tcp_input.c 139 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138413941404141414241434144414541464147414841494150415141524153415441554156415741584159416041614162416341644165416641674168416941704171417241734174417541764177417841794180418141824183418441854186418741884189419041914192419341944195419641974198419942004201420242034204420542064207420842094210421142124213421442154216421742184219422042214222422342244225422642274228422942304231423242334234423542364237423842394240424142424243424442454246424742484249425042514252425342544255425642574258425942604261426242634264426542664267426842694270427142724273427442754276427742784279428042814282428342844285428642874288428942904291429242934294429542964297429842994300430143024303430443054306430743084309431043114312431343144315431643174318431943204321432243234324432543264327432843294330433143324333433443354336433743384339434043414342434343444345434643474348434943504351435243534354435543564357435843594360436143624363436443654366436743684369437043714372437343744375437643774378437943804381438243834384438543864387438843894390439143924393439443954396439743984399440044014402440344044405440644074408440944104411441244134414441544164417441844194420442144224423442444254426442744284429443044314432443344344435443644374438443944404441444244434444444544464447444844494450445144524453445444554456445744584459446044614462446344644465446644674468446944704471447244734474447544764477447844794480448144824483448444854486448744884489449044914492449344944495449644974498449945004501450245034504450545064507450845094510451145124513451445154516451745184519452045214522452345244525452645274528452945304531453245334534453545364537453845394540454145424543454445454546454745484549455045514552455345544555455645574558455945604561456245634564456545664567456845694570457145724573457445754576457745784579458045814582458345844585458645874588458945904591459245934594459545964597459845994600460146024603460446054606460746084609461046114612461346144615461646174618461946204621462246234624462546264627462846294630463146324633463446354636463746384639464046414642464346444645464646474648464946504651465246534654465546564657465846594660466146624663466446654666466746684669467046714672467346744675467646774678467946804681468246834684468546864687468846894690469146924693469446954696469746984699470047014702470347044705470647074708470947104711471247134714471547164717471847194720472147224723472447254726472747284729473047314732473347344735473647374738473947404741474247434744474547464747474847494750475147524753475447554756475747584759476047614762476347644765476647674768476947704771477247734774477547764777477847794780478147824783478447854786478747884789479047914792479347944795479647974798479948004801480248034804480548064807480848094810481148124813481448154816481748184819482048214822482348244825482648274828482948304831483248334834483548364837483848394840484148424843484448454846484748484849485048514852485348544855485648574858485948604861486248634864486548664867486848694870487148724873487448754876487748784879488048814882488348844885488648874888488948904891489248934894489548964897489848994900490149024903490449054906490749084909491049114912491349144915
  1. /*
  2. * INET An implementation of the TCP/IP protocol suite for the LINUX
  3. * operating system. INET is implemented using the BSD Socket
  4. * interface as the means of communication with the user level.
  5. *
  6. * Implementation of the Transmission Control Protocol(TCP).
  7. *
  8. * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $
  9. *
  10. * Authors: Ross Biro
  11. * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  12. * Mark Evans, <evansmp@uhura.aston.ac.uk>
  13. * Corey Minyard <wf-rch!minyard@relay.EU.net>
  14. * Florian La Roche, <flla@stud.uni-sb.de>
  15. * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  16. * Linus Torvalds, <torvalds@cs.helsinki.fi>
  17. * Alan Cox, <gw4pts@gw4pts.ampr.org>
  18. * Matthew Dillon, <dillon@apollo.west.oic.com>
  19. * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  20. * Jorge Cwik, <jorge@laser.satlink.net>
  21. */
  22. /*
  23. * Changes:
  24. * Pedro Roque : Fast Retransmit/Recovery.
  25. * Two receive queues.
  26. * Retransmit queue handled by TCP.
  27. * Better retransmit timer handling.
  28. * New congestion avoidance.
  29. * Header prediction.
  30. * Variable renaming.
  31. *
  32. * Eric : Fast Retransmit.
  33. * Randy Scott : MSS option defines.
  34. * Eric Schenk : Fixes to slow start algorithm.
  35. * Eric Schenk : Yet another double ACK bug.
  36. * Eric Schenk : Delayed ACK bug fixes.
  37. * Eric Schenk : Floyd style fast retrans war avoidance.
  38. * David S. Miller : Don't allow zero congestion window.
  39. * Eric Schenk : Fix retransmitter so that it sends
  40. * next packet on ack of previous packet.
  41. * Andi Kleen : Moved open_request checking here
  42. * and process RSTs for open_requests.
  43. * Andi Kleen : Better prune_queue, and other fixes.
  44. * Andrey Savochkin: Fix RTT measurements in the presence of
  45. * timestamps.
  46. * Andrey Savochkin: Check sequence numbers correctly when
  47. * removing SACKs due to in sequence incoming
  48. * data segments.
  49. * Andi Kleen: Make sure we never ack data there is not
  50. * enough room for. Also make this condition
  51. * a fatal error if it might still happen.
  52. * Andi Kleen: Add tcp_measure_rcv_mss to make
  53. * connections with MSS<min(MTU,ann. MSS)
  54. * work without delayed acks.
  55. * Andi Kleen: Process packets with PSH set in the
  56. * fast path.
  57. * J Hadi Salim: ECN support
  58. * Andrei Gurtov,
  59. * Pasi Sarolahti,
  60. * Panu Kuhlberg: Experimental audit of TCP (re)transmission
  61. * engine. Lots of bugs are found.
  62. * Pasi Sarolahti: F-RTO for dealing with spurious RTOs
  63. */
  64. #include <linux/mm.h>
  65. #include <linux/module.h>
  66. #include <linux/sysctl.h>
  67. #include <net/tcp.h>
  68. #include <net/inet_common.h>
  69. #include <linux/ipsec.h>
  70. #include <asm/unaligned.h>
  71. #include <net/netdma.h>
  72. int sysctl_tcp_timestamps __read_mostly = 1;
  73. int sysctl_tcp_window_scaling __read_mostly = 1;
  74. int sysctl_tcp_sack __read_mostly = 1;
  75. int sysctl_tcp_fack __read_mostly = 1;
  76. int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  77. int sysctl_tcp_ecn __read_mostly;
  78. int sysctl_tcp_dsack __read_mostly = 1;
  79. int sysctl_tcp_app_win __read_mostly = 31;
  80. int sysctl_tcp_adv_win_scale __read_mostly = 2;
  81. int sysctl_tcp_stdurg __read_mostly;
  82. int sysctl_tcp_rfc1337 __read_mostly;
  83. int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  84. int sysctl_tcp_frto __read_mostly;
  85. int sysctl_tcp_frto_response __read_mostly;
  86. int sysctl_tcp_nometrics_save __read_mostly;
  87. int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
  88. int sysctl_tcp_abc __read_mostly;
  89. #define FLAG_DATA 0x01 /* Incoming frame contained data. */
  90. #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */
  91. #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */
  92. #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */
  93. #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */
  94. #define FLAG_DATA_SACKED 0x20 /* New SACK. */
  95. #define FLAG_ECE 0x40 /* ECE in this ACK */
  96. #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */
  97. #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/
  98. #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */
  99. #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED)
  100. #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
  101. #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE)
  102. #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED)
  103. #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0)
  104. #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2)
  105. #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4)
  106. #define IsSackFrto() (sysctl_tcp_frto == 0x2)
  107. #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
  108. /* Adapt the MSS value used to make delayed ack decision to the
  109. * real world.
  110. */
  111. static void tcp_measure_rcv_mss(struct sock *sk,
  112. const struct sk_buff *skb)
  113. {
  114. struct inet_connection_sock *icsk = inet_csk(sk);
  115. const unsigned int lss = icsk->icsk_ack.last_seg_size;
  116. unsigned int len;
  117. icsk->icsk_ack.last_seg_size = 0;
  118. /* skb->len may jitter because of SACKs, even if peer
  119. * sends good full-sized frames.
  120. */
  121. len = skb_shinfo(skb)->gso_size ?: skb->len;
  122. if (len >= icsk->icsk_ack.rcv_mss) {
  123. icsk->icsk_ack.rcv_mss = len;
  124. } else {
  125. /* Otherwise, we make more careful check taking into account,
  126. * that SACKs block is variable.
  127. *
  128. * "len" is invariant segment length, including TCP header.
  129. */
  130. len += skb->data - skb_transport_header(skb);
  131. if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) ||
  132. /* If PSH is not set, packet should be
  133. * full sized, provided peer TCP is not badly broken.
  134. * This observation (if it is correct 8)) allows
  135. * to handle super-low mtu links fairly.
  136. */
  137. (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
  138. !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
  139. /* Subtract also invariant (if peer is RFC compliant),
  140. * tcp header plus fixed timestamp option length.
  141. * Resulting "len" is MSS free of SACK jitter.
  142. */
  143. len -= tcp_sk(sk)->tcp_header_len;
  144. icsk->icsk_ack.last_seg_size = len;
  145. if (len == lss) {
  146. icsk->icsk_ack.rcv_mss = len;
  147. return;
  148. }
  149. }
  150. if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
  151. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
  152. icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
  153. }
  154. }
  155. static void tcp_incr_quickack(struct sock *sk)
  156. {
  157. struct inet_connection_sock *icsk = inet_csk(sk);
  158. unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
  159. if (quickacks==0)
  160. quickacks=2;
  161. if (quickacks > icsk->icsk_ack.quick)
  162. icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
  163. }
  164. void tcp_enter_quickack_mode(struct sock *sk)
  165. {
  166. struct inet_connection_sock *icsk = inet_csk(sk);
  167. tcp_incr_quickack(sk);
  168. icsk->icsk_ack.pingpong = 0;
  169. icsk->icsk_ack.ato = TCP_ATO_MIN;
  170. }
  171. /* Send ACKs quickly, if "quick" count is not exhausted
  172. * and the session is not interactive.
  173. */
  174. static inline int tcp_in_quickack_mode(const struct sock *sk)
  175. {
  176. const struct inet_connection_sock *icsk = inet_csk(sk);
  177. return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
  178. }
  179. /* Buffer size and advertised window tuning.
  180. *
  181. * 1. Tuning sk->sk_sndbuf, when connection enters established state.
  182. */
  183. static void tcp_fixup_sndbuf(struct sock *sk)
  184. {
  185. int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 +
  186. sizeof(struct sk_buff);
  187. if (sk->sk_sndbuf < 3 * sndmem)
  188. sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]);
  189. }
  190. /* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
  191. *
  192. * All tcp_full_space() is split to two parts: "network" buffer, allocated
  193. * forward and advertised in receiver window (tp->rcv_wnd) and
  194. * "application buffer", required to isolate scheduling/application
  195. * latencies from network.
  196. * window_clamp is maximal advertised window. It can be less than
  197. * tcp_full_space(), in this case tcp_full_space() - window_clamp
  198. * is reserved for "application" buffer. The less window_clamp is
  199. * the smoother our behaviour from viewpoint of network, but the lower
  200. * throughput and the higher sensitivity of the connection to losses. 8)
  201. *
  202. * rcv_ssthresh is more strict window_clamp used at "slow start"
  203. * phase to predict further behaviour of this connection.
  204. * It is used for two goals:
  205. * - to enforce header prediction at sender, even when application
  206. * requires some significant "application buffer". It is check #1.
  207. * - to prevent pruning of receive queue because of misprediction
  208. * of receiver window. Check #2.
  209. *
  210. * The scheme does not work when sender sends good segments opening
  211. * window and then starts to feed us spaghetti. But it should work
  212. * in common situations. Otherwise, we have to rely on queue collapsing.
  213. */
  214. /* Slow part of check#2. */
  215. static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
  216. {
  217. struct tcp_sock *tp = tcp_sk(sk);
  218. /* Optimize this! */
  219. int truesize = tcp_win_from_space(skb->truesize)/2;
  220. int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2;
  221. while (tp->rcv_ssthresh <= window) {
  222. if (truesize <= skb->len)
  223. return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
  224. truesize >>= 1;
  225. window >>= 1;
  226. }
  227. return 0;
  228. }
  229. static void tcp_grow_window(struct sock *sk,
  230. struct sk_buff *skb)
  231. {
  232. struct tcp_sock *tp = tcp_sk(sk);
  233. /* Check #1 */
  234. if (tp->rcv_ssthresh < tp->window_clamp &&
  235. (int)tp->rcv_ssthresh < tcp_space(sk) &&
  236. !tcp_memory_pressure) {
  237. int incr;
  238. /* Check #2. Increase window, if skb with such overhead
  239. * will fit to rcvbuf in future.
  240. */
  241. if (tcp_win_from_space(skb->truesize) <= skb->len)
  242. incr = 2*tp->advmss;
  243. else
  244. incr = __tcp_grow_window(sk, skb);
  245. if (incr) {
  246. tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp);
  247. inet_csk(sk)->icsk_ack.quick |= 1;
  248. }
  249. }
  250. }
  251. /* 3. Tuning rcvbuf, when connection enters established state. */
  252. static void tcp_fixup_rcvbuf(struct sock *sk)
  253. {
  254. struct tcp_sock *tp = tcp_sk(sk);
  255. int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff);
  256. /* Try to select rcvbuf so that 4 mss-sized segments
  257. * will fit to window and corresponding skbs will fit to our rcvbuf.
  258. * (was 3; 4 is minimum to allow fast retransmit to work.)
  259. */
  260. while (tcp_win_from_space(rcvmem) < tp->advmss)
  261. rcvmem += 128;
  262. if (sk->sk_rcvbuf < 4 * rcvmem)
  263. sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]);
  264. }
  265. /* 4. Try to fixup all. It is made immediately after connection enters
  266. * established state.
  267. */
  268. static void tcp_init_buffer_space(struct sock *sk)
  269. {
  270. struct tcp_sock *tp = tcp_sk(sk);
  271. int maxwin;
  272. if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
  273. tcp_fixup_rcvbuf(sk);
  274. if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
  275. tcp_fixup_sndbuf(sk);
  276. tp->rcvq_space.space = tp->rcv_wnd;
  277. maxwin = tcp_full_space(sk);
  278. if (tp->window_clamp >= maxwin) {
  279. tp->window_clamp = maxwin;
  280. if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
  281. tp->window_clamp = max(maxwin -
  282. (maxwin >> sysctl_tcp_app_win),
  283. 4 * tp->advmss);
  284. }
  285. /* Force reservation of one segment. */
  286. if (sysctl_tcp_app_win &&
  287. tp->window_clamp > 2 * tp->advmss &&
  288. tp->window_clamp + tp->advmss > maxwin)
  289. tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
  290. tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
  291. tp->snd_cwnd_stamp = tcp_time_stamp;
  292. }
  293. /* 5. Recalculate window clamp after socket hit its memory bounds. */
  294. static void tcp_clamp_window(struct sock *sk)
  295. {
  296. struct tcp_sock *tp = tcp_sk(sk);
  297. struct inet_connection_sock *icsk = inet_csk(sk);
  298. icsk->icsk_ack.quick = 0;
  299. if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
  300. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
  301. !tcp_memory_pressure &&
  302. atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) {
  303. sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
  304. sysctl_tcp_rmem[2]);
  305. }
  306. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
  307. tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss);
  308. }
  309. /* Initialize RCV_MSS value.
  310. * RCV_MSS is an our guess about MSS used by the peer.
  311. * We haven't any direct information about the MSS.
  312. * It's better to underestimate the RCV_MSS rather than overestimate.
  313. * Overestimations make us ACKing less frequently than needed.
  314. * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
  315. */
  316. void tcp_initialize_rcv_mss(struct sock *sk)
  317. {
  318. struct tcp_sock *tp = tcp_sk(sk);
  319. unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
  320. hint = min(hint, tp->rcv_wnd/2);
  321. hint = min(hint, TCP_MIN_RCVMSS);
  322. hint = max(hint, TCP_MIN_MSS);
  323. inet_csk(sk)->icsk_ack.rcv_mss = hint;
  324. }
  325. /* Receiver "autotuning" code.
  326. *
  327. * The algorithm for RTT estimation w/o timestamps is based on
  328. * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
  329. * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps>
  330. *
  331. * More detail on this code can be found at
  332. * <http://www.psc.edu/~jheffner/senior_thesis.ps>,
  333. * though this reference is out of date. A new paper
  334. * is pending.
  335. */
  336. static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
  337. {
  338. u32 new_sample = tp->rcv_rtt_est.rtt;
  339. long m = sample;
  340. if (m == 0)
  341. m = 1;
  342. if (new_sample != 0) {
  343. /* If we sample in larger samples in the non-timestamp
  344. * case, we could grossly overestimate the RTT especially
  345. * with chatty applications or bulk transfer apps which
  346. * are stalled on filesystem I/O.
  347. *
  348. * Also, since we are only going for a minimum in the
  349. * non-timestamp case, we do not smooth things out
  350. * else with timestamps disabled convergence takes too
  351. * long.
  352. */
  353. if (!win_dep) {
  354. m -= (new_sample >> 3);
  355. new_sample += m;
  356. } else if (m < new_sample)
  357. new_sample = m << 3;
  358. } else {
  359. /* No previous measure. */
  360. new_sample = m << 3;
  361. }
  362. if (tp->rcv_rtt_est.rtt != new_sample)
  363. tp->rcv_rtt_est.rtt = new_sample;
  364. }
  365. static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
  366. {
  367. if (tp->rcv_rtt_est.time == 0)
  368. goto new_measure;
  369. if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
  370. return;
  371. tcp_rcv_rtt_update(tp,
  372. jiffies - tp->rcv_rtt_est.time,
  373. 1);
  374. new_measure:
  375. tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
  376. tp->rcv_rtt_est.time = tcp_time_stamp;
  377. }
  378. static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb)
  379. {
  380. struct tcp_sock *tp = tcp_sk(sk);
  381. if (tp->rx_opt.rcv_tsecr &&
  382. (TCP_SKB_CB(skb)->end_seq -
  383. TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
  384. tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
  385. }
  386. /*
  387. * This function should be called every time data is copied to user space.
  388. * It calculates the appropriate TCP receive buffer space.
  389. */
  390. void tcp_rcv_space_adjust(struct sock *sk)
  391. {
  392. struct tcp_sock *tp = tcp_sk(sk);
  393. int time;
  394. int space;
  395. if (tp->rcvq_space.time == 0)
  396. goto new_measure;
  397. time = tcp_time_stamp - tp->rcvq_space.time;
  398. if (time < (tp->rcv_rtt_est.rtt >> 3) ||
  399. tp->rcv_rtt_est.rtt == 0)
  400. return;
  401. space = 2 * (tp->copied_seq - tp->rcvq_space.seq);
  402. space = max(tp->rcvq_space.space, space);
  403. if (tp->rcvq_space.space != space) {
  404. int rcvmem;
  405. tp->rcvq_space.space = space;
  406. if (sysctl_tcp_moderate_rcvbuf &&
  407. !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
  408. int new_clamp = space;
  409. /* Receive space grows, normalize in order to
  410. * take into account packet headers and sk_buff
  411. * structure overhead.
