struct sock *sk, struct sk_buff_head *list,
  3247. struct sk_buff *head, struct sk_buff *tail,
  3248. u32 start, u32 end)
  3249. {
  3250. struct sk_buff *skb;
  3251. /* First, check that queue is collapsible and find
  3252. * the point where collapsing can be useful. */
  3253. for (skb = head; skb != tail; ) {
  3254. /* No new bits? It is possible on ofo queue. */
  3255. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3256. struct sk_buff *next = skb->next;
  3257. __skb_unlink(skb, list);
  3258. __kfree_skb(skb);
  3259. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3260. skb = next;
  3261. continue;
  3262. }
  3263. /* The first skb to collapse is:
  3264. * - not SYN/FIN and
  3265. * - bloated or contains data before "start" or
  3266. * overlaps to the next one.
  3267. */
  3268. if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
  3269. (tcp_win_from_space(skb->truesize) > skb->len ||
  3270. before(TCP_SKB_CB(skb)->seq, start) ||
  3271. (skb->next != tail &&
  3272. TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
  3273. break;
  3274. /* Decided to skip this, advance start seq. */
  3275. start = TCP_SKB_CB(skb)->end_seq;
  3276. skb = skb->next;
  3277. }
  3278. if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
  3279. return;
  3280. while (before(start, end)) {
  3281. struct sk_buff *nskb;
  3282. int header = skb_headroom(skb);
  3283. int copy = SKB_MAX_ORDER(header, 0);
  3284. /* Too big header? This can happen with IPv6. */
  3285. if (copy < 0)
  3286. return;
  3287. if (end-start < copy)
  3288. copy = end-start;
  3289. nskb = alloc_skb(copy+header, GFP_ATOMIC);
  3290. if (!nskb)
  3291. return;
  3292. skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
  3293. skb_set_network_header(nskb, (skb_network_header(skb) -
  3294. skb->head));
  3295. skb_set_transport_header(nskb, (skb_transport_header(skb) -
  3296. skb->head));
  3297. skb_reserve(nskb, header);
  3298. memcpy(nskb->head, skb->head, header);
  3299. memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
  3300. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
  3301. __skb_insert(nskb, skb->prev, skb, list);
  3302. sk_stream_set_owner_r(nskb, sk);
  3303. /* Copy data, releasing collapsed skbs. */
  3304. while (copy > 0) {
  3305. int offset = start - TCP_SKB_CB(skb)->seq;
  3306. int size = TCP_SKB_CB(skb)->end_seq - start;
  3307. BUG_ON(offset < 0);
  3308. if (size > 0) {
  3309. size = min(copy, size);
  3310. if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
  3311. BUG();
  3312. TCP_SKB_CB(nskb)->end_seq += size;
  3313. copy -= size;
  3314. start += size;
  3315. }
  3316. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3317. struct sk_buff *next = skb->next;
  3318. __skb_unlink(skb, list);
  3319. __kfree_skb(skb);
  3320. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3321. skb = next;
  3322. if (skb == tail ||
  3323. tcp_hdr(skb)->syn ||
  3324. tcp_hdr(skb)->fin)
  3325. return;
  3326. }
  3327. }
  3328. }
  3329. }
  3330. /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
  3331. * and tcp_collapse() them until all the queue is collapsed.
  3332. */
  3333. static void tcp_collapse_ofo_queue(struct sock *sk)
  3334. {
  3335. struct tcp_sock *tp = tcp_sk(sk);
  3336. struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
  3337. struct sk_buff *head;
  3338. u32 start, end;
  3339. if (skb == NULL)
  3340. return;
  3341. start = TCP_SKB_CB(skb)->seq;
  3342. end = TCP_SKB_CB(skb)->end_seq;
  3343. head = skb;
  3344. for (;;) {
  3345. skb = skb->next;
  3346. /* Segment is terminated when we see gap or when
  3347. * we are at the end of all the queue. */
  3348. if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
  3349. after(TCP_SKB_CB(skb)->seq, end) ||
  3350. before(TCP_SKB_CB(skb)->end_seq, start)) {
  3351. tcp_collapse(sk, &tp->out_of_order_queue,
  3352. head, skb, start, end);
  3353. head = skb;
  3354. if (skb == (struct sk_buff *)&tp->out_of_order_queue)
  3355. break;
  3356. /* Start new segment */
  3357. start = TCP_SKB_CB(skb)->seq;
  3358. end = TCP_SKB_CB(skb)->end_seq;
  3359. } else {
  3360. if (before(TCP_SKB_CB(skb)->seq, start))
  3361. start = TCP_SKB_CB(skb)->seq;
  3362. if (after(TCP_SKB_CB(skb)->end_seq, end))
  3363. end = TCP_SKB_CB(skb)->end_seq;
  3364. }
  3365. }
  3366. }
  3367. /* Reduce allocated memory if we can, trying to get
  3368. * the socket within its memory limits again.
  3369. *
  3370. * Return less than zero if we should start dropping frames
  3371. * until the socket owning process reads some of the data
  3372. * to stabilize the situation.
