tcp_input.c 132 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138413941404141414241434144414541464147414841494150415141524153415441554156415741584159416041614162416341644165416641674168416941704171417241734174417541764177417841794180418141824183418441854186418741884189419041914192419341944195419641974198419942004201420242034204420542064207420842094210421142124213421442154216421742184219422042214222422342244225422642274228422942304231423242334234423542364237423842394240424142424243424442454246424742484249425042514252425342544255425642574258425942604261426242634264426542664267426842694270427142724273427442754276427742784279428042814282428342844285428642874288428942904291429242934294429542964297429842994300430143024303430443054306430743084309431043114312431343144315431643174318431943204321432243234324432543264327432843294330433143324333433443354336433743384339434043414342434343444345434643474348434943504351435243534354435543564357435843594360436143624363436443654366436743684369437043714372437343744375437643774378437943804381438243834384438543864387438843894390439143924393439443954396439743984399440044014402440344044405440644074408440944104411441244134414441544164417441844194420442144224423442444254426442744284429443044314432443344344435443644374438443944404441444244434444444544464447444844494450445144524453445444554456445744584459446044614462446344644465446644674468446944704471447244734474447544764477447844794480448144824483448444854486448744884489449044914492449344944495449644974498449945004501450245034504450545064507450845094510451145124513451445154516451745184519452045214522452345244525452645274528452945304531453245334534453545364537453845394540454145424543454445454546454745484549455045514552455345544555455645574558455945604561456245634564456545664567456845694570457145724573457445754576457745784579458045814582458345844585458645874588458945904591459245934594459545964597459845994600460146024603460446054606460746084609461046114612461346144615461646174618461946204621462246234624462546264627462846294630463146324633463446354636463746384639464046414642464346444645464646474648464946504651465246534654465546564657465846594660466146624663
  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 1.1.1.1 2007/06/12 07:27:14 eyryu 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_nometrics_save __read_mostly;
  86. int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
  87. int sysctl_tcp_abc __read_mostly;
  88. #define FLAG_DATA 0x01 /* Incoming frame contained data. */
  89. #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
  90. #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
  91. #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
  92. #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
  93. #define FLAG_DATA_SACKED 0x20 /* New SACK. */
  94. #define FLAG_ECE 0x40 /* ECE in this ACK */
  95. #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
  96. #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
  97. #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
  98. #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
  99. #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
  100. #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
  101. #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
  102. #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
  103. #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
  104. #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
  105. /* Adapt the MSS value used to make delayed ack decision to the
  106. * real world.
  107. */
  108. static void tcp_measure_rcv_mss(struct sock *sk,
  109. const struct sk_buff *skb)
  110. {
  111. struct inet_connection_sock *icsk = inet_csk(sk);
  112. const unsigned int lss = icsk->icsk_ack.last_seg_size;
  113. unsigned int len;
  114. icsk->icsk_ack.last_seg_size = 0;
  115. /* skb->len may jitter because of SACKs, even if peer
  116. * sends good full-sized frames.
  117. */
  118. len = skb_shinfo(skb)->gso_size ?: skb->len;
  119. if (len >= icsk->icsk_ack.rcv_mss) {
  120. icsk->icsk_ack.rcv_mss = len;
  121. } else {
  122. /* Otherwise, we make more careful check taking into account,
  123. * that SACKs block is variable.
  124. *
  125. * "len" is invariant segment length, including TCP header.
  126. */
  127. len += skb->data - skb->h.raw;
  128. if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
  129. /* If PSH is not set, packet should be
  130. * full sized, provided peer TCP is not badly broken.
  131. * This observation (if it is correct 8)) allows
  132. * to handle super-low mtu links fairly.
  133. */
  134. (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
  135. !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) {
  136. /* Subtract also invariant (if peer is RFC compliant),
  137. * tcp header plus fixed timestamp option length.
  138. * Resulting "len" is MSS free of SACK jitter.
  139. */
  140. len -= tcp_sk(sk)->tcp_header_len;
  141. icsk->icsk_ack.last_seg_size = len;
  142. if (len == lss) {
  143. icsk->icsk_ack.rcv_mss = len;
  144. return;
  145. }
  146. }
  147. if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
  148. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
  149. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
  150. }
  151. }
  152. static void tcp_incr_quickack(struct sock *sk)
  153. {
  154. struct inet_connection_sock *icsk = inet_csk(sk);
  155. unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
  156. if (quickacks==0)
  157. quickacks=2;
  158. if (quickacks > icsk->icsk_ack.quick)
  159. icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
  160. }
  161. void tcp_enter_quickack_mode(struct sock *sk)
  162. {
  163. struct inet_connection_sock *icsk = inet_csk(sk);
  164. tcp_incr_quickack(sk);
  165. icsk->icsk_ack.pingpong = 0;
  166. icsk->icsk_ack.ato = TCP_ATO_MIN;
  167. }
  168. /* Send ACKs quickly, if "quick" count is not exhausted
  169. * and the session is not interactive.
  170. */
  171. static inline int tcp_in_quickack_mode(const struct sock *sk)
  172. {
  173. const struct inet_connection_sock *icsk = inet_csk(sk);
  174. return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
  175. }
  176. /* Buffer size and advertised window tuning.
  177. *
  178. * 1. Tuning sk->sk_sndbuf, when connection enters established state.
  179. */
  180. static void tcp_fixup_sndbuf(struct sock *sk)
  181. {
  182. int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
  183. sizeof(struct sk_buff);
  184. if (sk->sk_sndbuf < 3 * sndmem)
  185. sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
  186. }
  187. /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
  188. *
  189. * All tcp_full_space() is split to two parts: "network" buffer, allocated
  190. * forward and advertised in receiver window (tp->rcv_wnd) and
  191. * "application buffer", required to isolate scheduling/application
  192. * latencies from network.
  193. * window_clamp is maximal advertised window. It can be less than
  194. * tcp_full_space(), in this case tcp_full_space() - window_clamp
  195. * is reserved for "application" buffer. The less window_clamp is
  196. * the smoother our behaviour from viewpoint of network, but the lower
  197. * throughput and the higher sensitivity of the connection to losses. 8)
  198. *
  199. * rcv_ssthresh is more strict window_clamp used at "slow start"
  200. * phase to predict further behaviour of this connection.
  201. * It is used for two goals:
  202. * - to enforce header prediction at sender, even when application
  203. * requires some significant "application buffer". It is check #1.
  204. * - to prevent pruning of receive queue because of misprediction
  205. * of receiver window. Check #2.
  206. *
  207. * The scheme does not work when sender sends good segments opening
  208. * window and then starts to feed us spaghetti. But it should work
  209. * in common situations. Otherwise, we have to rely on queue collapsing.
  210. */
  211. /* Slow part of check#2. */
  212. static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp,
  213. const struct sk_buff *skb)
  214. {
  215. /* Optimize this! */
  216. int truesize = tcp_win_from_space(skb->truesize)/2;
  217. int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
  218. while (tp->rcv_ssthresh <= window) {
  219. if (truesize <= skb->len)
  220. return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
  221. truesize >>= 1;
  222. window >>= 1;
  223. }
  224. return 0;
  225. }
  226. static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp,
  227. struct sk_buff *skb)
  228. {
  229. /* Check #1 */
  230. if (tp->rcv_ssthresh < tp->window_clamp &&
  231. (int)tp->rcv_ssthresh < tcp_space(sk) &&
  232. !tcp_memory_pressure) {
  233. int incr;
  234. /* Check #2. Increase window, if skb with such overhead
  235. * will fit to rcvbuf in future.
  236. */
  237. if (tcp_win_from_space(skb->truesize) <= skb->len)
  238. incr = 2*tp->advmss;
  239. else
  240. incr = __tcp_grow_window(sk, tp, skb);
  241. if (incr) {
  242. tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
  243. inet_csk(sk)->icsk_ack.quick |= 1;
  244. }
  245. }
  246. }
  247. /* 3. Tuning rcvbuf, when connection enters established state. */
  248. static void tcp_fixup_rcvbuf(struct sock *sk)
  249. {
  250. struct tcp_sock *tp = tcp_sk(sk);
  251. int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
  252. /* Try to select rcvbuf so that 4 mss-sized segments
  253. * will fit to window and corresponding skbs will fit to our rcvbuf.
  254. * (was 3; 4 is minimum to allow fast retransmit to work.)
  255. */
  256. while (tcp_win_from_space(rcvmem) < tp->advmss)
  257. rcvmem += 128;
  258. if (sk->sk_rcvbuf < 4 * rcvmem)
  259. sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
  260. }
  261. /* 4. Try to fixup all. It is made immediately after connection enters
  262. * established state.
  263. */
  264. static void tcp_init_buffer_space(struct sock *sk)
  265. {
  266. struct tcp_sock *tp = tcp_sk(sk);
  267. int maxwin;
  268. if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
  269. tcp_fixup_rcvbuf(sk);
  270. if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
  271. tcp_fixup_sndbuf(sk);
  272. tp->rcvq_space.space = tp->rcv_wnd;
  273. maxwin = tcp_full_space(sk);
  274. if (tp->window_clamp >= maxwin) {
  275. tp->window_clamp = maxwin;
  276. if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
  277. tp->window_clamp = max(maxwin -
  278. (maxwin >> sysctl_tcp_app_win),
  279. 4 * tp->advmss);
  280. }
  281. /* Force reservation of one segment. */
  282. if (sysctl_tcp_app_win &&
  283. tp->window_clamp > 2 * tp->advmss &&
  284. tp->window_clamp + tp->advmss > maxwin)
  285. tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
  286. tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
  287. tp->snd_cwnd_stamp = tcp_time_stamp;
  288. }
  289. /* 5. Recalculate window clamp after socket hit its memory bounds. */
  290. static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp)
  291. {
  292. struct inet_connection_sock *icsk = inet_csk(sk);
  293. icsk->icsk_ack.quick = 0;
  294. if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
  295. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
  296. !tcp_memory_pressure &&
  297. atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
  298. sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
  299. sysctl_tcp_rmem[2]);
  300. }
  301. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
  302. tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
  303. }
  304. /* Initialize RCV_MSS value.
  305. * RCV_MSS is an our guess about MSS used by the peer.
  306. * We haven't any direct information about the MSS.
  307. * It's better to underestimate the RCV_MSS rather than overestimate.
  308. * Overestimations make us ACKing less frequently than needed.
  309. * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
  310. */
  311. void tcp_initialize_rcv_mss(struct sock *sk)
  312. {
  313. struct tcp_sock *tp = tcp_sk(sk);
  314. unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
  315. hint = min(hint, tp->rcv_wnd/2);
  316. hint = min(hint, TCP_MIN_RCVMSS);
  317. hint = max(hint, TCP_MIN_MSS);
  318. inet_csk(sk)->icsk_ack.rcv_mss = hint;
  319. }
  320. /* Receiver "autotuning" code.
  321. *
  322. * The algorithm for RTT estimation w/o timestamps is based on
  323. * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
  324. * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
  325. *
  326. * More detail on this code can be found at
  327. * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
  328. * though this reference is out of date. A new paper
  329. * is pending.
  330. */
  331. static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
  332. {
  333. u32 new_sample = tp->rcv_rtt_est.rtt;
  334. long m = sample;
  335. if (m == 0)
  336. m = 1;
  337. if (new_sample != 0) {
  338. /* If we sample in larger samples in the non-timestamp
  339. * case, we could grossly overestimate the RTT especially
  340. * with chatty applications or bulk transfer apps which
  341. * are stalled on filesystem I/O.
  342. *
  343. * Also, since we are only going for a minimum in the
  344. * non-timestamp case, we do not smooth things out
  345. * else with timestamps disabled convergence takes too
  346. * long.
  347. */
  348. if (!win_dep) {
  349. m -= (new_sample >> 3);
  350. new_sample += m;
  351. } else if (m < new_sample)
  352. new_sample = m << 3;
  353. } else {
  354. /* No previous measure. */
  355. new_sample = m << 3;
  356. }
  357. if (tp->rcv_rtt_est.rtt != new_sample)
  358. tp->rcv_rtt_est.rtt = new_sample;
  359. }
  360. static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
  361. {
  362. if (tp->rcv_rtt_est.time == 0)
  363. goto new_measure;
  364. if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
  365. return;
  366. tcp_rcv_rtt_update(tp,
  367. jiffies - tp->rcv_rtt_est.time,
  368. 1);
  369. new_measure:
  370. tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
  371. tp->rcv_rtt_est.time = tcp_time_stamp;
  372. }
  373. static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
  374. {
  375. struct tcp_sock *tp = tcp_sk(sk);
  376. if (tp->rx_opt.rcv_tsecr &&
  377. (TCP_SKB_CB(skb)->end_seq -
  378. TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
  379. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
  380. }
  381. /*
  382. * This function should be called every time data is copied to user space.
  383. * It calculates the appropriate TCP receive buffer space.
  384. */
  385. void tcp_rcv_space_adjust(struct sock *sk)
  386. {
  387. struct tcp_sock *tp = tcp_sk(sk);
  388. int time;
  389. int space;
  390. if (tp->rcvq_space.time == 0)
  391. goto new_measure;
  392. time = tcp_time_stamp - tp->rcvq_space.time;
  393. if (time < (tp->rcv_rtt_est.rtt >> 3) ||
  394. tp->rcv_rtt_est.rtt == 0)
  395. return;
  396. space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
  397. space = max(tp->rcvq_space.space, space);
  398. if (tp->rcvq_space.space != space) {
  399. int rcvmem;
  400. tp->rcvq_space.space = space;
  401. if (sysctl_tcp_moderate_rcvbuf &&
  402. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
  403. int new_clamp = space;
  404. /* Receive space grows, normalize in order to
  405. * take into account packet headers and sk_buff
  406. * structure overhead.
  407. */
  408. space /= tp->advmss;
  409. if (!space)
  410. space = 1;
  411. rcvmem = (tp->advmss + MAX_TCP_HEADER +
  412. 16 + sizeof(struct sk_buff));
  413. while (tcp_win_from_space(rcvmem) < tp->advmss)
  414. rcvmem += 128;
  415. space *= rcvmem;
  416. space = min(space, sysctl_tcp_rmem[2]);
  417. if (space > sk->sk_rcvbuf) {
  418. sk->sk_rcvbuf = space;
  419. /* Make the window clamp follow along. */
  420. tp->window_clamp = new_clamp;
  421. }
  422. }
  423. }
  424. new_measure:
  425. tp->rcvq_space.seq = tp->copied_seq;
  426. tp->rcvq_space.time = tcp_time_stamp;
  427. }
  428. /* There is something which you must keep in mind when you analyze the
  429. * behavior of the tp->ato delayed ack timeout interval. When a
  430. * connection starts up, we want to ack as quickly as possible. The
  431. * problem is that "good" TCP's do slow start at the beginning of data
  432. * transmission. The means that until we send the first few ACK's the
  433. * sender will sit on his end and only queue most of his data, because
  434. * he can only send snd_cwnd unacked packets at any given time. For
  435. * each ACK we send, he increments snd_cwnd and transmits more of his
  436. * queue. -DaveM
  437. */
  438. static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
  439. {
  440. struct inet_connection_sock *icsk = inet_csk(sk);
  441. u32 now;
  442. inet_csk_schedule_ack(sk);
  443. tcp_measure_rcv_mss(sk, skb);
  444. tcp_rcv_rtt_measure(tp);
  445. now = tcp_time_stamp;
  446. if (!icsk->icsk_ack.ato) {
  447. /* The _first_ data packet received, initialize
  448. * delayed ACK engine.
  449. */
  450. tcp_incr_quickack(sk);
  451. icsk->icsk_ack.ato = TCP_ATO_MIN;
  452. } else {
  453. int m = now - icsk->icsk_ack.lrcvtime;
  454. if (m <= TCP_ATO_MIN/2) {
  455. /* The fastest case is the first. */
  456. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
  457. } else if (m < icsk->icsk_ack.ato) {
  458. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
  459. if (icsk->icsk_ack.ato > icsk->icsk_rto)
  460. icsk->icsk_ack.ato = icsk->icsk_rto;
  461. } else if (m > icsk->icsk_rto) {
  462. /* Too long gap. Apparently sender failed to
  463. * restart window, so that we send ACKs quickly.
  464. */
  465. tcp_incr_quickack(sk);
  466. sk_stream_mem_reclaim(sk);
  467. }
  468. }
  469. icsk->icsk_ack.lrcvtime = now;
  470. TCP_ECN_check_ce(tp, skb);
  471. if (skb->len >= 128)
  472. tcp_grow_window(sk, tp, skb);
  473. }
  474. /* Called to compute a smoothed rtt estimate. The data fed to this
  475. * routine either comes from timestamps, or from segments that were
  476. * known _not_ to have been retransmitted [see Karn/Partridge
  477. * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
  478. * piece by Van Jacobson.
  479. * NOTE: the next three routines used to be one big routine.
  480. * To save cycles in the RFC 1323 implementation it was better to break
  481. * it up into three procedures. -- erics
  482. */
  483. static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
  484. {
  485. struct tcp_sock *tp = tcp_sk(sk);
  486. long m = mrtt; /* RTT */
  487. /* The following amusing code comes from Jacobson's
  488. * article in SIGCOMM '88. Note that rtt and mdev
  489. * are scaled versions of rtt and mean deviation.
  490. * This is designed to be as fast as possible
  491. * m stands for "measurement".
  492. *
  493. * On a 1990 paper the rto value is changed to:
  494. * RTO = rtt + 4 * mdev
  495. *
  496. * Funny. This algorithm seems to be very broken.
  497. * These formulae increase RTO, when it should be decreased, increase
  498. * too slowly, when it should be increased quickly, decrease too quickly
  499. * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
  500. * does not matter how to _calculate_ it. Seems, it was trap
  501. * that VJ failed to avoid. 8)
  502. */
  503. if(m == 0)
  504. m = 1;
  505. if (tp->srtt != 0) {
  506. m -= (tp->srtt >> 3); /* m is now error in rtt est */
  507. tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
  508. if (m < 0) {
  509. m = -m; /* m is now abs(error) */
  510. m -= (tp->mdev >> 2); /* similar update on mdev */
  511. /* This is similar to one of Eifel findings.
  512. * Eifel blocks mdev updates when rtt decreases.
  513. * This solution is a bit different: we use finer gain
  514. * for mdev in this case (alpha*beta).
  515. * Like Eifel it also prevents growth of rto,
  516. * but also it limits too fast rto decreases,
  517. * happening in pure Eifel.
  518. */
  519. if (m > 0)
  520. m >>= 3;
  521. } else {
  522. m -= (tp->mdev >> 2); /* similar update on mdev */
  523. }
  524. tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
  525. if (tp->mdev > tp->mdev_max) {
  526. tp->mdev_max = tp->mdev;
  527. if (tp->mdev_max > tp->rttvar)
  528. tp->rttvar = tp->mdev_max;
  529. }
  530. if (after(tp->snd_una, tp->rtt_seq)) {
  531. if (tp->mdev_max < tp->rttvar)
  532. tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
  533. tp->rtt_seq = tp->snd_nxt;
  534. tp->mdev_max = TCP_RTO_MIN;
  535. }
  536. } else {
  537. /* no previous measure. */
  538. tp->srtt = m<<3; /* take the measured time to be rtt */
  539. tp->mdev = m<<1; /* make sure rto = 3*rtt */
  540. tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
  541. tp->rtt_seq = tp->snd_nxt;
  542. }
  543. }
  544. /* Calculate rto without backoff. This is the second half of Van Jacobson's
  545. * routine referred to above.