  412. */
  413. space /= tp->advmss;
  414. if (!space)
  415. space = 1;
  416. rcvmem = (tp->advmss + MAX_TCP_HEADER +
  417. 16 + sizeof(struct sk_buff));
  418. while (tcp_win_from_space(rcvmem) < tp->advmss)
  419. rcvmem += 128;
  420. space *= rcvmem;
  421. space = min(space, sysctl_tcp_rmem[2]);
  422. if (space > sk->sk_rcvbuf) {
  423. sk->sk_rcvbuf = space;
  424. /* Make the window clamp follow along. */
  425. tp->window_clamp = new_clamp;
  426. }
  427. }
  428. }
  429. new_measure:
  430. tp->rcvq_space.seq = tp->copied_seq;
  431. tp->rcvq_space.time = tcp_time_stamp;
  432. }
  433. /* There is something which you must keep in mind when you analyze the
  434. * behavior of the tp->ato delayed ack timeout interval. When a
  435. * connection starts up, we want to ack as quickly as possible. The
  436. * problem is that "good" TCP's do slow start at the beginning of data
  437. * transmission. The means that until we send the first few ACK's the
  438. * sender will sit on his end and only queue most of his data, because
  439. * he can only send snd_cwnd unacked packets at any given time. For
  440. * each ACK we send, he increments snd_cwnd and transmits more of his
  441. * queue. -DaveM
  442. */
  443. static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
  444. {
  445. struct tcp_sock *tp = tcp_sk(sk);
  446. struct inet_connection_sock *icsk = inet_csk(sk);
  447. u32 now;
  448. inet_csk_schedule_ack(sk);
  449. tcp_measure_rcv_mss(sk, skb);
  450. tcp_rcv_rtt_measure(tp);
  451. now = tcp_time_stamp;
  452. if (!icsk->icsk_ack.ato) {
  453. /* The _first_ data packet received, initialize
  454. * delayed ACK engine.
  455. */
  456. tcp_incr_quickack(sk);
  457. icsk->icsk_ack.ato = TCP_ATO_MIN;
  458. } else {
  459. int m = now - icsk->icsk_ack.lrcvtime;
  460. if (m <= TCP_ATO_MIN/2) {
  461. /* The fastest case is the first. */
  462. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
  463. } else if (m < icsk->icsk_ack.ato) {
  464. icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
  465. if (icsk->icsk_ack.ato > icsk->icsk_rto)
  466. icsk->icsk_ack.ato = icsk->icsk_rto;
  467. } else if (m > icsk->icsk_rto) {
  468. /* Too long gap. Apparently sender failed to
  469. * restart window, so that we send ACKs quickly.
  470. */
  471. tcp_incr_quickack(sk);
  472. sk_stream_mem_reclaim(sk);
  473. }
  474. }
  475. icsk->icsk_ack.lrcvtime = now;
  476. TCP_ECN_check_ce(tp, skb);
  477. if (skb->len >= 128)
  478. tcp_grow_window(sk, skb);
  479. }
  480. /* Called to compute a smoothed rtt estimate. The data fed to this
  481. * routine either comes from timestamps, or from segments that were
  482. * known _not_ to have been retransmitted [see Karn/Partridge
  483. * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
  484. * piece by Van Jacobson.
  485. * NOTE: the next three routines used to be one big routine.
  486. * To save cycles in the RFC 1323 implementation it was better to break
  487. * it up into three procedures. -- erics
  488. */
  489. static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt)
  490. {
  491. struct tcp_sock *tp = tcp_sk(sk);
  492. long m = mrtt; /* RTT */
  493. /* The following amusing code comes from Jacobson's
  494. * article in SIGCOMM '88. Note that rtt and mdev
  495. * are scaled versions of rtt and mean deviation.
  496. * This is designed to be as fast as possible
  497. * m stands for "measurement".
  498. *
  499. * On a 1990 paper the rto value is changed to:
  500. * RTO = rtt + 4 * mdev
  501. *
  502. * Funny. This algorithm seems to be very broken.
  503. * These formulae increase RTO, when it should be decreased, increase
  504. * too slowly, when it should be increased quickly, decrease too quickly
  505. * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
  506. * does not matter how to _calculate_ it. Seems, it was trap
  507. * that VJ failed to avoid. 8)
  508. */
  509. if (m == 0)
  510. m = 1;
  511. if (tp->srtt != 0) {
  512. m -= (tp->srtt >> 3); /* m is now error in rtt est */
  513. tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */
  514. if (m < 0) {
  515. m = -m; /* m is now abs(error) */
  516. m -= (tp->mdev >> 2); /* similar update on mdev */
  517. /* This is similar to one of Eifel findings.
  518. * Eifel blocks mdev updates when rtt decreases.
  519. * This solution is a bit different: we use finer gain
  520. * for mdev in this case (alpha*beta).
  521. * Like Eifel it also prevents growth of rto,
  522. * but also it limits too fast rto decreases,
  523. * happening in pure Eifel.
  524. */
  525. if (m > 0)
  526. m >>= 3;
  527. } else {
  528. m -= (tp->mdev >> 2); /* similar update on mdev */
  529. }
  530. tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */
  531. if (tp->mdev > tp->mdev_max) {
  532. tp->mdev_max = tp->mdev;
  533. if (tp->mdev_max > tp->rttvar)
  534. tp->rttvar = tp->mdev_max;
  535. }
  536. if (after(tp->snd_una, tp->rtt_seq)) {
  537. if (tp->mdev_max < tp->rttvar)
  538. tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2;
  539. tp->rtt_seq = tp->snd_nxt;
  540. tp->mdev_max = TCP_RTO_MIN;
  541. }
  542. } else {
  543. /* no previous measure. */
  544. tp->srtt = m<<3; /* take the measured time to be rtt */
  545. tp->mdev = m<<1; /* make sure rto = 3*rtt */
  546. tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
  547. tp->rtt_seq = tp->snd_nxt;
  548. }
  549. }
  550. /* Calculate rto without backoff. This is the second half of Van Jacobson's
  551. * routine referred to above.
  552. */
  553. static inline void tcp_set_rto(struct sock *sk)
  554. {
  555. const struct tcp_sock *tp = tcp_sk(sk);
  556. /* Old crap is replaced with new one. 8)
  557. *
  558. * More seriously:
  559. * 1. If rtt variance happened to be less 50msec, it is hallucination.
  560. * It cannot be less due to utterly erratic ACK generation made
  561. * at least by solaris and freebsd. "Erratic ACKs" has _nothing_
  562. * to do with delayed acks, because at cwnd>2 true delack timeout
  563. * is invisible. Actually, Linux-2.4 also generates erratic
  564. * ACKs in some circumstances.
  565. */
  566. inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar;
  567. /* 2. Fixups made earlier cannot be right.
  568. * If we do not estimate RTO correctly without them,
  569. * all the algo is pure shit and should be replaced
  570. * with correct one. It is exactly, which we pretend to do.
  571. */
  572. }
  573. /* NOTE: clamping at TCP_RTO_MIN is not required, current algo
  574. * guarantees that rto is higher.
  575. */
  576. static inline void tcp_bound_rto(struct sock *sk)
  577. {
  578. if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX)
  579. inet_csk(sk)->icsk_rto = TCP_RTO_MAX;
  580. }
  581. /* Save metrics learned by this TCP session.
  582. This function is called only, when TCP finishes successfully
  583. i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE.
  584. */
  585. void tcp_update_metrics(struct sock *sk)
  586. {
  587. struct tcp_sock *tp = tcp_sk(sk);
  588. struct dst_entry *dst = __sk_dst_get(sk);
  589. if (sysctl_tcp_nometrics_save)
  590. return;
  591. dst_confirm(dst);
  592. if (dst && (dst->flags&DST_HOST)) {
  593. const struct inet_connection_sock *icsk = inet_csk(sk);
  594. int m;
  595. if (icsk->icsk_backoff || !tp->srtt) {
  596. /* This session failed to estimate rtt. Why?
  597. * Probably, no packets returned in time.
  598. * Reset our results.
  599. */
  600. if (!(dst_metric_locked(dst, RTAX_RTT)))
  601. dst->metrics[RTAX_RTT-1] = 0;
  602. return;
  603. }
  604. m = dst_metric(dst, RTAX_RTT) - tp->srtt;
  605. /* If newly calculated rtt larger than stored one,
  606. * store new one. Otherwise, use EWMA. Remember,
  607. * rtt overestimation is always better than underestimation.
  608. */
  609. if (!(dst_metric_locked(dst, RTAX_RTT))) {
  610. if (m <= 0)
  611. dst->metrics[RTAX_RTT-1] = tp->srtt;
  612. else
  613. dst->metrics[RTAX_RTT-1] -= (m>>3);
  614. }
  615. if (!(dst_metric_locked(dst, RTAX_RTTVAR))) {
  616. if (m < 0)
  617. m = -m;
  618. /* Scale deviation to rttvar fixed point */
  619. m >>= 1;
  620. if (m < tp->mdev)
  621. m = tp->mdev;
  622. if (m >= dst_metric(dst, RTAX_RTTVAR))
  623. dst->metrics[RTAX_RTTVAR-1] = m;
  624. else
  625. dst->metrics[RTAX_RTTVAR-1] -=
  626. (dst->metrics[RTAX_RTTVAR-1] - m)>>2;
  627. }
  628. if (tp->snd_ssthresh >= 0xFFFF) {
  629. /* Slow start still did not finish. */
  630. if (dst_metric(dst, RTAX_SSTHRESH) &&
  631. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  632. (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH))
  633. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1;
  634. if (!dst_metric_locked(dst, RTAX_CWND) &&
  635. tp->snd_cwnd > dst_metric(dst, RTAX_CWND))
  636. dst->metrics[RTAX_CWND-1] = tp->snd_cwnd;
  637. } else if (tp->snd_cwnd > tp->snd_ssthresh &&
  638. icsk->icsk_ca_state == TCP_CA_Open) {
  639. /* Cong. avoidance phase, cwnd is reliable. */
  640. if (!dst_metric_locked(dst, RTAX_SSTHRESH))
  641. dst->metrics[RTAX_SSTHRESH-1] =
  642. max(tp->snd_cwnd >> 1, tp->snd_ssthresh);
  643. if (!dst_metric_locked(dst, RTAX_CWND))
  644. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1;
  645. } else {
  646. /* Else slow start did not finish, cwnd is non-sense,
  647. ssthresh may be also invalid.
  648. */
  649. if (!dst_metric_locked(dst, RTAX_CWND))
  650. dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1;
  651. if (dst->metrics[RTAX_SSTHRESH-1] &&
  652. !dst_metric_locked(dst, RTAX_SSTHRESH) &&
  653. tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1])
  654. dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh;
  655. }
  656. if (!dst_metric_locked(dst, RTAX_REORDERING)) {
  657. if (dst->metrics[RTAX_REORDERING-1] < tp->reordering &&
  658. tp->reordering != sysctl_tcp_reordering)
  659. dst->metrics[RTAX_REORDERING-1] = tp->reordering;
  660. }
  661. }
  662. }
  663. /* Numbers are taken from RFC2414. */
  664. __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst)
  665. {
  666. __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
  667. if (!cwnd) {
  668. if (tp->mss_cache > 1460)
  669. cwnd = 2;
  670. else
  671. cwnd = (tp->mss_cache > 1095) ? 3 : 4;
  672. }
  673. return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
  674. }
  675. /* Set slow start threshold and cwnd not falling to slow start */
  676. void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
  677. {
  678. struct tcp_sock *tp = tcp_sk(sk);
  679. const struct inet_connection_sock *icsk = inet_csk(sk);
  680. tp->prior_ssthresh = 0;
  681. tp->bytes_acked = 0;
  682. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  683. tp->undo_marker = 0;
  684. if (set_ssthresh)
  685. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  686. tp->snd_cwnd = min(tp->snd_cwnd,
  687. tcp_packets_in_flight(tp) + 1U);
  688. tp->snd_cwnd_cnt = 0;
  689. tp->high_seq = tp->snd_nxt;
  690. tp->snd_cwnd_stamp = tcp_time_stamp;
  691. TCP_ECN_queue_cwr(tp);
  692. tcp_set_ca_state(sk, TCP_CA_CWR);
  693. }
  694. }
  695. /* Initialize metrics on socket. */
  696. static void tcp_init_metrics(struct sock *sk)
  697. {
  698. struct tcp_sock *tp = tcp_sk(sk);
  699. struct dst_entry *dst = __sk_dst_get(sk);
  700. if (dst == NULL)
  701. goto reset;
  702. dst_confirm(dst);
  703. if (dst_metric_locked(dst, RTAX_CWND))
  704. tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND);
  705. if (dst_metric(dst, RTAX_SSTHRESH)) {
  706. tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH);
  707. if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
  708. tp->snd_ssthresh = tp->snd_cwnd_clamp;
  709. }
  710. if (dst_metric(dst, RTAX_REORDERING) &&
  711. tp->reordering != dst_metric(dst, RTAX_REORDERING)) {
  712. tp->rx_opt.sack_ok &= ~2;
  713. tp->reordering = dst_metric(dst, RTAX_REORDERING);
  714. }
  715. if (dst_metric(dst, RTAX_RTT) == 0)
  716. goto reset;
  717. if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3))
  718. goto reset;
  719. /* Initial rtt is determined from SYN,SYN-ACK.
  720. * The segment is small and rtt may appear much
  721. * less than real one. Use per-dst memory
  722. * to make it more realistic.
  723. *
  724. * A bit of theory. RTT is time passed after "normal" sized packet
  725. * is sent until it is ACKed. In normal circumstances sending small
  726. * packets force peer to delay ACKs and calculation is correct too.
  727. * The algorithm is adaptive and, provided we follow specs, it
  728. * NEVER underestimate RTT. BUT! If peer tries to make some clever
  729. * tricks sort of "quick acks" for time long enough to decrease RTT
  730. * to low value, and then abruptly stops to do it and starts to delay
  731. * ACKs, wait for troubles.
  732. */
  733. if (dst_metric(dst, RTAX_RTT) > tp->srtt) {
  734. tp->srtt = dst_metric(dst, RTAX_RTT);
  735. tp->rtt_seq = tp->snd_nxt;
  736. }
  737. if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) {
  738. tp->mdev = dst_metric(dst, RTAX_RTTVAR);
  739. tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN);
  740. }
  741. tcp_set_rto(sk);
  742. tcp_bound_rto(sk);
  743. if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp)
  744. goto reset;
  745. tp->snd_cwnd = tcp_init_cwnd(tp, dst);
  746. tp->snd_cwnd_stamp = tcp_time_stamp;
  747. return;
  748. reset:
  749. /* Play conservative. If timestamps are not
  750. * supported, TCP will fail to recalculate correct
  751. * rtt, if initial rto is too small. FORGET ALL AND RESET!
  752. */
  753. if (!tp->rx_opt.saw_tstamp && tp->srtt) {
  754. tp->srtt = 0;
  755. tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT;
  756. inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
  757. }
  758. }
  759. static void tcp_update_reordering(struct sock *sk, const int metric,
  760. const int ts)
  761. {
  762. struct tcp_sock *tp = tcp_sk(sk);
  763. if (metric > tp->reordering) {
  764. tp->reordering = min(TCP_MAX_REORDERING, metric);
  765. /* This exciting event is worth to be remembered. 8) */
  766. if (ts)
  767. NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER);
  768. else if (IsReno(tp))
  769. NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER);
  770. else if (IsFack(tp))
  771. NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER);
  772. else
  773. NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER);
  774. #if FASTRETRANS_DEBUG > 1
  775. printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n",
  776. tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
  777. tp->reordering,
  778. tp->fackets_out,
  779. tp->sacked_out,
  780. tp->undo_marker ? tp->undo_retrans : 0);
  781. #endif
  782. /* Disable FACK yet. */
  783. tp->rx_opt.sack_ok &= ~2;
  784. }
  785. }
  786. /* This procedure tags the retransmission queue when SACKs arrive.
  787. *
  788. * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
  789. * Packets in queue with these bits set are counted in variables
  790. * sacked_out, retrans_out and lost_out, correspondingly.
  791. *
  792. * Valid combinations are:
  793. * Tag InFlight Description
  794. * 0 1 - orig segment is in flight.
  795. * S 0 - nothing flies, orig reached receiver.
  796. * L 0 - nothing flies, orig lost by net.
  797. * R 2 - both orig and retransmit are in flight.
  798. * L|R 1 - orig is lost, retransmit is in flight.
  799. * S|R 1 - orig reached receiver, retrans is still in flight.
  800. * (L|S|R is logically valid, it could occur when L|R is sacked,
  801. * but it is equivalent to plain S and code short-curcuits it to S.
  802. * L|S is logically invalid, it would mean -1 packet in flight 8))
  803. *
  804. * These 6 states form finite state machine, controlled by the following events:
  805. * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
  806. * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
  807. * 3. Loss detection event of one of three flavors:
  808. * A. Scoreboard estimator decided the packet is lost.
  809. * A'. Reno "three dupacks" marks head of queue lost.
  810. * A''. Its FACK modfication, head until snd.fack is lost.
  811. * B. SACK arrives sacking data transmitted after never retransmitted
  812. * hole was sent out.
  813. * C. SACK arrives sacking SND.NXT at the moment, when the
  814. * segment was retransmitted.
  815. * 4. D-SACK added new rule: D-SACK changes any tag to S.
  816. *
  817. * It is pleasant to note, that state diagram turns out to be commutative,
  818. * so that we are allowed not to be bothered by order of our actions,
  819. * when multiple events arrive simultaneously. (see the function below).
  820. *
  821. * Reordering detection.
  822. * --------------------
  823. * Reordering metric is maximal distance, which a packet can be displaced
  824. * in packet stream. With SACKs we can estimate it:
  825. *
  826. * 1. SACK fills old hole and the corresponding segment was not
  827. * ever retransmitted -> reordering. Alas, we cannot use it
  828. * when segment was retransmitted.
  829. * 2. The last flaw is solved with D-SACK. D-SACK arrives
  830. * for retransmitted and already SACKed segment -> reordering..
  831. * Both of these heuristics are not used in Loss state, when we cannot
  832. * account for retransmits accurately.
  833. */
  834. static int
  835. tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una)
  836. {
  837. const struct inet_connection_sock *icsk = inet_csk(sk);
  838. struct tcp_sock *tp = tcp_sk(sk);
  839. unsigned char *ptr = (skb_transport_header(ack_skb) +
  840. TCP_SKB_CB(ack_skb)->sacked);
  841. struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2);
  842. struct sk_buff *cached_skb;
  843. int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3;
  844. int reord = tp->packets_out;
  845. int prior_fackets;
  846. u32 lost_retrans = 0;
  847. int flag = 0;
  848. int dup_sack = 0;
  849. int cached_fack_count;
  850. int i;
  851. int first_sack_index;
  852. if (!tp->sacked_out)
  853. tp->fackets_out = 0;
  854. prior_fackets = tp->fackets_out;
  855. /* Check for D-SACK. */
  856. if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) {
  857. dup_sack = 1;
  858. tp->rx_opt.sack_ok |= 4;
  859. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV);
  860. } else if (num_sacks > 1 &&
  861. !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) &&
  862. !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) {
  863. dup_sack = 1;
  864. tp->rx_opt.sack_ok |= 4;
  865. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV);
  866. }
  867. /* D-SACK for already forgotten data...
  868. * Do dumb counting. */
  869. if (dup_sack &&
  870. !after(ntohl(sp[0].end_seq), prior_snd_una) &&
  871. after(ntohl(sp[0].end_seq), tp->undo_marker))
  872. tp->undo_retrans--;
  873. /* Eliminate too old ACKs, but take into
  874. * account more or less fresh ones, they can
  875. * contain valid SACK info.