  3373. */
  3374. static int tcp_prune_queue(struct sock *sk)
  3375. {
  3376. struct tcp_sock *tp = tcp_sk(sk);
  3377. SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
  3378. NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
  3379. if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
  3380. tcp_clamp_window(sk);
  3381. else if (tcp_memory_pressure)
  3382. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
  3383. tcp_collapse_ofo_queue(sk);
  3384. tcp_collapse(sk, &sk->sk_receive_queue,
  3385. sk->sk_receive_queue.next,
  3386. (struct sk_buff*)&sk->sk_receive_queue,
  3387. tp->copied_seq, tp->rcv_nxt);
  3388. sk_stream_mem_reclaim(sk);
  3389. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3390. return 0;
  3391. /* Collapsing did not help, destructive actions follow.
  3392. * This must not ever occur. */
  3393. /* First, purge the out_of_order queue. */
  3394. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3395. NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
  3396. __skb_queue_purge(&tp->out_of_order_queue);
  3397. /* Reset SACK state. A conforming SACK implementation will
  3398. * do the same at a timeout based retransmit. When a connection
  3399. * is in a sad state like this, we care only about integrity
  3400. * of the connection not performance.
  3401. */
  3402. if (tp->rx_opt.sack_ok)
  3403. tcp_sack_reset(&tp->rx_opt);
  3404. sk_stream_mem_reclaim(sk);
  3405. }
  3406. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3407. return 0;
  3408. /* If we are really being abused, tell the caller to silently
  3409. * drop receive data on the floor. It will get retransmitted
  3410. * and hopefully then we'll have sufficient space.
  3411. */
  3412. NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
  3413. /* Massive buffer overcommit. */
  3414. tp->pred_flags = 0;
  3415. return -1;
  3416. }
  3417. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  3418. * As additional protections, we do not touch cwnd in retransmission phases,
  3419. * and if application hit its sndbuf limit recently.
  3420. */
  3421. void tcp_cwnd_application_limited(struct sock *sk)
  3422. {
  3423. struct tcp_sock *tp = tcp_sk(sk);
  3424. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  3425. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  3426. /* Limited by application or receiver window. */
  3427. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  3428. u32 win_used = max(tp->snd_cwnd_used, init_win);
  3429. if (win_used < tp->snd_cwnd) {
  3430. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  3431. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  3432. }
  3433. tp->snd_cwnd_used = 0;
  3434. }
  3435. tp->snd_cwnd_stamp = tcp_time_stamp;
  3436. }
  3437. static int tcp_should_expand_sndbuf(struct sock *sk)
  3438. {
  3439. struct tcp_sock *tp = tcp_sk(sk);
  3440. /* If the user specified a specific send buffer setting, do
  3441. * not modify it.
  3442. */
  3443. if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
  3444. return 0;
  3445. /* If we are under global TCP memory pressure, do not expand. */
  3446. if (tcp_memory_pressure)
  3447. return 0;
  3448. /* If we are under soft global TCP memory pressure, do not expand. */
  3449. if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
  3450. return 0;
  3451. /* If we filled the congestion window, do not expand. */
  3452. if (tp->packets_out >= tp->snd_cwnd)
  3453. return 0;
  3454. return 1;
  3455. }
  3456. /* When incoming ACK allowed to free some skb from write_queue,
  3457. * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
  3458. * on the exit from tcp input handler.
  3459. *
  3460. * PROBLEM: sndbuf expansion does not work well with largesend.
  3461. */
  3462. static void tcp_new_space(struct sock *sk)
  3463. {
  3464. struct tcp_sock *tp = tcp_sk(sk);
  3465. if (tcp_should_expand_sndbuf(sk)) {
  3466. int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
  3467. MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
  3468. demanded = max_t(unsigned int, tp->snd_cwnd,
  3469. tp->reordering + 1);
  3470. sndmem *= 2*demanded;
  3471. if (sndmem > sk->sk_sndbuf)
  3472. sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
  3473. tp->snd_cwnd_stamp = tcp_time_stamp;
  3474. }
  3475. sk->sk_write_space(sk);
  3476. }
  3477. static void tcp_check_space(struct sock *sk)
  3478. {
  3479. if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
  3480. sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
  3481. if (sk->sk_socket &&
  3482. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
  3483. tcp_new_space(sk);
  3484. }
  3485. }
  3486. static inline void tcp_data_snd_check(struct sock *sk)
  3487. {
  3488. tcp_push_pending_frames(sk);
  3489. tcp_check_space(sk);
  3490. }
  3491. /*
  3492. * Check if sending an ack is needed.
  3493. */
  3494. static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
  3495. {
  3496. struct tcp_sock *tp = tcp_sk(sk);
  3497. /* More than one full frame received... */
  3498. if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
  3499. /* ... and right edge of window advances far enough.
  3500. * (tcp_recvmsg() will send ACK otherwise). Or...
  3501. */
  3502. && __tcp_select_window(sk) >= tp->rcv_wnd) ||
  3503. /* We ACK each frame or... */
  3504. tcp_in_quickack_mode(sk) ||
  3505. /* We have out of order data. */
  3506. (ofo_possible &&
  3507. skb_peek(&tp->out_of_order_queue))) {
  3508. /* Then ack it now */
  3509. tcp_send_ack(sk);
  3510. } else {
  3511. /* Else, send delayed ack. */
  3512. tcp_send_delayed_ack(sk);
  3513. }
  3514. }
  3515. static inline void tcp_ack_snd_check(struct sock *sk)
  3516. {
  3517. if (!inet_csk_ack_scheduled(sk)) {
  3518. /* We sent a data segment already. */
  3519. return;
  3520. }
  3521. __tcp_ack_snd_check(sk, 1);
  3522. }
  3523. /*
  3524. * This routine is only called when we have urgent data
  3525. * signaled. Its the 'slow' part of tcp_urg. It could be
  3526. * moved inline now as tcp_urg is only called from one
  3527. * place. We handle URGent data wrong. We have to - as
  3528. * BSD still doesn't use the correction from RFC961.