  546. */
  547. static inline void tcp_set_rto(struct sock *sk)
  548. {
  549. const struct tcp_sock *tp = tcp_sk(sk);
  550. /* Old crap is replaced with new one. 8)
  551. *
  552. * More seriously:
  553. * 1. If rtt variance happened to be less 50msec, it is hallucination.
  554. * It cannot be less due to utterly erratic ACK generation made
  555. * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
  556. * to do with delayed acks, because at cwnd>2 true delack timeout
  557. * is invisible. Actually, Linux-2.4 also generates erratic
  558. * ACKs in some circumstances.
  559. */
  560. inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
  561. /* 2. Fixups made earlier cannot be right.
  562. * If we do not estimate RTO correctly without them,
  563. * all the algo is pure shit and should be replaced
  564. * with correct one. It is exactly, which we pretend to do.
  565. */
  566. }
  567. /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
  568. * guarantees that rto is higher.
  569. */
  570. static inline void tcp_bound_rto(struct sock *sk)
  571. {
  572. if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
  573. inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
  574. }
  575. /* Save metrics learned by this TCP session.
  576. This function is called only, when TCP finishes successfully
  577. i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
  578. */
  579. void tcp_update_metrics(struct sock *sk)
  580. {
  581. struct tcp_sock *tp = tcp_sk(sk);
  582. struct dst_entry *dst = __sk_dst_get(sk);
  583. if (sysctl_tcp_nometrics_save)
  584. return;
  585. dst_confirm(dst);
  586. if (dst && (dst->flags&DST_HOST)) {
  587. const struct inet_connection_sock *icsk = inet_csk(sk);
  588. int m;
  589. if (icsk->icsk_backoff || !tp->srtt) {
  590. /* This session failed to estimate rtt. Why?
  591. * Probably, no packets returned in time.
  592. * Reset our results.
  593. */
  594. if (!(dst_metric_locked(dst, RTAX_RTT)))
  595. dst->metrics[RTAX_RTT-1] = 0;
  596. return;
  597. }
  598. m = dst_metric(dst, RTAX_RTT) - tp->srtt;
  599. /* If newly calculated rtt larger than stored one,
  600. * store new one. Otherwise, use EWMA. Remember,
  601. * rtt overestimation is always better than underestimation.
  602. */
  603. if (!(dst_metric_locked(dst, RTAX_RTT))) {
  604. if (m <= 0)
  605. dst->metrics[RTAX_RTT-1] = tp->srtt;
  606. else
  607. dst->metrics[RTAX_RTT-1] -= (m>>3);
  608. }
  609. if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
  610. if (m < 0)
  611. m = -m;
  612. /* Scale deviation to rttvar fixed point */
  613. m >>= 1;
  614. if (m < tp->mdev)
  615. m = tp->mdev;
  616. if (m >= dst_metric(dst, RTAX_RTTVAR))
  617. dst->metrics[RTAX_RTTVAR-1] = m;
  618. else
  619. dst->metrics[RTAX_RTTVAR-1] -=
  620. (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
  621. }
  622. if (tp->snd_ssthresh >= 0xFFFF) {
  623. /* Slow start still did not finish. */
  624. if (dst_metric(dst, RTAX_SSTHRESH) &&
  625. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  626. (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
  627. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
  628. if (!dst_metric_locked(dst, RTAX_CWND) &&
  629. tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
  630. dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
  631. } else if (tp->snd_cwnd > tp->snd_ssthresh &&
  632. icsk->icsk_ca_state == TCP_CA_Open) {
  633. /* Cong. avoidance phase, cwnd is reliable. */
  634. if (!dst_metric_locked(dst, RTAX_SSTHRESH))
  635. dst->metrics[RTAX_SSTHRESH-1] =
  636. max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
  637. if (!dst_metric_locked(dst, RTAX_CWND))
  638. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
  639. } else {
  640. /* Else slow start did not finish, cwnd is non-sense,
  641. ssthresh may be also invalid.
  642. */
  643. if (!dst_metric_locked(dst, RTAX_CWND))
  644. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
  645. if (dst->metrics[RTAX_SSTHRESH-1] &&
  646. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  647. tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
  648. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
  649. }
  650. if (!dst_metric_locked(dst, RTAX_REORDERING)) {
  651. if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
  652. tp->reordering != sysctl_tcp_reordering)
  653. dst->metrics[RTAX_REORDERING-1] = tp->reordering;
  654. }
  655. }
  656. }
  657. /* Numbers are taken from RFC2414. */
  658. __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
  659. {
  660. __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
  661. if (!cwnd) {
  662. if (tp->mss_cache > 1460)
  663. cwnd = 2;
  664. else
  665. cwnd = (tp->mss_cache > 1095) ? 3 : 4;
  666. }
  667. return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
  668. }
  669. /* Set slow start threshold and cwnd not falling to slow start */
  670. void tcp_enter_cwr(struct sock *sk)
  671. {
  672. struct tcp_sock *tp = tcp_sk(sk);
  673. tp->prior_ssthresh = 0;
  674. tp->bytes_acked = 0;
  675. if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
  676. tp->undo_marker = 0;
  677. tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
  678. tp->snd_cwnd = min(tp->snd_cwnd,
  679. tcp_packets_in_flight(tp) + 1U);
  680. tp->snd_cwnd_cnt = 0;
  681. tp->high_seq = tp->snd_nxt;
  682. tp->snd_cwnd_stamp = tcp_time_stamp;
  683. TCP_ECN_queue_cwr(tp);
  684. tcp_set_ca_state(sk, TCP_CA_CWR);
  685. }
  686. }
  687. /* Initialize metrics on socket. */
  688. static void tcp_init_metrics(struct sock *sk)
  689. {
  690. struct tcp_sock *tp = tcp_sk(sk);
  691. struct dst_entry *dst = __sk_dst_get(sk);
  692. if (dst == NULL)
  693. goto reset;
  694. dst_confirm(dst);
  695. if (dst_metric_locked(dst, RTAX_CWND))
  696. tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
  697. if (dst_metric(dst, RTAX_SSTHRESH)) {
  698. tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
  699. if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
  700. tp->snd_ssthresh = tp->snd_cwnd_clamp;
  701. }
  702. if (dst_metric(dst, RTAX_REORDERING) &&
  703. tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
  704. tp->rx_opt.sack_ok &= ~2;
  705. tp->reordering = dst_metric(dst, RTAX_REORDERING);
  706. }
  707. if (dst_metric(dst, RTAX_RTT) == 0)
  708. goto reset;
  709. if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
  710. goto reset;
  711. /* Initial rtt is determined from SYN,SYN-ACK.
  712. * The segment is small and rtt may appear much
  713. * less than real one. Use per-dst memory
  714. * to make it more realistic.
  715. *
  716. * A bit of theory. RTT is time passed after "normal" sized packet
  717. * is sent until it is ACKed. In normal circumstances sending small
  718. * packets force peer to delay ACKs and calculation is correct too.
  719. * The algorithm is adaptive and, provided we follow specs, it
  720. * NEVER underestimate RTT. BUT! If peer tries to make some clever
  721. * tricks sort of "quick acks" for time long enough to decrease RTT
  722. * to low value, and then abruptly stops to do it and starts to delay
  723. * ACKs, wait for troubles.
  724. */
  725. if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
  726. tp->srtt = dst_metric(dst, RTAX_RTT);
  727. tp->rtt_seq = tp->snd_nxt;
  728. }
  729. if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
  730. tp->mdev = dst_metric(dst, RTAX_RTTVAR);
  731. tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
  732. }
  733. tcp_set_rto(sk);
  734. tcp_bound_rto(sk);
  735. if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
  736. goto reset;
  737. tp->snd_cwnd = tcp_init_cwnd(tp, dst);
  738. tp->snd_cwnd_stamp = tcp_time_stamp;
  739. return;
  740. reset:
  741. /* Play conservative. If timestamps are not
  742. * supported, TCP will fail to recalculate correct
  743. * rtt, if initial rto is too small. FORGET ALL AND RESET!
  744. */
  745. if (!tp->rx_opt.saw_tstamp && tp->srtt) {
  746. tp->srtt = 0;
  747. tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
  748. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  749. }
  750. }
  751. static void tcp_update_reordering(struct sock *sk, const int metric,
  752. const int ts)
  753. {
  754. struct tcp_sock *tp = tcp_sk(sk);
  755. if (metric > tp->reordering) {
  756. tp->reordering = min(TCP_MAX_REORDERING, metric);
  757. /* This exciting event is worth to be remembered. 8) */
  758. if (ts)
  759. NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
  760. else if (IsReno(tp))
  761. NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
  762. else if (IsFack(tp))
  763. NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
  764. else
  765. NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
  766. #if FASTRETRANS_DEBUG > 1
  767. printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
  768. tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
  769. tp->reordering,
  770. tp->fackets_out,
  771. tp->sacked_out,
  772. tp->undo_marker ? tp->undo_retrans : 0);
  773. #endif
  774. /* Disable FACK yet. */
  775. tp->rx_opt.sack_ok &= ~2;
  776. }
  777. }
  778. /* This procedure tags the retransmission queue when SACKs arrive.
  779. *
  780. * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
  781. * Packets in queue with these bits set are counted in variables
  782. * sacked_out, retrans_out and lost_out, correspondingly.
  783. *
  784. * Valid combinations are:
  785. * Tag InFlight Description
  786. * 0 1 - orig segment is in flight.
  787. * S 0 - nothing flies, orig reached receiver.
  788. * L 0 - nothing flies, orig lost by net.
  789. * R 2 - both orig and retransmit are in flight.
  790. * L|R 1 - orig is lost, retransmit is in flight.
  791. * S|R 1 - orig reached receiver, retrans is still in flight.
  792. * (L|S|R is logically valid, it could occur when L|R is sacked,
  793. * but it is equivalent to plain S and code short-curcuits it to S.
  794. * L|S is logically invalid, it would mean -1 packet in flight 8))
  795. *
  796. * These 6 states form finite state machine, controlled by the following events:
  797. * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
  798. * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
  799. * 3. Loss detection event of one of three flavors:
  800. * A. Scoreboard estimator decided the packet is lost.
  801. * A'. Reno "three dupacks" marks head of queue lost.
  802. * A''. Its FACK modfication, head until snd.fack is lost.
  803. * B. SACK arrives sacking data transmitted after never retransmitted
  804. * hole was sent out.
  805. * C. SACK arrives sacking SND.NXT at the moment, when the
  806. * segment was retransmitted.
  807. * 4. D-SACK added new rule: D-SACK changes any tag to S.
  808. *
  809. * It is pleasant to note, that state diagram turns out to be commutative,
  810. * so that we are allowed not to be bothered by order of our actions,
  811. * when multiple events arrive simultaneously. (see the function below).
  812. *
  813. * Reordering detection.
  814. * --------------------
  815. * Reordering metric is maximal distance, which a packet can be displaced
  816. * in packet stream. With SACKs we can estimate it:
  817. *
  818. * 1. SACK fills old hole and the corresponding segment was not
  819. * ever retransmitted -> reordering. Alas, we cannot use it
  820. * when segment was retransmitted.
  821. * 2. The last flaw is solved with D-SACK. D-SACK arrives
  822. * for retransmitted and already SACKed segment -> reordering..
  823. * Both of these heuristics are not used in Loss state, when we cannot
  824. * account for retransmits accurately.
  825. */
  826. static int
  827. tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
  828. {
  829. const struct inet_connection_sock *icsk = inet_csk(sk);
  830. struct tcp_sock *tp = tcp_sk(sk);
  831. unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked;
  832. struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
  833. struct sk_buff *cached_skb;
  834. int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
  835. int reord = tp->packets_out;
  836. int prior_fackets;
  837. u32 lost_retrans = 0;
  838. int flag = 0;
  839. int dup_sack = 0;
  840. int cached_fack_count;
  841. int i;
  842. int first_sack_index;
  843. if (!tp->sacked_out)
  844. tp->fackets_out = 0;
  845. prior_fackets = tp->fackets_out;
  846. /* Check for D-SACK. */
  847. if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
  848. dup_sack = 1;
  849. tp->rx_opt.sack_ok |= 4;
  850. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
  851. } else if (num_sacks > 1 &&
  852. !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
  853. !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
  854. dup_sack = 1;
  855. tp->rx_opt.sack_ok |= 4;
  856. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
  857. }
  858. /* D-SACK for already forgotten data...
  859. * Do dumb counting. */
  860. if (dup_sack &&
  861. !after(ntohl(sp[0].end_seq), prior_snd_una) &&
  862. after(ntohl(sp[0].end_seq), tp->undo_marker))
  863. tp->undo_retrans--;
  864. /* Eliminate too old ACKs, but take into
  865. * account more or less fresh ones, they can
  866. * contain valid SACK info.
  867. */
  868. if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
  869. return 0;
  870. /* SACK fastpath:
  871. * if the only SACK change is the increase of the end_seq of
  872. * the first block then only apply that SACK block
  873. * and use retrans queue hinting otherwise slowpath */
  874. flag = 1;
  875. for (i = 0; i < num_sacks; i++) {
  876. __be32 start_seq = sp[i].start_seq;
  877. __be32 end_seq = sp[i].end_seq;
  878. if (i == 0) {
  879. if (tp->recv_sack_cache[i].start_seq != start_seq)
  880. flag = 0;
  881. } else {
  882. if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
  883. (tp->recv_sack_cache[i].end_seq != end_seq))
  884. flag = 0;
  885. }
  886. tp->recv_sack_cache[i].start_seq = start_seq;
  887. tp->recv_sack_cache[i].end_seq = end_seq;
  888. }
  889. /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
  890. for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
  891. tp->recv_sack_cache[i].start_seq = 0;
  892. tp->recv_sack_cache[i].end_seq = 0;
  893. }
  894. first_sack_index = 0;
  895. if (flag)
  896. num_sacks = 1;
  897. else {
  898. int j;
  899. tp->fastpath_skb_hint = NULL;
  900. /* order SACK blocks to allow in order walk of the retrans queue */
  901. for (i = num_sacks-1; i > 0; i--) {
  902. for (j = 0; j < i; j++){
  903. if (after(ntohl(sp[j].start_seq),
  904. ntohl(sp[j+1].start_seq))){
  905. struct tcp_sack_block_wire tmp;
  906. tmp = sp[j];
  907. sp[j] = sp[j+1];
  908. sp[j+1] = tmp;
  909. /* Track where the first SACK block goes to */
  910. if (j == first_sack_index)
  911. first_sack_index = j+1;
  912. }
  913. }
  914. }
  915. }
  916. /* clear flag as used for different purpose in following code */
  917. flag = 0;
  918. /* Use SACK fastpath hint if valid */
  919. cached_skb = tp->fastpath_skb_hint;
  920. cached_fack_count = tp->fastpath_cnt_hint;
  921. if (!cached_skb) {
  922. cached_skb = sk->sk_write_queue.next;
  923. cached_fack_count = 0;
  924. }
  925. for (i=0; i<num_sacks; i++, sp++) {
  926. struct sk_buff *skb;
  927. __u32 start_seq = ntohl(sp->start_seq);
  928. __u32 end_seq = ntohl(sp->end_seq);
  929. int fack_count;
  930. skb = cached_skb;
  931. fack_count = cached_fack_count;
  932. /* Event "B" in the comment above. */
  933. if (after(end_seq, tp->high_seq))
  934. flag |= FLAG_DATA_LOST;
  935. sk_stream_for_retrans_queue_from(skb, sk) {
  936. int in_sack, pcount;
  937. u8 sacked;
  938. cached_skb = skb;
  939. cached_fack_count = fack_count;
  940. if (i == first_sack_index) {
  941. tp->fastpath_skb_hint = skb;
  942. tp->fastpath_cnt_hint = fack_count;
  943. }
  944. /* The retransmission queue is always in order, so
  945. * we can short-circuit the walk early.
  946. */
  947. if (!before(TCP_SKB_CB(skb)->seq, end_seq))
  948. break;
  949. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  950. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  951. pcount = tcp_skb_pcount(skb);
  952. if (pcount > 1 && !in_sack &&
  953. after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
  954. unsigned int pkt_len;
  955. in_sack = !after(start_seq,
  956. TCP_SKB_CB(skb)->seq);
  957. if (!in_sack)
  958. pkt_len = (start_seq -
  959. TCP_SKB_CB(skb)->seq);
  960. else
  961. pkt_len = (end_seq -
  962. TCP_SKB_CB(skb)->seq);
  963. if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
  964. break;
  965. pcount = tcp_skb_pcount(skb);
  966. }
  967. fack_count += pcount;
  968. sacked = TCP_SKB_CB(skb)->sacked;
  969. /* Account D-SACK for retransmitted packet. */
  970. if ((dup_sack && in_sack) &&
  971. (sacked & TCPCB_RETRANS) &&
  972. after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
  973. tp->undo_retrans--;
  974. /* The frame is ACKed. */
  975. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
  976. if (sacked&TCPCB_RETRANS) {
  977. if ((dup_sack && in_sack) &&
  978. (sacked&TCPCB_SACKED_ACKED))
  979. reord = min(fack_count, reord);
  980. } else {
  981. /* If it was in a hole, we detected reordering. */
  982. if (fack_count < prior_fackets &&
  983. !(sacked&TCPCB_SACKED_ACKED))
  984. reord = min(fack_count, reord);
  985. }
  986. /* Nothing to do; acked frame is about to be dropped. */
  987. continue;
  988. }
  989. if ((sacked&TCPCB_SACKED_RETRANS) &&
  990. after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
  991. (!lost_retrans || after(end_seq, lost_retrans)))
  992. lost_retrans = end_seq;
  993. if (!in_sack)
  994. continue;
  995. if (!(sacked&TCPCB_SACKED_ACKED)) {
  996. if (sacked & TCPCB_SACKED_RETRANS) {
  997. /* If the segment is not tagged as lost,
  998. * we do not clear RETRANS, believing
  999. * that retransmission is still in flight.
  1000. */
  1001. if (sacked & TCPCB_LOST) {
  1002. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1003. tp->lost_out -= tcp_skb_pcount(skb);
  1004. tp->retrans_out -= tcp_skb_pcount(skb);
  1005. /* clear lost hint */
  1006. tp->retransmit_skb_hint = NULL;
  1007. }
  1008. } else {
  1009. /* New sack for not retransmitted frame,
  1010. * which was in hole. It is reordering.
  1011. */
  1012. if (!(sacked & TCPCB_RETRANS) &&
  1013. fack_count < prior_fackets)
  1014. reord = min(fack_count, reord);
  1015. if (sacked & TCPCB_LOST) {
  1016. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1017. tp->lost_out -= tcp_skb_pcount(skb);
  1018. /* clear lost hint */
  1019. tp->retransmit_skb_hint = NULL;
  1020. }
  1021. }
  1022. TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
  1023. flag |= FLAG_DATA_SACKED;
  1024. tp->sacked_out += tcp_skb_pcount(skb);
  1025. if (fack_count > tp->fackets_out)
  1026. tp->fackets_out = fack_count;
  1027. } else {
  1028. if (dup_sack && (sacked&TCPCB_RETRANS))
  1029. reord = min(fack_count, reord);
  1030. }
  1031. /* D-SACK. We can detect redundant retransmission
  1032. * in S|R and plain R frames and clear it.
  1033. * undo_retrans is decreased above, L|R frames
  1034. * are accounted above as well.