  876. */
  877. if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
  878. return 0;
  879. /* SACK fastpath:
  880. * if the only SACK change is the increase of the end_seq of
  881. * the first block then only apply that SACK block
  882. * and use retrans queue hinting otherwise slowpath */
  883. flag = 1;
  884. for (i = 0; i < num_sacks; i++) {
  885. __be32 start_seq = sp[i].start_seq;
  886. __be32 end_seq = sp[i].end_seq;
  887. if (i == 0) {
  888. if (tp->recv_sack_cache[i].start_seq != start_seq)
  889. flag = 0;
  890. } else {
  891. if ((tp->recv_sack_cache[i].start_seq != start_seq) ||
  892. (tp->recv_sack_cache[i].end_seq != end_seq))
  893. flag = 0;
  894. }
  895. tp->recv_sack_cache[i].start_seq = start_seq;
  896. tp->recv_sack_cache[i].end_seq = end_seq;
  897. }
  898. /* Clear the rest of the cache sack blocks so they won't match mistakenly. */
  899. for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) {
  900. tp->recv_sack_cache[i].start_seq = 0;
  901. tp->recv_sack_cache[i].end_seq = 0;
  902. }
  903. first_sack_index = 0;
  904. if (flag)
  905. num_sacks = 1;
  906. else {
  907. int j;
  908. tp->fastpath_skb_hint = NULL;
  909. /* order SACK blocks to allow in order walk of the retrans queue */
  910. for (i = num_sacks-1; i > 0; i--) {
  911. for (j = 0; j < i; j++){
  912. if (after(ntohl(sp[j].start_seq),
  913. ntohl(sp[j+1].start_seq))){
  914. struct tcp_sack_block_wire tmp;
  915. tmp = sp[j];
  916. sp[j] = sp[j+1];
  917. sp[j+1] = tmp;
  918. /* Track where the first SACK block goes to */
  919. if (j == first_sack_index)
  920. first_sack_index = j+1;
  921. }
  922. }
  923. }
  924. }
  925. /* clear flag as used for different purpose in following code */
  926. flag = 0;
  927. /* Use SACK fastpath hint if valid */
  928. cached_skb = tp->fastpath_skb_hint;
  929. cached_fack_count = tp->fastpath_cnt_hint;
  930. if (!cached_skb) {
  931. cached_skb = tcp_write_queue_head(sk);
  932. cached_fack_count = 0;
  933. }
  934. for (i=0; i<num_sacks; i++, sp++) {
  935. struct sk_buff *skb;
  936. __u32 start_seq = ntohl(sp->start_seq);
  937. __u32 end_seq = ntohl(sp->end_seq);
  938. int fack_count;
  939. skb = cached_skb;
  940. fack_count = cached_fack_count;
  941. /* Event "B" in the comment above. */
  942. if (after(end_seq, tp->high_seq))
  943. flag |= FLAG_DATA_LOST;
  944. tcp_for_write_queue_from(skb, sk) {
  945. int in_sack, pcount;
  946. u8 sacked;
  947. if (skb == tcp_send_head(sk))
  948. break;
  949. cached_skb = skb;
  950. cached_fack_count = fack_count;
  951. if (i == first_sack_index) {
  952. tp->fastpath_skb_hint = skb;
  953. tp->fastpath_cnt_hint = fack_count;
  954. }
  955. /* The retransmission queue is always in order, so
  956. * we can short-circuit the walk early.
  957. */
  958. if (!before(TCP_SKB_CB(skb)->seq, end_seq))
  959. break;
  960. in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
  961. !before(end_seq, TCP_SKB_CB(skb)->end_seq);
  962. pcount = tcp_skb_pcount(skb);
  963. if (pcount > 1 && !in_sack &&
  964. after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
  965. unsigned int pkt_len;
  966. in_sack = !after(start_seq,
  967. TCP_SKB_CB(skb)->seq);
  968. if (!in_sack)
  969. pkt_len = (start_seq -
  970. TCP_SKB_CB(skb)->seq);
  971. else
  972. pkt_len = (end_seq -
  973. TCP_SKB_CB(skb)->seq);
  974. if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size))
  975. break;
  976. pcount = tcp_skb_pcount(skb);
  977. }
  978. fack_count += pcount;
  979. sacked = TCP_SKB_CB(skb)->sacked;
  980. /* Account D-SACK for retransmitted packet. */
  981. if ((dup_sack && in_sack) &&
  982. (sacked & TCPCB_RETRANS) &&
  983. after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker))
  984. tp->undo_retrans--;
  985. /* The frame is ACKed. */
  986. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) {
  987. if (sacked&TCPCB_RETRANS) {
  988. if ((dup_sack && in_sack) &&
  989. (sacked&TCPCB_SACKED_ACKED))
  990. reord = min(fack_count, reord);
  991. } else {
  992. /* If it was in a hole, we detected reordering. */
  993. if (fack_count < prior_fackets &&
  994. !(sacked&TCPCB_SACKED_ACKED))
  995. reord = min(fack_count, reord);
  996. }
  997. /* Nothing to do; acked frame is about to be dropped. */
  998. continue;
  999. }
  1000. if ((sacked&TCPCB_SACKED_RETRANS) &&
  1001. after(end_seq, TCP_SKB_CB(skb)->ack_seq) &&
  1002. (!lost_retrans || after(end_seq, lost_retrans)))
  1003. lost_retrans = end_seq;
  1004. if (!in_sack)
  1005. continue;
  1006. if (!(sacked&TCPCB_SACKED_ACKED)) {
  1007. if (sacked & TCPCB_SACKED_RETRANS) {
  1008. /* If the segment is not tagged as lost,
  1009. * we do not clear RETRANS, believing
  1010. * that retransmission is still in flight.
  1011. */
  1012. if (sacked & TCPCB_LOST) {
  1013. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1014. tp->lost_out -= tcp_skb_pcount(skb);
  1015. tp->retrans_out -= tcp_skb_pcount(skb);
  1016. /* clear lost hint */
  1017. tp->retransmit_skb_hint = NULL;
  1018. }
  1019. } else {
  1020. /* New sack for not retransmitted frame,
  1021. * which was in hole. It is reordering.
  1022. */
  1023. if (!(sacked & TCPCB_RETRANS) &&
  1024. fack_count < prior_fackets)
  1025. reord = min(fack_count, reord);
  1026. if (sacked & TCPCB_LOST) {
  1027. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1028. tp->lost_out -= tcp_skb_pcount(skb);
  1029. /* clear lost hint */
  1030. tp->retransmit_skb_hint = NULL;
  1031. }
  1032. /* SACK enhanced F-RTO detection.
  1033. * Set flag if and only if non-rexmitted
  1034. * segments below frto_highmark are
  1035. * SACKed (RFC4138; Appendix B).
  1036. * Clearing correct due to in-order walk
  1037. */
  1038. if (after(end_seq, tp->frto_highmark)) {
  1039. flag &= ~FLAG_ONLY_ORIG_SACKED;
  1040. } else {
  1041. if (!(sacked & TCPCB_RETRANS))
  1042. flag |= FLAG_ONLY_ORIG_SACKED;
  1043. }
  1044. }
  1045. TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED;
  1046. flag |= FLAG_DATA_SACKED;
  1047. tp->sacked_out += tcp_skb_pcount(skb);
  1048. if (fack_count > tp->fackets_out)
  1049. tp->fackets_out = fack_count;
  1050. } else {
  1051. if (dup_sack && (sacked&TCPCB_RETRANS))
  1052. reord = min(fack_count, reord);
  1053. }
  1054. /* D-SACK. We can detect redundant retransmission
  1055. * in S|R and plain R frames and clear it.
  1056. * undo_retrans is decreased above, L|R frames
  1057. * are accounted above as well.
  1058. */
  1059. if (dup_sack &&
  1060. (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) {
  1061. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1062. tp->retrans_out -= tcp_skb_pcount(skb);
  1063. tp->retransmit_skb_hint = NULL;
  1064. }
  1065. }
  1066. }
  1067. /* Check for lost retransmit. This superb idea is
  1068. * borrowed from "ratehalving". Event "C".
  1069. * Later note: FACK people cheated me again 8),
  1070. * we have to account for reordering! Ugly,
  1071. * but should help.
  1072. */
  1073. if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) {
  1074. struct sk_buff *skb;
  1075. tcp_for_write_queue(skb, sk) {
  1076. if (skb == tcp_send_head(sk))
  1077. break;
  1078. if (after(TCP_SKB_CB(skb)->seq, lost_retrans))
  1079. break;
  1080. if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
  1081. continue;
  1082. if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) &&
  1083. after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) &&
  1084. (IsFack(tp) ||
  1085. !before(lost_retrans,
  1086. TCP_SKB_CB(skb)->ack_seq + tp->reordering *
  1087. tp->mss_cache))) {
  1088. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1089. tp->retrans_out -= tcp_skb_pcount(skb);
  1090. /* clear lost hint */
  1091. tp->retransmit_skb_hint = NULL;
  1092. if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) {
  1093. tp->lost_out += tcp_skb_pcount(skb);
  1094. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1095. flag |= FLAG_DATA_SACKED;
  1096. NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT);
  1097. }
  1098. }
  1099. }
  1100. }
  1101. tp->left_out = tp->sacked_out + tp->lost_out;
  1102. if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss &&
  1103. (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark)))
  1104. tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0);
  1105. #if FASTRETRANS_DEBUG > 0
  1106. BUG_TRAP((int)tp->sacked_out >= 0);
  1107. BUG_TRAP((int)tp->lost_out >= 0);
  1108. BUG_TRAP((int)tp->retrans_out >= 0);
  1109. BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0);
  1110. #endif
  1111. return flag;
  1112. }
  1113. /* F-RTO can only be used if TCP has never retransmitted anything other than
  1114. * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here)
  1115. */
  1116. int tcp_use_frto(struct sock *sk)
  1117. {
  1118. const struct tcp_sock *tp = tcp_sk(sk);
  1119. struct sk_buff *skb;
  1120. if (!sysctl_tcp_frto)
  1121. return 0;
  1122. if (IsSackFrto())
  1123. return 1;
  1124. /* Avoid expensive walking of rexmit queue if possible */
  1125. if (tp->retrans_out > 1)
  1126. return 0;
  1127. skb = tcp_write_queue_head(sk);
  1128. skb = tcp_write_queue_next(sk, skb); /* Skips head */
  1129. tcp_for_write_queue_from(skb, sk) {
  1130. if (skb == tcp_send_head(sk))
  1131. break;
  1132. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1133. return 0;
  1134. /* Short-circuit when first non-SACKed skb has been checked */
  1135. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED))
  1136. break;
  1137. }
  1138. return 1;
  1139. }
  1140. /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO
  1141. * recovery a bit and use heuristics in tcp_process_frto() to detect if
  1142. * the RTO was spurious. Only clear SACKED_RETRANS of the head here to
  1143. * keep retrans_out counting accurate (with SACK F-RTO, other than head
  1144. * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS
  1145. * bits are handled if the Loss state is really to be entered (in
  1146. * tcp_enter_frto_loss).
  1147. *
  1148. * Do like tcp_enter_loss() would; when RTO expires the second time it
  1149. * does:
  1150. * "Reduce ssthresh if it has not yet been made inside this window."
  1151. */
  1152. void tcp_enter_frto(struct sock *sk)
  1153. {
  1154. const struct inet_connection_sock *icsk = inet_csk(sk);
  1155. struct tcp_sock *tp = tcp_sk(sk);
  1156. struct sk_buff *skb;
  1157. if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) ||
  1158. tp->snd_una == tp->high_seq ||
  1159. ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) &&
  1160. !icsk->icsk_retransmits)) {
  1161. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1162. /* Our state is too optimistic in ssthresh() call because cwnd
  1163. * is not reduced until tcp_enter_frto_loss() when previous FRTO
  1164. * recovery has not yet completed. Pattern would be this: RTO,
  1165. * Cumulative ACK, RTO (2xRTO for the same segment does not end
  1166. * up here twice).
  1167. * RFC4138 should be more specific on what to do, even though
  1168. * RTO is quite unlikely to occur after the first Cumulative ACK
  1169. * due to back-off and complexity of triggering events ...
  1170. */
  1171. if (tp->frto_counter) {
  1172. u32 stored_cwnd;
  1173. stored_cwnd = tp->snd_cwnd;
  1174. tp->snd_cwnd = 2;
  1175. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1176. tp->snd_cwnd = stored_cwnd;
  1177. } else {
  1178. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1179. }
  1180. /* ... in theory, cong.control module could do "any tricks" in
  1181. * ssthresh(), which means that ca_state, lost bits and lost_out
  1182. * counter would have to be faked before the call occurs. We
  1183. * consider that too expensive, unlikely and hacky, so modules
  1184. * using these in ssthresh() must deal these incompatibility
  1185. * issues if they receives CA_EVENT_FRTO and frto_counter != 0
  1186. */
  1187. tcp_ca_event(sk, CA_EVENT_FRTO);
  1188. }
  1189. tp->undo_marker = tp->snd_una;
  1190. tp->undo_retrans = 0;
  1191. skb = tcp_write_queue_head(sk);
  1192. if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  1193. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  1194. tp->retrans_out -= tcp_skb_pcount(skb);
  1195. }
  1196. tcp_sync_left_out(tp);
  1197. /* Earlier loss recovery underway (see RFC4138; Appendix B).
  1198. * The last condition is necessary at least in tp->frto_counter case.
  1199. */
  1200. if (IsSackFrto() && (tp->frto_counter ||
  1201. ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) &&
  1202. after(tp->high_seq, tp->snd_una)) {
  1203. tp->frto_highmark = tp->high_seq;
  1204. } else {
  1205. tp->frto_highmark = tp->snd_nxt;
  1206. }
  1207. tcp_set_ca_state(sk, TCP_CA_Disorder);
  1208. tp->high_seq = tp->snd_nxt;
  1209. tp->frto_counter = 1;
  1210. }
  1211. /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO,
  1212. * which indicates that we should follow the traditional RTO recovery,
  1213. * i.e. mark everything lost and do go-back-N retransmission.
  1214. */
  1215. static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag)
  1216. {
  1217. struct tcp_sock *tp = tcp_sk(sk);
  1218. struct sk_buff *skb;
  1219. int cnt = 0;
  1220. tp->sacked_out = 0;
  1221. tp->lost_out = 0;
  1222. tp->fackets_out = 0;
  1223. tp->retrans_out = 0;
  1224. tcp_for_write_queue(skb, sk) {
  1225. if (skb == tcp_send_head(sk))
  1226. break;
  1227. cnt += tcp_skb_pcount(skb);
  1228. /*
  1229. * Count the retransmission made on RTO correctly (only when
  1230. * waiting for the first ACK and did not get it)...
  1231. */
  1232. if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) {
  1233. tp->retrans_out += tcp_skb_pcount(skb);
  1234. /* ...enter this if branch just for the first segment */
  1235. flag |= FLAG_DATA_ACKED;
  1236. } else {
  1237. TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
  1238. }
  1239. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
  1240. /* Do not mark those segments lost that were
  1241. * forward transmitted after RTO
  1242. */
  1243. if (!after(TCP_SKB_CB(skb)->end_seq,
  1244. tp->frto_highmark)) {
  1245. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1246. tp->lost_out += tcp_skb_pcount(skb);
  1247. }
  1248. } else {
  1249. tp->sacked_out += tcp_skb_pcount(skb);
  1250. tp->fackets_out = cnt;
  1251. }
  1252. }
  1253. tcp_sync_left_out(tp);
  1254. tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments;
  1255. tp->snd_cwnd_cnt = 0;
  1256. tp->snd_cwnd_stamp = tcp_time_stamp;
  1257. tp->undo_marker = 0;
  1258. tp->frto_counter = 0;
  1259. tp->reordering = min_t(unsigned int, tp->reordering,
  1260. sysctl_tcp_reordering);
  1261. tcp_set_ca_state(sk, TCP_CA_Loss);
  1262. tp->high_seq = tp->frto_highmark;
  1263. TCP_ECN_queue_cwr(tp);
  1264. clear_all_retrans_hints(tp);
  1265. }
  1266. void tcp_clear_retrans(struct tcp_sock *tp)
  1267. {
  1268. tp->left_out = 0;
  1269. tp->retrans_out = 0;
  1270. tp->fackets_out = 0;
  1271. tp->sacked_out = 0;
  1272. tp->lost_out = 0;
  1273. tp->undo_marker = 0;
  1274. tp->undo_retrans = 0;
  1275. }
  1276. /* Enter Loss state. If "how" is not zero, forget all SACK information
  1277. * and reset tags completely, otherwise preserve SACKs. If receiver
  1278. * dropped its ofo queue, we will know this due to reneging detection.
  1279. */
  1280. void tcp_enter_loss(struct sock *sk, int how)
  1281. {
  1282. const struct inet_connection_sock *icsk = inet_csk(sk);
  1283. struct tcp_sock *tp = tcp_sk(sk);
  1284. struct sk_buff *skb;
  1285. int cnt = 0;
  1286. /* Reduce ssthresh if it has not yet been made inside this window. */
  1287. if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq ||
  1288. (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
  1289. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1290. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1291. tcp_ca_event(sk, CA_EVENT_LOSS);
  1292. }
  1293. tp->snd_cwnd = 1;
  1294. tp->snd_cwnd_cnt = 0;
  1295. tp->snd_cwnd_stamp = tcp_time_stamp;
  1296. tp->bytes_acked = 0;
  1297. tcp_clear_retrans(tp);
  1298. /* Push undo marker, if it was plain RTO and nothing
  1299. * was retransmitted. */
  1300. if (!how)
  1301. tp->undo_marker = tp->snd_una;
  1302. tcp_for_write_queue(skb, sk) {
  1303. if (skb == tcp_send_head(sk))
  1304. break;
  1305. cnt += tcp_skb_pcount(skb);
  1306. if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS)
  1307. tp->undo_marker = 0;
  1308. TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
  1309. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
  1310. TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
  1311. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1312. tp->lost_out += tcp_skb_pcount(skb);
  1313. } else {
  1314. tp->sacked_out += tcp_skb_pcount(skb);
  1315. tp->fackets_out = cnt;
  1316. }
  1317. }
  1318. tcp_sync_left_out(tp);
  1319. tp->reordering = min_t(unsigned int, tp->reordering,
  1320. sysctl_tcp_reordering);
  1321. tcp_set_ca_state(sk, TCP_CA_Loss);
  1322. tp->high_seq = tp->snd_nxt;
  1323. TCP_ECN_queue_cwr(tp);
  1324. /* Abort FRTO algorithm if one is in progress */
  1325. tp->frto_counter = 0;
  1326. clear_all_retrans_hints(tp);
  1327. }
  1328. static int tcp_check_sack_reneging(struct sock *sk)
  1329. {
  1330. struct sk_buff *skb;
  1331. /* If ACK arrived pointing to a remembered SACK,
  1332. * it means that our remembered SACKs do not reflect
  1333. * real state of receiver i.e.
  1334. * receiver _host_ is heavily congested (or buggy).
  1335. * Do processing similar to RTO timeout.
  1336. */
  1337. if ((skb = tcp_write_queue_head(sk)) != NULL &&
  1338. (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
  1339. struct inet_connection_sock *icsk = inet_csk(sk);
  1340. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING);
  1341. tcp_enter_loss(sk, 1);
  1342. icsk->icsk_retransmits++;
  1343. tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
  1344. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  1345. icsk->icsk_rto, TCP_RTO_MAX);
  1346. return 1;
  1347. }
  1348. return 0;
  1349. }
  1350. static inline int tcp_fackets_out(struct tcp_sock *tp)
  1351. {
  1352. return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out;
  1353. }
  1354. static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb)
  1355. {
  1356. return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto);
  1357. }
  1358. static inline int tcp_head_timedout(struct sock *sk)
  1359. {
  1360. struct tcp_sock *tp = tcp_sk(sk);
  1361. return tp->packets_out &&
  1362. tcp_skb_timedout(sk, tcp_write_queue_head(sk));
  1363. }
  1364. /* Linux NewReno/SACK/FACK/ECN state machine.