  3529. * For 1003.1g we should support a new option TCP_STDURG to permit
  3530. * either form (or just set the sysctl tcp_stdurg).
  3531. */
  3532. static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
  3533. {
  3534. struct tcp_sock *tp = tcp_sk(sk);
  3535. u32 ptr = ntohs(th->urg_ptr);
  3536. if (ptr && !sysctl_tcp_stdurg)
  3537. ptr--;
  3538. ptr += ntohl(th->seq);
  3539. /* Ignore urgent data that we've already seen and read. */
  3540. if (after(tp->copied_seq, ptr))
  3541. return;
  3542. /* Do not replay urg ptr.
  3543. *
  3544. * NOTE: interesting situation not covered by specs.
  3545. * Misbehaving sender may send urg ptr, pointing to segment,
  3546. * which we already have in ofo queue. We are not able to fetch
  3547. * such data and will stay in TCP_URG_NOTYET until will be eaten
  3548. * by recvmsg(). Seems, we are not obliged to handle such wicked
  3549. * situations. But it is worth to think about possibility of some
  3550. * DoSes using some hypothetical application level deadlock.
  3551. */
  3552. if (before(ptr, tp->rcv_nxt))
  3553. return;
  3554. /* Do we already have a newer (or duplicate) urgent pointer? */
  3555. if (tp->urg_data && !after(ptr, tp->urg_seq))
  3556. return;
  3557. /* Tell the world about our new urgent pointer. */
  3558. sk_send_sigurg(sk);
  3559. /* We may be adding urgent data when the last byte read was
  3560. * urgent. To do this requires some care. We cannot just ignore
  3561. * tp->copied_seq since we would read the last urgent byte again
  3562. * as data, nor can we alter copied_seq until this data arrives
  3563. * or we break the semantics of SIOCATMARK (and thus sockatmark())
  3564. *
  3565. * NOTE. Double Dutch. Rendering to plain English: author of comment
  3566. * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
  3567. * and expect that both A and B disappear from stream. This is _wrong_.
  3568. * Though this happens in BSD with high probability, this is occasional.
  3569. * Any application relying on this is buggy. Note also, that fix "works"
  3570. * only in this artificial test. Insert some normal data between A and B and we will
  3571. * decline of BSD again. Verdict: it is better to remove to trap
  3572. * buggy users.
  3573. */
  3574. if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
  3575. !sock_flag(sk, SOCK_URGINLINE) &&
  3576. tp->copied_seq != tp->rcv_nxt) {
  3577. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  3578. tp->copied_seq++;
  3579. if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
  3580. __skb_unlink(skb, &sk->sk_receive_queue);
  3581. __kfree_skb(skb);
  3582. }
  3583. }
  3584. tp->urg_data = TCP_URG_NOTYET;
  3585. tp->urg_seq = ptr;
  3586. /* Disable header prediction. */
  3587. tp->pred_flags = 0;
  3588. }
  3589. /* This is the 'fast' part of urgent handling. */
  3590. static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
  3591. {
  3592. struct tcp_sock *tp = tcp_sk(sk);
  3593. /* Check if we get a new urgent pointer - normally not. */
  3594. if (th->urg)
  3595. tcp_check_urg(sk,th);
  3596. /* Do we wait for any urgent data? - normally not... */
  3597. if (tp->urg_data == TCP_URG_NOTYET) {
  3598. u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
  3599. th->syn;
  3600. /* Is the urgent pointer pointing into this packet? */
  3601. if (ptr < skb->len) {
  3602. u8 tmp;
  3603. if (skb_copy_bits(skb, ptr, &tmp, 1))
  3604. BUG();
  3605. tp->urg_data = TCP_URG_VALID | tmp;
  3606. if (!sock_flag(sk, SOCK_DEAD))
  3607. sk->sk_data_ready(sk, 0);
  3608. }
  3609. }
  3610. }
  3611. static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
  3612. {
  3613. struct tcp_sock *tp = tcp_sk(sk);
  3614. int chunk = skb->len - hlen;
  3615. int err;
  3616. local_bh_enable();
  3617. if (skb_csum_unnecessary(skb))
  3618. err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
  3619. else
  3620. err = skb_copy_and_csum_datagram_iovec(skb, hlen,
  3621. tp->ucopy.iov);
  3622. if (!err) {
  3623. tp->ucopy.len -= chunk;
  3624. tp->copied_seq += chunk;
  3625. tcp_rcv_space_adjust(sk);
  3626. }
  3627. local_bh_disable();
  3628. return err;
  3629. }
  3630. static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3631. {
  3632. __sum16 result;
  3633. if (sock_owned_by_user(sk)) {
  3634. local_bh_enable();
  3635. result = __tcp_checksum_complete(skb);
  3636. local_bh_disable();
  3637. } else {
  3638. result = __tcp_checksum_complete(skb);
  3639. }
  3640. return result;
  3641. }
  3642. static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3643. {
  3644. return !skb_csum_unnecessary(skb) &&
  3645. __tcp_checksum_complete_user(sk, skb);
  3646. }
  3647. #ifdef CONFIG_NET_DMA
  3648. static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
  3649. {
  3650. struct tcp_sock *tp = tcp_sk(sk);
  3651. int chunk = skb->len - hlen;
  3652. int dma_cookie;
  3653. int copied_early = 0;
  3654. if (tp->ucopy.wakeup)
  3655. return 0;
  3656. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  3657. tp->ucopy.dma_chan = get_softnet_dma();
  3658. if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
  3659. dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
  3660. skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
  3661. if (dma_cookie < 0)
  3662. goto out;
  3663. tp->ucopy.dma_cookie = dma_cookie;
  3664. copied_early = 1;
  3665. tp->ucopy.len -= chunk;
  3666. tp->copied_seq += chunk;
  3667. tcp_rcv_space_adjust(sk);
  3668. if ((tp->ucopy.len == 0) ||
  3669. (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
  3670. (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
  3671. tp->ucopy.wakeup = 1;
  3672. sk->sk_data_ready(sk, 0);
  3673. }
  3674. } else if (chunk > 0) {
  3675. tp->ucopy.wakeup = 1;
  3676. sk->sk_data_ready(sk, 0);
  3677. }
  3678. out:
  3679. return copied_early;
  3680. }
  3681. #endif /* CONFIG_NET_DMA */
  3682. /*
  3683. * TCP receive function for the ESTABLISHED state.