  1035. */
  1036. if (dup_sack &&
  1037. (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
  1038. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1039. tp->retrans_out -= tcp_skb_pcount(skb);
  1040. tp->retransmit_skb_hint = NULL;
  1041. }
  1042. }
  1043. }
  1044. /* Check for lost retransmit. This superb idea is
  1045. * borrowed from "ratehalving". Event "C".
  1046. * Later note: FACK people cheated me again 8),
  1047. * we have to account for reordering! Ugly,
  1048. * but should help.
  1049. */
  1050. if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
  1051. struct sk_buff *skb;
  1052. sk_stream_for_retrans_queue(skb, sk) {
  1053. if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
  1054. break;
  1055. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1056. continue;
  1057. if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
  1058. after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
  1059. (IsFack(tp) ||
  1060. !before(lost_retrans,
  1061. TCP_SKB_CB(skb)->ack_seq + tp->reordering *
  1062. tp->mss_cache))) {
  1063. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1064. tp->retrans_out -= tcp_skb_pcount(skb);
  1065. /* clear lost hint */
  1066. tp->retransmit_skb_hint = NULL;
  1067. if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  1068. tp->lost_out += tcp_skb_pcount(skb);
  1069. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1070. flag |= FLAG_DATA_SACKED;
  1071. NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
  1072. }
  1073. }
  1074. }
  1075. }
  1076. tp->left_out = tp->sacked_out + tp->lost_out;
  1077. if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss)
  1078. tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
  1079. #if FASTRETRANS_DEBUG > 0
  1080. BUG_TRAP((int)tp->sacked_out >= 0);
  1081. BUG_TRAP((int)tp->lost_out >= 0);
  1082. BUG_TRAP((int)tp->retrans_out >= 0);
  1083. BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
  1084. #endif
  1085. return flag;
  1086. }
  1087. /* RTO occurred, but do not yet enter loss state. Instead, transmit two new
  1088. * segments to see from the next ACKs whether any data was really missing.
  1089. * If the RTO was spurious, new ACKs should arrive.
  1090. */
  1091. void tcp_enter_frto(struct sock *sk)
  1092. {
  1093. const struct inet_connection_sock *icsk = inet_csk(sk);
  1094. struct tcp_sock *tp = tcp_sk(sk);
  1095. struct sk_buff *skb;
  1096. tp->frto_counter = 1;
  1097. if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
  1098. tp->snd_una == tp->high_seq ||
  1099. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1100. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1101. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1102. tcp_ca_event(sk, CA_EVENT_FRTO);
  1103. }
  1104. /* Have to clear retransmission markers here to keep the bookkeeping
  1105. * in shape, even though we are not yet in Loss state.
  1106. * If something was really lost, it is eventually caught up
  1107. * in tcp_enter_frto_loss.
  1108. */
  1109. tp->retrans_out = 0;
  1110. tp->undo_marker = tp->snd_una;
  1111. tp->undo_retrans = 0;
  1112. sk_stream_for_retrans_queue(skb, sk) {
  1113. TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS;
  1114. }
  1115. tcp_sync_left_out(tp);
  1116. tcp_set_ca_state(sk, TCP_CA_Open);
  1117. tp->frto_highmark = tp->snd_nxt;
  1118. }
  1119. /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
  1120. * which indicates that we should follow the traditional RTO recovery,
  1121. * i.e. mark everything lost and do go-back-N retransmission.
  1122. */
  1123. static void tcp_enter_frto_loss(struct sock *sk)
  1124. {
  1125. struct tcp_sock *tp = tcp_sk(sk);
  1126. struct sk_buff *skb;
  1127. int cnt = 0;
  1128. tp->sacked_out = 0;
  1129. tp->lost_out = 0;
  1130. tp->fackets_out = 0;
  1131. sk_stream_for_retrans_queue(skb, sk) {
  1132. cnt += tcp_skb_pcount(skb);
  1133. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1134. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
  1135. /* Do not mark those segments lost that were
  1136. * forward transmitted after RTO
  1137. */
  1138. if (!after(TCP_SKB_CB(skb)->end_seq,
  1139. tp->frto_highmark)) {
  1140. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1141. tp->lost_out += tcp_skb_pcount(skb);
  1142. }
  1143. } else {
  1144. tp->sacked_out += tcp_skb_pcount(skb);
  1145. tp->fackets_out = cnt;
  1146. }
  1147. }
  1148. tcp_sync_left_out(tp);
  1149. tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1;
  1150. tp->snd_cwnd_cnt = 0;
  1151. tp->snd_cwnd_stamp = tcp_time_stamp;
  1152. tp->undo_marker = 0;
  1153. tp->frto_counter = 0;
  1154. tp->reordering = min_t(unsigned int, tp->reordering,
  1155. sysctl_tcp_reordering);
  1156. tcp_set_ca_state(sk, TCP_CA_Loss);
  1157. tp->high_seq = tp->frto_highmark;
  1158. TCP_ECN_queue_cwr(tp);
  1159. clear_all_retrans_hints(tp);
  1160. }
  1161. void tcp_clear_retrans(struct tcp_sock *tp)
  1162. {
  1163. tp->left_out = 0;
  1164. tp->retrans_out = 0;
  1165. tp->fackets_out = 0;
  1166. tp->sacked_out = 0;
  1167. tp->lost_out = 0;
  1168. tp->undo_marker = 0;
  1169. tp->undo_retrans = 0;
  1170. }
  1171. /* Enter Loss state. If "how" is not zero, forget all SACK information
  1172. * and reset tags completely, otherwise preserve SACKs. If receiver
  1173. * dropped its ofo queue, we will know this due to reneging detection.
  1174. */
  1175. void tcp_enter_loss(struct sock *sk, int how)
  1176. {
  1177. const struct inet_connection_sock *icsk = inet_csk(sk);
  1178. struct tcp_sock *tp = tcp_sk(sk);
  1179. struct sk_buff *skb;
  1180. int cnt = 0;
  1181. /* Reduce ssthresh if it has not yet been made inside this window. */
  1182. if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
  1183. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1184. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1185. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1186. tcp_ca_event(sk, CA_EVENT_LOSS);
  1187. }
  1188. tp->snd_cwnd = 1;
  1189. tp->snd_cwnd_cnt = 0;
  1190. tp->snd_cwnd_stamp = tcp_time_stamp;
  1191. tp->bytes_acked = 0;
  1192. tcp_clear_retrans(tp);
  1193. /* Push undo marker, if it was plain RTO and nothing
  1194. * was retransmitted. */
  1195. if (!how)
  1196. tp->undo_marker = tp->snd_una;
  1197. sk_stream_for_retrans_queue(skb, sk) {
  1198. cnt += tcp_skb_pcount(skb);
  1199. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1200. tp->undo_marker = 0;
  1201. TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
  1202. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
  1203. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
  1204. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1205. tp->lost_out += tcp_skb_pcount(skb);
  1206. } else {
  1207. tp->sacked_out += tcp_skb_pcount(skb);
  1208. tp->fackets_out = cnt;
  1209. }
  1210. }
  1211. tcp_sync_left_out(tp);
  1212. tp->reordering = min_t(unsigned int, tp->reordering,
  1213. sysctl_tcp_reordering);
  1214. tcp_set_ca_state(sk, TCP_CA_Loss);
  1215. tp->high_seq = tp->snd_nxt;
  1216. TCP_ECN_queue_cwr(tp);
  1217. clear_all_retrans_hints(tp);
  1218. }
  1219. static int tcp_check_sack_reneging(struct sock *sk)
  1220. {
  1221. struct sk_buff *skb;
  1222. /* If ACK arrived pointing to a remembered SACK,
  1223. * it means that our remembered SACKs do not reflect
  1224. * real state of receiver i.e.
  1225. * receiver _host_ is heavily congested (or buggy).
  1226. * Do processing similar to RTO timeout.
  1227. */
  1228. if ((skb = skb_peek(&sk->sk_write_queue)) != NULL &&
  1229. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
  1230. struct inet_connection_sock *icsk = inet_csk(sk);
  1231. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
  1232. tcp_enter_loss(sk, 1);
  1233. icsk->icsk_retransmits++;
  1234. tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue));
  1235. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1236. icsk->icsk_rto, TCP_RTO_MAX);
  1237. return 1;
  1238. }
  1239. return 0;
  1240. }
  1241. static inline int tcp_fackets_out(struct tcp_sock *tp)
  1242. {
  1243. return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
  1244. }
  1245. static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
  1246. {
  1247. return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
  1248. }
  1249. static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp)
  1250. {
  1251. return tp->packets_out &&
  1252. tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue));
  1253. }
  1254. /* Linux NewReno/SACK/FACK/ECN state machine.
  1255. * --------------------------------------
  1256. *
  1257. * "Open" Normal state, no dubious events, fast path.
  1258. * "Disorder" In all the respects it is "Open",
  1259. * but requires a bit more attention. It is entered when
  1260. * we see some SACKs or dupacks. It is split of "Open"
  1261. * mainly to move some processing from fast path to slow one.
  1262. * "CWR" CWND was reduced due to some Congestion Notification event.
  1263. * It can be ECN, ICMP source quench, local device congestion.
  1264. * "Recovery" CWND was reduced, we are fast-retransmitting.
  1265. * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
  1266. *
  1267. * tcp_fastretrans_alert() is entered:
  1268. * - each incoming ACK, if state is not "Open"
  1269. * - when arrived ACK is unusual, namely:
  1270. * * SACK
  1271. * * Duplicate ACK.
  1272. * * ECN ECE.
  1273. *
  1274. * Counting packets in flight is pretty simple.
  1275. *
  1276. * in_flight = packets_out - left_out + retrans_out
  1277. *
  1278. * packets_out is SND.NXT-SND.UNA counted in packets.
  1279. *
  1280. * retrans_out is number of retransmitted segments.
  1281. *
  1282. * left_out is number of segments left network, but not ACKed yet.
  1283. *
  1284. * left_out = sacked_out + lost_out
  1285. *
  1286. * sacked_out: Packets, which arrived to receiver out of order
  1287. * and hence not ACKed. With SACKs this number is simply
  1288. * amount of SACKed data. Even without SACKs
  1289. * it is easy to give pretty reliable estimate of this number,
  1290. * counting duplicate ACKs.
  1291. *
  1292. * lost_out: Packets lost by network. TCP has no explicit
  1293. * "loss notification" feedback from network (for now).
  1294. * It means that this number can be only _guessed_.
  1295. * Actually, it is the heuristics to predict lossage that
  1296. * distinguishes different algorithms.
  1297. *
  1298. * F.e. after RTO, when all the queue is considered as lost,
  1299. * lost_out = packets_out and in_flight = retrans_out.
  1300. *
  1301. * Essentially, we have now two algorithms counting
  1302. * lost packets.
  1303. *
  1304. * FACK: It is the simplest heuristics. As soon as we decided
  1305. * that something is lost, we decide that _all_ not SACKed
  1306. * packets until the most forward SACK are lost. I.e.
  1307. * lost_out = fackets_out - sacked_out and left_out = fackets_out.
  1308. * It is absolutely correct estimate, if network does not reorder
  1309. * packets. And it loses any connection to reality when reordering
  1310. * takes place. We use FACK by default until reordering
  1311. * is suspected on the path to this destination.
  1312. *
  1313. * NewReno: when Recovery is entered, we assume that one segment
  1314. * is lost (classic Reno). While we are in Recovery and
  1315. * a partial ACK arrives, we assume that one more packet
  1316. * is lost (NewReno). This heuristics are the same in NewReno
  1317. * and SACK.
  1318. *
  1319. * Imagine, that's all! Forget about all this shamanism about CWND inflation
  1320. * deflation etc. CWND is real congestion window, never inflated, changes
  1321. * only according to classic VJ rules.
  1322. *
  1323. * Really tricky (and requiring careful tuning) part of algorithm
  1324. * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
  1325. * The first determines the moment _when_ we should reduce CWND and,
  1326. * hence, slow down forward transmission. In fact, it determines the moment
  1327. * when we decide that hole is caused by loss, rather than by a reorder.
  1328. *
  1329. * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
  1330. * holes, caused by lost packets.
  1331. *
  1332. * And the most logically complicated part of algorithm is undo
  1333. * heuristics. We detect false retransmits due to both too early
  1334. * fast retransmit (reordering) and underestimated RTO, analyzing
  1335. * timestamps and D-SACKs. When we detect that some segments were
  1336. * retransmitted by mistake and CWND reduction was wrong, we undo
  1337. * window reduction and abort recovery phase. This logic is hidden
  1338. * inside several functions named tcp_try_undo_<something>.
  1339. */
  1340. /* This function decides, when we should leave Disordered state
  1341. * and enter Recovery phase, reducing congestion window.
  1342. *
  1343. * Main question: may we further continue forward transmission
  1344. * with the same cwnd?
  1345. */
  1346. static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp)
  1347. {
  1348. __u32 packets_out;
  1349. /* Trick#1: The loss is proven. */
  1350. if (tp->lost_out)
  1351. return 1;
  1352. /* Not-A-Trick#2 : Classic rule... */
  1353. if (tcp_fackets_out(tp) > tp->reordering)
  1354. return 1;
  1355. /* Trick#3 : when we use RFC2988 timer restart, fast
  1356. * retransmit can be triggered by timeout of queue head.
  1357. */
  1358. if (tcp_head_timedout(sk, tp))
  1359. return 1;
  1360. /* Trick#4: It is still not OK... But will it be useful to delay
  1361. * recovery more?
  1362. */
  1363. packets_out = tp->packets_out;
  1364. if (packets_out <= tp->reordering &&
  1365. tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
  1366. !tcp_may_send_now(sk, tp)) {
  1367. /* We have nothing to send. This connection is limited
  1368. * either by receiver window or by application.
  1369. */
  1370. return 1;
  1371. }
  1372. return 0;
  1373. }
  1374. /* If we receive more dupacks than we expected counting segments
  1375. * in assumption of absent reordering, interpret this as reordering.
  1376. * The only another reason could be bug in receiver TCP.
  1377. */
  1378. static void tcp_check_reno_reordering(struct sock *sk, const int addend)
  1379. {
  1380. struct tcp_sock *tp = tcp_sk(sk);
  1381. u32 holes;
  1382. holes = max(tp->lost_out, 1U);
  1383. holes = min(holes, tp->packets_out);
  1384. if ((tp->sacked_out + holes) > tp->packets_out) {
  1385. tp->sacked_out = tp->packets_out - holes;
  1386. tcp_update_reordering(sk, tp->packets_out + addend, 0);
  1387. }
  1388. }
  1389. /* Emulate SACKs for SACKless connection: account for a new dupack. */
  1390. static void tcp_add_reno_sack(struct sock *sk)
  1391. {
  1392. struct tcp_sock *tp = tcp_sk(sk);
  1393. tp->sacked_out++;
  1394. tcp_check_reno_reordering(sk, 0);
  1395. tcp_sync_left_out(tp);
  1396. }
  1397. /* Account for ACK, ACKing some data in Reno Recovery phase. */
  1398. static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked)
  1399. {
  1400. if (acked > 0) {
  1401. /* One ACK acked hole. The rest eat duplicate ACKs. */
  1402. if (acked-1 >= tp->sacked_out)
  1403. tp->sacked_out = 0;
  1404. else
  1405. tp->sacked_out -= acked-1;
  1406. }
  1407. tcp_check_reno_reordering(sk, acked);
  1408. tcp_sync_left_out(tp);
  1409. }
  1410. static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
  1411. {
  1412. tp->sacked_out = 0;
  1413. tp->left_out = tp->lost_out;
  1414. }
  1415. /* Mark head of queue up as lost. */
  1416. static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp,
  1417. int packets, u32 high_seq)
  1418. {
  1419. struct sk_buff *skb;
  1420. int cnt;
  1421. BUG_TRAP(packets <= tp->packets_out);
  1422. if (tp->lost_skb_hint) {
  1423. skb = tp->lost_skb_hint;
  1424. cnt = tp->lost_cnt_hint;
  1425. } else {
  1426. skb = sk->sk_write_queue.next;
  1427. cnt = 0;
  1428. }
  1429. sk_stream_for_retrans_queue_from(skb, sk) {
  1430. /* TODO: do this better */
  1431. /* this is not the most efficient way to do this... */
  1432. tp->lost_skb_hint = skb;
  1433. tp->lost_cnt_hint = cnt;
  1434. cnt += tcp_skb_pcount(skb);
  1435. if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
  1436. break;
  1437. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1438. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1439. tp->lost_out += tcp_skb_pcount(skb);
  1440. /* clear xmit_retransmit_queue hints
  1441. * if this is beyond hint */
  1442. if(tp->retransmit_skb_hint != NULL &&
  1443. before(TCP_SKB_CB(skb)->seq,
  1444. TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) {
  1445. tp->retransmit_skb_hint = NULL;
  1446. }
  1447. }
  1448. }
  1449. tcp_sync_left_out(tp);
  1450. }
  1451. /* Account newly detected lost packet(s) */
  1452. static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp)
  1453. {
  1454. if (IsFack(tp)) {
  1455. int lost = tp->fackets_out - tp->reordering;
  1456. if (lost <= 0)
  1457. lost = 1;
  1458. tcp_mark_head_lost(sk, tp, lost, tp->high_seq);
  1459. } else {
  1460. tcp_mark_head_lost(sk, tp, 1, tp->high_seq);
  1461. }
  1462. /* New heuristics: it is possible only after we switched
  1463. * to restart timer each time when something is ACKed.
  1464. * Hence, we can detect timed out packets during fast
  1465. * retransmit without falling to slow start.
  1466. */
  1467. if (!IsReno(tp) && tcp_head_timedout(sk, tp)) {
  1468. struct sk_buff *skb;
  1469. skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
  1470. : sk->sk_write_queue.next;
  1471. sk_stream_for_retrans_queue_from(skb, sk) {
  1472. if (!tcp_skb_timedout(sk, skb))
  1473. break;
  1474. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1475. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1476. tp->lost_out += tcp_skb_pcount(skb);
  1477. /* clear xmit_retrans hint */
  1478. if (tp->retransmit_skb_hint &&
  1479. before(TCP_SKB_CB(skb)->seq,
  1480. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  1481. tp->retransmit_skb_hint = NULL;
  1482. }
  1483. }
  1484. tp->scoreboard_skb_hint = skb;
  1485. tcp_sync_left_out(tp);
  1486. }
  1487. }
  1488. /* CWND moderation, preventing bursts due to too big ACKs
  1489. * in dubious situations.
  1490. */
  1491. static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
  1492. {
  1493. tp->snd_cwnd = min(tp->snd_cwnd,
  1494. tcp_packets_in_flight(tp)+tcp_max_burst(tp));
  1495. tp->snd_cwnd_stamp = tcp_time_stamp;
  1496. }
  1497. /* Lower bound on congestion window is slow start threshold
  1498. * unless congestion avoidance choice decides to overide it.
  1499. */
  1500. static inline u32 tcp_cwnd_min(const struct sock *sk)
  1501. {
  1502. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1503. return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
  1504. }
  1505. /* Decrease cwnd each second ack. */
  1506. static void tcp_cwnd_down(struct sock *sk)
  1507. {
  1508. struct tcp_sock *tp = tcp_sk(sk);
  1509. int decr = tp->snd_cwnd_cnt + 1;
  1510. tp->snd_cwnd_cnt = decr&1;
  1511. decr >>= 1;
  1512. if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
  1513. tp->snd_cwnd -= decr;
  1514. tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
  1515. tp->snd_cwnd_stamp = tcp_time_stamp;
  1516. }
  1517. /* Nothing was retransmitted or returned timestamp is less
  1518. * than timestamp of the first retransmission.