  1365. * --------------------------------------
  1366. *
  1367. * "Open" Normal state, no dubious events, fast path.
  1368. * "Disorder" In all the respects it is "Open",
  1369. * but requires a bit more attention. It is entered when
  1370. * we see some SACKs or dupacks. It is split of "Open"
  1371. * mainly to move some processing from fast path to slow one.
  1372. * "CWR" CWND was reduced due to some Congestion Notification event.
  1373. * It can be ECN, ICMP source quench, local device congestion.
  1374. * "Recovery" CWND was reduced, we are fast-retransmitting.
  1375. * "Loss" CWND was reduced due to RTO timeout or SACK reneging.
  1376. *
  1377. * tcp_fastretrans_alert() is entered:
  1378. * - each incoming ACK, if state is not "Open"
  1379. * - when arrived ACK is unusual, namely:
  1380. * * SACK
  1381. * * Duplicate ACK.
  1382. * * ECN ECE.
  1383. *
  1384. * Counting packets in flight is pretty simple.
  1385. *
  1386. * in_flight = packets_out - left_out + retrans_out
  1387. *
  1388. * packets_out is SND.NXT-SND.UNA counted in packets.
  1389. *
  1390. * retrans_out is number of retransmitted segments.
  1391. *
  1392. * left_out is number of segments left network, but not ACKed yet.
  1393. *
  1394. * left_out = sacked_out + lost_out
  1395. *
  1396. * sacked_out: Packets, which arrived to receiver out of order
  1397. * and hence not ACKed. With SACKs this number is simply
  1398. * amount of SACKed data. Even without SACKs
  1399. * it is easy to give pretty reliable estimate of this number,
  1400. * counting duplicate ACKs.
  1401. *
  1402. * lost_out: Packets lost by network. TCP has no explicit
  1403. * "loss notification" feedback from network (for now).
  1404. * It means that this number can be only _guessed_.
  1405. * Actually, it is the heuristics to predict lossage that
  1406. * distinguishes different algorithms.
  1407. *
  1408. * F.e. after RTO, when all the queue is considered as lost,
  1409. * lost_out = packets_out and in_flight = retrans_out.
  1410. *
  1411. * Essentially, we have now two algorithms counting
  1412. * lost packets.
  1413. *
  1414. * FACK: It is the simplest heuristics. As soon as we decided
  1415. * that something is lost, we decide that _all_ not SACKed
  1416. * packets until the most forward SACK are lost. I.e.
  1417. * lost_out = fackets_out - sacked_out and left_out = fackets_out.
  1418. * It is absolutely correct estimate, if network does not reorder
  1419. * packets. And it loses any connection to reality when reordering
  1420. * takes place. We use FACK by default until reordering
  1421. * is suspected on the path to this destination.
  1422. *
  1423. * NewReno: when Recovery is entered, we assume that one segment
  1424. * is lost (classic Reno). While we are in Recovery and
  1425. * a partial ACK arrives, we assume that one more packet
  1426. * is lost (NewReno). This heuristics are the same in NewReno
  1427. * and SACK.
  1428. *
  1429. * Imagine, that's all! Forget about all this shamanism about CWND inflation
  1430. * deflation etc. CWND is real congestion window, never inflated, changes
  1431. * only according to classic VJ rules.
  1432. *
  1433. * Really tricky (and requiring careful tuning) part of algorithm
  1434. * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
  1435. * The first determines the moment _when_ we should reduce CWND and,
  1436. * hence, slow down forward transmission. In fact, it determines the moment
  1437. * when we decide that hole is caused by loss, rather than by a reorder.
  1438. *
  1439. * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
  1440. * holes, caused by lost packets.
  1441. *
  1442. * And the most logically complicated part of algorithm is undo
  1443. * heuristics. We detect false retransmits due to both too early
  1444. * fast retransmit (reordering) and underestimated RTO, analyzing
  1445. * timestamps and D-SACKs. When we detect that some segments were
  1446. * retransmitted by mistake and CWND reduction was wrong, we undo
  1447. * window reduction and abort recovery phase. This logic is hidden
  1448. * inside several functions named tcp_try_undo_<something>.
  1449. */
  1450. /* This function decides, when we should leave Disordered state
  1451. * and enter Recovery phase, reducing congestion window.
  1452. *
  1453. * Main question: may we further continue forward transmission
  1454. * with the same cwnd?
  1455. */
  1456. static int tcp_time_to_recover(struct sock *sk)
  1457. {
  1458. struct tcp_sock *tp = tcp_sk(sk);
  1459. __u32 packets_out;
  1460. /* Do not perform any recovery during FRTO algorithm */
  1461. if (tp->frto_counter)
  1462. return 0;
  1463. /* Trick#1: The loss is proven. */
  1464. if (tp->lost_out)
  1465. return 1;
  1466. /* Not-A-Trick#2 : Classic rule... */
  1467. if (tcp_fackets_out(tp) > tp->reordering)
  1468. return 1;
  1469. /* Trick#3 : when we use RFC2988 timer restart, fast
  1470. * retransmit can be triggered by timeout of queue head.
  1471. */
  1472. if (tcp_head_timedout(sk))
  1473. return 1;
  1474. /* Trick#4: It is still not OK... But will it be useful to delay
  1475. * recovery more?
  1476. */
  1477. packets_out = tp->packets_out;
  1478. if (packets_out <= tp->reordering &&
  1479. tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
  1480. !tcp_may_send_now(sk)) {
  1481. /* We have nothing to send. This connection is limited
  1482. * either by receiver window or by application.
  1483. */
  1484. return 1;
  1485. }
  1486. return 0;
  1487. }
  1488. /* If we receive more dupacks than we expected counting segments
  1489. * in assumption of absent reordering, interpret this as reordering.
  1490. * The only another reason could be bug in receiver TCP.
  1491. */
  1492. static void tcp_check_reno_reordering(struct sock *sk, const int addend)
  1493. {
  1494. struct tcp_sock *tp = tcp_sk(sk);
  1495. u32 holes;
  1496. holes = max(tp->lost_out, 1U);
  1497. holes = min(holes, tp->packets_out);
  1498. if ((tp->sacked_out + holes) > tp->packets_out) {
  1499. tp->sacked_out = tp->packets_out - holes;
  1500. tcp_update_reordering(sk, tp->packets_out + addend, 0);
  1501. }
  1502. }
  1503. /* Emulate SACKs for SACKless connection: account for a new dupack. */
  1504. static void tcp_add_reno_sack(struct sock *sk)
  1505. {
  1506. struct tcp_sock *tp = tcp_sk(sk);
  1507. tp->sacked_out++;
  1508. tcp_check_reno_reordering(sk, 0);
  1509. tcp_sync_left_out(tp);
  1510. }
  1511. /* Account for ACK, ACKing some data in Reno Recovery phase. */
  1512. static void tcp_remove_reno_sacks(struct sock *sk, int acked)
  1513. {
  1514. struct tcp_sock *tp = tcp_sk(sk);
  1515. if (acked > 0) {
  1516. /* One ACK acked hole. The rest eat duplicate ACKs. */
  1517. if (acked-1 >= tp->sacked_out)
  1518. tp->sacked_out = 0;
  1519. else
  1520. tp->sacked_out -= acked-1;
  1521. }
  1522. tcp_check_reno_reordering(sk, acked);
  1523. tcp_sync_left_out(tp);
  1524. }
  1525. static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
  1526. {
  1527. tp->sacked_out = 0;
  1528. tp->left_out = tp->lost_out;
  1529. }
  1530. /* Mark head of queue up as lost. */
  1531. static void tcp_mark_head_lost(struct sock *sk,
  1532. int packets, u32 high_seq)
  1533. {
  1534. struct tcp_sock *tp = tcp_sk(sk);
  1535. struct sk_buff *skb;
  1536. int cnt;
  1537. BUG_TRAP(packets <= tp->packets_out);
  1538. if (tp->lost_skb_hint) {
  1539. skb = tp->lost_skb_hint;
  1540. cnt = tp->lost_cnt_hint;
  1541. } else {
  1542. skb = tcp_write_queue_head(sk);
  1543. cnt = 0;
  1544. }
  1545. tcp_for_write_queue_from(skb, sk) {
  1546. if (skb == tcp_send_head(sk))
  1547. break;
  1548. /* TODO: do this better */
  1549. /* this is not the most efficient way to do this... */
  1550. tp->lost_skb_hint = skb;
  1551. tp->lost_cnt_hint = cnt;
  1552. cnt += tcp_skb_pcount(skb);
  1553. if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq))
  1554. break;
  1555. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1556. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1557. tp->lost_out += tcp_skb_pcount(skb);
  1558. /* clear xmit_retransmit_queue hints
  1559. * if this is beyond hint */
  1560. if (tp->retransmit_skb_hint != NULL &&
  1561. before(TCP_SKB_CB(skb)->seq,
  1562. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  1563. tp->retransmit_skb_hint = NULL;
  1564. }
  1565. }
  1566. tcp_sync_left_out(tp);
  1567. }
  1568. /* Account newly detected lost packet(s) */
  1569. static void tcp_update_scoreboard(struct sock *sk)
  1570. {
  1571. struct tcp_sock *tp = tcp_sk(sk);
  1572. if (IsFack(tp)) {
  1573. int lost = tp->fackets_out - tp->reordering;
  1574. if (lost <= 0)
  1575. lost = 1;
  1576. tcp_mark_head_lost(sk, lost, tp->high_seq);
  1577. } else {
  1578. tcp_mark_head_lost(sk, 1, tp->high_seq);
  1579. }
  1580. /* New heuristics: it is possible only after we switched
  1581. * to restart timer each time when something is ACKed.
  1582. * Hence, we can detect timed out packets during fast
  1583. * retransmit without falling to slow start.
  1584. */
  1585. if (!IsReno(tp) && tcp_head_timedout(sk)) {
  1586. struct sk_buff *skb;
  1587. skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint
  1588. : tcp_write_queue_head(sk);
  1589. tcp_for_write_queue_from(skb, sk) {
  1590. if (skb == tcp_send_head(sk))
  1591. break;
  1592. if (!tcp_skb_timedout(sk, skb))
  1593. break;
  1594. if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) {
  1595. TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
  1596. tp->lost_out += tcp_skb_pcount(skb);
  1597. /* clear xmit_retrans hint */
  1598. if (tp->retransmit_skb_hint &&
  1599. before(TCP_SKB_CB(skb)->seq,
  1600. TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
  1601. tp->retransmit_skb_hint = NULL;
  1602. }
  1603. }
  1604. tp->scoreboard_skb_hint = skb;
  1605. tcp_sync_left_out(tp);
  1606. }
  1607. }
  1608. /* CWND moderation, preventing bursts due to too big ACKs
  1609. * in dubious situations.
  1610. */
  1611. static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
  1612. {
  1613. tp->snd_cwnd = min(tp->snd_cwnd,
  1614. tcp_packets_in_flight(tp)+tcp_max_burst(tp));
  1615. tp->snd_cwnd_stamp = tcp_time_stamp;
  1616. }
  1617. /* Lower bound on congestion window is slow start threshold
  1618. * unless congestion avoidance choice decides to overide it.
  1619. */
  1620. static inline u32 tcp_cwnd_min(const struct sock *sk)
  1621. {
  1622. const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
  1623. return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh;
  1624. }
  1625. /* Decrease cwnd each second ack. */
  1626. static void tcp_cwnd_down(struct sock *sk)
  1627. {
  1628. struct tcp_sock *tp = tcp_sk(sk);
  1629. int decr = tp->snd_cwnd_cnt + 1;
  1630. tp->snd_cwnd_cnt = decr&1;
  1631. decr >>= 1;
  1632. if (decr && tp->snd_cwnd > tcp_cwnd_min(sk))
  1633. tp->snd_cwnd -= decr;
  1634. tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1);
  1635. tp->snd_cwnd_stamp = tcp_time_stamp;
  1636. }
  1637. /* Nothing was retransmitted or returned timestamp is less
  1638. * than timestamp of the first retransmission.
  1639. */
  1640. static inline int tcp_packet_delayed(struct tcp_sock *tp)
  1641. {
  1642. return !tp->retrans_stamp ||
  1643. (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  1644. (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0);
  1645. }
  1646. /* Undo procedures. */
  1647. #if FASTRETRANS_DEBUG > 1
  1648. static void DBGUNDO(struct sock *sk, const char *msg)
  1649. {
  1650. struct tcp_sock *tp = tcp_sk(sk);
  1651. struct inet_sock *inet = inet_sk(sk);
  1652. printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n",
  1653. msg,
  1654. NIPQUAD(inet->daddr), ntohs(inet->dport),
  1655. tp->snd_cwnd, tp->left_out,
  1656. tp->snd_ssthresh, tp->prior_ssthresh,
  1657. tp->packets_out);
  1658. }
  1659. #else
  1660. #define DBGUNDO(x...) do { } while (0)
  1661. #endif
  1662. static void tcp_undo_cwr(struct sock *sk, const int undo)
  1663. {
  1664. struct tcp_sock *tp = tcp_sk(sk);
  1665. if (tp->prior_ssthresh) {
  1666. const struct inet_connection_sock *icsk = inet_csk(sk);
  1667. if (icsk->icsk_ca_ops->undo_cwnd)
  1668. tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
  1669. else
  1670. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1);
  1671. if (undo && tp->prior_ssthresh > tp->snd_ssthresh) {
  1672. tp->snd_ssthresh = tp->prior_ssthresh;
  1673. TCP_ECN_withdraw_cwr(tp);
  1674. }
  1675. } else {
  1676. tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
  1677. }
  1678. tcp_moderate_cwnd(tp);
  1679. tp->snd_cwnd_stamp = tcp_time_stamp;
  1680. /* There is something screwy going on with the retrans hints after
  1681. an undo */
  1682. clear_all_retrans_hints(tp);
  1683. }
  1684. static inline int tcp_may_undo(struct tcp_sock *tp)
  1685. {
  1686. return tp->undo_marker &&
  1687. (!tp->undo_retrans || tcp_packet_delayed(tp));
  1688. }
  1689. /* People celebrate: "We love our President!" */
  1690. static int tcp_try_undo_recovery(struct sock *sk)
  1691. {
  1692. struct tcp_sock *tp = tcp_sk(sk);
  1693. if (tcp_may_undo(tp)) {
  1694. /* Happy end! We did not retransmit anything
  1695. * or our original transmission succeeded.
  1696. */
  1697. DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
  1698. tcp_undo_cwr(sk, 1);
  1699. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
  1700. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1701. else
  1702. NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO);
  1703. tp->undo_marker = 0;
  1704. }
  1705. if (tp->snd_una == tp->high_seq && IsReno(tp)) {
  1706. /* Hold old state until something *above* high_seq
  1707. * is ACKed. For Reno it is MUST to prevent false
  1708. * fast retransmits (RFC2582). SACK TCP is safe. */
  1709. tcp_moderate_cwnd(tp);
  1710. return 1;
  1711. }
  1712. tcp_set_ca_state(sk, TCP_CA_Open);
  1713. return 0;
  1714. }
  1715. /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
  1716. static void tcp_try_undo_dsack(struct sock *sk)
  1717. {
  1718. struct tcp_sock *tp = tcp_sk(sk);
  1719. if (tp->undo_marker && !tp->undo_retrans) {
  1720. DBGUNDO(sk, "D-SACK");
  1721. tcp_undo_cwr(sk, 1);
  1722. tp->undo_marker = 0;
  1723. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO);
  1724. }
  1725. }
  1726. /* Undo during fast recovery after partial ACK. */
  1727. static int tcp_try_undo_partial(struct sock *sk, int acked)
  1728. {
  1729. struct tcp_sock *tp = tcp_sk(sk);
  1730. /* Partial ACK arrived. Force Hoe's retransmit. */
  1731. int failed = IsReno(tp) || tp->fackets_out>tp->reordering;
  1732. if (tcp_may_undo(tp)) {
  1733. /* Plain luck! Hole if filled with delayed
  1734. * packet, rather than with a retransmit.
  1735. */
  1736. if (tp->retrans_out == 0)
  1737. tp->retrans_stamp = 0;
  1738. tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
  1739. DBGUNDO(sk, "Hoe");
  1740. tcp_undo_cwr(sk, 0);
  1741. NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO);
  1742. /* So... Do not make Hoe's retransmit yet.
  1743. * If the first packet was delayed, the rest
  1744. * ones are most probably delayed as well.
  1745. */
  1746. failed = 0;
  1747. }
  1748. return failed;
  1749. }
  1750. /* Undo during loss recovery after partial ACK. */
  1751. static int tcp_try_undo_loss(struct sock *sk)
  1752. {
  1753. struct tcp_sock *tp = tcp_sk(sk);
  1754. if (tcp_may_undo(tp)) {
  1755. struct sk_buff *skb;
  1756. tcp_for_write_queue(skb, sk) {
  1757. if (skb == tcp_send_head(sk))
  1758. break;
  1759. TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
  1760. }
  1761. clear_all_retrans_hints(tp);
  1762. DBGUNDO(sk, "partial loss");
  1763. tp->lost_out = 0;
  1764. tp->left_out = tp->sacked_out;
  1765. tcp_undo_cwr(sk, 1);
  1766. NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO);
  1767. inet_csk(sk)->icsk_retransmits = 0;
  1768. tp->undo_marker = 0;
  1769. if (!IsReno(tp))
  1770. tcp_set_ca_state(sk, TCP_CA_Open);
  1771. return 1;
  1772. }
  1773. return 0;
  1774. }
  1775. static inline void tcp_complete_cwr(struct sock *sk)
  1776. {
  1777. struct tcp_sock *tp = tcp_sk(sk);
  1778. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  1779. tp->snd_cwnd_stamp = tcp_time_stamp;
  1780. tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
  1781. }
  1782. static void tcp_try_to_open(struct sock *sk, int flag)
  1783. {
  1784. struct tcp_sock *tp = tcp_sk(sk);
  1785. tcp_sync_left_out(tp);
  1786. if (tp->retrans_out == 0)
  1787. tp->retrans_stamp = 0;
  1788. if (flag&FLAG_ECE)
  1789. tcp_enter_cwr(sk, 1);
  1790. if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
  1791. int state = TCP_CA_Open;
  1792. if (tp->left_out || tp->retrans_out || tp->undo_marker)
  1793. state = TCP_CA_Disorder;
  1794. if (inet_csk(sk)->icsk_ca_state != state) {
  1795. tcp_set_ca_state(sk, state);
  1796. tp->high_seq = tp->snd_nxt;
  1797. }
  1798. tcp_moderate_cwnd(tp);
  1799. } else {
  1800. tcp_cwnd_down(sk);
  1801. }
  1802. }
  1803. static void tcp_mtup_probe_failed(struct sock *sk)
  1804. {
  1805. struct inet_connection_sock *icsk = inet_csk(sk);
  1806. icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
  1807. icsk->icsk_mtup.probe_size = 0;
  1808. }
  1809. static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb)
  1810. {
  1811. struct tcp_sock *tp = tcp_sk(sk);
  1812. struct inet_connection_sock *icsk = inet_csk(sk);
  1813. /* FIXME: breaks with very large cwnd */
  1814. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1815. tp->snd_cwnd = tp->snd_cwnd *
  1816. tcp_mss_to_mtu(sk, tp->mss_cache) /
  1817. icsk->icsk_mtup.probe_size;
  1818. tp->snd_cwnd_cnt = 0;
  1819. tp->snd_cwnd_stamp = tcp_time_stamp;
  1820. tp->rcv_ssthresh = tcp_current_ssthresh(sk);
  1821. icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
  1822. icsk->icsk_mtup.probe_size = 0;
  1823. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  1824. }
  1825. /* Process an event, which can update packets-in-flight not trivially.