  3684. *
  3685. * It is split into a fast path and a slow path. The fast path is
  3686. * disabled when:
  3687. * - A zero window was announced from us - zero window probing
  3688. * is only handled properly in the slow path.
  3689. * - Out of order segments arrived.
  3690. * - Urgent data is expected.
  3691. * - There is no buffer space left
  3692. * - Unexpected TCP flags/window values/header lengths are received
  3693. * (detected by checking the TCP header against pred_flags)
  3694. * - Data is sent in both directions. Fast path only supports pure senders
  3695. * or pure receivers (this means either the sequence number or the ack
  3696. * value must stay constant)
  3697. * - Unexpected TCP option.
  3698. *
  3699. * When these conditions are not satisfied it drops into a standard
  3700. * receive procedure patterned after RFC793 to handle all cases.
  3701. * The first three cases are guaranteed by proper pred_flags setting,
  3702. * the rest is checked inline. Fast processing is turned on in
  3703. * tcp_data_queue when everything is OK.
  3704. */
  3705. int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  3706. struct tcphdr *th, unsigned len)
  3707. {
  3708. struct tcp_sock *tp = tcp_sk(sk);
  3709. /*
  3710. * Header prediction.
  3711. * The code loosely follows the one in the famous
  3712. * "30 instruction TCP receive" Van Jacobson mail.
  3713. *
  3714. * Van's trick is to deposit buffers into socket queue
  3715. * on a device interrupt, to call tcp_recv function
  3716. * on the receive process context and checksum and copy
  3717. * the buffer to user space. smart...
  3718. *
  3719. * Our current scheme is not silly either but we take the
  3720. * extra cost of the net_bh soft interrupt processing...
  3721. * We do checksum and copy also but from device to kernel.
  3722. */
  3723. tp->rx_opt.saw_tstamp = 0;
  3724. /* pred_flags is 0xS?10 << 16 + snd_wnd
  3725. * if header_prediction is to be made
  3726. * 'S' will always be tp->tcp_header_len >> 2
  3727. * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
  3728. * turn it off (when there are holes in the receive
  3729. * space for instance)
  3730. * PSH flag is ignored.
  3731. */
  3732. if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
  3733. TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3734. int tcp_header_len = tp->tcp_header_len;
  3735. /* Timestamp header prediction: tcp_header_len
  3736. * is automatically equal to th->doff*4 due to pred_flags
  3737. * match.
  3738. */
  3739. /* Check timestamp */
  3740. if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
  3741. __be32 *ptr = (__be32 *)(th + 1);
  3742. /* No? Slow path! */
  3743. if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  3744. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
  3745. goto slow_path;
  3746. tp->rx_opt.saw_tstamp = 1;
  3747. ++ptr;
  3748. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  3749. ++ptr;
  3750. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  3751. /* If PAWS failed, check it more carefully in slow path */
  3752. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
  3753. goto slow_path;
  3754. /* DO NOT update ts_recent here, if checksum fails
  3755. * and timestamp was corrupted part, it will result
  3756. * in a hung connection since we will drop all
  3757. * future packets due to the PAWS test.
  3758. */
  3759. }
  3760. if (len <= tcp_header_len) {
  3761. /* Bulk data transfer: sender */
  3762. if (len == tcp_header_len) {
  3763. /* Predicted packet is in window by definition.
  3764. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3765. * Hence, check seq<=rcv_wup reduces to:
  3766. */
  3767. if (tcp_header_len ==
  3768. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3769. tp->rcv_nxt == tp->rcv_wup)
  3770. tcp_store_ts_recent(tp);
  3771. /* We know that such packets are checksummed
  3772. * on entry.