  1519. */
  1520. static inline int tcp_packet_delayed(struct tcp_sock *tp)
  1521. {
  1522. return !tp->retrans_stamp ||
  1523. (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  1524. (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
  1525. }
  1526. /* Undo procedures. */
  1527. #if FASTRETRANS_DEBUG > 1
  1528. static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg)
  1529. {
  1530. struct inet_sock *inet = inet_sk(sk);
  1531. printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
  1532. msg,
  1533. NIPQUAD(inet->daddr), ntohs(inet->dport),
  1534. tp->snd_cwnd, tp->left_out,
  1535. tp->snd_ssthresh, tp->prior_ssthresh,
  1536. tp->packets_out);
  1537. }
  1538. #else
  1539. #define DBGUNDO(x...) do { } while (0)
  1540. #endif
  1541. static void tcp_undo_cwr(struct sock *sk, const int undo)
  1542. {
  1543. struct tcp_sock *tp = tcp_sk(sk);
  1544. if (tp->prior_ssthresh) {
  1545. const struct inet_connection_sock *icsk = inet_csk(sk);
  1546. if (icsk->icsk_ca_ops->undo_cwnd)
  1547. tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
  1548. else
  1549. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
  1550. if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
  1551. tp->snd_ssthresh = tp->prior_ssthresh;
  1552. TCP_ECN_withdraw_cwr(tp);
  1553. }
  1554. } else {
  1555. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
  1556. }
  1557. tcp_moderate_cwnd(tp);
  1558. tp->snd_cwnd_stamp = tcp_time_stamp;
  1559. /* There is something screwy going on with the retrans hints after
  1560. an undo */
  1561. clear_all_retrans_hints(tp);
  1562. }
  1563. static inline int tcp_may_undo(struct tcp_sock *tp)
  1564. {
  1565. return tp->undo_marker &&
  1566. (!tp->undo_retrans || tcp_packet_delayed(tp));
  1567. }
  1568. /* People celebrate: "We love our President!" */
  1569. static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp)
  1570. {
  1571. if (tcp_may_undo(tp)) {
  1572. /* Happy end! We did not retransmit anything
  1573. * or our original transmission succeeded.
  1574. */
  1575. DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
  1576. tcp_undo_cwr(sk, 1);
  1577. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
  1578. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1579. else
  1580. NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
  1581. tp->undo_marker = 0;
  1582. }
  1583. if (tp->snd_una == tp->high_seq && IsReno(tp)) {
  1584. /* Hold old state until something *above* high_seq
  1585. * is ACKed. For Reno it is MUST to prevent false
  1586. * fast retransmits (RFC2582). SACK TCP is safe. */
  1587. tcp_moderate_cwnd(tp);
  1588. return 1;
  1589. }
  1590. tcp_set_ca_state(sk, TCP_CA_Open);
  1591. return 0;
  1592. }
  1593. /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
  1594. static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp)
  1595. {
  1596. if (tp->undo_marker && !tp->undo_retrans) {
  1597. DBGUNDO(sk, tp, "D-SACK");
  1598. tcp_undo_cwr(sk, 1);
  1599. tp->undo_marker = 0;
  1600. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
  1601. }
  1602. }
  1603. /* Undo during fast recovery after partial ACK. */
  1604. static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp,
  1605. int acked)
  1606. {
  1607. /* Partial ACK arrived. Force Hoe's retransmit. */
  1608. int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
  1609. if (tcp_may_undo(tp)) {
  1610. /* Plain luck! Hole if filled with delayed
  1611. * packet, rather than with a retransmit.
  1612. */
  1613. if (tp->retrans_out == 0)
  1614. tp->retrans_stamp = 0;
  1615. tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
  1616. DBGUNDO(sk, tp, "Hoe");
  1617. tcp_undo_cwr(sk, 0);
  1618. NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
  1619. /* So... Do not make Hoe's retransmit yet.
  1620. * If the first packet was delayed, the rest
  1621. * ones are most probably delayed as well.
  1622. */
  1623. failed = 0;
  1624. }
  1625. return failed;
  1626. }
  1627. /* Undo during loss recovery after partial ACK. */
  1628. static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp)
  1629. {
  1630. if (tcp_may_undo(tp)) {
  1631. struct sk_buff *skb;
  1632. sk_stream_for_retrans_queue(skb, sk) {
  1633. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1634. }
  1635. clear_all_retrans_hints(tp);
  1636. DBGUNDO(sk, tp, "partial loss");
  1637. tp->lost_out = 0;
  1638. tp->left_out = tp->sacked_out;
  1639. tcp_undo_cwr(sk, 1);
  1640. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1641. inet_csk(sk)->icsk_retransmits = 0;
  1642. tp->undo_marker = 0;
  1643. if (!IsReno(tp))
  1644. tcp_set_ca_state(sk, TCP_CA_Open);
  1645. return 1;
  1646. }
  1647. return 0;
  1648. }
  1649. static inline void tcp_complete_cwr(struct sock *sk)
  1650. {
  1651. struct tcp_sock *tp = tcp_sk(sk);
  1652. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  1653. tp->snd_cwnd_stamp = tcp_time_stamp;
  1654. tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
  1655. }
  1656. static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag)
  1657. {
  1658. tp->left_out = tp->sacked_out;
  1659. if (tp->retrans_out == 0)
  1660. tp->retrans_stamp = 0;
  1661. if (flag&FLAG_ECE)
  1662. tcp_enter_cwr(sk);
  1663. if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
  1664. int state = TCP_CA_Open;
  1665. if (tp->left_out || tp->retrans_out || tp->undo_marker)
  1666. state = TCP_CA_Disorder;
  1667. if (inet_csk(sk)->icsk_ca_state != state) {
  1668. tcp_set_ca_state(sk, state);
  1669. tp->high_seq = tp->snd_nxt;
  1670. }
  1671. tcp_moderate_cwnd(tp);
  1672. } else {
  1673. tcp_cwnd_down(sk);
  1674. }
  1675. }
  1676. static void tcp_mtup_probe_failed(struct sock *sk)
  1677. {
  1678. struct inet_connection_sock *icsk = inet_csk(sk);
  1679. icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
  1680. icsk->icsk_mtup.probe_size = 0;
  1681. }
  1682. static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
  1683. {
  1684. struct tcp_sock *tp = tcp_sk(sk);
  1685. struct inet_connection_sock *icsk = inet_csk(sk);
  1686. /* FIXME: breaks with very large cwnd */
  1687. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1688. tp->snd_cwnd = tp->snd_cwnd *
  1689. tcp_mss_to_mtu(sk, tp->mss_cache) /
  1690. icsk->icsk_mtup.probe_size;
  1691. tp->snd_cwnd_cnt = 0;
  1692. tp->snd_cwnd_stamp = tcp_time_stamp;
  1693. tp->rcv_ssthresh = tcp_current_ssthresh(sk);
  1694. icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
  1695. icsk->icsk_mtup.probe_size = 0;
  1696. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  1697. }
  1698. /* Process an event, which can update packets-in-flight not trivially.
  1699. * Main goal of this function is to calculate new estimate for left_out,
  1700. * taking into account both packets sitting in receiver's buffer and
  1701. * packets lost by network.
  1702. *
  1703. * Besides that it does CWND reduction, when packet loss is detected
  1704. * and changes state of machine.
  1705. *
  1706. * It does _not_ decide what to send, it is made in function
  1707. * tcp_xmit_retransmit_queue().
  1708. */
  1709. static void
  1710. tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
  1711. int prior_packets, int flag)
  1712. {
  1713. struct inet_connection_sock *icsk = inet_csk(sk);
  1714. struct tcp_sock *tp = tcp_sk(sk);
  1715. int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
  1716. /* Some technical things:
  1717. * 1. Reno does not count dupacks (sacked_out) automatically. */
  1718. if (!tp->packets_out)
  1719. tp->sacked_out = 0;
  1720. /* 2. SACK counts snd_fack in packets inaccurately. */
  1721. if (tp->sacked_out == 0)
  1722. tp->fackets_out = 0;
  1723. /* Now state machine starts.
  1724. * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
  1725. if (flag&FLAG_ECE)
  1726. tp->prior_ssthresh = 0;
  1727. /* B. In all the states check for reneging SACKs. */
  1728. if (tp->sacked_out && tcp_check_sack_reneging(sk))
  1729. return;
  1730. /* C. Process data loss notification, provided it is valid. */
  1731. if ((flag&FLAG_DATA_LOST) &&
  1732. before(tp->snd_una, tp->high_seq) &&
  1733. icsk->icsk_ca_state != TCP_CA_Open &&
  1734. tp->fackets_out > tp->reordering) {
  1735. tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq);
  1736. NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
  1737. }
  1738. /* D. Synchronize left_out to current state. */
  1739. tcp_sync_left_out(tp);
  1740. /* E. Check state exit conditions. State can be terminated
  1741. * when high_seq is ACKed. */
  1742. if (icsk->icsk_ca_state == TCP_CA_Open) {
  1743. if (!sysctl_tcp_frto)
  1744. BUG_TRAP(tp->retrans_out == 0);
  1745. tp->retrans_stamp = 0;
  1746. } else if (!before(tp->snd_una, tp->high_seq)) {
  1747. switch (icsk->icsk_ca_state) {
  1748. case TCP_CA_Loss:
  1749. icsk->icsk_retransmits = 0;
  1750. if (tcp_try_undo_recovery(sk, tp))
  1751. return;
  1752. break;
  1753. case TCP_CA_CWR:
  1754. /* CWR is to be held something *above* high_seq
  1755. * is ACKed for CWR bit to reach receiver. */
  1756. if (tp->snd_una != tp->high_seq) {
  1757. tcp_complete_cwr(sk);
  1758. tcp_set_ca_state(sk, TCP_CA_Open);
  1759. }
  1760. break;
  1761. case TCP_CA_Disorder:
  1762. tcp_try_undo_dsack(sk, tp);
  1763. if (!tp->undo_marker ||
  1764. /* For SACK case do not Open to allow to undo
  1765. * catching for all duplicate ACKs. */
  1766. IsReno(tp) || tp->snd_una != tp->high_seq) {
  1767. tp->undo_marker = 0;
  1768. tcp_set_ca_state(sk, TCP_CA_Open);
  1769. }
  1770. break;
  1771. case TCP_CA_Recovery:
  1772. if (IsReno(tp))
  1773. tcp_reset_reno_sack(tp);
  1774. if (tcp_try_undo_recovery(sk, tp))
  1775. return;
  1776. tcp_complete_cwr(sk);
  1777. break;
  1778. }
  1779. }
  1780. /* F. Process state. */
  1781. switch (icsk->icsk_ca_state) {
  1782. case TCP_CA_Recovery:
  1783. if (prior_snd_una == tp->snd_una) {
  1784. if (IsReno(tp) && is_dupack)
  1785. tcp_add_reno_sack(sk);
  1786. } else {
  1787. int acked = prior_packets - tp->packets_out;
  1788. if (IsReno(tp))
  1789. tcp_remove_reno_sacks(sk, tp, acked);
  1790. is_dupack = tcp_try_undo_partial(sk, tp, acked);
  1791. }
  1792. break;
  1793. case TCP_CA_Loss:
  1794. if (flag&FLAG_DATA_ACKED)
  1795. icsk->icsk_retransmits = 0;
  1796. if (!tcp_try_undo_loss(sk, tp)) {
  1797. tcp_moderate_cwnd(tp);
  1798. tcp_xmit_retransmit_queue(sk);
  1799. return;
  1800. }
  1801. if (icsk->icsk_ca_state != TCP_CA_Open)
  1802. return;
  1803. /* Loss is undone; fall through to processing in Open state. */
  1804. default:
  1805. if (IsReno(tp)) {
  1806. if (tp->snd_una != prior_snd_una)
  1807. tcp_reset_reno_sack(tp);
  1808. if (is_dupack)
  1809. tcp_add_reno_sack(sk);
  1810. }
  1811. if (icsk->icsk_ca_state == TCP_CA_Disorder)
  1812. tcp_try_undo_dsack(sk, tp);
  1813. if (!tcp_time_to_recover(sk, tp)) {
  1814. tcp_try_to_open(sk, tp, flag);
  1815. return;
  1816. }
  1817. /* MTU probe failure: don't reduce cwnd */
  1818. if (icsk->icsk_ca_state < TCP_CA_CWR &&
  1819. icsk->icsk_mtup.probe_size &&
  1820. tp->snd_una == tp->mtu_probe.probe_seq_start) {
  1821. tcp_mtup_probe_failed(sk);
  1822. /* Restores the reduction we did in tcp_mtup_probe() */
  1823. tp->snd_cwnd++;
  1824. tcp_simple_retransmit(sk);
  1825. return;
  1826. }
  1827. /* Otherwise enter Recovery state */
  1828. if (IsReno(tp))
  1829. NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
  1830. else
  1831. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
  1832. tp->high_seq = tp->snd_nxt;
  1833. tp->prior_ssthresh = 0;
  1834. tp->undo_marker = tp->snd_una;
  1835. tp->undo_retrans = tp->retrans_out;
  1836. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  1837. if (!(flag&FLAG_ECE))
  1838. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1839. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1840. TCP_ECN_queue_cwr(tp);
  1841. }
  1842. tp->bytes_acked = 0;
  1843. tp->snd_cwnd_cnt = 0;
  1844. tcp_set_ca_state(sk, TCP_CA_Recovery);
  1845. }
  1846. if (is_dupack || tcp_head_timedout(sk, tp))
  1847. tcp_update_scoreboard(sk, tp);
  1848. tcp_cwnd_down(sk);
  1849. tcp_xmit_retransmit_queue(sk);
  1850. }
  1851. /* Read draft-ietf-tcplw-high-performance before mucking
  1852. * with this code. (Supersedes RFC1323)
  1853. */
  1854. static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
  1855. {
  1856. /* RTTM Rule: A TSecr value received in a segment is used to
  1857. * update the averaged RTT measurement only if the segment
  1858. * acknowledges some new data, i.e., only if it advances the
  1859. * left edge of the send window.
  1860. *
  1861. * See draft-ietf-tcplw-high-performance-00, section 3.3.
  1862. * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
  1863. *
  1864. * Changed: reset backoff as soon as we see the first valid sample.
  1865. * If we do not, we get strongly overestimated rto. With timestamps
  1866. * samples are accepted even from very old segments: f.e., when rtt=1
  1867. * increases to 8, we retransmit 5 times and after 8 seconds delayed
  1868. * answer arrives rto becomes 120 seconds! If at least one of segments
  1869. * in window is lost... Voila. --ANK (010210)
  1870. */
  1871. struct tcp_sock *tp = tcp_sk(sk);
  1872. const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
  1873. tcp_rtt_estimator(sk, seq_rtt);
  1874. tcp_set_rto(sk);
  1875. inet_csk(sk)->icsk_backoff = 0;
  1876. tcp_bound_rto(sk);
  1877. }
  1878. static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
  1879. {
  1880. /* We don't have a timestamp. Can only use
  1881. * packets that are not retransmitted to determine
  1882. * rtt estimates. Also, we must not reset the
  1883. * backoff for rto until we get a non-retransmitted
  1884. * packet. This allows us to deal with a situation
  1885. * where the network delay has increased suddenly.
  1886. * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
  1887. */
  1888. if (flag & FLAG_RETRANS_DATA_ACKED)
  1889. return;
  1890. tcp_rtt_estimator(sk, seq_rtt);
  1891. tcp_set_rto(sk);
  1892. inet_csk(sk)->icsk_backoff = 0;
  1893. tcp_bound_rto(sk);
  1894. }
  1895. static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
  1896. const s32 seq_rtt)
  1897. {
  1898. const struct tcp_sock *tp = tcp_sk(sk);
  1899. /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
  1900. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  1901. tcp_ack_saw_tstamp(sk, flag);
  1902. else if (seq_rtt >= 0)
  1903. tcp_ack_no_tstamp(sk, seq_rtt, flag);
  1904. }
  1905. static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
  1906. u32 in_flight, int good)
  1907. {
  1908. const struct inet_connection_sock *icsk = inet_csk(sk);
  1909. icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
  1910. tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
  1911. }
  1912. /* Restart timer after forward progress on connection.
  1913. * RFC2988 recommends to restart timer to now+rto.
  1914. */
  1915. static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp)
  1916. {
  1917. if (!tp->packets_out) {
  1918. inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
  1919. } else {
  1920. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  1921. }
  1922. }
  1923. static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
  1924. __u32 now, __s32 *seq_rtt)
  1925. {
  1926. struct tcp_sock *tp = tcp_sk(sk);
  1927. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  1928. __u32 seq = tp->snd_una;
  1929. __u32 packets_acked;
  1930. int acked = 0;
  1931. /* If we get here, the whole TSO packet has not been
  1932. * acked.
  1933. */
  1934. BUG_ON(!after(scb->end_seq, seq));
  1935. packets_acked = tcp_skb_pcount(skb);
  1936. if (tcp_trim_head(sk, skb, seq - scb->seq))
  1937. return 0;
  1938. packets_acked -= tcp_skb_pcount(skb);
  1939. if (packets_acked) {
  1940. __u8 sacked = scb->sacked;
  1941. acked |= FLAG_DATA_ACKED;
  1942. if (sacked) {
  1943. if (sacked & TCPCB_RETRANS) {
  1944. if (sacked & TCPCB_SACKED_RETRANS)
  1945. tp->retrans_out -= packets_acked;
  1946. acked |= FLAG_RETRANS_DATA_ACKED;
  1947. *seq_rtt = -1;
  1948. } else if (*seq_rtt < 0)
  1949. *seq_rtt = now - scb->when;
  1950. if (sacked & TCPCB_SACKED_ACKED)
  1951. tp->sacked_out -= packets_acked;
  1952. if (sacked & TCPCB_LOST)
  1953. tp->lost_out -= packets_acked;
  1954. if (sacked & TCPCB_URG) {
  1955. if (tp->urg_mode &&
  1956. !before(seq, tp->snd_up))
  1957. tp->urg_mode = 0;
  1958. }
  1959. } else if (*seq_rtt < 0)
  1960. *seq_rtt = now - scb->when;
  1961. if (tp->fackets_out) {
  1962. __u32 dval = min(tp->fackets_out, packets_acked);
  1963. tp->fackets_out -= dval;
  1964. }
  1965. tp->packets_out -= packets_acked;
  1966. BUG_ON(tcp_skb_pcount(skb) == 0);
  1967. BUG_ON(!before(scb->seq, scb->end_seq));
  1968. }
  1969. return acked;
  1970. }
  1971. static u32 tcp_usrtt(struct timeval *tv)
  1972. {
  1973. struct timeval now;
  1974. do_gettimeofday(&now);
  1975. return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
  1976. }
  1977. /* Remove acknowledged frames from the retransmission queue. */
  1978. static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
  1979. {
  1980. struct tcp_sock *tp = tcp_sk(sk);
  1981. const struct inet_connection_sock *icsk = inet_csk(sk);
  1982. struct sk_buff *skb;
  1983. __u32 now = tcp_time_stamp;
  1984. int acked = 0;
  1985. __s32 seq_rtt = -1;
  1986. u32 pkts_acked = 0;
  1987. void (*rtt_sample)(struct sock *sk, u32 usrtt)
  1988. = icsk->icsk_ca_ops->rtt_sample;
  1989. struct timeval tv = { .tv_sec = 0, .tv_usec = 0 };
  1990. while ((skb = skb_peek(&sk->sk_write_queue)) &&
  1991. skb != sk->sk_send_head) {
  1992. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  1993. __u8 sacked = scb->sacked;
  1994. /* If our packet is before the ack sequence we can
  1995. * discard it as it's confirmed to have arrived at
  1996. * the other end.