  1826. * Main goal of this function is to calculate new estimate for left_out,
  1827. * taking into account both packets sitting in receiver's buffer and
  1828. * packets lost by network.
  1829. *
  1830. * Besides that it does CWND reduction, when packet loss is detected
  1831. * and changes state of machine.
  1832. *
  1833. * It does _not_ decide what to send, it is made in function
  1834. * tcp_xmit_retransmit_queue().
  1835. */
  1836. static void
  1837. tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una,
  1838. int prior_packets, int flag)
  1839. {
  1840. struct inet_connection_sock *icsk = inet_csk(sk);
  1841. struct tcp_sock *tp = tcp_sk(sk);
  1842. int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP));
  1843. /* Some technical things:
  1844. * 1. Reno does not count dupacks (sacked_out) automatically. */
  1845. if (!tp->packets_out)
  1846. tp->sacked_out = 0;
  1847. /* 2. SACK counts snd_fack in packets inaccurately. */
  1848. if (tp->sacked_out == 0)
  1849. tp->fackets_out = 0;
  1850. /* Now state machine starts.
  1851. * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
  1852. if (flag&FLAG_ECE)
  1853. tp->prior_ssthresh = 0;
  1854. /* B. In all the states check for reneging SACKs. */
  1855. if (tp->sacked_out && tcp_check_sack_reneging(sk))
  1856. return;
  1857. /* C. Process data loss notification, provided it is valid. */
  1858. if ((flag&FLAG_DATA_LOST) &&
  1859. before(tp->snd_una, tp->high_seq) &&
  1860. icsk->icsk_ca_state != TCP_CA_Open &&
  1861. tp->fackets_out > tp->reordering) {
  1862. tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq);
  1863. NET_INC_STATS_BH(LINUX_MIB_TCPLOSS);
  1864. }
  1865. /* D. Synchronize left_out to current state. */
  1866. tcp_sync_left_out(tp);
  1867. /* E. Check state exit conditions. State can be terminated
  1868. * when high_seq is ACKed. */
  1869. if (icsk->icsk_ca_state == TCP_CA_Open) {
  1870. BUG_TRAP(tp->retrans_out == 0);
  1871. tp->retrans_stamp = 0;
  1872. } else if (!before(tp->snd_una, tp->high_seq)) {
  1873. switch (icsk->icsk_ca_state) {
  1874. case TCP_CA_Loss:
  1875. icsk->icsk_retransmits = 0;
  1876. if (tcp_try_undo_recovery(sk))
  1877. return;
  1878. break;
  1879. case TCP_CA_CWR:
  1880. /* CWR is to be held something *above* high_seq
  1881. * is ACKed for CWR bit to reach receiver. */
  1882. if (tp->snd_una != tp->high_seq) {
  1883. tcp_complete_cwr(sk);
  1884. tcp_set_ca_state(sk, TCP_CA_Open);
  1885. }
  1886. break;
  1887. case TCP_CA_Disorder:
  1888. tcp_try_undo_dsack(sk);
  1889. if (!tp->undo_marker ||
  1890. /* For SACK case do not Open to allow to undo
  1891. * catching for all duplicate ACKs. */
  1892. IsReno(tp) || tp->snd_una != tp->high_seq) {
  1893. tp->undo_marker = 0;
  1894. tcp_set_ca_state(sk, TCP_CA_Open);
  1895. }
  1896. break;
  1897. case TCP_CA_Recovery:
  1898. if (IsReno(tp))
  1899. tcp_reset_reno_sack(tp);
  1900. if (tcp_try_undo_recovery(sk))
  1901. return;
  1902. tcp_complete_cwr(sk);
  1903. break;
  1904. }
  1905. }
  1906. /* F. Process state. */
  1907. switch (icsk->icsk_ca_state) {
  1908. case TCP_CA_Recovery:
  1909. if (prior_snd_una == tp->snd_una) {
  1910. if (IsReno(tp) && is_dupack)
  1911. tcp_add_reno_sack(sk);
  1912. } else {
  1913. int acked = prior_packets - tp->packets_out;
  1914. if (IsReno(tp))
  1915. tcp_remove_reno_sacks(sk, acked);
  1916. is_dupack = tcp_try_undo_partial(sk, acked);
  1917. }
  1918. break;
  1919. case TCP_CA_Loss:
  1920. if (flag&FLAG_DATA_ACKED)
  1921. icsk->icsk_retransmits = 0;
  1922. if (!tcp_try_undo_loss(sk)) {
  1923. tcp_moderate_cwnd(tp);
  1924. tcp_xmit_retransmit_queue(sk);
  1925. return;
  1926. }
  1927. if (icsk->icsk_ca_state != TCP_CA_Open)
  1928. return;
  1929. /* Loss is undone; fall through to processing in Open state. */
  1930. default:
  1931. if (IsReno(tp)) {
  1932. if (tp->snd_una != prior_snd_una)
  1933. tcp_reset_reno_sack(tp);
  1934. if (is_dupack)
  1935. tcp_add_reno_sack(sk);
  1936. }
  1937. if (icsk->icsk_ca_state == TCP_CA_Disorder)
  1938. tcp_try_undo_dsack(sk);
  1939. if (!tcp_time_to_recover(sk)) {
  1940. tcp_try_to_open(sk, flag);
  1941. return;
  1942. }
  1943. /* MTU probe failure: don't reduce cwnd */
  1944. if (icsk->icsk_ca_state < TCP_CA_CWR &&
  1945. icsk->icsk_mtup.probe_size &&
  1946. tp->snd_una == tp->mtu_probe.probe_seq_start) {
  1947. tcp_mtup_probe_failed(sk);
  1948. /* Restores the reduction we did in tcp_mtup_probe() */
  1949. tp->snd_cwnd++;
  1950. tcp_simple_retransmit(sk);
  1951. return;
  1952. }
  1953. /* Otherwise enter Recovery state */
  1954. if (IsReno(tp))
  1955. NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY);
  1956. else
  1957. NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY);
  1958. tp->high_seq = tp->snd_nxt;
  1959. tp->prior_ssthresh = 0;
  1960. tp->undo_marker = tp->snd_una;
  1961. tp->undo_retrans = tp->retrans_out;
  1962. if (icsk->icsk_ca_state < TCP_CA_CWR) {
  1963. if (!(flag&FLAG_ECE))
  1964. tp->prior_ssthresh = tcp_current_ssthresh(sk);
  1965. tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
  1966. TCP_ECN_queue_cwr(tp);
  1967. }
  1968. tp->bytes_acked = 0;
  1969. tp->snd_cwnd_cnt = 0;
  1970. tcp_set_ca_state(sk, TCP_CA_Recovery);
  1971. }
  1972. if (is_dupack || tcp_head_timedout(sk))
  1973. tcp_update_scoreboard(sk);
  1974. tcp_cwnd_down(sk);
  1975. tcp_xmit_retransmit_queue(sk);
  1976. }
  1977. /* Read draft-ietf-tcplw-high-performance before mucking
  1978. * with this code. (Supersedes RFC1323)
  1979. */
  1980. static void tcp_ack_saw_tstamp(struct sock *sk, int flag)
  1981. {
  1982. /* RTTM Rule: A TSecr value received in a segment is used to
  1983. * update the averaged RTT measurement only if the segment
  1984. * acknowledges some new data, i.e., only if it advances the
  1985. * left edge of the send window.
  1986. *
  1987. * See draft-ietf-tcplw-high-performance-00, section 3.3.
  1988. * 1998/04/10 Andrey V. Savochkin <saw@msu.ru>
  1989. *
  1990. * Changed: reset backoff as soon as we see the first valid sample.
  1991. * If we do not, we get strongly overestimated rto. With timestamps
  1992. * samples are accepted even from very old segments: f.e., when rtt=1
  1993. * increases to 8, we retransmit 5 times and after 8 seconds delayed
  1994. * answer arrives rto becomes 120 seconds! If at least one of segments
  1995. * in window is lost... Voila. --ANK (010210)
  1996. */
  1997. struct tcp_sock *tp = tcp_sk(sk);
  1998. const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
  1999. tcp_rtt_estimator(sk, seq_rtt);
  2000. tcp_set_rto(sk);
  2001. inet_csk(sk)->icsk_backoff = 0;
  2002. tcp_bound_rto(sk);
  2003. }
  2004. static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag)
  2005. {
  2006. /* We don't have a timestamp. Can only use
  2007. * packets that are not retransmitted to determine
  2008. * rtt estimates. Also, we must not reset the
  2009. * backoff for rto until we get a non-retransmitted
  2010. * packet. This allows us to deal with a situation
  2011. * where the network delay has increased suddenly.
  2012. * I.e. Karn's algorithm. (SIGCOMM '87, p5.)
  2013. */
  2014. if (flag & FLAG_RETRANS_DATA_ACKED)
  2015. return;
  2016. tcp_rtt_estimator(sk, seq_rtt);
  2017. tcp_set_rto(sk);
  2018. inet_csk(sk)->icsk_backoff = 0;
  2019. tcp_bound_rto(sk);
  2020. }
  2021. static inline void tcp_ack_update_rtt(struct sock *sk, const int flag,
  2022. const s32 seq_rtt)
  2023. {
  2024. const struct tcp_sock *tp = tcp_sk(sk);
  2025. /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */
  2026. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
  2027. tcp_ack_saw_tstamp(sk, flag);
  2028. else if (seq_rtt >= 0)
  2029. tcp_ack_no_tstamp(sk, seq_rtt, flag);
  2030. }
  2031. static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
  2032. u32 in_flight, int good)
  2033. {
  2034. const struct inet_connection_sock *icsk = inet_csk(sk);
  2035. icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good);
  2036. tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
  2037. }
  2038. /* Restart timer after forward progress on connection.
  2039. * RFC2988 recommends to restart timer to now+rto.
  2040. */
  2041. static void tcp_ack_packets_out(struct sock *sk)
  2042. {
  2043. struct tcp_sock *tp = tcp_sk(sk);
  2044. if (!tp->packets_out) {
  2045. inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
  2046. } else {
  2047. inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
  2048. }
  2049. }
  2050. static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb,
  2051. __u32 now, __s32 *seq_rtt)
  2052. {
  2053. struct tcp_sock *tp = tcp_sk(sk);
  2054. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2055. __u32 seq = tp->snd_una;
  2056. __u32 packets_acked;
  2057. int acked = 0;
  2058. /* If we get here, the whole TSO packet has not been
  2059. * acked.
  2060. */
  2061. BUG_ON(!after(scb->end_seq, seq));
  2062. packets_acked = tcp_skb_pcount(skb);
  2063. if (tcp_trim_head(sk, skb, seq - scb->seq))
  2064. return 0;
  2065. packets_acked -= tcp_skb_pcount(skb);
  2066. if (packets_acked) {
  2067. __u8 sacked = scb->sacked;
  2068. acked |= FLAG_DATA_ACKED;
  2069. if (sacked) {
  2070. if (sacked & TCPCB_RETRANS) {
  2071. if (sacked & TCPCB_SACKED_RETRANS)
  2072. tp->retrans_out -= packets_acked;
  2073. acked |= FLAG_RETRANS_DATA_ACKED;
  2074. *seq_rtt = -1;
  2075. } else if (*seq_rtt < 0)
  2076. *seq_rtt = now - scb->when;
  2077. if (sacked & TCPCB_SACKED_ACKED)
  2078. tp->sacked_out -= packets_acked;
  2079. if (sacked & TCPCB_LOST)
  2080. tp->lost_out -= packets_acked;
  2081. if (sacked & TCPCB_URG) {
  2082. if (tp->urg_mode &&
  2083. !before(seq, tp->snd_up))
  2084. tp->urg_mode = 0;
  2085. }
  2086. } else if (*seq_rtt < 0)
  2087. *seq_rtt = now - scb->when;
  2088. if (tp->fackets_out) {
  2089. __u32 dval = min(tp->fackets_out, packets_acked);
  2090. tp->fackets_out -= dval;
  2091. }
  2092. tp->packets_out -= packets_acked;
  2093. BUG_ON(tcp_skb_pcount(skb) == 0);
  2094. BUG_ON(!before(scb->seq, scb->end_seq));
  2095. }
  2096. return acked;
  2097. }
  2098. /* Remove acknowledged frames from the retransmission queue. */
  2099. static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p)
  2100. {
  2101. struct tcp_sock *tp = tcp_sk(sk);
  2102. const struct inet_connection_sock *icsk = inet_csk(sk);
  2103. struct sk_buff *skb;
  2104. __u32 now = tcp_time_stamp;
  2105. int acked = 0;
  2106. int prior_packets = tp->packets_out;
  2107. __s32 seq_rtt = -1;
  2108. ktime_t last_ackt = net_invalid_timestamp();
  2109. while ((skb = tcp_write_queue_head(sk)) &&
  2110. skb != tcp_send_head(sk)) {
  2111. struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  2112. __u8 sacked = scb->sacked;
  2113. /* If our packet is before the ack sequence we can
  2114. * discard it as it's confirmed to have arrived at
  2115. * the other end.
  2116. */
  2117. if (after(scb->end_seq, tp->snd_una)) {
  2118. if (tcp_skb_pcount(skb) > 1 &&
  2119. after(tp->snd_una, scb->seq))
  2120. acked |= tcp_tso_acked(sk, skb,
  2121. now, &seq_rtt);
  2122. break;
  2123. }
  2124. /* Initial outgoing SYN's get put onto the write_queue
  2125. * just like anything else we transmit. It is not
  2126. * true data, and if we misinform our callers that
  2127. * this ACK acks real data, we will erroneously exit
  2128. * connection startup slow start one packet too
  2129. * quickly. This is severely frowned upon behavior.
  2130. */
  2131. if (!(scb->flags & TCPCB_FLAG_SYN)) {
  2132. acked |= FLAG_DATA_ACKED;
  2133. } else {
  2134. acked |= FLAG_SYN_ACKED;
  2135. tp->retrans_stamp = 0;
  2136. }
  2137. /* MTU probing checks */
  2138. if (icsk->icsk_mtup.probe_size) {
  2139. if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) {
  2140. tcp_mtup_probe_success(sk, skb);
  2141. }
  2142. }
  2143. if (sacked) {
  2144. if (sacked & TCPCB_RETRANS) {
  2145. if (sacked & TCPCB_SACKED_RETRANS)
  2146. tp->retrans_out -= tcp_skb_pcount(skb);
  2147. acked |= FLAG_RETRANS_DATA_ACKED;
  2148. seq_rtt = -1;
  2149. } else if (seq_rtt < 0) {
  2150. seq_rtt = now - scb->when;
  2151. last_ackt = skb->tstamp;
  2152. }
  2153. if (sacked & TCPCB_SACKED_ACKED)
  2154. tp->sacked_out -= tcp_skb_pcount(skb);
  2155. if (sacked & TCPCB_LOST)
  2156. tp->lost_out -= tcp_skb_pcount(skb);
  2157. if (sacked & TCPCB_URG) {
  2158. if (tp->urg_mode &&
  2159. !before(scb->end_seq, tp->snd_up))
  2160. tp->urg_mode = 0;
  2161. }
  2162. } else if (seq_rtt < 0) {
  2163. seq_rtt = now - scb->when;
  2164. last_ackt = skb->tstamp;
  2165. }
  2166. tcp_dec_pcount_approx(&tp->fackets_out, skb);
  2167. tcp_packets_out_dec(tp, skb);
  2168. tcp_unlink_write_queue(skb, sk);
  2169. sk_stream_free_skb(sk, skb);
  2170. clear_all_retrans_hints(tp);
  2171. }
  2172. if (acked&FLAG_ACKED) {
  2173. u32 pkts_acked = prior_packets - tp->packets_out;
  2174. const struct tcp_congestion_ops *ca_ops
  2175. = inet_csk(sk)->icsk_ca_ops;
  2176. tcp_ack_update_rtt(sk, acked, seq_rtt);
  2177. tcp_ack_packets_out(sk);
  2178. /* Is the ACK triggering packet unambiguous? */
  2179. if (acked & FLAG_RETRANS_DATA_ACKED)
  2180. last_ackt = net_invalid_timestamp();
  2181. if (ca_ops->pkts_acked)
  2182. ca_ops->pkts_acked(sk, pkts_acked, last_ackt);
  2183. }
  2184. #if FASTRETRANS_DEBUG > 0
  2185. BUG_TRAP((int)tp->sacked_out >= 0);
  2186. BUG_TRAP((int)tp->lost_out >= 0);
  2187. BUG_TRAP((int)tp->retrans_out >= 0);
  2188. if (!tp->packets_out && tp->rx_opt.sack_ok) {
  2189. const struct inet_connection_sock *icsk = inet_csk(sk);
  2190. if (tp->lost_out) {
  2191. printk(KERN_DEBUG "Leak l=%u %d\n",
  2192. tp->lost_out, icsk->icsk_ca_state);
  2193. tp->lost_out = 0;
  2194. }
  2195. if (tp->sacked_out) {
  2196. printk(KERN_DEBUG "Leak s=%u %d\n",
  2197. tp->sacked_out, icsk->icsk_ca_state);
  2198. tp->sacked_out = 0;
  2199. }
  2200. if (tp->retrans_out) {
  2201. printk(KERN_DEBUG "Leak r=%u %d\n",
  2202. tp->retrans_out, icsk->icsk_ca_state);
  2203. tp->retrans_out = 0;
  2204. }
  2205. }
  2206. #endif
  2207. *seq_rtt_p = seq_rtt;
  2208. return acked;
  2209. }
  2210. static void tcp_ack_probe(struct sock *sk)
  2211. {
  2212. const struct tcp_sock *tp = tcp_sk(sk);
  2213. struct inet_connection_sock *icsk = inet_csk(sk);
  2214. /* Was it a usable window open? */
  2215. if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
  2216. tp->snd_una + tp->snd_wnd)) {
  2217. icsk->icsk_backoff = 0;
  2218. inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
  2219. /* Socket must be waked up by subsequent tcp_data_snd_check().
  2220. * This function is not for random using!
  2221. */
  2222. } else {
  2223. inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
  2224. min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
  2225. TCP_RTO_MAX);
  2226. }
  2227. }
  2228. static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag)
  2229. {
  2230. return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
  2231. inet_csk(sk)->icsk_ca_state != TCP_CA_Open);
  2232. }
  2233. static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag)
  2234. {
  2235. const struct tcp_sock *tp = tcp_sk(sk);
  2236. return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) &&
  2237. !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR));
  2238. }
  2239. /* Check that window update is acceptable.
  2240. * The function assumes that snd_una<=ack<=snd_next.
  2241. */
  2242. static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack,
  2243. const u32 ack_seq, const u32 nwin)
  2244. {
  2245. return (after(ack, tp->snd_una) ||
  2246. after(ack_seq, tp->snd_wl1) ||
  2247. (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd));
  2248. }
  2249. /* Update our send window.