  3773. */
  3774. tcp_ack(sk, skb, 0);
  3775. __kfree_skb(skb);
  3776. tcp_data_snd_check(sk);
  3777. return 0;
  3778. } else { /* Header too small */
  3779. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3780. goto discard;
  3781. }
  3782. } else {
  3783. int eaten = 0;
  3784. int copied_early = 0;
  3785. if (tp->copied_seq == tp->rcv_nxt &&
  3786. len - tcp_header_len <= tp->ucopy.len) {
  3787. #ifdef CONFIG_NET_DMA
  3788. if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
  3789. copied_early = 1;
  3790. eaten = 1;
  3791. }
  3792. #endif
  3793. if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
  3794. __set_current_state(TASK_RUNNING);
  3795. if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
  3796. eaten = 1;
  3797. }
  3798. if (eaten) {
  3799. /* Predicted packet is in window by definition.
  3800. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3801. * Hence, check seq<=rcv_wup reduces to:
  3802. */
  3803. if (tcp_header_len ==
  3804. (sizeof(struct tcphdr) +
  3805. TCPOLEN_TSTAMP_ALIGNED) &&
  3806. tp->rcv_nxt == tp->rcv_wup)
  3807. tcp_store_ts_recent(tp);
  3808. tcp_rcv_rtt_measure_ts(sk, skb);
  3809. __skb_pull(skb, tcp_header_len);
  3810. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3811. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
  3812. }
  3813. if (copied_early)
  3814. tcp_cleanup_rbuf(sk, skb->len);
  3815. }
  3816. if (!eaten) {
  3817. if (tcp_checksum_complete_user(sk, skb))
  3818. goto csum_error;
  3819. /* Predicted packet is in window by definition.
  3820. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3821. * Hence, check seq<=rcv_wup reduces to:
  3822. */
  3823. if (tcp_header_len ==
  3824. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3825. tp->rcv_nxt == tp->rcv_wup)
  3826. tcp_store_ts_recent(tp);
  3827. tcp_rcv_rtt_measure_ts(sk, skb);
  3828. if ((int)skb->truesize > sk->sk_forward_alloc)
  3829. goto step5;
  3830. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
  3831. /* Bulk data transfer: receiver */
  3832. __skb_pull(skb,tcp_header_len);
  3833. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3834. sk_stream_set_owner_r(skb, sk);
  3835. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3836. }
  3837. tcp_event_data_recv(sk, skb);
  3838. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
  3839. /* Well, only one small jumplet in fast path... */
  3840. tcp_ack(sk, skb, FLAG_DATA);
  3841. tcp_data_snd_check(sk);
  3842. if (!inet_csk_ack_scheduled(sk))
  3843. goto no_ack;
  3844. }
  3845. __tcp_ack_snd_check(sk, 0);
  3846. no_ack:
  3847. #ifdef CONFIG_NET_DMA
  3848. if (copied_early)
  3849. __skb_queue_tail(&sk->sk_async_wait_queue, skb);
  3850. else
  3851. #endif
  3852. if (eaten)
  3853. __kfree_skb(skb);
  3854. else
  3855. sk->sk_data_ready(sk, 0);
  3856. return 0;
  3857. }
  3858. }
  3859. slow_path:
  3860. if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
  3861. goto csum_error;
  3862. /*
  3863. * RFC1323: H1. Apply PAWS check first.
  3864. */
  3865. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  3866. tcp_paws_discard(sk, skb)) {
  3867. if (!th->rst) {
  3868. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  3869. tcp_send_dupack(sk, skb);
  3870. goto discard;
  3871. }
  3872. /* Resets are accepted even if PAWS failed.
  3873. ts_recent update must be made after we are sure
  3874. that the packet is in window.
  3875. */
  3876. }
  3877. /*
  3878. * Standard slow path.
  3879. */
  3880. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  3881. /* RFC793, page 37: "In all states except SYN-SENT, all reset
  3882. * (RST) segments are validated by checking their SEQ-fields."
  3883. * And page 69: "If an incoming segment is not acceptable,
  3884. * an acknowledgment should be sent in reply (unless the RST bit
  3885. * is set, if so drop the segment and return)".
  3886. */
  3887. if (!th->rst)
  3888. tcp_send_dupack(sk, skb);
  3889. goto discard;
  3890. }
  3891. if (th->rst) {
  3892. tcp_reset(sk);
  3893. goto discard;
  3894. }
  3895. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  3896. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3897. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3898. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  3899. tcp_reset(sk);
  3900. return 1;
  3901. }
  3902. step5:
  3903. if (th->ack)
  3904. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3905. tcp_rcv_rtt_measure_ts(sk, skb);
  3906. /* Process urgent data. */
  3907. tcp_urg(sk, skb, th);
  3908. /* step 7: process the segment text */
  3909. tcp_data_queue(sk, skb);
  3910. tcp_data_snd_check(sk);
  3911. tcp_ack_snd_check(sk);
  3912. return 0;
  3913. csum_error:
  3914. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3915. discard:
  3916. __kfree_skb(skb);
  3917. return 0;
  3918. }
  3919. static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
  3920. struct tcphdr *th, unsigned len)
  3921. {
  3922. struct tcp_sock *tp = tcp_sk(sk);
  3923. struct inet_connection_sock *icsk = inet_csk(sk);
  3924. int saved_clamp = tp->rx_opt.mss_clamp;
  3925. tcp_parse_options(skb, &tp->rx_opt, 0);
  3926. if (th->ack) {
  3927. /* rfc793:
  3928. * "If the state is SYN-SENT then
  3929. * first check the ACK bit
  3930. * If the ACK bit is set
  3931. * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
  3932. * a reset (unless the RST bit is set, if so drop
  3933. * the segment and return)"
  3934. *
  3935. * We do not send data with SYN, so that RFC-correct
  3936. * test reduces to:
  3937. */
  3938. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
  3939. goto reset_and_undo;
  3940. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  3941. !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
  3942. tcp_time_stamp)) {
  3943. NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
  3944. goto reset_and_undo;
  3945. }
  3946. /* Now ACK is acceptable.