  1997. */
  1998. if (after(scb->end_seq, tp->snd_una)) {
  1999. if (tcp_skb_pcount(skb) > 1 &&
  2000. after(tp->snd_una, scb->seq))
  2001. acked |= tcp_tso_acked(sk, skb,
  2002. now, &seq_rtt);
  2003. break;
  2004. }
  2005. /* Initial outgoing SYN's get put onto the write_queue
  2006. * just like anything else we transmit. It is not
  2007. * true data, and if we misinform our callers that
  2008. * this ACK acks real data, we will erroneously exit
  2009. * connection startup slow start one packet too
  2010. * quickly. This is severely frowned upon behavior.
  2011. */
  2012. if (!(scb->flags & TCPCB_FLAG_SYN)) {
  2013. acked |= FLAG_DATA_ACKED;
  2014. ++pkts_acked;
  2015. } else {
  2016. acked |= FLAG_SYN_ACKED;
  2017. tp->retrans_stamp = 0;
  2018. }
  2019. /* MTU probing checks */
  2020. if (icsk->icsk_mtup.probe_size) {
  2021. if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
  2022. tcp_mtup_probe_success(sk, skb);
  2023. }
  2024. }
  2025. if (sacked) {
  2026. if (sacked & TCPCB_RETRANS) {
  2027. if(sacked & TCPCB_SACKED_RETRANS)
  2028. tp->retrans_out -= tcp_skb_pcount(skb);
  2029. acked |= FLAG_RETRANS_DATA_ACKED;
  2030. seq_rtt = -1;
  2031. } else if (seq_rtt < 0) {
  2032. seq_rtt = now - scb->when;
  2033. skb_get_timestamp(skb, &tv);
  2034. }
  2035. if (sacked & TCPCB_SACKED_ACKED)
  2036. tp->sacked_out -= tcp_skb_pcount(skb);
  2037. if (sacked & TCPCB_LOST)
  2038. tp->lost_out -= tcp_skb_pcount(skb);
  2039. if (sacked & TCPCB_URG) {
  2040. if (tp->urg_mode &&
  2041. !before(scb->end_seq, tp->snd_up))
  2042. tp->urg_mode = 0;
  2043. }
  2044. } else if (seq_rtt < 0) {
  2045. seq_rtt = now - scb->when;
  2046. skb_get_timestamp(skb, &tv);
  2047. }
  2048. tcp_dec_pcount_approx(&tp->fackets_out, skb);
  2049. tcp_packets_out_dec(tp, skb);
  2050. __skb_unlink(skb, &sk->sk_write_queue);
  2051. sk_stream_free_skb(sk, skb);
  2052. clear_all_retrans_hints(tp);
  2053. }
  2054. if (acked&FLAG_ACKED) {
  2055. tcp_ack_update_rtt(sk, acked, seq_rtt);
  2056. tcp_ack_packets_out(sk, tp);
  2057. if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
  2058. (*rtt_sample)(sk, tcp_usrtt(&tv));
  2059. if (icsk->icsk_ca_ops->pkts_acked)
  2060. icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
  2061. }
  2062. #if FASTRETRANS_DEBUG > 0
  2063. BUG_TRAP((int)tp->sacked_out >= 0);
  2064. BUG_TRAP((int)tp->lost_out >= 0);
  2065. BUG_TRAP((int)tp->retrans_out >= 0);
  2066. if (!tp->packets_out && tp->rx_opt.sack_ok) {
  2067. const struct inet_connection_sock *icsk = inet_csk(sk);
  2068. if (tp->lost_out) {
  2069. printk(KERN_DEBUG "Leak l=%u %d\n",
  2070. tp->lost_out, icsk->icsk_ca_state);
  2071. tp->lost_out = 0;
  2072. }
  2073. if (tp->sacked_out) {
  2074. printk(KERN_DEBUG "Leak s=%u %d\n",
  2075. tp->sacked_out, icsk->icsk_ca_state);
  2076. tp->sacked_out = 0;
  2077. }
  2078. if (tp->retrans_out) {
  2079. printk(KERN_DEBUG "Leak r=%u %d\n",
  2080. tp->retrans_out, icsk->icsk_ca_state);
  2081. tp->retrans_out = 0;
  2082. }
  2083. }
  2084. #endif
  2085. *seq_rtt_p = seq_rtt;
  2086. return acked;
  2087. }
  2088. static void tcp_ack_probe(struct sock *sk)
  2089. {
  2090. const struct tcp_sock *tp = tcp_sk(sk);
  2091. struct inet_connection_sock *icsk = inet_csk(sk);
  2092. /* Was it a usable window open? */
  2093. if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq,
  2094. tp->snd_una + tp->snd_wnd)) {
  2095. icsk->icsk_backoff = 0;
  2096. inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
  2097. /* Socket must be waked up by subsequent tcp_data_snd_check().
  2098. * This function is not for random using!
  2099. */
  2100. } else {
  2101. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2102. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2103. TCP_RTO_MAX);
  2104. }
  2105. }
  2106. static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
  2107. {
  2108. return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
  2109. inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
  2110. }
  2111. static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
  2112. {
  2113. const struct tcp_sock *tp = tcp_sk(sk);
  2114. return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
  2115. !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
  2116. }
  2117. /* Check that window update is acceptable.
  2118. * The function assumes that snd_una<=ack<=snd_next.
  2119. */
  2120. static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
  2121. const u32 ack_seq, const u32 nwin)
  2122. {
  2123. return (after(ack, tp->snd_una) ||
  2124. after(ack_seq, tp->snd_wl1) ||
  2125. (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
  2126. }
  2127. /* Update our send window.
  2128. *
  2129. * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
  2130. * and in FreeBSD. NetBSD's one is even worse.) is wrong.
  2131. */
  2132. static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp,
  2133. struct sk_buff *skb, u32 ack, u32 ack_seq)
  2134. {
  2135. int flag = 0;
  2136. u32 nwin = ntohs(skb->h.th->window);
  2137. if (likely(!skb->h.th->syn))
  2138. nwin <<= tp->rx_opt.snd_wscale;
  2139. if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
  2140. flag |= FLAG_WIN_UPDATE;
  2141. tcp_update_wl(tp, ack, ack_seq);
  2142. if (tp->snd_wnd != nwin) {
  2143. tp->snd_wnd = nwin;
  2144. /* Note, it is the only place, where
  2145. * fast path is recovered for sending TCP.
  2146. */
  2147. tp->pred_flags = 0;
  2148. tcp_fast_path_check(sk, tp);
  2149. if (nwin > tp->max_window) {
  2150. tp->max_window = nwin;
  2151. tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
  2152. }
  2153. }
  2154. }
  2155. tp->snd_una = ack;
  2156. return flag;
  2157. }
  2158. static void tcp_process_frto(struct sock *sk, u32 prior_snd_una)
  2159. {
  2160. struct tcp_sock *tp = tcp_sk(sk);
  2161. tcp_sync_left_out(tp);
  2162. if (tp->snd_una == prior_snd_una ||
  2163. !before(tp->snd_una, tp->frto_highmark)) {
  2164. /* RTO was caused by loss, start retransmitting in
  2165. * go-back-N slow start
  2166. */
  2167. tcp_enter_frto_loss(sk);
  2168. return;
  2169. }
  2170. if (tp->frto_counter == 1) {
  2171. /* First ACK after RTO advances the window: allow two new
  2172. * segments out.
  2173. */
  2174. tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
  2175. } else {
  2176. /* Also the second ACK after RTO advances the window.
  2177. * The RTO was likely spurious. Reduce cwnd and continue
  2178. * in congestion avoidance
  2179. */
  2180. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  2181. tcp_moderate_cwnd(tp);
  2182. }
  2183. /* F-RTO affects on two new ACKs following RTO.
  2184. * At latest on third ACK the TCP behavior is back to normal.
  2185. */
  2186. tp->frto_counter = (tp->frto_counter + 1) % 3;
  2187. }
  2188. /* This routine deals with incoming acks, but not outgoing ones. */
  2189. static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
  2190. {
  2191. struct inet_connection_sock *icsk = inet_csk(sk);
  2192. struct tcp_sock *tp = tcp_sk(sk);
  2193. u32 prior_snd_una = tp->snd_una;
  2194. u32 ack_seq = TCP_SKB_CB(skb)->seq;
  2195. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2196. u32 prior_in_flight;
  2197. s32 seq_rtt;
  2198. int prior_packets;
  2199. /* If the ack is newer than sent or older than previous acks
  2200. * then we can probably ignore it.
  2201. */
  2202. if (after(ack, tp->snd_nxt))
  2203. goto uninteresting_ack;
  2204. if (before(ack, prior_snd_una))
  2205. goto old_ack;
  2206. if (sysctl_tcp_abc) {
  2207. if (icsk->icsk_ca_state < TCP_CA_CWR)
  2208. tp->bytes_acked += ack - prior_snd_una;
  2209. else if (icsk->icsk_ca_state == TCP_CA_Loss)
  2210. /* we assume just one segment left network */
  2211. tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
  2212. }
  2213. if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
  2214. /* Window is constant, pure forward advance.
  2215. * No more checks are required.
  2216. * Note, we use the fact that SND.UNA>=SND.WL2.
  2217. */
  2218. tcp_update_wl(tp, ack, ack_seq);
  2219. tp->snd_una = ack;
  2220. flag |= FLAG_WIN_UPDATE;
  2221. tcp_ca_event(sk, CA_EVENT_FAST_ACK);
  2222. NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
  2223. } else {
  2224. if (ack_seq != TCP_SKB_CB(skb)->end_seq)
  2225. flag |= FLAG_DATA;
  2226. else
  2227. NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
  2228. flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq);
  2229. if (TCP_SKB_CB(skb)->sacked)
  2230. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2231. if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th))
  2232. flag |= FLAG_ECE;
  2233. tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
  2234. }
  2235. /* We passed data and got it acked, remove any soft error
  2236. * log. Something worked...
  2237. */
  2238. sk->sk_err_soft = 0;
  2239. tp->rcv_tstamp = tcp_time_stamp;
  2240. prior_packets = tp->packets_out;
  2241. if (!prior_packets)
  2242. goto no_queue;
  2243. prior_in_flight = tcp_packets_in_flight(tp);
  2244. /* See if we can take anything off of the retransmit queue. */
  2245. flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
  2246. if (tp->frto_counter)
  2247. tcp_process_frto(sk, prior_snd_una);
  2248. if (tcp_ack_is_dubious(sk, flag)) {
  2249. /* Advance CWND, if state allows this. */
  2250. if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag))
  2251. tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0);
  2252. tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
  2253. } else {
  2254. if ((flag & FLAG_DATA_ACKED))
  2255. tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
  2256. }
  2257. if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
  2258. dst_confirm(sk->sk_dst_cache);
  2259. return 1;
  2260. no_queue:
  2261. icsk->icsk_probes_out = 0;
  2262. /* If this ack opens up a zero window, clear backoff. It was
  2263. * being used to time the probes, and is probably far higher than
  2264. * it needs to be for normal retransmission.
  2265. */
  2266. if (sk->sk_send_head)
  2267. tcp_ack_probe(sk);
  2268. return 1;
  2269. old_ack:
  2270. if (TCP_SKB_CB(skb)->sacked)
  2271. tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2272. uninteresting_ack:
  2273. SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  2274. return 0;
  2275. }
  2276. /* Look for tcp options. Normally only called on SYN and SYNACK packets.
  2277. * But, this can also be called on packets in the established flow when
  2278. * the fast version below fails.
  2279. */
  2280. void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
  2281. {
  2282. unsigned char *ptr;
  2283. struct tcphdr *th = skb->h.th;
  2284. int length=(th->doff*4)-sizeof(struct tcphdr);
  2285. ptr = (unsigned char *)(th + 1);
  2286. opt_rx->saw_tstamp = 0;
  2287. while(length>0) {
  2288. int opcode=*ptr++;
  2289. int opsize;
  2290. switch (opcode) {
  2291. case TCPOPT_EOL:
  2292. return;
  2293. case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
  2294. length--;
  2295. continue;
  2296. default:
  2297. opsize=*ptr++;
  2298. if (opsize < 2) /* "silly options" */
  2299. return;
  2300. if (opsize > length)
  2301. return; /* don't parse partial options */
  2302. switch(opcode) {
  2303. case TCPOPT_MSS:
  2304. if(opsize==TCPOLEN_MSS && th->syn && !estab) {
  2305. u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
  2306. if (in_mss) {
  2307. if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
  2308. in_mss = opt_rx->user_mss;
  2309. opt_rx->mss_clamp = in_mss;
  2310. }
  2311. }
  2312. break;
  2313. case TCPOPT_WINDOW:
  2314. if(opsize==TCPOLEN_WINDOW && th->syn && !estab)
  2315. if (sysctl_tcp_window_scaling) {
  2316. __u8 snd_wscale = *(__u8 *) ptr;
  2317. opt_rx->wscale_ok = 1;
  2318. if (snd_wscale > 14) {
  2319. if(net_ratelimit())
  2320. printk(KERN_INFO "tcp_parse_options: Illegal window "
  2321. "scaling value %d >14 received.\n",
  2322. snd_wscale);
  2323. snd_wscale = 14;
  2324. }
  2325. opt_rx->snd_wscale = snd_wscale;
  2326. }
  2327. break;
  2328. case TCPOPT_TIMESTAMP:
  2329. if(opsize==TCPOLEN_TIMESTAMP) {
  2330. if ((estab && opt_rx->tstamp_ok) ||
  2331. (!estab && sysctl_tcp_timestamps)) {
  2332. opt_rx->saw_tstamp = 1;
  2333. opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
  2334. opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
  2335. }
  2336. }
  2337. break;
  2338. case TCPOPT_SACK_PERM:
  2339. if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
  2340. if (sysctl_tcp_sack) {
  2341. opt_rx->sack_ok = 1;
  2342. tcp_sack_reset(opt_rx);
  2343. }
  2344. }
  2345. break;
  2346. case TCPOPT_SACK:
  2347. if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
  2348. !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
  2349. opt_rx->sack_ok) {
  2350. TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
  2351. }
  2352. #ifdef CONFIG_TCP_MD5SIG
  2353. case TCPOPT_MD5SIG:
  2354. /*
  2355. * The MD5 Hash has already been
  2356. * checked (see tcp_v{4,6}_do_rcv()).
  2357. */
  2358. break;
  2359. #endif
  2360. };
  2361. ptr+=opsize-2;
  2362. length-=opsize;
  2363. };
  2364. }
  2365. }
  2366. /* Fast parse options. This hopes to only see timestamps.
  2367. * If it is wrong it falls back on tcp_parse_options().
  2368. */
  2369. static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
  2370. struct tcp_sock *tp)
  2371. {
  2372. if (th->doff == sizeof(struct tcphdr)>>2) {
  2373. tp->rx_opt.saw_tstamp = 0;
  2374. return 0;
  2375. } else if (tp->rx_opt.tstamp_ok &&
  2376. th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
  2377. __be32 *ptr = (__be32 *)(th + 1);
  2378. if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  2379. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
  2380. tp->rx_opt.saw_tstamp = 1;
  2381. ++ptr;
  2382. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  2383. ++ptr;
  2384. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  2385. return 1;
  2386. }
  2387. }
  2388. tcp_parse_options(skb, &tp->rx_opt, 1);
  2389. return 1;
  2390. }
  2391. static inline void tcp_store_ts_recent(struct tcp_sock *tp)
  2392. {
  2393. tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
  2394. tp->rx_opt.ts_recent_stamp = xtime.tv_sec;
  2395. }
  2396. static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
  2397. {
  2398. if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
  2399. /* PAWS bug workaround wrt. ACK frames, the PAWS discard
  2400. * extra check below makes sure this can only happen
  2401. * for pure ACK frames. -DaveM
  2402. *
  2403. * Not only, also it occurs for expired timestamps.
  2404. */
  2405. if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
  2406. xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
  2407. tcp_store_ts_recent(tp);
  2408. }
  2409. }
  2410. /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
  2411. *
  2412. * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
  2413. * it can pass through stack. So, the following predicate verifies that
  2414. * this segment is not used for anything but congestion avoidance or
  2415. * fast retransmit. Moreover, we even are able to eliminate most of such
  2416. * second order effects, if we apply some small "replay" window (~RTO)
  2417. * to timestamp space.
  2418. *
  2419. * All these measures still do not guarantee that we reject wrapped ACKs
  2420. * on networks with high bandwidth, when sequence space is recycled fastly,
  2421. * but it guarantees that such events will be very rare and do not affect
  2422. * connection seriously. This doesn't look nice, but alas, PAWS is really
  2423. * buggy extension.
  2424. *
  2425. * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
  2426. * states that events when retransmit arrives after original data are rare.
  2427. * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
  2428. * the biggest problem on large power networks even with minor reordering.
  2429. * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
  2430. * up to bandwidth of 18Gigabit/sec. 8) ]
  2431. */
  2432. static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
  2433. {
  2434. struct tcp_sock *tp = tcp_sk(sk);
  2435. struct tcphdr *th = skb->h.th;
  2436. u32 seq = TCP_SKB_CB(skb)->seq;
  2437. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2438. return (/* 1. Pure ACK with correct sequence number. */
  2439. (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
  2440. /* 2. ... and duplicate ACK. */
  2441. ack == tp->snd_una &&
  2442. /* 3. ... and does not update window. */
  2443. !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
  2444. /* 4. ... and sits in replay window. */
  2445. (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
  2446. }
  2447. static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
  2448. {
  2449. const struct tcp_sock *tp = tcp_sk(sk);
  2450. return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
  2451. xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
  2452. !tcp_disordered_ack(sk, skb));
  2453. }
  2454. /* Check segment sequence number for validity.
  2455. *
  2456. * Segment controls are considered valid, if the segment
  2457. * fits to the window after truncation to the window. Acceptability
  2458. * of data (and SYN, FIN, of course) is checked separately.
  2459. * See tcp_data_queue(), for example.
  2460. *
  2461. * Also, controls (RST is main one) are accepted using RCV.WUP instead
  2462. * of RCV.NXT. Peer still did not advance his SND.UNA when we
  2463. * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
  2464. * (borrowed from freebsd)
  2465. */
  2466. static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2467. {
  2468. return !before(end_seq, tp->rcv_wup) &&
  2469. !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
  2470. }
  2471. /* When we get a reset we do this. */
  2472. static void tcp_reset(struct sock *sk)
  2473. {
  2474. /* We want the right error as BSD sees it (and indeed as we do). */
  2475. switch (sk->sk_state) {
  2476. case TCP_SYN_SENT:
  2477. sk->sk_err = ECONNREFUSED;
  2478. break;
  2479. case TCP_CLOSE_WAIT:
  2480. sk->sk_err = EPIPE;
  2481. break;
  2482. case TCP_CLOSE:
  2483. return;
  2484. default:
  2485. sk->sk_err = ECONNRESET;
  2486. }
  2487. if (!sock_flag(sk, SOCK_DEAD))
  2488. sk->sk_error_report(sk);
  2489. tcp_done(sk);
  2490. }
  2491. /*
  2492. * Process the FIN bit. This now behaves as it is supposed to work
  2493. * and the FIN takes effect when it is validly part of sequence
  2494. * space. Not before when we get holes.