  2250. *
  2251. * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
  2252. * and in FreeBSD. NetBSD's one is even worse.) is wrong.
  2253. */
  2254. static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack,
  2255. u32 ack_seq)
  2256. {
  2257. struct tcp_sock *tp = tcp_sk(sk);
  2258. int flag = 0;
  2259. u32 nwin = ntohs(tcp_hdr(skb)->window);
  2260. if (likely(!tcp_hdr(skb)->syn))
  2261. nwin <<= tp->rx_opt.snd_wscale;
  2262. if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
  2263. flag |= FLAG_WIN_UPDATE;
  2264. tcp_update_wl(tp, ack, ack_seq);
  2265. if (tp->snd_wnd != nwin) {
  2266. tp->snd_wnd = nwin;
  2267. /* Note, it is the only place, where
  2268. * fast path is recovered for sending TCP.
  2269. */
  2270. tp->pred_flags = 0;
  2271. tcp_fast_path_check(sk);
  2272. if (nwin > tp->max_window) {
  2273. tp->max_window = nwin;
  2274. tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
  2275. }
  2276. }
  2277. }
  2278. tp->snd_una = ack;
  2279. return flag;
  2280. }
  2281. /* A very conservative spurious RTO response algorithm: reduce cwnd and
  2282. * continue in congestion avoidance.
  2283. */
  2284. static void tcp_conservative_spur_to_response(struct tcp_sock *tp)
  2285. {
  2286. tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh);
  2287. tp->snd_cwnd_cnt = 0;
  2288. TCP_ECN_queue_cwr(tp);
  2289. tcp_moderate_cwnd(tp);
  2290. }
  2291. /* A conservative spurious RTO response algorithm: reduce cwnd using
  2292. * rate halving and continue in congestion avoidance.
  2293. */
  2294. static void tcp_ratehalving_spur_to_response(struct sock *sk)
  2295. {
  2296. tcp_enter_cwr(sk, 0);
  2297. }
  2298. static void tcp_undo_spur_to_response(struct sock *sk, int flag)
  2299. {
  2300. if (flag&FLAG_ECE)
  2301. tcp_ratehalving_spur_to_response(sk);
  2302. else
  2303. tcp_undo_cwr(sk, 1);
  2304. }
  2305. /* F-RTO spurious RTO detection algorithm (RFC4138)
  2306. *
  2307. * F-RTO affects during two new ACKs following RTO (well, almost, see inline
  2308. * comments). State (ACK number) is kept in frto_counter. When ACK advances
  2309. * window (but not to or beyond highest sequence sent before RTO):
  2310. * On First ACK, send two new segments out.
  2311. * On Second ACK, RTO was likely spurious. Do spurious response (response
  2312. * algorithm is not part of the F-RTO detection algorithm
  2313. * given in RFC4138 but can be selected separately).
  2314. * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss
  2315. * and TCP falls back to conventional RTO recovery. F-RTO allows overriding
  2316. * of Nagle, this is done using frto_counter states 2 and 3, when a new data
  2317. * segment of any size sent during F-RTO, state 2 is upgraded to 3.
  2318. *
  2319. * Rationale: if the RTO was spurious, new ACKs should arrive from the
  2320. * original window even after we transmit two new data segments.
  2321. *
  2322. * SACK version:
  2323. * on first step, wait until first cumulative ACK arrives, then move to
  2324. * the second step. In second step, the next ACK decides.
  2325. *
  2326. * F-RTO is implemented (mainly) in four functions:
  2327. * - tcp_use_frto() is used to determine if TCP is can use F-RTO
  2328. * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is
  2329. * called when tcp_use_frto() showed green light
  2330. * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm
  2331. * - tcp_enter_frto_loss() is called if there is not enough evidence
  2332. * to prove that the RTO is indeed spurious. It transfers the control
  2333. * from F-RTO to the conventional RTO recovery
  2334. */
  2335. static int tcp_process_frto(struct sock *sk, u32 prior_snd_una, int flag)
  2336. {
  2337. struct tcp_sock *tp = tcp_sk(sk);
  2338. tcp_sync_left_out(tp);
  2339. /* Duplicate the behavior from Loss state (fastretrans_alert) */
  2340. if (flag&FLAG_DATA_ACKED)
  2341. inet_csk(sk)->icsk_retransmits = 0;
  2342. if (!before(tp->snd_una, tp->frto_highmark)) {
  2343. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag);
  2344. return 1;
  2345. }
  2346. if (!IsSackFrto() || IsReno(tp)) {
  2347. /* RFC4138 shortcoming in step 2; should also have case c):
  2348. * ACK isn't duplicate nor advances window, e.g., opposite dir
  2349. * data, winupdate
  2350. */
  2351. if ((tp->snd_una == prior_snd_una) && (flag&FLAG_NOT_DUP) &&
  2352. !(flag&FLAG_FORWARD_PROGRESS))
  2353. return 1;
  2354. if (!(flag&FLAG_DATA_ACKED)) {
  2355. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3),
  2356. flag);
  2357. return 1;
  2358. }
  2359. } else {
  2360. if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) {
  2361. /* Prevent sending of new data. */
  2362. tp->snd_cwnd = min(tp->snd_cwnd,
  2363. tcp_packets_in_flight(tp));
  2364. return 1;
  2365. }
  2366. if ((tp->frto_counter >= 2) &&
  2367. (!(flag&FLAG_FORWARD_PROGRESS) ||
  2368. ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) {
  2369. /* RFC4138 shortcoming (see comment above) */
  2370. if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP))
  2371. return 1;
  2372. tcp_enter_frto_loss(sk, 3, flag);
  2373. return 1;
  2374. }
  2375. }
  2376. if (tp->frto_counter == 1) {
  2377. /* Sending of the next skb must be allowed or no FRTO */
  2378. if (!tcp_send_head(sk) ||
  2379. after(TCP_SKB_CB(tcp_send_head(sk))->end_seq,
  2380. tp->snd_una + tp->snd_wnd)) {
  2381. tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3),
  2382. flag);
  2383. return 1;
  2384. }
  2385. tp->snd_cwnd = tcp_packets_in_flight(tp) + 2;
  2386. tp->frto_counter = 2;
  2387. return 1;
  2388. } else {
  2389. switch (sysctl_tcp_frto_response) {
  2390. case 2:
  2391. tcp_undo_spur_to_response(sk, flag);
  2392. break;
  2393. case 1:
  2394. tcp_conservative_spur_to_response(tp);
  2395. break;
  2396. default:
  2397. tcp_ratehalving_spur_to_response(sk);
  2398. break;
  2399. }
  2400. tp->frto_counter = 0;
  2401. }
  2402. return 0;
  2403. }
  2404. /* This routine deals with incoming acks, but not outgoing ones. */
  2405. static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag)
  2406. {
  2407. struct inet_connection_sock *icsk = inet_csk(sk);
  2408. struct tcp_sock *tp = tcp_sk(sk);
  2409. u32 prior_snd_una = tp->snd_una;
  2410. u32 ack_seq = TCP_SKB_CB(skb)->seq;
  2411. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2412. u32 prior_in_flight;
  2413. s32 seq_rtt;
  2414. int prior_packets;
  2415. int frto_cwnd = 0;
  2416. /* If the ack is newer than sent or older than previous acks
  2417. * then we can probably ignore it.
  2418. */
  2419. if (after(ack, tp->snd_nxt))
  2420. goto uninteresting_ack;
  2421. if (before(ack, prior_snd_una))
  2422. goto old_ack;
  2423. if (sysctl_tcp_abc) {
  2424. if (icsk->icsk_ca_state < TCP_CA_CWR)
  2425. tp->bytes_acked += ack - prior_snd_una;
  2426. else if (icsk->icsk_ca_state == TCP_CA_Loss)
  2427. /* we assume just one segment left network */
  2428. tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache);
  2429. }
  2430. if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
  2431. /* Window is constant, pure forward advance.
  2432. * No more checks are required.
  2433. * Note, we use the fact that SND.UNA>=SND.WL2.
  2434. */
  2435. tcp_update_wl(tp, ack, ack_seq);
  2436. tp->snd_una = ack;
  2437. flag |= FLAG_WIN_UPDATE;
  2438. tcp_ca_event(sk, CA_EVENT_FAST_ACK);
  2439. NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS);
  2440. } else {
  2441. if (ack_seq != TCP_SKB_CB(skb)->end_seq)
  2442. flag |= FLAG_DATA;
  2443. else
  2444. NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS);
  2445. flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
  2446. if (TCP_SKB_CB(skb)->sacked)
  2447. flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2448. if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
  2449. flag |= FLAG_ECE;
  2450. tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
  2451. }
  2452. /* We passed data and got it acked, remove any soft error
  2453. * log. Something worked...
  2454. */
  2455. sk->sk_err_soft = 0;
  2456. tp->rcv_tstamp = tcp_time_stamp;
  2457. prior_packets = tp->packets_out;
  2458. if (!prior_packets)
  2459. goto no_queue;
  2460. prior_in_flight = tcp_packets_in_flight(tp);
  2461. /* See if we can take anything off of the retransmit queue. */
  2462. flag |= tcp_clean_rtx_queue(sk, &seq_rtt);
  2463. if (tp->frto_counter)
  2464. frto_cwnd = tcp_process_frto(sk, prior_snd_una, flag);
  2465. if (tcp_ack_is_dubious(sk, flag)) {
  2466. /* Advance CWND, if state allows this. */
  2467. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd &&
  2468. tcp_may_raise_cwnd(sk, flag))
  2469. tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0);
  2470. tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag);
  2471. } else {
  2472. if ((flag & FLAG_DATA_ACKED) && !frto_cwnd)
  2473. tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1);
  2474. }
  2475. if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP))
  2476. dst_confirm(sk->sk_dst_cache);
  2477. return 1;
  2478. no_queue:
  2479. icsk->icsk_probes_out = 0;
  2480. /* If this ack opens up a zero window, clear backoff. It was
  2481. * being used to time the probes, and is probably far higher than
  2482. * it needs to be for normal retransmission.
  2483. */
  2484. if (tcp_send_head(sk))
  2485. tcp_ack_probe(sk);
  2486. return 1;
  2487. old_ack:
  2488. if (TCP_SKB_CB(skb)->sacked)
  2489. tcp_sacktag_write_queue(sk, skb, prior_snd_una);
  2490. uninteresting_ack:
  2491. SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
  2492. return 0;
  2493. }
  2494. /* Look for tcp options. Normally only called on SYN and SYNACK packets.
  2495. * But, this can also be called on packets in the established flow when
  2496. * the fast version below fails.
  2497. */
  2498. void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab)
  2499. {
  2500. unsigned char *ptr;
  2501. struct tcphdr *th = tcp_hdr(skb);
  2502. int length=(th->doff*4)-sizeof(struct tcphdr);
  2503. ptr = (unsigned char *)(th + 1);
  2504. opt_rx->saw_tstamp = 0;
  2505. while (length > 0) {
  2506. int opcode=*ptr++;
  2507. int opsize;
  2508. switch (opcode) {
  2509. case TCPOPT_EOL:
  2510. return;
  2511. case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
  2512. length--;
  2513. continue;
  2514. default:
  2515. opsize=*ptr++;
  2516. if (opsize < 2) /* "silly options" */
  2517. return;
  2518. if (opsize > length)
  2519. return; /* don't parse partial options */
  2520. switch (opcode) {
  2521. case TCPOPT_MSS:
  2522. if (opsize==TCPOLEN_MSS && th->syn && !estab) {
  2523. u16 in_mss = ntohs(get_unaligned((__be16 *)ptr));
  2524. if (in_mss) {
  2525. if (opt_rx->user_mss && opt_rx->user_mss < in_mss)
  2526. in_mss = opt_rx->user_mss;
  2527. opt_rx->mss_clamp = in_mss;
  2528. }
  2529. }
  2530. break;
  2531. case TCPOPT_WINDOW:
  2532. if (opsize==TCPOLEN_WINDOW && th->syn && !estab)
  2533. if (sysctl_tcp_window_scaling) {
  2534. __u8 snd_wscale = *(__u8 *) ptr;
  2535. opt_rx->wscale_ok = 1;
  2536. if (snd_wscale > 14) {
  2537. if (net_ratelimit())
  2538. printk(KERN_INFO "tcp_parse_options: Illegal window "
  2539. "scaling value %d >14 received.\n",
  2540. snd_wscale);
  2541. snd_wscale = 14;
  2542. }
  2543. opt_rx->snd_wscale = snd_wscale;
  2544. }
  2545. break;
  2546. case TCPOPT_TIMESTAMP:
  2547. if (opsize==TCPOLEN_TIMESTAMP) {
  2548. if ((estab && opt_rx->tstamp_ok) ||
  2549. (!estab && sysctl_tcp_timestamps)) {
  2550. opt_rx->saw_tstamp = 1;
  2551. opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr));
  2552. opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4)));
  2553. }
  2554. }
  2555. break;
  2556. case TCPOPT_SACK_PERM:
  2557. if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) {
  2558. if (sysctl_tcp_sack) {
  2559. opt_rx->sack_ok = 1;
  2560. tcp_sack_reset(opt_rx);
  2561. }
  2562. }
  2563. break;
  2564. case TCPOPT_SACK:
  2565. if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
  2566. !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
  2567. opt_rx->sack_ok) {
  2568. TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
  2569. }
  2570. break;
  2571. #ifdef CONFIG_TCP_MD5SIG
  2572. case TCPOPT_MD5SIG:
  2573. /*
  2574. * The MD5 Hash has already been
  2575. * checked (see tcp_v{4,6}_do_rcv()).
  2576. */
  2577. break;
  2578. #endif
  2579. }
  2580. ptr+=opsize-2;
  2581. length-=opsize;
  2582. }
  2583. }
  2584. }
  2585. /* Fast parse options. This hopes to only see timestamps.
  2586. * If it is wrong it falls back on tcp_parse_options().
  2587. */
  2588. static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th,
  2589. struct tcp_sock *tp)
  2590. {
  2591. if (th->doff == sizeof(struct tcphdr)>>2) {
  2592. tp->rx_opt.saw_tstamp = 0;
  2593. return 0;
  2594. } else if (tp->rx_opt.tstamp_ok &&
  2595. th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) {
  2596. __be32 *ptr = (__be32 *)(th + 1);
  2597. if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  2598. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
  2599. tp->rx_opt.saw_tstamp = 1;
  2600. ++ptr;
  2601. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  2602. ++ptr;
  2603. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  2604. return 1;
  2605. }
  2606. }
  2607. tcp_parse_options(skb, &tp->rx_opt, 1);
  2608. return 1;
  2609. }
  2610. static inline void tcp_store_ts_recent(struct tcp_sock *tp)
  2611. {
  2612. tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
  2613. tp->rx_opt.ts_recent_stamp = get_seconds();
  2614. }
  2615. static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
  2616. {
  2617. if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
  2618. /* PAWS bug workaround wrt. ACK frames, the PAWS discard
  2619. * extra check below makes sure this can only happen
  2620. * for pure ACK frames. -DaveM
  2621. *
  2622. * Not only, also it occurs for expired timestamps.
  2623. */
  2624. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 ||
  2625. get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS)
  2626. tcp_store_ts_recent(tp);
  2627. }
  2628. }
  2629. /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
  2630. *
  2631. * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
  2632. * it can pass through stack. So, the following predicate verifies that
  2633. * this segment is not used for anything but congestion avoidance or
  2634. * fast retransmit. Moreover, we even are able to eliminate most of such
  2635. * second order effects, if we apply some small "replay" window (~RTO)
  2636. * to timestamp space.
  2637. *
  2638. * All these measures still do not guarantee that we reject wrapped ACKs
  2639. * on networks with high bandwidth, when sequence space is recycled fastly,
  2640. * but it guarantees that such events will be very rare and do not affect
  2641. * connection seriously. This doesn't look nice, but alas, PAWS is really
  2642. * buggy extension.
  2643. *
  2644. * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
  2645. * states that events when retransmit arrives after original data are rare.
  2646. * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
  2647. * the biggest problem on large power networks even with minor reordering.
  2648. * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
  2649. * up to bandwidth of 18Gigabit/sec. 8) ]
  2650. */
  2651. static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
  2652. {
  2653. struct tcp_sock *tp = tcp_sk(sk);
  2654. struct tcphdr *th = tcp_hdr(skb);
  2655. u32 seq = TCP_SKB_CB(skb)->seq;
  2656. u32 ack = TCP_SKB_CB(skb)->ack_seq;
  2657. return (/* 1. Pure ACK with correct sequence number. */
  2658. (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
  2659. /* 2. ... and duplicate ACK. */
  2660. ack == tp->snd_una &&
  2661. /* 3. ... and does not update window. */
  2662. !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
  2663. /* 4. ... and sits in replay window. */
  2664. (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
  2665. }
  2666. static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb)
  2667. {
  2668. const struct tcp_sock *tp = tcp_sk(sk);
  2669. return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW &&
  2670. get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS &&
  2671. !tcp_disordered_ack(sk, skb));
  2672. }
  2673. /* Check segment sequence number for validity.
  2674. *
  2675. * Segment controls are considered valid, if the segment
  2676. * fits to the window after truncation to the window. Acceptability
  2677. * of data (and SYN, FIN, of course) is checked separately.
  2678. * See tcp_data_queue(), for example.
  2679. *
  2680. * Also, controls (RST is main one) are accepted using RCV.WUP instead
  2681. * of RCV.NXT. Peer still did not advance his SND.UNA when we
  2682. * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
  2683. * (borrowed from freebsd)
  2684. */
  2685. static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2686. {
  2687. return !before(end_seq, tp->rcv_wup) &&
  2688. !after(seq, tp->rcv_nxt + tcp_receive_window(tp));
  2689. }
  2690. /* When we get a reset we do this. */
  2691. static void tcp_reset(struct sock *sk)
  2692. {
  2693. /* We want the right error as BSD sees it (and indeed as we do). */
  2694. switch (sk->sk_state) {
  2695. case TCP_SYN_SENT:
  2696. sk->sk_err = ECONNREFUSED;
  2697. break;
  2698. case TCP_CLOSE_WAIT:
  2699. sk->sk_err = EPIPE;
  2700. break;
  2701. case TCP_CLOSE:
  2702. return;
  2703. default:
  2704. sk->sk_err = ECONNRESET;
  2705. }
  2706. if (!sock_flag(sk, SOCK_DEAD))
  2707. sk->sk_error_report(sk);
  2708. tcp_done(sk);
  2709. }
  2710. /*
  2711. * Process the FIN bit. This now behaves as it is supposed to work
  2712. * and the FIN takes effect when it is validly part of sequence
  2713. * space. Not before when we get holes.
  2714. *
  2715. * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
  2716. * (and thence onto LAST-ACK and finally, CLOSE, we never enter
  2717. * TIME-WAIT)
  2718. *
  2719. * If we are in FINWAIT-1, a received FIN indicates simultaneous
  2720. * close and we go into CLOSING (and later onto TIME-WAIT)
  2721. *
  2722. * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
  2723. */
  2724. static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th)
  2725. {
  2726. struct tcp_sock *tp = tcp_sk(sk);
  2727. inet_csk_schedule_ack(sk);
  2728. sk->sk_shutdown |= RCV_SHUTDOWN;
  2729. sock_set_flag(sk, SOCK_DONE);
  2730. switch (sk->sk_state) {
  2731. case TCP_SYN_RECV:
  2732. case TCP_ESTABLISHED:
  2733. /* Move to CLOSE_WAIT */
  2734. tcp_set_state(sk, TCP_CLOSE_WAIT);
  2735. inet_csk(sk)->icsk_ack.pingpong = 1;
  2736. break;
  2737. case TCP_CLOSE_WAIT:
  2738. case TCP_CLOSING:
  2739. /* Received a retransmission of the FIN, do
  2740. * nothing.