  3947. *
  3948. * "If the RST bit is set
  3949. * If the ACK was acceptable then signal the user "error:
  3950. * connection reset", drop the segment, enter CLOSED state,
  3951. * delete TCB, and return."
  3952. */
  3953. if (th->rst) {
  3954. tcp_reset(sk);
  3955. goto discard;
  3956. }
  3957. /* rfc793:
  3958. * "fifth, if neither of the SYN or RST bits is set then
  3959. * drop the segment and return."
  3960. *
  3961. * See note below!
  3962. * --ANK(990513)
  3963. */
  3964. if (!th->syn)
  3965. goto discard_and_undo;
  3966. /* rfc793:
  3967. * "If the SYN bit is on ...
  3968. * are acceptable then ...
  3969. * (our SYN has been ACKed), change the connection
  3970. * state to ESTABLISHED..."
  3971. */
  3972. TCP_ECN_rcv_synack(tp, th);
  3973. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  3974. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3975. /* Ok.. it's good. Set up sequence numbers and
  3976. * move to established.
  3977. */
  3978. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  3979. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  3980. /* RFC1323: The window in SYN & SYN/ACK segments is
  3981. * never scaled.
  3982. */
  3983. tp->snd_wnd = ntohs(th->window);
  3984. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
  3985. if (!tp->rx_opt.wscale_ok) {
  3986. tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
  3987. tp->window_clamp = min(tp->window_clamp, 65535U);
  3988. }
  3989. if (tp->rx_opt.saw_tstamp) {
  3990. tp->rx_opt.tstamp_ok = 1;
  3991. tp->tcp_header_len =
  3992. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  3993. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  3994. tcp_store_ts_recent(tp);
  3995. } else {
  3996. tp->tcp_header_len = sizeof(struct tcphdr);
  3997. }
  3998. if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
  3999. tp->rx_opt.sack_ok |= 2;
  4000. tcp_mtup_init(sk);
  4001. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4002. tcp_initialize_rcv_mss(sk);
  4003. /* Remember, tcp_poll() does not lock socket!
  4004. * Change state from SYN-SENT only after copied_seq
  4005. * is initialized. */
  4006. tp->copied_seq = tp->rcv_nxt;
  4007. smp_mb();
  4008. tcp_set_state(sk, TCP_ESTABLISHED);
  4009. security_inet_conn_established(sk, skb);
  4010. /* Make sure socket is routed, for correct metrics. */
  4011. icsk->icsk_af_ops->rebuild_header(sk);
  4012. tcp_init_metrics(sk);
  4013. tcp_init_congestion_control(sk);
  4014. /* Prevent spurious tcp_cwnd_restart() on first data
  4015. * packet.
  4016. */
  4017. tp->lsndtime = tcp_time_stamp;
  4018. tcp_init_buffer_space(sk);
  4019. if (sock_flag(sk, SOCK_KEEPOPEN))
  4020. inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
  4021. if (!tp->rx_opt.snd_wscale)
  4022. __tcp_fast_path_on(tp, tp->snd_wnd);
  4023. else
  4024. tp->pred_flags = 0;
  4025. if (!sock_flag(sk, SOCK_DEAD)) {
  4026. sk->sk_state_change(sk);
  4027. sk_wake_async(sk, 0, POLL_OUT);
  4028. }
  4029. if (sk->sk_write_pending ||
  4030. icsk->icsk_accept_queue.rskq_defer_accept ||
  4031. icsk->icsk_ack.pingpong) {
  4032. /* Save one ACK. Data will be ready after
  4033. * several ticks, if write_pending is set.
  4034. *
  4035. * It may be deleted, but with this feature tcpdumps
  4036. * look so _wonderfully_ clever, that I was not able
  4037. * to stand against the temptation 8) --ANK
  4038. */
  4039. inet_csk_schedule_ack(sk);
  4040. icsk->icsk_ack.lrcvtime = tcp_time_stamp;
  4041. icsk->icsk_ack.ato = TCP_ATO_MIN;
  4042. tcp_incr_quickack(sk);
  4043. tcp_enter_quickack_mode(sk);
  4044. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  4045. TCP_DELACK_MAX, TCP_RTO_MAX);
  4046. discard:
  4047. __kfree_skb(skb);
  4048. return 0;
  4049. } else {
  4050. tcp_send_ack(sk);
  4051. }
  4052. return -1;
  4053. }
  4054. /* No ACK in the segment */
  4055. if (th->rst) {
  4056. /* rfc793:
  4057. * "If the RST bit is set
  4058. *
  4059. * Otherwise (no ACK) drop the segment and return."
  4060. */
  4061. goto discard_and_undo;
  4062. }
  4063. /* PAWS check. */
  4064. if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
  4065. goto discard_and_undo;
  4066. if (th->syn) {
  4067. /* We see SYN without ACK. It is attempt of
  4068. * simultaneous connect with crossed SYNs.