  2495. *
  2496. * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
  2497. * (and thence onto LAST-ACK and finally, CLOSE, we never enter
  2498. * TIME-WAIT)
  2499. *
  2500. * If we are in FINWAIT-1, a received FIN indicates simultaneous
  2501. * close and we go into CLOSING (and later onto TIME-WAIT)
  2502. *
  2503. * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
  2504. */
  2505. static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
  2506. {
  2507. struct tcp_sock *tp = tcp_sk(sk);
  2508. inet_csk_schedule_ack(sk);
  2509. sk->sk_shutdown |= RCV_SHUTDOWN;
  2510. sock_set_flag(sk, SOCK_DONE);
  2511. switch (sk->sk_state) {
  2512. case TCP_SYN_RECV:
  2513. case TCP_ESTABLISHED:
  2514. /* Move to CLOSE_WAIT */
  2515. tcp_set_state(sk, TCP_CLOSE_WAIT);
  2516. inet_csk(sk)->icsk_ack.pingpong = 1;
  2517. break;
  2518. case TCP_CLOSE_WAIT:
  2519. case TCP_CLOSING:
  2520. /* Received a retransmission of the FIN, do
  2521. * nothing.
  2522. */
  2523. break;
  2524. case TCP_LAST_ACK:
  2525. /* RFC793: Remain in the LAST-ACK state. */
  2526. break;
  2527. case TCP_FIN_WAIT1:
  2528. /* This case occurs when a simultaneous close
  2529. * happens, we must ack the received FIN and
  2530. * enter the CLOSING state.
  2531. */
  2532. tcp_send_ack(sk);
  2533. tcp_set_state(sk, TCP_CLOSING);
  2534. break;
  2535. case TCP_FIN_WAIT2:
  2536. /* Received a FIN -- send ACK and enter TIME_WAIT. */
  2537. tcp_send_ack(sk);
  2538. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  2539. break;
  2540. default:
  2541. /* Only TCP_LISTEN and TCP_CLOSE are left, in these
  2542. * cases we should never reach this piece of code.
  2543. */
  2544. printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
  2545. __FUNCTION__, sk->sk_state);
  2546. break;
  2547. };
  2548. /* It _is_ possible, that we have something out-of-order _after_ FIN.
  2549. * Probably, we should reset in this case. For now drop them.
  2550. */
  2551. __skb_queue_purge(&tp->out_of_order_queue);
  2552. if (tp->rx_opt.sack_ok)
  2553. tcp_sack_reset(&tp->rx_opt);
  2554. sk_stream_mem_reclaim(sk);
  2555. if (!sock_flag(sk, SOCK_DEAD)) {
  2556. sk->sk_state_change(sk);
  2557. /* Do not send POLL_HUP for half duplex close. */
  2558. if (sk->sk_shutdown == SHUTDOWN_MASK ||
  2559. sk->sk_state == TCP_CLOSE)
  2560. sk_wake_async(sk, 1, POLL_HUP);
  2561. else
  2562. sk_wake_async(sk, 1, POLL_IN);
  2563. }
  2564. }
  2565. static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
  2566. {
  2567. if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
  2568. if (before(seq, sp->start_seq))
  2569. sp->start_seq = seq;
  2570. if (after(end_seq, sp->end_seq))
  2571. sp->end_seq = end_seq;
  2572. return 1;
  2573. }
  2574. return 0;
  2575. }
  2576. static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2577. {
  2578. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2579. if (before(seq, tp->rcv_nxt))
  2580. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
  2581. else
  2582. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
  2583. tp->rx_opt.dsack = 1;
  2584. tp->duplicate_sack[0].start_seq = seq;
  2585. tp->duplicate_sack[0].end_seq = end_seq;
  2586. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
  2587. }
  2588. }
  2589. static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2590. {
  2591. if (!tp->rx_opt.dsack)
  2592. tcp_dsack_set(tp, seq, end_seq);
  2593. else
  2594. tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
  2595. }
  2596. static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
  2597. {
  2598. struct tcp_sock *tp = tcp_sk(sk);
  2599. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  2600. before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  2601. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  2602. tcp_enter_quickack_mode(sk);
  2603. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2604. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  2605. if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
  2606. end_seq = tp->rcv_nxt;
  2607. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
  2608. }
  2609. }
  2610. tcp_send_ack(sk);
  2611. }
  2612. /* These routines update the SACK block as out-of-order packets arrive or
  2613. * in-order packets close up the sequence space.
  2614. */
  2615. static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
  2616. {
  2617. int this_sack;
  2618. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2619. struct tcp_sack_block *swalk = sp+1;
  2620. /* See if the recent change to the first SACK eats into
  2621. * or hits the sequence space of other SACK blocks, if so coalesce.
  2622. */
  2623. for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
  2624. if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
  2625. int i;
  2626. /* Zap SWALK, by moving every further SACK up by one slot.
  2627. * Decrease num_sacks.
  2628. */
  2629. tp->rx_opt.num_sacks--;
  2630. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2631. for(i=this_sack; i < tp->rx_opt.num_sacks; i++)
  2632. sp[i] = sp[i+1];
  2633. continue;
  2634. }
  2635. this_sack++, swalk++;
  2636. }
  2637. }
  2638. static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
  2639. {
  2640. __u32 tmp;
  2641. tmp = sack1->start_seq;
  2642. sack1->start_seq = sack2->start_seq;
  2643. sack2->start_seq = tmp;
  2644. tmp = sack1->end_seq;
  2645. sack1->end_seq = sack2->end_seq;
  2646. sack2->end_seq = tmp;
  2647. }
  2648. static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
  2649. {
  2650. struct tcp_sock *tp = tcp_sk(sk);
  2651. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2652. int cur_sacks = tp->rx_opt.num_sacks;
  2653. int this_sack;
  2654. if (!cur_sacks)
  2655. goto new_sack;
  2656. for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
  2657. if (tcp_sack_extend(sp, seq, end_seq)) {
  2658. /* Rotate this_sack to the first one. */
  2659. for (; this_sack>0; this_sack--, sp--)
  2660. tcp_sack_swap(sp, sp-1);
  2661. if (cur_sacks > 1)
  2662. tcp_sack_maybe_coalesce(tp);
  2663. return;
  2664. }
  2665. }
  2666. /* Could not find an adjacent existing SACK, build a new one,
  2667. * put it at the front, and shift everyone else down. We
  2668. * always know there is at least one SACK present already here.
  2669. *
  2670. * If the sack array is full, forget about the last one.
  2671. */
  2672. if (this_sack >= 4) {
  2673. this_sack--;
  2674. tp->rx_opt.num_sacks--;
  2675. sp--;
  2676. }
  2677. for(; this_sack > 0; this_sack--, sp--)
  2678. *sp = *(sp-1);
  2679. new_sack:
  2680. /* Build the new head SACK, and we're done. */
  2681. sp->start_seq = seq;
  2682. sp->end_seq = end_seq;
  2683. tp->rx_opt.num_sacks++;
  2684. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2685. }
  2686. /* RCV.NXT advances, some SACKs should be eaten. */
  2687. static void tcp_sack_remove(struct tcp_sock *tp)
  2688. {
  2689. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2690. int num_sacks = tp->rx_opt.num_sacks;
  2691. int this_sack;
  2692. /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
  2693. if (skb_queue_empty(&tp->out_of_order_queue)) {
  2694. tp->rx_opt.num_sacks = 0;
  2695. tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
  2696. return;
  2697. }
  2698. for(this_sack = 0; this_sack < num_sacks; ) {
  2699. /* Check if the start of the sack is covered by RCV.NXT. */
  2700. if (!before(tp->rcv_nxt, sp->start_seq)) {
  2701. int i;
  2702. /* RCV.NXT must cover all the block! */
  2703. BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
  2704. /* Zap this SACK, by moving forward any other SACKS. */
  2705. for (i=this_sack+1; i < num_sacks; i++)
  2706. tp->selective_acks[i-1] = tp->selective_acks[i];
  2707. num_sacks--;
  2708. continue;
  2709. }
  2710. this_sack++;
  2711. sp++;
  2712. }
  2713. if (num_sacks != tp->rx_opt.num_sacks) {
  2714. tp->rx_opt.num_sacks = num_sacks;
  2715. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2716. }
  2717. }
  2718. /* This one checks to see if we can put data from the
  2719. * out_of_order queue into the receive_queue.
  2720. */
  2721. static void tcp_ofo_queue(struct sock *sk)
  2722. {
  2723. struct tcp_sock *tp = tcp_sk(sk);
  2724. __u32 dsack_high = tp->rcv_nxt;
  2725. struct sk_buff *skb;
  2726. while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
  2727. if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  2728. break;
  2729. if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
  2730. __u32 dsack = dsack_high;
  2731. if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
  2732. dsack_high = TCP_SKB_CB(skb)->end_seq;
  2733. tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
  2734. }
  2735. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  2736. SOCK_DEBUG(sk, "ofo packet was already received \n");
  2737. __skb_unlink(skb, &tp->out_of_order_queue);
  2738. __kfree_skb(skb);
  2739. continue;
  2740. }
  2741. SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
  2742. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  2743. TCP_SKB_CB(skb)->end_seq);
  2744. __skb_unlink(skb, &tp->out_of_order_queue);
  2745. __skb_queue_tail(&sk->sk_receive_queue, skb);
  2746. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  2747. if(skb->h.th->fin)
  2748. tcp_fin(skb, sk, skb->h.th);
  2749. }
  2750. }
  2751. static int tcp_prune_queue(struct sock *sk);
  2752. static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
  2753. {
  2754. struct tcphdr *th = skb->h.th;
  2755. struct tcp_sock *tp = tcp_sk(sk);
  2756. int eaten = -1;
  2757. if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
  2758. goto drop;
  2759. __skb_pull(skb, th->doff*4);
  2760. TCP_ECN_accept_cwr(tp, skb);
  2761. if (tp->rx_opt.dsack) {
  2762. tp->rx_opt.dsack = 0;
  2763. tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
  2764. 4 - tp->rx_opt.tstamp_ok);
  2765. }
  2766. /* Queue data for delivery to the user.
  2767. * Packets in sequence go to the receive queue.
  2768. * Out of sequence packets to the out_of_order_queue.
  2769. */
  2770. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  2771. if (tcp_receive_window(tp) == 0)
  2772. goto out_of_window;
  2773. /* Ok. In sequence. In window. */
  2774. if (tp->ucopy.task == current &&
  2775. tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
  2776. sock_owned_by_user(sk) && !tp->urg_data) {
  2777. int chunk = min_t(unsigned int, skb->len,
  2778. tp->ucopy.len);
  2779. __set_current_state(TASK_RUNNING);
  2780. local_bh_enable();
  2781. if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
  2782. tp->ucopy.len -= chunk;
  2783. tp->copied_seq += chunk;
  2784. eaten = (chunk == skb->len && !th->fin);
  2785. tcp_rcv_space_adjust(sk);
  2786. }
  2787. local_bh_disable();
  2788. }
  2789. if (eaten <= 0) {
  2790. queue_and_out:
  2791. if (eaten < 0 &&
  2792. (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  2793. !sk_stream_rmem_schedule(sk, skb))) {
  2794. if (tcp_prune_queue(sk) < 0 ||
  2795. !sk_stream_rmem_schedule(sk, skb))
  2796. goto drop;
  2797. }
  2798. sk_stream_set_owner_r(skb, sk);
  2799. __skb_queue_tail(&sk->sk_receive_queue, skb);
  2800. }
  2801. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  2802. if(skb->len)
  2803. tcp_event_data_recv(sk, tp, skb);
  2804. if(th->fin)
  2805. tcp_fin(skb, sk, th);
  2806. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  2807. tcp_ofo_queue(sk);
  2808. /* RFC2581. 4.2. SHOULD send immediate ACK, when
  2809. * gap in queue is filled.
  2810. */
  2811. if (skb_queue_empty(&tp->out_of_order_queue))
  2812. inet_csk(sk)->icsk_ack.pingpong = 0;
  2813. }
  2814. if (tp->rx_opt.num_sacks)
  2815. tcp_sack_remove(tp);
  2816. tcp_fast_path_check(sk, tp);
  2817. if (eaten > 0)
  2818. __kfree_skb(skb);
  2819. else if (!sock_flag(sk, SOCK_DEAD))
  2820. sk->sk_data_ready(sk, 0);
  2821. return;
  2822. }
  2823. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  2824. /* A retransmit, 2nd most common case. Force an immediate ack. */
  2825. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  2826. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  2827. out_of_window:
  2828. tcp_enter_quickack_mode(sk);
  2829. inet_csk_schedule_ack(sk);
  2830. drop:
  2831. __kfree_skb(skb);
  2832. return;
  2833. }
  2834. /* Out of window. F.e. zero window probe. */
  2835. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
  2836. goto out_of_window;
  2837. tcp_enter_quickack_mode(sk);
  2838. if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  2839. /* Partial packet, seq < rcv_next < end_seq */
  2840. SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
  2841. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  2842. TCP_SKB_CB(skb)->end_seq);
  2843. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
  2844. /* If window is closed, drop tail of packet. But after
  2845. * remembering D-SACK for its head made in previous line.
  2846. */
  2847. if (!tcp_receive_window(tp))
  2848. goto out_of_window;
  2849. goto queue_and_out;
  2850. }
  2851. TCP_ECN_check_ce(tp, skb);
  2852. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  2853. !sk_stream_rmem_schedule(sk, skb)) {
  2854. if (tcp_prune_queue(sk) < 0 ||
  2855. !sk_stream_rmem_schedule(sk, skb))
  2856. goto drop;
  2857. }
  2858. /* Disable header prediction. */
  2859. tp->pred_flags = 0;
  2860. inet_csk_schedule_ack(sk);
  2861. SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
  2862. tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  2863. sk_stream_set_owner_r(skb, sk);
  2864. if (!skb_peek(&tp->out_of_order_queue)) {
  2865. /* Initial out of order segment, build 1 SACK. */
  2866. if (tp->rx_opt.sack_ok) {
  2867. tp->rx_opt.num_sacks = 1;
  2868. tp->rx_opt.dsack = 0;
  2869. tp->rx_opt.eff_sacks = 1;
  2870. tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
  2871. tp->selective_acks[0].end_seq =
  2872. TCP_SKB_CB(skb)->end_seq;
  2873. }
  2874. __skb_queue_head(&tp->out_of_order_queue,skb);
  2875. } else {
  2876. struct sk_buff *skb1 = tp->out_of_order_queue.prev;
  2877. u32 seq = TCP_SKB_CB(skb)->seq;
  2878. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  2879. if (seq == TCP_SKB_CB(skb1)->end_seq) {
  2880. __skb_append(skb1, skb, &tp->out_of_order_queue);
  2881. if (!tp->rx_opt.num_sacks ||
  2882. tp->selective_acks[0].end_seq != seq)
  2883. goto add_sack;
  2884. /* Common case: data arrive in order after hole. */
  2885. tp->selective_acks[0].end_seq = end_seq;
  2886. return;
  2887. }
  2888. /* Find place to insert this segment. */
  2889. do {
  2890. if (!after(TCP_SKB_CB(skb1)->seq, seq))
  2891. break;
  2892. } while ((skb1 = skb1->prev) !=
  2893. (struct sk_buff*)&tp->out_of_order_queue);
  2894. /* Do skb overlap to previous one? */
  2895. if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
  2896. before(seq, TCP_SKB_CB(skb1)->end_seq)) {
  2897. if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  2898. /* All the bits are present. Drop. */
  2899. __kfree_skb(skb);
  2900. tcp_dsack_set(tp, seq, end_seq);
  2901. goto add_sack;
  2902. }
  2903. if (after(seq, TCP_SKB_CB(skb1)->seq)) {
  2904. /* Partial overlap. */
  2905. tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
  2906. } else {
  2907. skb1 = skb1->prev;
  2908. }
  2909. }
  2910. __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
  2911. /* And clean segments covered by new one as whole. */
  2912. while ((skb1 = skb->next) !=
  2913. (struct sk_buff*)&tp->out_of_order_queue &&
  2914. after(end_seq, TCP_SKB_CB(skb1)->seq)) {
  2915. if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  2916. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
  2917. break;
  2918. }
  2919. __skb_unlink(skb1, &tp->out_of_order_queue);
  2920. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
  2921. __kfree_skb(skb1);
  2922. }
  2923. add_sack:
  2924. if (tp->rx_opt.sack_ok)
  2925. tcp_sack_new_ofo_skb(sk, seq, end_seq);
  2926. }
  2927. }
  2928. /* Collapse contiguous sequence of skbs head..tail with
  2929. * sequence numbers start..end.
  2930. * Segments with FIN/SYN are not collapsed (only because this
  2931. * simplifies code)
  2932. */
  2933. static void
  2934. tcp_collapse(struct sock *sk, struct sk_buff_head *list,
  2935. struct sk_buff *head, struct sk_buff *tail,
  2936. u32 start, u32 end)
  2937. {
  2938. struct sk_buff *skb;
  2939. /* First, check that queue is collapsible and find
  2940. * the point where collapsing can be useful. */
  2941. for (skb = head; skb != tail; ) {
  2942. /* No new bits? It is possible on ofo queue. */
  2943. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  2944. struct sk_buff *next = skb->next;
  2945. __skb_unlink(skb, list);
  2946. __kfree_skb(skb);
  2947. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  2948. skb = next;
  2949. continue;
  2950. }
  2951. /* The first skb to collapse is:
  2952. * - not SYN/FIN and
  2953. * - bloated or contains data before "start" or
  2954. * overlaps to the next one.
  2955. */
  2956. if (!skb->h.th->syn && !skb->h.th->fin &&
  2957. (tcp_win_from_space(skb->truesize) > skb->len ||
  2958. before(TCP_SKB_CB(skb)->seq, start) ||
  2959. (skb->next != tail &&
  2960. TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
  2961. break;
  2962. /* Decided to skip this, advance start seq. */
  2963. start = TCP_SKB_CB(skb)->end_seq;
  2964. skb = skb->next;
  2965. }
  2966. if (skb == tail || skb->h.th->syn || skb->h.th->fin)
  2967. return;
  2968. while (before(start, end)) {
  2969. struct sk_buff *nskb;
  2970. int header = skb_headroom(skb);
  2971. int copy = SKB_MAX_ORDER(header, 0);
  2972. /* Too big header? This can happen with IPv6. */
  2973. if (copy < 0)
  2974. return;
  2975. if (end-start < copy)
  2976. copy = end-start;
  2977. nskb = alloc_skb(copy+header, GFP_ATOMIC);
  2978. if (!nskb)
  2979. return;
  2980. skb_reserve(nskb, header);
  2981. memcpy(nskb->head, skb->head, header);
  2982. nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head);
  2983. nskb->h.raw = nskb->head + (skb->h.raw-skb->head);
  2984. nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head);
  2985. memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
  2986. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
  2987. __skb_insert(nskb, skb->prev, skb, list);
  2988. sk_stream_set_owner_r(nskb, sk);
  2989. /* Copy data, releasing collapsed skbs. */
  2990. while (copy > 0) {
  2991. int offset = start - TCP_SKB_CB(skb)->seq;
  2992. int size = TCP_SKB_CB(skb)->end_seq - start;
  2993. BUG_ON(offset < 0);
  2994. if (size > 0) {
  2995. size = min(copy, size);
  2996. if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
  2997. BUG();
  2998. TCP_SKB_CB(nskb)->end_seq += size;
  2999. copy -= size;
  3000. start += size;
  3001. }
  3002. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3003. struct sk_buff *next = skb->next;
  3004. __skb_unlink(skb, list);
  3005. __kfree_skb(skb);
  3006. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3007. skb = next;
  3008. if (skb == tail || skb->h.th->syn || skb->h.th->fin)
  3009. return;
  3010. }
  3011. }
  3012. }
  3013. }
  3014. /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
  3015. * and tcp_collapse() them until all the queue is collapsed.