  2741. */
  2742. break;
  2743. case TCP_LAST_ACK:
  2744. /* RFC793: Remain in the LAST-ACK state. */
  2745. break;
  2746. case TCP_FIN_WAIT1:
  2747. /* This case occurs when a simultaneous close
  2748. * happens, we must ack the received FIN and
  2749. * enter the CLOSING state.
  2750. */
  2751. tcp_send_ack(sk);
  2752. tcp_set_state(sk, TCP_CLOSING);
  2753. break;
  2754. case TCP_FIN_WAIT2:
  2755. /* Received a FIN -- send ACK and enter TIME_WAIT. */
  2756. tcp_send_ack(sk);
  2757. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  2758. break;
  2759. default:
  2760. /* Only TCP_LISTEN and TCP_CLOSE are left, in these
  2761. * cases we should never reach this piece of code.
  2762. */
  2763. printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n",
  2764. __FUNCTION__, sk->sk_state);
  2765. break;
  2766. }
  2767. /* It _is_ possible, that we have something out-of-order _after_ FIN.
  2768. * Probably, we should reset in this case. For now drop them.
  2769. */
  2770. __skb_queue_purge(&tp->out_of_order_queue);
  2771. if (tp->rx_opt.sack_ok)
  2772. tcp_sack_reset(&tp->rx_opt);
  2773. sk_stream_mem_reclaim(sk);
  2774. if (!sock_flag(sk, SOCK_DEAD)) {
  2775. sk->sk_state_change(sk);
  2776. /* Do not send POLL_HUP for half duplex close. */
  2777. if (sk->sk_shutdown == SHUTDOWN_MASK ||
  2778. sk->sk_state == TCP_CLOSE)
  2779. sk_wake_async(sk, 1, POLL_HUP);
  2780. else
  2781. sk_wake_async(sk, 1, POLL_IN);
  2782. }
  2783. }
  2784. static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq)
  2785. {
  2786. if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
  2787. if (before(seq, sp->start_seq))
  2788. sp->start_seq = seq;
  2789. if (after(end_seq, sp->end_seq))
  2790. sp->end_seq = end_seq;
  2791. return 1;
  2792. }
  2793. return 0;
  2794. }
  2795. static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2796. {
  2797. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2798. if (before(seq, tp->rcv_nxt))
  2799. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT);
  2800. else
  2801. NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT);
  2802. tp->rx_opt.dsack = 1;
  2803. tp->duplicate_sack[0].start_seq = seq;
  2804. tp->duplicate_sack[0].end_seq = end_seq;
  2805. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok);
  2806. }
  2807. }
  2808. static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq)
  2809. {
  2810. if (!tp->rx_opt.dsack)
  2811. tcp_dsack_set(tp, seq, end_seq);
  2812. else
  2813. tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
  2814. }
  2815. static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb)
  2816. {
  2817. struct tcp_sock *tp = tcp_sk(sk);
  2818. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  2819. before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  2820. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  2821. tcp_enter_quickack_mode(sk);
  2822. if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) {
  2823. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  2824. if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
  2825. end_seq = tp->rcv_nxt;
  2826. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq);
  2827. }
  2828. }
  2829. tcp_send_ack(sk);
  2830. }
  2831. /* These routines update the SACK block as out-of-order packets arrive or
  2832. * in-order packets close up the sequence space.
  2833. */
  2834. static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
  2835. {
  2836. int this_sack;
  2837. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2838. struct tcp_sack_block *swalk = sp+1;
  2839. /* See if the recent change to the first SACK eats into
  2840. * or hits the sequence space of other SACK blocks, if so coalesce.
  2841. */
  2842. for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) {
  2843. if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
  2844. int i;
  2845. /* Zap SWALK, by moving every further SACK up by one slot.
  2846. * Decrease num_sacks.
  2847. */
  2848. tp->rx_opt.num_sacks--;
  2849. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2850. for (i=this_sack; i < tp->rx_opt.num_sacks; i++)
  2851. sp[i] = sp[i+1];
  2852. continue;
  2853. }
  2854. this_sack++, swalk++;
  2855. }
  2856. }
  2857. static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2)
  2858. {
  2859. __u32 tmp;
  2860. tmp = sack1->start_seq;
  2861. sack1->start_seq = sack2->start_seq;
  2862. sack2->start_seq = tmp;
  2863. tmp = sack1->end_seq;
  2864. sack1->end_seq = sack2->end_seq;
  2865. sack2->end_seq = tmp;
  2866. }
  2867. static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
  2868. {
  2869. struct tcp_sock *tp = tcp_sk(sk);
  2870. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2871. int cur_sacks = tp->rx_opt.num_sacks;
  2872. int this_sack;
  2873. if (!cur_sacks)
  2874. goto new_sack;
  2875. for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) {
  2876. if (tcp_sack_extend(sp, seq, end_seq)) {
  2877. /* Rotate this_sack to the first one. */
  2878. for (; this_sack>0; this_sack--, sp--)
  2879. tcp_sack_swap(sp, sp-1);
  2880. if (cur_sacks > 1)
  2881. tcp_sack_maybe_coalesce(tp);
  2882. return;
  2883. }
  2884. }
  2885. /* Could not find an adjacent existing SACK, build a new one,
  2886. * put it at the front, and shift everyone else down. We
  2887. * always know there is at least one SACK present already here.
  2888. *
  2889. * If the sack array is full, forget about the last one.
  2890. */
  2891. if (this_sack >= 4) {
  2892. this_sack--;
  2893. tp->rx_opt.num_sacks--;
  2894. sp--;
  2895. }
  2896. for (; this_sack > 0; this_sack--, sp--)
  2897. *sp = *(sp-1);
  2898. new_sack:
  2899. /* Build the new head SACK, and we're done. */
  2900. sp->start_seq = seq;
  2901. sp->end_seq = end_seq;
  2902. tp->rx_opt.num_sacks++;
  2903. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2904. }
  2905. /* RCV.NXT advances, some SACKs should be eaten. */
  2906. static void tcp_sack_remove(struct tcp_sock *tp)
  2907. {
  2908. struct tcp_sack_block *sp = &tp->selective_acks[0];
  2909. int num_sacks = tp->rx_opt.num_sacks;
  2910. int this_sack;
  2911. /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
  2912. if (skb_queue_empty(&tp->out_of_order_queue)) {
  2913. tp->rx_opt.num_sacks = 0;
  2914. tp->rx_opt.eff_sacks = tp->rx_opt.dsack;
  2915. return;
  2916. }
  2917. for (this_sack = 0; this_sack < num_sacks; ) {
  2918. /* Check if the start of the sack is covered by RCV.NXT. */
  2919. if (!before(tp->rcv_nxt, sp->start_seq)) {
  2920. int i;
  2921. /* RCV.NXT must cover all the block! */
  2922. BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq));
  2923. /* Zap this SACK, by moving forward any other SACKS. */
  2924. for (i=this_sack+1; i < num_sacks; i++)
  2925. tp->selective_acks[i-1] = tp->selective_acks[i];
  2926. num_sacks--;
  2927. continue;
  2928. }
  2929. this_sack++;
  2930. sp++;
  2931. }
  2932. if (num_sacks != tp->rx_opt.num_sacks) {
  2933. tp->rx_opt.num_sacks = num_sacks;
  2934. tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok);
  2935. }
  2936. }
  2937. /* This one checks to see if we can put data from the
  2938. * out_of_order queue into the receive_queue.
  2939. */
  2940. static void tcp_ofo_queue(struct sock *sk)
  2941. {
  2942. struct tcp_sock *tp = tcp_sk(sk);
  2943. __u32 dsack_high = tp->rcv_nxt;
  2944. struct sk_buff *skb;
  2945. while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
  2946. if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  2947. break;
  2948. if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
  2949. __u32 dsack = dsack_high;
  2950. if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
  2951. dsack_high = TCP_SKB_CB(skb)->end_seq;
  2952. tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack);
  2953. }
  2954. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  2955. SOCK_DEBUG(sk, "ofo packet was already received \n");
  2956. __skb_unlink(skb, &tp->out_of_order_queue);
  2957. __kfree_skb(skb);
  2958. continue;
  2959. }
  2960. SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
  2961. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  2962. TCP_SKB_CB(skb)->end_seq);
  2963. __skb_unlink(skb, &tp->out_of_order_queue);
  2964. __skb_queue_tail(&sk->sk_receive_queue, skb);
  2965. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  2966. if (tcp_hdr(skb)->fin)
  2967. tcp_fin(skb, sk, tcp_hdr(skb));
  2968. }
  2969. }
  2970. static int tcp_prune_queue(struct sock *sk);
  2971. static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
  2972. {
  2973. struct tcphdr *th = tcp_hdr(skb);
  2974. struct tcp_sock *tp = tcp_sk(sk);
  2975. int eaten = -1;
  2976. if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
  2977. goto drop;
  2978. __skb_pull(skb, th->doff*4);
  2979. TCP_ECN_accept_cwr(tp, skb);
  2980. if (tp->rx_opt.dsack) {
  2981. tp->rx_opt.dsack = 0;
  2982. tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks,
  2983. 4 - tp->rx_opt.tstamp_ok);
  2984. }
  2985. /* Queue data for delivery to the user.
  2986. * Packets in sequence go to the receive queue.
  2987. * Out of sequence packets to the out_of_order_queue.
  2988. */
  2989. if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  2990. if (tcp_receive_window(tp) == 0)
  2991. goto out_of_window;
  2992. /* Ok. In sequence. In window. */
  2993. if (tp->ucopy.task == current &&
  2994. tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
  2995. sock_owned_by_user(sk) && !tp->urg_data) {
  2996. int chunk = min_t(unsigned int, skb->len,
  2997. tp->ucopy.len);
  2998. __set_current_state(TASK_RUNNING);
  2999. local_bh_enable();
  3000. if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
  3001. tp->ucopy.len -= chunk;
  3002. tp->copied_seq += chunk;
  3003. eaten = (chunk == skb->len && !th->fin);
  3004. tcp_rcv_space_adjust(sk);
  3005. }
  3006. local_bh_disable();
  3007. }
  3008. if (eaten <= 0) {
  3009. queue_and_out:
  3010. if (eaten < 0 &&
  3011. (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3012. !sk_stream_rmem_schedule(sk, skb))) {
  3013. if (tcp_prune_queue(sk) < 0 ||
  3014. !sk_stream_rmem_schedule(sk, skb))
  3015. goto drop;
  3016. }
  3017. sk_stream_set_owner_r(skb, sk);
  3018. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3019. }
  3020. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3021. if (skb->len)
  3022. tcp_event_data_recv(sk, skb);
  3023. if (th->fin)
  3024. tcp_fin(skb, sk, th);
  3025. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3026. tcp_ofo_queue(sk);
  3027. /* RFC2581. 4.2. SHOULD send immediate ACK, when
  3028. * gap in queue is filled.
  3029. */
  3030. if (skb_queue_empty(&tp->out_of_order_queue))
  3031. inet_csk(sk)->icsk_ack.pingpong = 0;
  3032. }
  3033. if (tp->rx_opt.num_sacks)
  3034. tcp_sack_remove(tp);
  3035. tcp_fast_path_check(sk);
  3036. if (eaten > 0)
  3037. __kfree_skb(skb);
  3038. else if (!sock_flag(sk, SOCK_DEAD))
  3039. sk->sk_data_ready(sk, 0);
  3040. return;
  3041. }
  3042. if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
  3043. /* A retransmit, 2nd most common case. Force an immediate ack. */
  3044. NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST);
  3045. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3046. out_of_window:
  3047. tcp_enter_quickack_mode(sk);
  3048. inet_csk_schedule_ack(sk);
  3049. drop:
  3050. __kfree_skb(skb);
  3051. return;
  3052. }
  3053. /* Out of window. F.e. zero window probe. */
  3054. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
  3055. goto out_of_window;
  3056. tcp_enter_quickack_mode(sk);
  3057. if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3058. /* Partial packet, seq < rcv_next < end_seq */
  3059. SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
  3060. tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
  3061. TCP_SKB_CB(skb)->end_seq);
  3062. tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
  3063. /* If window is closed, drop tail of packet. But after
  3064. * remembering D-SACK for its head made in previous line.
  3065. */
  3066. if (!tcp_receive_window(tp))
  3067. goto out_of_window;
  3068. goto queue_and_out;
  3069. }
  3070. TCP_ECN_check_ce(tp, skb);
  3071. if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
  3072. !sk_stream_rmem_schedule(sk, skb)) {
  3073. if (tcp_prune_queue(sk) < 0 ||
  3074. !sk_stream_rmem_schedule(sk, skb))
  3075. goto drop;
  3076. }
  3077. /* Disable header prediction. */
  3078. tp->pred_flags = 0;
  3079. inet_csk_schedule_ack(sk);
  3080. SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
  3081. tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
  3082. sk_stream_set_owner_r(skb, sk);
  3083. if (!skb_peek(&tp->out_of_order_queue)) {
  3084. /* Initial out of order segment, build 1 SACK. */
  3085. if (tp->rx_opt.sack_ok) {
  3086. tp->rx_opt.num_sacks = 1;
  3087. tp->rx_opt.dsack = 0;
  3088. tp->rx_opt.eff_sacks = 1;
  3089. tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
  3090. tp->selective_acks[0].end_seq =
  3091. TCP_SKB_CB(skb)->end_seq;
  3092. }
  3093. __skb_queue_head(&tp->out_of_order_queue,skb);
  3094. } else {
  3095. struct sk_buff *skb1 = tp->out_of_order_queue.prev;
  3096. u32 seq = TCP_SKB_CB(skb)->seq;
  3097. u32 end_seq = TCP_SKB_CB(skb)->end_seq;
  3098. if (seq == TCP_SKB_CB(skb1)->end_seq) {
  3099. __skb_append(skb1, skb, &tp->out_of_order_queue);
  3100. if (!tp->rx_opt.num_sacks ||
  3101. tp->selective_acks[0].end_seq != seq)
  3102. goto add_sack;
  3103. /* Common case: data arrive in order after hole. */
  3104. tp->selective_acks[0].end_seq = end_seq;
  3105. return;
  3106. }
  3107. /* Find place to insert this segment. */
  3108. do {
  3109. if (!after(TCP_SKB_CB(skb1)->seq, seq))
  3110. break;
  3111. } while ((skb1 = skb1->prev) !=
  3112. (struct sk_buff*)&tp->out_of_order_queue);
  3113. /* Do skb overlap to previous one? */
  3114. if (skb1 != (struct sk_buff*)&tp->out_of_order_queue &&
  3115. before(seq, TCP_SKB_CB(skb1)->end_seq)) {
  3116. if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3117. /* All the bits are present. Drop. */
  3118. __kfree_skb(skb);
  3119. tcp_dsack_set(tp, seq, end_seq);
  3120. goto add_sack;
  3121. }
  3122. if (after(seq, TCP_SKB_CB(skb1)->seq)) {
  3123. /* Partial overlap. */
  3124. tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq);
  3125. } else {
  3126. skb1 = skb1->prev;
  3127. }
  3128. }
  3129. __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue);
  3130. /* And clean segments covered by new one as whole. */
  3131. while ((skb1 = skb->next) !=
  3132. (struct sk_buff*)&tp->out_of_order_queue &&
  3133. after(end_seq, TCP_SKB_CB(skb1)->seq)) {
  3134. if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
  3135. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq);
  3136. break;
  3137. }
  3138. __skb_unlink(skb1, &tp->out_of_order_queue);
  3139. tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq);
  3140. __kfree_skb(skb1);
  3141. }
  3142. add_sack:
  3143. if (tp->rx_opt.sack_ok)
  3144. tcp_sack_new_ofo_skb(sk, seq, end_seq);
  3145. }
  3146. }
  3147. /* Collapse contiguous sequence of skbs head..tail with
  3148. * sequence numbers start..end.
  3149. * Segments with FIN/SYN are not collapsed (only because this
  3150. * simplifies code)
  3151. */
  3152. static void
  3153. tcp_collapse(struct sock *sk, struct sk_buff_head *list,
  3154. struct sk_buff *head, struct sk_buff *tail,
  3155. u32 start, u32 end)
  3156. {
  3157. struct sk_buff *skb;
  3158. /* First, check that queue is collapsible and find
  3159. * the point where collapsing can be useful. */
  3160. for (skb = head; skb != tail; ) {
  3161. /* No new bits? It is possible on ofo queue. */
  3162. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3163. struct sk_buff *next = skb->next;
  3164. __skb_unlink(skb, list);
  3165. __kfree_skb(skb);
  3166. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3167. skb = next;
  3168. continue;
  3169. }
  3170. /* The first skb to collapse is:
  3171. * - not SYN/FIN and
  3172. * - bloated or contains data before "start" or
  3173. * overlaps to the next one.
  3174. */
  3175. if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
  3176. (tcp_win_from_space(skb->truesize) > skb->len ||
  3177. before(TCP_SKB_CB(skb)->seq, start) ||
  3178. (skb->next != tail &&
  3179. TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq)))
  3180. break;
  3181. /* Decided to skip this, advance start seq. */
  3182. start = TCP_SKB_CB(skb)->end_seq;
  3183. skb = skb->next;
  3184. }
  3185. if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
  3186. return;
  3187. while (before(start, end)) {
  3188. struct sk_buff *nskb;
  3189. int header = skb_headroom(skb);
  3190. int copy = SKB_MAX_ORDER(header, 0);
  3191. /* Too big header? This can happen with IPv6. */
  3192. if (copy < 0)
  3193. return;
  3194. if (end-start < copy)
  3195. copy = end-start;
  3196. nskb = alloc_skb(copy+header, GFP_ATOMIC);
  3197. if (!nskb)
  3198. return;
  3199. skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
  3200. skb_set_network_header(nskb, (skb_network_header(skb) -
  3201. skb->head));
  3202. skb_set_transport_header(nskb, (skb_transport_header(skb) -
  3203. skb->head));
  3204. skb_reserve(nskb, header);
  3205. memcpy(nskb->head, skb->head, header);
  3206. memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
  3207. TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
  3208. __skb_insert(nskb, skb->prev, skb, list);
  3209. sk_stream_set_owner_r(nskb, sk);
  3210. /* Copy data, releasing collapsed skbs. */
  3211. while (copy > 0) {
  3212. int offset = start - TCP_SKB_CB(skb)->seq;
  3213. int size = TCP_SKB_CB(skb)->end_seq - start;
  3214. BUG_ON(offset < 0);
  3215. if (size > 0) {
  3216. size = min(copy, size);
  3217. if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
  3218. BUG();
  3219. TCP_SKB_CB(nskb)->end_seq += size;
  3220. copy -= size;
  3221. start += size;
  3222. }
  3223. if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
  3224. struct sk_buff *next = skb->next;
  3225. __skb_unlink(skb, list);
  3226. __kfree_skb(skb);
  3227. NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED);
  3228. skb = next;
  3229. if (skb == tail ||
  3230. tcp_hdr(skb)->syn ||
  3231. tcp_hdr(skb)->fin)
  3232. return;
  3233. }
  3234. }
  3235. }
  3236. }
  3237. /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
  3238. * and tcp_collapse() them until all the queue is collapsed.