  4069. * Particularly, it can be connect to self.
  4070. */
  4071. tcp_set_state(sk, TCP_SYN_RECV);
  4072. if (tp->rx_opt.saw_tstamp) {
  4073. tp->rx_opt.tstamp_ok = 1;
  4074. tcp_store_ts_recent(tp);
  4075. tp->tcp_header_len =
  4076. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  4077. } else {
  4078. tp->tcp_header_len = sizeof(struct tcphdr);
  4079. }
  4080. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  4081. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  4082. /* RFC1323: The window in SYN & SYN/ACK segments is
  4083. * never scaled.
  4084. */
  4085. tp->snd_wnd = ntohs(th->window);
  4086. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  4087. tp->max_window = tp->snd_wnd;
  4088. TCP_ECN_rcv_syn(tp, th);
  4089. tcp_mtup_init(sk);
  4090. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  4091. tcp_initialize_rcv_mss(sk);
  4092. tcp_send_synack(sk);
  4093. #if 0
  4094. /* Note, we could accept data and URG from this segment.
  4095. * There are no obstacles to make this.
  4096. *
  4097. * However, if we ignore data in ACKless segments sometimes,
  4098. * we have no reasons to accept it sometimes.
  4099. * Also, seems the code doing it in step6 of tcp_rcv_state_process
  4100. * is not flawless. So, discard packet for sanity.
  4101. * Uncomment this return to process the data.
  4102. */
  4103. return -1;
  4104. #else
  4105. goto discard;
  4106. #endif
  4107. }
  4108. /* "fifth, if neither of the SYN or RST bits is set then
  4109. * drop the segment and return."
  4110. */
  4111. discard_and_undo:
  4112. tcp_clear_options(&tp->rx_opt);
  4113. tp->rx_opt.mss_clamp = saved_clamp;
  4114. goto discard;
  4115. reset_and_undo:
  4116. tcp_clear_options(&tp->rx_opt);
  4117. tp->rx_opt.mss_clamp = saved_clamp;
  4118. return 1;
  4119. }
  4120. /*
  4121. * This function implements the receiving procedure of RFC 793 for
  4122. * all states except ESTABLISHED and TIME_WAIT.
  4123. * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
  4124. * address independent.
  4125. */
  4126. int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
  4127. struct tcphdr *th, unsigned len)
  4128. {
  4129. struct tcp_sock *tp = tcp_sk(sk);
  4130. struct inet_connection_sock *icsk = inet_csk(sk);
  4131. int queued = 0;
  4132. tp->rx_opt.saw_tstamp = 0;
  4133. switch (sk->sk_state) {
  4134. case TCP_CLOSE:
  4135. goto discard;
  4136. case TCP_LISTEN:
  4137. if (th->ack)
  4138. return 1;
  4139. if (th->rst)
  4140. goto discard;
  4141. if (th->syn) {
  4142. if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
  4143. return 1;
  4144. /* Now we have several options: In theory there is
  4145. * nothing else in the frame. KA9Q has an option to
  4146. * send data with the syn, BSD accepts data with the
  4147. * syn up to the [to be] advertised window and
  4148. * Solaris 2.1 gives you a protocol error. For now
  4149. * we just ignore it, that fits the spec precisely
  4150. * and avoids incompatibilities. It would be nice in
  4151. * future to drop through and process the data.
  4152. *
  4153. * Now that TTCP is starting to be used we ought to
  4154. * queue this data.
  4155. * But, this leaves one open to an easy denial of
  4156. * service attack, and SYN cookies can't defend
  4157. * against this problem. So, we drop the data
  4158. * in the interest of security over speed unless
  4159. * it's still in use.
  4160. */
  4161. kfree_skb(skb);
  4162. return 0;
  4163. }
  4164. goto discard;
  4165. case TCP_SYN_SENT:
  4166. queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
  4167. if (queued >= 0)
  4168. return queued;
  4169. /* Do step6 onward by hand. */
  4170. tcp_urg(sk, skb, th);
  4171. __kfree_skb(skb);
  4172. tcp_data_snd_check(sk);
  4173. return 0;
  4174. }
  4175. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  4176. tcp_paws_discard(sk, skb)) {
  4177. if (!th->rst) {
  4178. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  4179. tcp_send_dupack(sk, skb);
  4180. goto discard;
  4181. }
  4182. /* Reset is accepted even if it did not pass PAWS. */
  4183. }
  4184. /* step 1: check sequence number */
  4185. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  4186. if (!th->rst)
  4187. tcp_send_dupack(sk, skb);
  4188. goto discard;
  4189. }
  4190. /* step 2: check RST bit */
  4191. if (th->rst) {
  4192. tcp_reset(sk);
  4193. goto discard;
  4194. }
  4195. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  4196. /* step 3: check security and precedence [ignored] */
  4197. /* step 4:
  4198. *
  4199. * Check for a SYN in window.
  4200. */
  4201. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  4202. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  4203. tcp_reset(sk);
  4204. return 1;
  4205. }
  4206. /* step 5: check the ACK field */
  4207. if (th->ack) {
  4208. int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
  4209. switch (sk->sk_state) {
  4210. case TCP_SYN_RECV:
  4211. if (acceptable) {
  4212. tp->copied_seq = tp->rcv_nxt;
  4213. smp_mb();
  4214. tcp_set_state(sk, TCP_ESTABLISHED);
  4215. sk->sk_state_change(sk);
  4216. /* Note, that this wakeup is only for marginal
  4217. * crossed SYN case. Passively open sockets
  4218. * are not waked up, because sk->sk_sleep ==
  4219. * NULL and sk->sk_socket == NULL.