  3016. */
  3017. static void tcp_collapse_ofo_queue(struct sock *sk)
  3018. {
  3019. struct tcp_sock *tp = tcp_sk(sk);
  3020. struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
  3021. struct sk_buff *head;
  3022. u32 start, end;
  3023. if (skb == NULL)
  3024. return;
  3025. start = TCP_SKB_CB(skb)->seq;
  3026. end = TCP_SKB_CB(skb)->end_seq;
  3027. head = skb;
  3028. for (;;) {
  3029. skb = skb->next;
  3030. /* Segment is terminated when we see gap or when
  3031. * we are at the end of all the queue. */
  3032. if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
  3033. after(TCP_SKB_CB(skb)->seq, end) ||
  3034. before(TCP_SKB_CB(skb)->end_seq, start)) {
  3035. tcp_collapse(sk, &tp->out_of_order_queue,
  3036. head, skb, start, end);
  3037. head = skb;
  3038. if (skb == (struct sk_buff *)&tp->out_of_order_queue)
  3039. break;
  3040. /* Start new segment */
  3041. start = TCP_SKB_CB(skb)->seq;
  3042. end = TCP_SKB_CB(skb)->end_seq;
  3043. } else {
  3044. if (before(TCP_SKB_CB(skb)->seq, start))
  3045. start = TCP_SKB_CB(skb)->seq;
  3046. if (after(TCP_SKB_CB(skb)->end_seq, end))
  3047. end = TCP_SKB_CB(skb)->end_seq;
  3048. }
  3049. }
  3050. }
  3051. /* Reduce allocated memory if we can, trying to get
  3052. * the socket within its memory limits again.
  3053. *
  3054. * Return less than zero if we should start dropping frames
  3055. * until the socket owning process reads some of the data
  3056. * to stabilize the situation.
  3057. */
  3058. static int tcp_prune_queue(struct sock *sk)
  3059. {
  3060. struct tcp_sock *tp = tcp_sk(sk);
  3061. SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
  3062. NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
  3063. if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
  3064. tcp_clamp_window(sk, tp);
  3065. else if (tcp_memory_pressure)
  3066. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
  3067. tcp_collapse_ofo_queue(sk);
  3068. tcp_collapse(sk, &sk->sk_receive_queue,
  3069. sk->sk_receive_queue.next,
  3070. (struct sk_buff*)&sk->sk_receive_queue,
  3071. tp->copied_seq, tp->rcv_nxt);
  3072. sk_stream_mem_reclaim(sk);
  3073. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3074. return 0;
  3075. /* Collapsing did not help, destructive actions follow.
  3076. * This must not ever occur. */
  3077. /* First, purge the out_of_order queue. */
  3078. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3079. NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
  3080. __skb_queue_purge(&tp->out_of_order_queue);
  3081. /* Reset SACK state. A conforming SACK implementation will
  3082. * do the same at a timeout based retransmit. When a connection
  3083. * is in a sad state like this, we care only about integrity
  3084. * of the connection not performance.
  3085. */
  3086. if (tp->rx_opt.sack_ok)
  3087. tcp_sack_reset(&tp->rx_opt);
  3088. sk_stream_mem_reclaim(sk);
  3089. }
  3090. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3091. return 0;
  3092. /* If we are really being abused, tell the caller to silently
  3093. * drop receive data on the floor. It will get retransmitted
  3094. * and hopefully then we'll have sufficient space.
  3095. */
  3096. NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
  3097. /* Massive buffer overcommit. */
  3098. tp->pred_flags = 0;
  3099. return -1;
  3100. }
  3101. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  3102. * As additional protections, we do not touch cwnd in retransmission phases,
  3103. * and if application hit its sndbuf limit recently.
  3104. */
  3105. void tcp_cwnd_application_limited(struct sock *sk)
  3106. {
  3107. struct tcp_sock *tp = tcp_sk(sk);
  3108. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  3109. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  3110. /* Limited by application or receiver window. */
  3111. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  3112. u32 win_used = max(tp->snd_cwnd_used, init_win);
  3113. if (win_used < tp->snd_cwnd) {
  3114. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  3115. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  3116. }
  3117. tp->snd_cwnd_used = 0;
  3118. }
  3119. tp->snd_cwnd_stamp = tcp_time_stamp;
  3120. }
  3121. static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp)
  3122. {
  3123. /* If the user specified a specific send buffer setting, do
  3124. * not modify it.
  3125. */
  3126. if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
  3127. return 0;
  3128. /* If we are under global TCP memory pressure, do not expand. */
  3129. if (tcp_memory_pressure)
  3130. return 0;
  3131. /* If we are under soft global TCP memory pressure, do not expand. */
  3132. if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
  3133. return 0;
  3134. /* If we filled the congestion window, do not expand. */
  3135. if (tp->packets_out >= tp->snd_cwnd)
  3136. return 0;
  3137. return 1;
  3138. }
  3139. /* When incoming ACK allowed to free some skb from write_queue,
  3140. * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
  3141. * on the exit from tcp input handler.
  3142. *
  3143. * PROBLEM: sndbuf expansion does not work well with largesend.
  3144. */
  3145. static void tcp_new_space(struct sock *sk)
  3146. {
  3147. struct tcp_sock *tp = tcp_sk(sk);
  3148. if (tcp_should_expand_sndbuf(sk, tp)) {
  3149. int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
  3150. MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
  3151. demanded = max_t(unsigned int, tp->snd_cwnd,
  3152. tp->reordering + 1);
  3153. sndmem *= 2*demanded;
  3154. if (sndmem > sk->sk_sndbuf)
  3155. sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
  3156. tp->snd_cwnd_stamp = tcp_time_stamp;
  3157. }
  3158. sk->sk_write_space(sk);
  3159. }
  3160. static void tcp_check_space(struct sock *sk)
  3161. {
  3162. if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
  3163. sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
  3164. if (sk->sk_socket &&
  3165. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
  3166. tcp_new_space(sk);
  3167. }
  3168. }
  3169. static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp)
  3170. {
  3171. tcp_push_pending_frames(sk, tp);
  3172. tcp_check_space(sk);
  3173. }
  3174. /*
  3175. * Check if sending an ack is needed.
  3176. */
  3177. static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
  3178. {
  3179. struct tcp_sock *tp = tcp_sk(sk);
  3180. /* More than one full frame received... */
  3181. if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
  3182. /* ... and right edge of window advances far enough.
  3183. * (tcp_recvmsg() will send ACK otherwise). Or...
  3184. */
  3185. && __tcp_select_window(sk) >= tp->rcv_wnd) ||
  3186. /* We ACK each frame or... */
  3187. tcp_in_quickack_mode(sk) ||
  3188. /* We have out of order data. */
  3189. (ofo_possible &&
  3190. skb_peek(&tp->out_of_order_queue))) {
  3191. /* Then ack it now */
  3192. tcp_send_ack(sk);
  3193. } else {
  3194. /* Else, send delayed ack. */
  3195. tcp_send_delayed_ack(sk);
  3196. }
  3197. }
  3198. static inline void tcp_ack_snd_check(struct sock *sk)
  3199. {
  3200. if (!inet_csk_ack_scheduled(sk)) {
  3201. /* We sent a data segment already. */
  3202. return;
  3203. }
  3204. __tcp_ack_snd_check(sk, 1);
  3205. }
  3206. /*
  3207. * This routine is only called when we have urgent data
  3208. * signaled. Its the 'slow' part of tcp_urg. It could be
  3209. * moved inline now as tcp_urg is only called from one
  3210. * place. We handle URGent data wrong. We have to - as
  3211. * BSD still doesn't use the correction from RFC961.
  3212. * For 1003.1g we should support a new option TCP_STDURG to permit
  3213. * either form (or just set the sysctl tcp_stdurg).
  3214. */
  3215. static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
  3216. {
  3217. struct tcp_sock *tp = tcp_sk(sk);
  3218. u32 ptr = ntohs(th->urg_ptr);
  3219. if (ptr && !sysctl_tcp_stdurg)
  3220. ptr--;
  3221. ptr += ntohl(th->seq);
  3222. /* Ignore urgent data that we've already seen and read. */
  3223. if (after(tp->copied_seq, ptr))
  3224. return;
  3225. /* Do not replay urg ptr.
  3226. *
  3227. * NOTE: interesting situation not covered by specs.
  3228. * Misbehaving sender may send urg ptr, pointing to segment,
  3229. * which we already have in ofo queue. We are not able to fetch
  3230. * such data and will stay in TCP_URG_NOTYET until will be eaten
  3231. * by recvmsg(). Seems, we are not obliged to handle such wicked
  3232. * situations. But it is worth to think about possibility of some
  3233. * DoSes using some hypothetical application level deadlock.
  3234. */
  3235. if (before(ptr, tp->rcv_nxt))
  3236. return;
  3237. /* Do we already have a newer (or duplicate) urgent pointer? */
  3238. if (tp->urg_data && !after(ptr, tp->urg_seq))
  3239. return;
  3240. /* Tell the world about our new urgent pointer. */
  3241. sk_send_sigurg(sk);
  3242. /* We may be adding urgent data when the last byte read was
  3243. * urgent. To do this requires some care. We cannot just ignore
  3244. * tp->copied_seq since we would read the last urgent byte again
  3245. * as data, nor can we alter copied_seq until this data arrives
  3246. * or we break the semantics of SIOCATMARK (and thus sockatmark())
  3247. *
  3248. * NOTE. Double Dutch. Rendering to plain English: author of comment
  3249. * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
  3250. * and expect that both A and B disappear from stream. This is _wrong_.
  3251. * Though this happens in BSD with high probability, this is occasional.
  3252. * Any application relying on this is buggy. Note also, that fix "works"
  3253. * only in this artificial test. Insert some normal data between A and B and we will
  3254. * decline of BSD again. Verdict: it is better to remove to trap
  3255. * buggy users.
  3256. */
  3257. if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
  3258. !sock_flag(sk, SOCK_URGINLINE) &&
  3259. tp->copied_seq != tp->rcv_nxt) {
  3260. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  3261. tp->copied_seq++;
  3262. if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
  3263. __skb_unlink(skb, &sk->sk_receive_queue);
  3264. __kfree_skb(skb);
  3265. }
  3266. }
  3267. tp->urg_data = TCP_URG_NOTYET;
  3268. tp->urg_seq = ptr;
  3269. /* Disable header prediction. */
  3270. tp->pred_flags = 0;
  3271. }
  3272. /* This is the 'fast' part of urgent handling. */
  3273. static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
  3274. {
  3275. struct tcp_sock *tp = tcp_sk(sk);
  3276. /* Check if we get a new urgent pointer - normally not. */
  3277. if (th->urg)
  3278. tcp_check_urg(sk,th);
  3279. /* Do we wait for any urgent data? - normally not... */
  3280. if (tp->urg_data == TCP_URG_NOTYET) {
  3281. u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
  3282. th->syn;
  3283. /* Is the urgent pointer pointing into this packet? */
  3284. if (ptr < skb->len) {
  3285. u8 tmp;
  3286. if (skb_copy_bits(skb, ptr, &tmp, 1))
  3287. BUG();
  3288. tp->urg_data = TCP_URG_VALID | tmp;
  3289. if (!sock_flag(sk, SOCK_DEAD))
  3290. sk->sk_data_ready(sk, 0);
  3291. }
  3292. }
  3293. }
  3294. static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
  3295. {
  3296. struct tcp_sock *tp = tcp_sk(sk);
  3297. int chunk = skb->len - hlen;
  3298. int err;
  3299. local_bh_enable();
  3300. if (skb->ip_summed==CHECKSUM_UNNECESSARY)
  3301. err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
  3302. else
  3303. err = skb_copy_and_csum_datagram_iovec(skb, hlen,
  3304. tp->ucopy.iov);
  3305. if (!err) {
  3306. tp->ucopy.len -= chunk;
  3307. tp->copied_seq += chunk;
  3308. tcp_rcv_space_adjust(sk);
  3309. }
  3310. local_bh_disable();
  3311. return err;
  3312. }
  3313. static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3314. {
  3315. __sum16 result;
  3316. if (sock_owned_by_user(sk)) {
  3317. local_bh_enable();
  3318. result = __tcp_checksum_complete(skb);
  3319. local_bh_disable();
  3320. } else {
  3321. result = __tcp_checksum_complete(skb);
  3322. }
  3323. return result;
  3324. }
  3325. static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3326. {
  3327. return skb->ip_summed != CHECKSUM_UNNECESSARY &&
  3328. __tcp_checksum_complete_user(sk, skb);
  3329. }
  3330. #ifdef CONFIG_NET_DMA
  3331. static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
  3332. {
  3333. struct tcp_sock *tp = tcp_sk(sk);
  3334. int chunk = skb->len - hlen;
  3335. int dma_cookie;
  3336. int copied_early = 0;
  3337. if (tp->ucopy.wakeup)
  3338. return 0;
  3339. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  3340. tp->ucopy.dma_chan = get_softnet_dma();
  3341. if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) {
  3342. dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
  3343. skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
  3344. if (dma_cookie < 0)
  3345. goto out;
  3346. tp->ucopy.dma_cookie = dma_cookie;
  3347. copied_early = 1;
  3348. tp->ucopy.len -= chunk;
  3349. tp->copied_seq += chunk;
  3350. tcp_rcv_space_adjust(sk);
  3351. if ((tp->ucopy.len == 0) ||
  3352. (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) ||
  3353. (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
  3354. tp->ucopy.wakeup = 1;
  3355. sk->sk_data_ready(sk, 0);
  3356. }
  3357. } else if (chunk > 0) {
  3358. tp->ucopy.wakeup = 1;
  3359. sk->sk_data_ready(sk, 0);
  3360. }
  3361. out:
  3362. return copied_early;
  3363. }
  3364. #endif /* CONFIG_NET_DMA */
  3365. /*
  3366. * TCP receive function for the ESTABLISHED state.
  3367. *
  3368. * It is split into a fast path and a slow path. The fast path is
  3369. * disabled when:
  3370. * - A zero window was announced from us - zero window probing
  3371. * is only handled properly in the slow path.
  3372. * - Out of order segments arrived.
  3373. * - Urgent data is expected.
  3374. * - There is no buffer space left
  3375. * - Unexpected TCP flags/window values/header lengths are received
  3376. * (detected by checking the TCP header against pred_flags)
  3377. * - Data is sent in both directions. Fast path only supports pure senders
  3378. * or pure receivers (this means either the sequence number or the ack
  3379. * value must stay constant)
  3380. * - Unexpected TCP option.
  3381. *
  3382. * When these conditions are not satisfied it drops into a standard
  3383. * receive procedure patterned after RFC793 to handle all cases.
  3384. * The first three cases are guaranteed by proper pred_flags setting,
  3385. * the rest is checked inline. Fast processing is turned on in
  3386. * tcp_data_queue when everything is OK.
  3387. */
  3388. int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  3389. struct tcphdr *th, unsigned len)
  3390. {
  3391. struct tcp_sock *tp = tcp_sk(sk);
  3392. /*
  3393. * Header prediction.
  3394. * The code loosely follows the one in the famous
  3395. * "30 instruction TCP receive" Van Jacobson mail.
  3396. *
  3397. * Van's trick is to deposit buffers into socket queue
  3398. * on a device interrupt, to call tcp_recv function
  3399. * on the receive process context and checksum and copy
  3400. * the buffer to user space. smart...
  3401. *
  3402. * Our current scheme is not silly either but we take the
  3403. * extra cost of the net_bh soft interrupt processing...
  3404. * We do checksum and copy also but from device to kernel.
  3405. */
  3406. tp->rx_opt.saw_tstamp = 0;
  3407. /* pred_flags is 0xS?10 << 16 + snd_wnd
  3408. * if header_prediction is to be made
  3409. * 'S' will always be tp->tcp_header_len >> 2
  3410. * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
  3411. * turn it off (when there are holes in the receive
  3412. * space for instance)
  3413. * PSH flag is ignored.
  3414. */
  3415. if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
  3416. TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3417. int tcp_header_len = tp->tcp_header_len;
  3418. /* Timestamp header prediction: tcp_header_len
  3419. * is automatically equal to th->doff*4 due to pred_flags
  3420. * match.
  3421. */
  3422. /* Check timestamp */
  3423. if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
  3424. __be32 *ptr = (__be32 *)(th + 1);
  3425. /* No? Slow path! */
  3426. if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  3427. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
  3428. goto slow_path;
  3429. tp->rx_opt.saw_tstamp = 1;
  3430. ++ptr;
  3431. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  3432. ++ptr;
  3433. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  3434. /* If PAWS failed, check it more carefully in slow path */
  3435. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
  3436. goto slow_path;
  3437. /* DO NOT update ts_recent here, if checksum fails
  3438. * and timestamp was corrupted part, it will result
  3439. * in a hung connection since we will drop all
  3440. * future packets due to the PAWS test.
  3441. */
  3442. }
  3443. if (len <= tcp_header_len) {
  3444. /* Bulk data transfer: sender */
  3445. if (len == tcp_header_len) {
  3446. /* Predicted packet is in window by definition.
  3447. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3448. * Hence, check seq<=rcv_wup reduces to:
  3449. */
  3450. if (tcp_header_len ==
  3451. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3452. tp->rcv_nxt == tp->rcv_wup)
  3453. tcp_store_ts_recent(tp);
  3454. /* We know that such packets are checksummed
  3455. * on entry.
  3456. */
  3457. tcp_ack(sk, skb, 0);
  3458. __kfree_skb(skb);
  3459. tcp_data_snd_check(sk, tp);
  3460. return 0;
  3461. } else { /* Header too small */
  3462. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3463. goto discard;
  3464. }
  3465. } else {
  3466. int eaten = 0;
  3467. int copied_early = 0;
  3468. if (tp->copied_seq == tp->rcv_nxt &&
  3469. len - tcp_header_len <= tp->ucopy.len) {
  3470. #ifdef CONFIG_NET_DMA
  3471. if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
  3472. copied_early = 1;
  3473. eaten = 1;
  3474. }
  3475. #endif
  3476. if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
  3477. __set_current_state(TASK_RUNNING);
  3478. if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
  3479. eaten = 1;
  3480. }
  3481. if (eaten) {
  3482. /* Predicted packet is in window by definition.
  3483. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3484. * Hence, check seq<=rcv_wup reduces to:
  3485. */
  3486. if (tcp_header_len ==
  3487. (sizeof(struct tcphdr) +
  3488. TCPOLEN_TSTAMP_ALIGNED) &&
  3489. tp->rcv_nxt == tp->rcv_wup)
  3490. tcp_store_ts_recent(tp);
  3491. tcp_rcv_rtt_measure_ts(sk, skb);
  3492. __skb_pull(skb, tcp_header_len);
  3493. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3494. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
  3495. }
  3496. if (copied_early)
  3497. tcp_cleanup_rbuf(sk, skb->len);
  3498. }
  3499. if (!eaten) {
  3500. if (tcp_checksum_complete_user(sk, skb))
  3501. goto csum_error;
  3502. /* Predicted packet is in window by definition.
  3503. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3504. * Hence, check seq<=rcv_wup reduces to:
  3505. */
  3506. if (tcp_header_len ==
  3507. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3508. tp->rcv_nxt == tp->rcv_wup)
  3509. tcp_store_ts_recent(tp);
  3510. tcp_rcv_rtt_measure_ts(sk, skb);
  3511. if ((int)skb->truesize > sk->sk_forward_alloc)
  3512. goto step5;
  3513. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
  3514. /* Bulk data transfer: receiver */
  3515. __skb_pull(skb,tcp_header_len);
  3516. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3517. sk_stream_set_owner_r(skb, sk);
  3518. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3519. }
  3520. tcp_event_data_recv(sk, tp, skb);
  3521. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
  3522. /* Well, only one small jumplet in fast path... */
  3523. tcp_ack(sk, skb, FLAG_DATA);
  3524. tcp_data_snd_check(sk, tp);
  3525. if (!inet_csk_ack_scheduled(sk))
  3526. goto no_ack;
  3527. }
  3528. __tcp_ack_snd_check(sk, 0);
  3529. no_ack:
  3530. #ifdef CONFIG_NET_DMA
  3531. if (copied_early)
  3532. __skb_queue_tail(&sk->sk_async_wait_queue, skb);
  3533. else
  3534. #endif
  3535. if (eaten)
  3536. __kfree_skb(skb);
  3537. else
  3538. sk->sk_data_ready(sk, 0);
  3539. return 0;
  3540. }
  3541. }
  3542. slow_path:
  3543. if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
  3544. goto csum_error;
  3545. /*
  3546. * RFC1323: H1. Apply PAWS check first.
  3547. */
  3548. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  3549. tcp_paws_discard(sk, skb)) {
  3550. if (!th->rst) {
  3551. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  3552. tcp_send_dupack(sk, skb);
  3553. goto discard;
  3554. }
  3555. /* Resets are accepted even if PAWS failed.
  3556. ts_recent update must be made after we are sure
  3557. that the packet is in window.
  3558. */
  3559. }
  3560. /*
  3561. * Standard slow path.
  3562. */
  3563. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  3564. /* RFC793, page 37: "In all states except SYN-SENT, all reset
  3565. * (RST) segments are validated by checking their SEQ-fields."
  3566. * And page 69: "If an incoming segment is not acceptable,
  3567. * an acknowledgment should be sent in reply (unless the RST bit
  3568. * is set, if so drop the segment and return)".
  3569. */
  3570. if (!th->rst)
  3571. tcp_send_dupack(sk, skb);
  3572. goto discard;
  3573. }
  3574. if(th->rst) {
  3575. tcp_reset(sk);
  3576. goto discard;
  3577. }
  3578. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  3579. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3580. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3581. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  3582. tcp_reset(sk);
  3583. return 1;
  3584. }
  3585. step5:
  3586. if(th->ack)
  3587. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3588. tcp_rcv_rtt_measure_ts(sk, skb);
  3589. /* Process urgent data. */
  3590. tcp_urg(sk, skb, th);
  3591. /* step 7: process the segment text */
  3592. tcp_data_queue(sk, skb);
  3593. tcp_data_snd_check(sk, tp);
  3594. tcp_ack_snd_check(sk);
  3595. return 0;
  3596. csum_error:
  3597. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3598. discard:
  3599. __kfree_skb(skb);
  3600. return 0;
  3601. }
  3602. static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
  3603. struct tcphdr *th, unsigned len)
  3604. {
  3605. struct tcp_sock *tp = tcp_sk(sk);
  3606. struct inet_connection_sock *icsk = inet_csk(sk);
  3607. int saved_clamp = tp->rx_opt.mss_clamp;
  3608. tcp_parse_options(skb, &tp->rx_opt, 0);
  3609. if (th->ack) {
  3610. /* rfc793:
  3611. * "If the state is SYN-SENT then
  3612. * first check the ACK bit
  3613. * If the ACK bit is set
  3614. * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
  3615. * a reset (unless the RST bit is set, if so drop
  3616. * the segment and return)"
  3617. *
  3618. * We do not send data with SYN, so that RFC-correct
  3619. * test reduces to:
  3620. */
  3621. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
  3622. goto reset_and_undo;
  3623. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  3624. !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
  3625. tcp_time_stamp)) {
  3626. NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
  3627. goto reset_and_undo;
  3628. }
  3629. /* Now ACK is acceptable.
  3630. *
  3631. * "If the RST bit is set
  3632. * If the ACK was acceptable then signal the user "error:
  3633. * connection reset", drop the segment, enter CLOSED state,
  3634. * delete TCB, and return."
  3635. */
  3636. if (th->rst) {
  3637. tcp_reset(sk);
  3638. goto discard;
  3639. }
  3640. /* rfc793:
  3641. * "fifth, if neither of the SYN or RST bits is set then
  3642. * drop the segment and return."
  3643. *
  3644. * See note below!
  3645. * --ANK(990513)
  3646. */
  3647. if (!th->syn)
  3648. goto discard_and_undo;
  3649. /* rfc793:
  3650. * "If the SYN bit is on ...
  3651. * are acceptable then ...
  3652. * (our SYN has been ACKed), change the connection
  3653. * state to ESTABLISHED..."
  3654. */
  3655. TCP_ECN_rcv_synack(tp, th);
  3656. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  3657. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3658. /* Ok.. it's good. Set up sequence numbers and
  3659. * move to established.
  3660. */
  3661. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  3662. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  3663. /* RFC1323: The window in SYN & SYN/ACK segments is
  3664. * never scaled.
  3665. */
  3666. tp->snd_wnd = ntohs(th->window);
  3667. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
  3668. if (!tp->rx_opt.wscale_ok) {
  3669. tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
  3670. tp->window_clamp = min(tp->window_clamp, 65535U);
  3671. }
  3672. if (tp->rx_opt.saw_tstamp) {
  3673. tp->rx_opt.tstamp_ok = 1;
  3674. tp->tcp_header_len =
  3675. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  3676. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  3677. tcp_store_ts_recent(tp);
  3678. } else {
  3679. tp->tcp_header_len = sizeof(struct tcphdr);
  3680. }
  3681. if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
  3682. tp->rx_opt.sack_ok |= 2;
  3683. tcp_mtup_init(sk);
  3684. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  3685. tcp_initialize_rcv_mss(sk);
  3686. /* Remember, tcp_poll() does not lock socket!
  3687. * Change state from SYN-SENT only after copied_seq
  3688. * is initialized. */
  3689. tp->copied_seq = tp->rcv_nxt;
  3690. smp_mb();
  3691. tcp_set_state(sk, TCP_ESTABLISHED);
  3692. security_inet_conn_established(sk, skb);
  3693. /* Make sure socket is routed, for correct metrics. */
  3694. icsk->icsk_af_ops->rebuild_header(sk);
  3695. tcp_init_metrics(sk);
  3696. tcp_init_congestion_control(sk);
  3697. /* Prevent spurious tcp_cwnd_restart() on first data
  3698. * packet.
  3699. */
  3700. tp->lsndtime = tcp_time_stamp;
  3701. tcp_init_buffer_space(sk);
  3702. if (sock_flag(sk, SOCK_KEEPOPEN))
  3703. inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
  3704. if (!tp->rx_opt.snd_wscale)
  3705. __tcp_fast_path_on(tp, tp->snd_wnd);
  3706. else
  3707. tp->pred_flags = 0;
  3708. if (!sock_flag(sk, SOCK_DEAD)) {
  3709. sk->sk_state_change(sk);
  3710. sk_wake_async(sk, 0, POLL_OUT);
  3711. }
  3712. if (sk->sk_write_pending ||
  3713. icsk->icsk_accept_queue.rskq_defer_accept ||
  3714. icsk->icsk_ack.pingpong) {
  3715. /* Save one ACK. Data will be ready after
  3716. * several ticks, if write_pending is set.
  3717. *
  3718. * It may be deleted, but with this feature tcpdumps
  3719. * look so _wonderfully_ clever, that I was not able
  3720. * to stand against the temptation 8) --ANK
  3721. */
  3722. inet_csk_schedule_ack(sk);
  3723. icsk->icsk_ack.lrcvtime = tcp_time_stamp;
  3724. icsk->icsk_ack.ato = TCP_ATO_MIN;
  3725. tcp_incr_quickack(sk);
  3726. tcp_enter_quickack_mode(sk);
  3727. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  3728. TCP_DELACK_MAX, TCP_RTO_MAX);
  3729. discard:
  3730. __kfree_skb(skb);
  3731. return 0;
  3732. } else {
  3733. tcp_send_ack(sk);
  3734. }
  3735. return -1;
  3736. }
  3737. /* No ACK in the segment */
  3738. if (th->rst) {
  3739. /* rfc793:
  3740. * "If the RST bit is set
  3741. *
  3742. * Otherwise (no ACK) drop the segment and return."
  3743. */
  3744. goto discard_and_undo;
  3745. }
  3746. /* PAWS check. */
  3747. if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
  3748. goto discard_and_undo;
  3749. if (th->syn) {
  3750. /* We see SYN without ACK. It is attempt of
  3751. * simultaneous connect with crossed SYNs.
  3752. * Particularly, it can be connect to self.
  3753. */
  3754. tcp_set_state(sk, TCP_SYN_RECV);
  3755. if (tp->rx_opt.saw_tstamp) {
  3756. tp->rx_opt.tstamp_ok = 1;
  3757. tcp_store_ts_recent(tp);
  3758. tp->tcp_header_len =
  3759. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  3760. } else {
  3761. tp->tcp_header_len = sizeof(struct tcphdr);
  3762. }
  3763. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  3764. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  3765. /* RFC1323: The window in SYN & SYN/ACK segments is
  3766. * never scaled.
  3767. */
  3768. tp->snd_wnd = ntohs(th->window);
  3769. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  3770. tp->max_window = tp->snd_wnd;
  3771. TCP_ECN_rcv_syn(tp, th);
  3772. tcp_mtup_init(sk);
  3773. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  3774. tcp_initialize_rcv_mss(sk);
  3775. tcp_send_synack(sk);
  3776. #if 0
  3777. /* Note, we could accept data and URG from this segment.
  3778. * There are no obstacles to make this.
  3779. *
  3780. * However, if we ignore data in ACKless segments sometimes,
  3781. * we have no reasons to accept it sometimes.
  3782. * Also, seems the code doing it in step6 of tcp_rcv_state_process
  3783. * is not flawless. So, discard packet for sanity.
  3784. * Uncomment this return to process the data.
  3785. */
  3786. return -1;
  3787. #else
  3788. goto discard;
  3789. #endif
  3790. }
  3791. /* "fifth, if neither of the SYN or RST bits is set then
  3792. * drop the segment and return."
  3793. */
  3794. discard_and_undo:
  3795. tcp_clear_options(&tp->rx_opt);
  3796. tp->rx_opt.mss_clamp = saved_clamp;
  3797. goto discard;
  3798. reset_and_undo:
  3799. tcp_clear_options(&tp->rx_opt);
  3800. tp->rx_opt.mss_clamp = saved_clamp;
  3801. return 1;
  3802. }
  3803. /*
  3804. * This function implements the receiving procedure of RFC 793 for
  3805. * all states except ESTABLISHED and TIME_WAIT.
  3806. * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
  3807. * address independent.
  3808. */
  3809. int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
  3810. struct tcphdr *th, unsigned len)
  3811. {
  3812. struct tcp_sock *tp = tcp_sk(sk);
  3813. struct inet_connection_sock *icsk = inet_csk(sk);
  3814. int queued = 0;
  3815. tp->rx_opt.saw_tstamp = 0;
  3816. switch (sk->sk_state) {
  3817. case TCP_CLOSE:
  3818. goto discard;
  3819. case TCP_LISTEN:
  3820. if(th->ack)
  3821. return 1;
  3822. if(th->rst)
  3823. goto discard;
  3824. if(th->syn) {
  3825. if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
  3826. return 1;
  3827. /* Now we have several options: In theory there is
  3828. * nothing else in the frame. KA9Q has an option to
  3829. * send data with the syn, BSD accepts data with the
  3830. * syn up to the [to be] advertised window and
  3831. * Solaris 2.1 gives you a protocol error. For now
  3832. * we just ignore it, that fits the spec precisely
  3833. * and avoids incompatibilities. It would be nice in
  3834. * future to drop through and process the data.
  3835. *
  3836. * Now that TTCP is starting to be used we ought to
  3837. * queue this data.
  3838. * But, this leaves one open to an easy denial of
  3839. * service attack, and SYN cookies can't defend
  3840. * against this problem. So, we drop the data
  3841. * in the interest of security over speed unless
  3842. * it's still in use.
  3843. */
  3844. kfree_skb(skb);
  3845. return 0;
  3846. }
  3847. goto discard;
  3848. case TCP_SYN_SENT:
  3849. queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
  3850. if (queued >= 0)
  3851. return queued;
  3852. /* Do step6 onward by hand. */
  3853. tcp_urg(sk, skb, th);
  3854. __kfree_skb(skb);
  3855. tcp_data_snd_check(sk, tp);
  3856. return 0;
  3857. }
  3858. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  3859. tcp_paws_discard(sk, skb)) {
  3860. if (!th->rst) {
  3861. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  3862. tcp_send_dupack(sk, skb);
  3863. goto discard;
  3864. }
  3865. /* Reset is accepted even if it did not pass PAWS. */
  3866. }
  3867. /* step 1: check sequence number */
  3868. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  3869. if (!th->rst)
  3870. tcp_send_dupack(sk, skb);
  3871. goto discard;
  3872. }
  3873. /* step 2: check RST bit */
  3874. if(th->rst) {
  3875. tcp_reset(sk);
  3876. goto discard;
  3877. }
  3878. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  3879. /* step 3: check security and precedence [ignored] */
  3880. /* step 4:
  3881. *
  3882. * Check for a SYN in window.
  3883. */
  3884. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3885. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  3886. tcp_reset(sk);
  3887. return 1;
  3888. }
  3889. /* step 5: check the ACK field */
  3890. if (th->ack) {
  3891. int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
  3892. switch(sk->sk_state) {
  3893. case TCP_SYN_RECV:
  3894. if (acceptable) {
  3895. tp->copied_seq = tp->rcv_nxt;
  3896. smp_mb();
  3897. tcp_set_state(sk, TCP_ESTABLISHED);
  3898. sk->sk_state_change(sk);
  3899. /* Note, that this wakeup is only for marginal
  3900. * crossed SYN case. Passively open sockets
  3901. * are not waked up, because sk->sk_sleep ==
  3902. * NULL and sk->sk_socket == NULL.
  3903. */
  3904. if (sk->sk_socket) {
  3905. sk_wake_async(sk,0,POLL_OUT);
  3906. }
  3907. tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
  3908. tp->snd_wnd = ntohs(th->window) <<
  3909. tp->rx_opt.snd_wscale;
  3910. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
  3911. TCP_SKB_CB(skb)->seq);
  3912. /* tcp_ack considers this ACK as duplicate
  3913. * and does not calculate rtt.
  3914. * Fix it at least with timestamps.
  3915. */
  3916. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  3917. !tp->srtt)
  3918. tcp_ack_saw_tstamp(sk, 0);
  3919. if (tp->rx_opt.tstamp_ok)
  3920. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  3921. /* Make sure socket is routed, for
  3922. * correct metrics.
  3923. */
  3924. icsk->icsk_af_ops->rebuild_header(sk);
  3925. tcp_init_metrics(sk);
  3926. tcp_init_congestion_control(sk);
  3927. /* Prevent spurious tcp_cwnd_restart() on
  3928. * first data packet.
  3929. */
  3930. tp->lsndtime = tcp_time_stamp;
  3931. tcp_mtup_init(sk);
  3932. tcp_initialize_rcv_mss(sk);
  3933. tcp_init_buffer_space(sk);
  3934. tcp_fast_path_on(tp);
  3935. } else {
  3936. return 1;
  3937. }
  3938. break;
  3939. case TCP_FIN_WAIT1:
  3940. if (tp->snd_una == tp->write_seq) {
  3941. tcp_set_state(sk, TCP_FIN_WAIT2);
  3942. sk->sk_shutdown |= SEND_SHUTDOWN;
  3943. dst_confirm(sk->sk_dst_cache);
  3944. if (!sock_flag(sk, SOCK_DEAD))
  3945. /* Wake up lingering close() */
  3946. sk->sk_state_change(sk);
  3947. else {
  3948. int tmo;
  3949. if (tp->linger2 < 0 ||
  3950. (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  3951. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
  3952. tcp_done(sk);
  3953. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  3954. return 1;
  3955. }
  3956. tmo = tcp_fin_time(sk);
  3957. if (tmo > TCP_TIMEWAIT_LEN) {
  3958. inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
  3959. } else if (th->fin || sock_owned_by_user(sk)) {
  3960. /* Bad case. We could lose such FIN otherwise.
  3961. * It is not a big problem, but it looks confusing
  3962. * and not so rare event. We still can lose it now,
  3963. * if it spins in bh_lock_sock(), but it is really
  3964. * marginal case.
  3965. */
  3966. inet_csk_reset_keepalive_timer(sk, tmo);
  3967. } else {
  3968. tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
  3969. goto discard;
  3970. }
  3971. }
  3972. }
  3973. break;
  3974. case TCP_CLOSING:
  3975. if (tp->snd_una == tp->write_seq) {
  3976. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  3977. goto discard;
  3978. }
  3979. break;
  3980. case TCP_LAST_ACK:
  3981. if (tp->snd_una == tp->write_seq) {
  3982. tcp_update_metrics(sk);
  3983. tcp_done(sk);
  3984. goto discard;
  3985. }
  3986. break;
  3987. }
  3988. } else
  3989. goto discard;
  3990. /* step 6: check the URG bit */
  3991. tcp_urg(sk, skb, th);
  3992. /* step 7: process the segment text */
  3993. switch (sk->sk_state) {
  3994. case TCP_CLOSE_WAIT:
  3995. case TCP_CLOSING:
  3996. case TCP_LAST_ACK:
  3997. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  3998. break;
  3999. case TCP_FIN_WAIT1:
  4000. case TCP_FIN_WAIT2:
  4001. /* RFC 793 says to queue data in these states,
  4002. * RFC 1122 says we MUST send a reset.
  4003. * BSD 4.4 also does reset.
  4004. */
  4005. if (sk->sk_shutdown & RCV_SHUTDOWN) {
  4006. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4007. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
  4008. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4009. tcp_reset(sk);
  4010. return 1;
  4011. }
  4012. }
  4013. /* Fall through */
  4014. case TCP_ESTABLISHED:
  4015. tcp_data_queue(sk, skb);
  4016. queued = 1;
  4017. break;
  4018. }
  4019. /* tcp_data could move socket to TIME-WAIT */
  4020. if (sk->sk_state != TCP_CLOSE) {
  4021. tcp_data_snd_check(sk, tp);
  4022. tcp_ack_snd_check(sk);
  4023. }
  4024. if (!queued) {
  4025. discard:
  4026. __kfree_skb(skb);
  4027. }
  4028. return 0;
  4029. }
  4030. EXPORT_SYMBOL(sysctl_tcp_ecn);
  4031. EXPORT_SYMBOL(sysctl_tcp_reordering);
  4032. EXPORT_SYMBOL(tcp_parse_options);
  4033. EXPORT_SYMBOL(tcp_rcv_established);
  4034. EXPORT_SYMBOL(tcp_rcv_state_process);
  4035. EXPORT_SYMBOL(tcp_initialize_rcv_mss);