  4220. */
  4221. if (sk->sk_socket) {
  4222. sk_wake_async(sk,0,POLL_OUT);
  4223. }
  4224. tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
  4225. tp->snd_wnd = ntohs(th->window) <<
  4226. tp->rx_opt.snd_wscale;
  4227. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
  4228. TCP_SKB_CB(skb)->seq);
  4229. /* tcp_ack considers this ACK as duplicate
  4230. * and does not calculate rtt.
  4231. * Fix it at least with timestamps.
  4232. */
  4233. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  4234. !tp->srtt)
  4235. tcp_ack_saw_tstamp(sk, 0);
  4236. if (tp->rx_opt.tstamp_ok)
  4237. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4238. /* Make sure socket is routed, for
  4239. * correct metrics.
  4240. */
  4241. icsk->icsk_af_ops->rebuild_header(sk);
  4242. tcp_init_metrics(sk);
  4243. tcp_init_congestion_control(sk);
  4244. /* Prevent spurious tcp_cwnd_restart() on
  4245. * first data packet.
  4246. */
  4247. tp->lsndtime = tcp_time_stamp;
  4248. tcp_mtup_init(sk);
  4249. tcp_initialize_rcv_mss(sk);
  4250. tcp_init_buffer_space(sk);
  4251. tcp_fast_path_on(tp);
  4252. } else {
  4253. return 1;
  4254. }
  4255. break;
  4256. case TCP_FIN_WAIT1:
  4257. if (tp->snd_una == tp->write_seq) {
  4258. tcp_set_state(sk, TCP_FIN_WAIT2);
  4259. sk->sk_shutdown |= SEND_SHUTDOWN;
  4260. dst_confirm(sk->sk_dst_cache);
  4261. if (!sock_flag(sk, SOCK_DEAD))
  4262. /* Wake up lingering close() */
  4263. sk->sk_state_change(sk);
  4264. else {
  4265. int tmo;
  4266. if (tp->linger2 < 0 ||
  4267. (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4268. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
  4269. tcp_done(sk);
  4270. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4271. return 1;
  4272. }
  4273. tmo = tcp_fin_time(sk);
  4274. if (tmo > TCP_TIMEWAIT_LEN) {
  4275. inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
  4276. } else if (th->fin || sock_owned_by_user(sk)) {
  4277. /* Bad case. We could lose such FIN otherwise.
  4278. * It is not a big problem, but it looks confusing
  4279. * and not so rare event. We still can lose it now,
  4280. * if it spins in bh_lock_sock(), but it is really
  4281. * marginal case.
  4282. */
  4283. inet_csk_reset_keepalive_timer(sk, tmo);
  4284. } else {
  4285. tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
  4286. goto discard;
  4287. }
  4288. }
  4289. }
  4290. break;
  4291. case TCP_CLOSING:
  4292. if (tp->snd_una == tp->write_seq) {
  4293. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  4294. goto discard;
  4295. }
  4296. break;
  4297. case TCP_LAST_ACK:
  4298. if (tp->snd_una == tp->write_seq) {
  4299. tcp_update_metrics(sk);
  4300. tcp_done(sk);
  4301. goto discard;
  4302. }
  4303. break;
  4304. }
  4305. } else
  4306. goto discard;
  4307. /* step 6: check the URG bit */
  4308. tcp_urg(sk, skb, th);
  4309. /* step 7: process the segment text */
  4310. switch (sk->sk_state) {
  4311. case TCP_CLOSE_WAIT:
  4312. case TCP_CLOSING:
  4313. case TCP_LAST_ACK:
  4314. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  4315. break;
  4316. case TCP_FIN_WAIT1:
  4317. case TCP_FIN_WAIT2:
  4318. /* RFC 793 says to queue data in these states,
  4319. * RFC 1122 says we MUST send a reset.
  4320. * BSD 4.4 also does reset.
  4321. */
  4322. if (sk->sk_shutdown & RCV_SHUTDOWN) {
  4323. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4324. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
  4325. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4326. tcp_reset(sk);
  4327. return 1;
  4328. }
  4329. }
  4330. /* Fall through */
  4331. case TCP_ESTABLISHED:
  4332. tcp_data_queue(sk, skb);
  4333. queued = 1;
  4334. break;
  4335. }
  4336. /* tcp_data could move socket to TIME-WAIT */
  4337. if (sk->sk_state != TCP_CLOSE) {
  4338. tcp_data_snd_check(sk);
  4339. tcp_ack_snd_check(sk);
  4340. }
  4341. if (!queued) {
  4342. discard:
  4343. __kfree_skb(skb);
  4344. }
  4345. return 0;
  4346. }
  4347. EXPORT_SYMBOL(sysctl_tcp_ecn);
  4348. EXPORT_SYMBOL(sysctl_tcp_reordering);
  4349. EXPORT_SYMBOL(tcp_parse_options);
  4350. EXPORT_SYMBOL(tcp_rcv_established);
  4351. EXPORT_SYMBOL(tcp_rcv_state_process);
  4352. EXPORT_SYMBOL(tcp_initialize_rcv_mss);