  3239. */
  3240. static void tcp_collapse_ofo_queue(struct sock *sk)
  3241. {
  3242. struct tcp_sock *tp = tcp_sk(sk);
  3243. struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
  3244. struct sk_buff *head;
  3245. u32 start, end;
  3246. if (skb == NULL)
  3247. return;
  3248. start = TCP_SKB_CB(skb)->seq;
  3249. end = TCP_SKB_CB(skb)->end_seq;
  3250. head = skb;
  3251. for (;;) {
  3252. skb = skb->next;
  3253. /* Segment is terminated when we see gap or when
  3254. * we are at the end of all the queue. */
  3255. if (skb == (struct sk_buff *)&tp->out_of_order_queue ||
  3256. after(TCP_SKB_CB(skb)->seq, end) ||
  3257. before(TCP_SKB_CB(skb)->end_seq, start)) {
  3258. tcp_collapse(sk, &tp->out_of_order_queue,
  3259. head, skb, start, end);
  3260. head = skb;
  3261. if (skb == (struct sk_buff *)&tp->out_of_order_queue)
  3262. break;
  3263. /* Start new segment */
  3264. start = TCP_SKB_CB(skb)->seq;
  3265. end = TCP_SKB_CB(skb)->end_seq;
  3266. } else {
  3267. if (before(TCP_SKB_CB(skb)->seq, start))
  3268. start = TCP_SKB_CB(skb)->seq;
  3269. if (after(TCP_SKB_CB(skb)->end_seq, end))
  3270. end = TCP_SKB_CB(skb)->end_seq;
  3271. }
  3272. }
  3273. }
  3274. /* Reduce allocated memory if we can, trying to get
  3275. * the socket within its memory limits again.
  3276. *
  3277. * Return less than zero if we should start dropping frames
  3278. * until the socket owning process reads some of the data
  3279. * to stabilize the situation.
  3280. */
  3281. static int tcp_prune_queue(struct sock *sk)
  3282. {
  3283. struct tcp_sock *tp = tcp_sk(sk);
  3284. SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
  3285. NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED);
  3286. if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
  3287. tcp_clamp_window(sk);
  3288. else if (tcp_memory_pressure)
  3289. tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
  3290. tcp_collapse_ofo_queue(sk);
  3291. tcp_collapse(sk, &sk->sk_receive_queue,
  3292. sk->sk_receive_queue.next,
  3293. (struct sk_buff*)&sk->sk_receive_queue,
  3294. tp->copied_seq, tp->rcv_nxt);
  3295. sk_stream_mem_reclaim(sk);
  3296. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3297. return 0;
  3298. /* Collapsing did not help, destructive actions follow.
  3299. * This must not ever occur. */
  3300. /* First, purge the out_of_order queue. */
  3301. if (!skb_queue_empty(&tp->out_of_order_queue)) {
  3302. NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED);
  3303. __skb_queue_purge(&tp->out_of_order_queue);
  3304. /* Reset SACK state. A conforming SACK implementation will
  3305. * do the same at a timeout based retransmit. When a connection
  3306. * is in a sad state like this, we care only about integrity
  3307. * of the connection not performance.
  3308. */
  3309. if (tp->rx_opt.sack_ok)
  3310. tcp_sack_reset(&tp->rx_opt);
  3311. sk_stream_mem_reclaim(sk);
  3312. }
  3313. if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
  3314. return 0;
  3315. /* If we are really being abused, tell the caller to silently
  3316. * drop receive data on the floor. It will get retransmitted
  3317. * and hopefully then we'll have sufficient space.
  3318. */
  3319. NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED);
  3320. /* Massive buffer overcommit. */
  3321. tp->pred_flags = 0;
  3322. return -1;
  3323. }
  3324. /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
  3325. * As additional protections, we do not touch cwnd in retransmission phases,
  3326. * and if application hit its sndbuf limit recently.
  3327. */
  3328. void tcp_cwnd_application_limited(struct sock *sk)
  3329. {
  3330. struct tcp_sock *tp = tcp_sk(sk);
  3331. if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
  3332. sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
  3333. /* Limited by application or receiver window. */
  3334. u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
  3335. u32 win_used = max(tp->snd_cwnd_used, init_win);
  3336. if (win_used < tp->snd_cwnd) {
  3337. tp->snd_ssthresh = tcp_current_ssthresh(sk);
  3338. tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
  3339. }
  3340. tp->snd_cwnd_used = 0;
  3341. }
  3342. tp->snd_cwnd_stamp = tcp_time_stamp;
  3343. }
  3344. static int tcp_should_expand_sndbuf(struct sock *sk)
  3345. {
  3346. struct tcp_sock *tp = tcp_sk(sk);
  3347. /* If the user specified a specific send buffer setting, do
  3348. * not modify it.
  3349. */
  3350. if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
  3351. return 0;
  3352. /* If we are under global TCP memory pressure, do not expand. */
  3353. if (tcp_memory_pressure)
  3354. return 0;
  3355. /* If we are under soft global TCP memory pressure, do not expand. */
  3356. if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0])
  3357. return 0;
  3358. /* If we filled the congestion window, do not expand. */
  3359. if (tp->packets_out >= tp->snd_cwnd)
  3360. return 0;
  3361. return 1;
  3362. }
  3363. /* When incoming ACK allowed to free some skb from write_queue,
  3364. * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
  3365. * on the exit from tcp input handler.
  3366. *
  3367. * PROBLEM: sndbuf expansion does not work well with largesend.
  3368. */
  3369. static void tcp_new_space(struct sock *sk)
  3370. {
  3371. struct tcp_sock *tp = tcp_sk(sk);
  3372. if (tcp_should_expand_sndbuf(sk)) {
  3373. int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
  3374. MAX_TCP_HEADER + 16 + sizeof(struct sk_buff),
  3375. demanded = max_t(unsigned int, tp->snd_cwnd,
  3376. tp->reordering + 1);
  3377. sndmem *= 2*demanded;
  3378. if (sndmem > sk->sk_sndbuf)
  3379. sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
  3380. tp->snd_cwnd_stamp = tcp_time_stamp;
  3381. }
  3382. sk->sk_write_space(sk);
  3383. }
  3384. static void tcp_check_space(struct sock *sk)
  3385. {
  3386. if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
  3387. sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
  3388. if (sk->sk_socket &&
  3389. test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
  3390. tcp_new_space(sk);
  3391. }
  3392. }
  3393. static inline void tcp_data_snd_check(struct sock *sk)
  3394. {
  3395. tcp_push_pending_frames(sk);
  3396. tcp_check_space(sk);
  3397. }
  3398. /*
  3399. * Check if sending an ack is needed.
  3400. */
  3401. static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
  3402. {
  3403. struct tcp_sock *tp = tcp_sk(sk);
  3404. /* More than one full frame received... */
  3405. if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss
  3406. /* ... and right edge of window advances far enough.
  3407. * (tcp_recvmsg() will send ACK otherwise). Or...
  3408. */
  3409. && __tcp_select_window(sk) >= tp->rcv_wnd) ||
  3410. /* We ACK each frame or... */
  3411. tcp_in_quickack_mode(sk) ||
  3412. /* We have out of order data. */
  3413. (ofo_possible &&
  3414. skb_peek(&tp->out_of_order_queue))) {
  3415. /* Then ack it now */
  3416. tcp_send_ack(sk);
  3417. } else {
  3418. /* Else, send delayed ack. */
  3419. tcp_send_delayed_ack(sk);
  3420. }
  3421. }
  3422. static inline void tcp_ack_snd_check(struct sock *sk)
  3423. {
  3424. if (!inet_csk_ack_scheduled(sk)) {
  3425. /* We sent a data segment already. */
  3426. return;
  3427. }
  3428. __tcp_ack_snd_check(sk, 1);
  3429. }
  3430. /*
  3431. * This routine is only called when we have urgent data
  3432. * signaled. Its the 'slow' part of tcp_urg. It could be
  3433. * moved inline now as tcp_urg is only called from one
  3434. * place. We handle URGent data wrong. We have to - as
  3435. * BSD still doesn't use the correction from RFC961.
  3436. * For 1003.1g we should support a new option TCP_STDURG to permit
  3437. * either form (or just set the sysctl tcp_stdurg).
  3438. */
  3439. static void tcp_check_urg(struct sock * sk, struct tcphdr * th)
  3440. {
  3441. struct tcp_sock *tp = tcp_sk(sk);
  3442. u32 ptr = ntohs(th->urg_ptr);
  3443. if (ptr && !sysctl_tcp_stdurg)
  3444. ptr--;
  3445. ptr += ntohl(th->seq);
  3446. /* Ignore urgent data that we've already seen and read. */
  3447. if (after(tp->copied_seq, ptr))
  3448. return;
  3449. /* Do not replay urg ptr.
  3450. *
  3451. * NOTE: interesting situation not covered by specs.
  3452. * Misbehaving sender may send urg ptr, pointing to segment,
  3453. * which we already have in ofo queue. We are not able to fetch
  3454. * such data and will stay in TCP_URG_NOTYET until will be eaten
  3455. * by recvmsg(). Seems, we are not obliged to handle such wicked
  3456. * situations. But it is worth to think about possibility of some
  3457. * DoSes using some hypothetical application level deadlock.
  3458. */
  3459. if (before(ptr, tp->rcv_nxt))
  3460. return;
  3461. /* Do we already have a newer (or duplicate) urgent pointer? */
  3462. if (tp->urg_data && !after(ptr, tp->urg_seq))
  3463. return;
  3464. /* Tell the world about our new urgent pointer. */
  3465. sk_send_sigurg(sk);
  3466. /* We may be adding urgent data when the last byte read was
  3467. * urgent. To do this requires some care. We cannot just ignore
  3468. * tp->copied_seq since we would read the last urgent byte again
  3469. * as data, nor can we alter copied_seq until this data arrives
  3470. * or we break the semantics of SIOCATMARK (and thus sockatmark())
  3471. *
  3472. * NOTE. Double Dutch. Rendering to plain English: author of comment
  3473. * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB);
  3474. * and expect that both A and B disappear from stream. This is _wrong_.
  3475. * Though this happens in BSD with high probability, this is occasional.
  3476. * Any application relying on this is buggy. Note also, that fix "works"
  3477. * only in this artificial test. Insert some normal data between A and B and we will
  3478. * decline of BSD again. Verdict: it is better to remove to trap
  3479. * buggy users.
  3480. */
  3481. if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
  3482. !sock_flag(sk, SOCK_URGINLINE) &&
  3483. tp->copied_seq != tp->rcv_nxt) {
  3484. struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
  3485. tp->copied_seq++;
  3486. if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
  3487. __skb_unlink(skb, &sk->sk_receive_queue);
  3488. __kfree_skb(skb);
  3489. }
  3490. }
  3491. tp->urg_data = TCP_URG_NOTYET;
  3492. tp->urg_seq = ptr;
  3493. /* Disable header prediction. */
  3494. tp->pred_flags = 0;
  3495. }
  3496. /* This is the 'fast' part of urgent handling. */
  3497. static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th)
  3498. {
  3499. struct tcp_sock *tp = tcp_sk(sk);
  3500. /* Check if we get a new urgent pointer - normally not. */
  3501. if (th->urg)
  3502. tcp_check_urg(sk,th);
  3503. /* Do we wait for any urgent data? - normally not... */
  3504. if (tp->urg_data == TCP_URG_NOTYET) {
  3505. u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
  3506. th->syn;
  3507. /* Is the urgent pointer pointing into this packet? */
  3508. if (ptr < skb->len) {
  3509. u8 tmp;
  3510. if (skb_copy_bits(skb, ptr, &tmp, 1))
  3511. BUG();
  3512. tp->urg_data = TCP_URG_VALID | tmp;
  3513. if (!sock_flag(sk, SOCK_DEAD))
  3514. sk->sk_data_ready(sk, 0);
  3515. }
  3516. }
  3517. }
  3518. static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
  3519. {
  3520. struct tcp_sock *tp = tcp_sk(sk);
  3521. int chunk = skb->len - hlen;
  3522. int err;
  3523. local_bh_enable();
  3524. if (skb_csum_unnecessary(skb))
  3525. err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
  3526. else
  3527. err = skb_copy_and_csum_datagram_iovec(skb, hlen,
  3528. tp->ucopy.iov);
  3529. if (!err) {
  3530. tp->ucopy.len -= chunk;
  3531. tp->copied_seq += chunk;
  3532. tcp_rcv_space_adjust(sk);
  3533. }
  3534. local_bh_disable();
  3535. return err;
  3536. }
  3537. static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3538. {
  3539. __sum16 result;
  3540. if (sock_owned_by_user(sk)) {
  3541. local_bh_enable();
  3542. result = __tcp_checksum_complete(skb);
  3543. local_bh_disable();
  3544. } else {
  3545. result = __tcp_checksum_complete(skb);
  3546. }
  3547. return result;
  3548. }
  3549. static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb)
  3550. {
  3551. return !skb_csum_unnecessary(skb) &&
  3552. __tcp_checksum_complete_user(sk, skb);
  3553. }
  3554. #ifdef CONFIG_NET_DMA
  3555. static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen)
  3556. {
  3557. struct tcp_sock *tp = tcp_sk(sk);
  3558. int chunk = skb->len - hlen;
  3559. int dma_cookie;
  3560. int copied_early = 0;
  3561. if (tp->ucopy.wakeup)
  3562. return 0;
  3563. if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
  3564. tp->ucopy.dma_chan = get_softnet_dma();
  3565. if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
  3566. dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
  3567. skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list);
  3568. if (dma_cookie < 0)
  3569. goto out;
  3570. tp->ucopy.dma_cookie = dma_cookie;
  3571. copied_early = 1;
  3572. tp->ucopy.len -= chunk;
  3573. tp->copied_seq += chunk;
  3574. tcp_rcv_space_adjust(sk);
  3575. if ((tp->ucopy.len == 0) ||
  3576. (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
  3577. (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
  3578. tp->ucopy.wakeup = 1;
  3579. sk->sk_data_ready(sk, 0);
  3580. }
  3581. } else if (chunk > 0) {
  3582. tp->ucopy.wakeup = 1;
  3583. sk->sk_data_ready(sk, 0);
  3584. }
  3585. out:
  3586. return copied_early;
  3587. }
  3588. #endif /* CONFIG_NET_DMA */
  3589. /*
  3590. * TCP receive function for the ESTABLISHED state.
  3591. *
  3592. * It is split into a fast path and a slow path. The fast path is
  3593. * disabled when:
  3594. * - A zero window was announced from us - zero window probing
  3595. * is only handled properly in the slow path.
  3596. * - Out of order segments arrived.
  3597. * - Urgent data is expected.
  3598. * - There is no buffer space left
  3599. * - Unexpected TCP flags/window values/header lengths are received
  3600. * (detected by checking the TCP header against pred_flags)
  3601. * - Data is sent in both directions. Fast path only supports pure senders
  3602. * or pure receivers (this means either the sequence number or the ack
  3603. * value must stay constant)
  3604. * - Unexpected TCP option.
  3605. *
  3606. * When these conditions are not satisfied it drops into a standard
  3607. * receive procedure patterned after RFC793 to handle all cases.
  3608. * The first three cases are guaranteed by proper pred_flags setting,
  3609. * the rest is checked inline. Fast processing is turned on in
  3610. * tcp_data_queue when everything is OK.
  3611. */
  3612. int tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
  3613. struct tcphdr *th, unsigned len)
  3614. {
  3615. struct tcp_sock *tp = tcp_sk(sk);
  3616. /*
  3617. * Header prediction.
  3618. * The code loosely follows the one in the famous
  3619. * "30 instruction TCP receive" Van Jacobson mail.
  3620. *
  3621. * Van's trick is to deposit buffers into socket queue
  3622. * on a device interrupt, to call tcp_recv function
  3623. * on the receive process context and checksum and copy
  3624. * the buffer to user space. smart...
  3625. *
  3626. * Our current scheme is not silly either but we take the
  3627. * extra cost of the net_bh soft interrupt processing...
  3628. * We do checksum and copy also but from device to kernel.
  3629. */
  3630. tp->rx_opt.saw_tstamp = 0;
  3631. /* pred_flags is 0xS?10 << 16 + snd_wnd
  3632. * if header_prediction is to be made
  3633. * 'S' will always be tp->tcp_header_len >> 2
  3634. * '?' will be 0 for the fast path, otherwise pred_flags is 0 to
  3635. * turn it off (when there are holes in the receive
  3636. * space for instance)
  3637. * PSH flag is ignored.
  3638. */
  3639. if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
  3640. TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
  3641. int tcp_header_len = tp->tcp_header_len;
  3642. /* Timestamp header prediction: tcp_header_len
  3643. * is automatically equal to th->doff*4 due to pred_flags
  3644. * match.
  3645. */
  3646. /* Check timestamp */
  3647. if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
  3648. __be32 *ptr = (__be32 *)(th + 1);
  3649. /* No? Slow path! */
  3650. if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
  3651. | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP))
  3652. goto slow_path;
  3653. tp->rx_opt.saw_tstamp = 1;
  3654. ++ptr;
  3655. tp->rx_opt.rcv_tsval = ntohl(*ptr);
  3656. ++ptr;
  3657. tp->rx_opt.rcv_tsecr = ntohl(*ptr);
  3658. /* If PAWS failed, check it more carefully in slow path */
  3659. if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
  3660. goto slow_path;
  3661. /* DO NOT update ts_recent here, if checksum fails
  3662. * and timestamp was corrupted part, it will result
  3663. * in a hung connection since we will drop all
  3664. * future packets due to the PAWS test.
  3665. */
  3666. }
  3667. if (len <= tcp_header_len) {
  3668. /* Bulk data transfer: sender */
  3669. if (len == tcp_header_len) {
  3670. /* Predicted packet is in window by definition.
  3671. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3672. * Hence, check seq<=rcv_wup reduces to:
  3673. */
  3674. if (tcp_header_len ==
  3675. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3676. tp->rcv_nxt == tp->rcv_wup)
  3677. tcp_store_ts_recent(tp);
  3678. /* We know that such packets are checksummed
  3679. * on entry.
  3680. */
  3681. tcp_ack(sk, skb, 0);
  3682. __kfree_skb(skb);
  3683. tcp_data_snd_check(sk);
  3684. return 0;
  3685. } else { /* Header too small */
  3686. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3687. goto discard;
  3688. }
  3689. } else {
  3690. int eaten = 0;
  3691. int copied_early = 0;
  3692. if (tp->copied_seq == tp->rcv_nxt &&
  3693. len - tcp_header_len <= tp->ucopy.len) {
  3694. #ifdef CONFIG_NET_DMA
  3695. if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
  3696. copied_early = 1;
  3697. eaten = 1;
  3698. }
  3699. #endif
  3700. if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) {
  3701. __set_current_state(TASK_RUNNING);
  3702. if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
  3703. eaten = 1;
  3704. }
  3705. if (eaten) {
  3706. /* Predicted packet is in window by definition.
  3707. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3708. * Hence, check seq<=rcv_wup reduces to:
  3709. */
  3710. if (tcp_header_len ==
  3711. (sizeof(struct tcphdr) +
  3712. TCPOLEN_TSTAMP_ALIGNED) &&
  3713. tp->rcv_nxt == tp->rcv_wup)
  3714. tcp_store_ts_recent(tp);
  3715. tcp_rcv_rtt_measure_ts(sk, skb);
  3716. __skb_pull(skb, tcp_header_len);
  3717. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3718. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER);
  3719. }
  3720. if (copied_early)
  3721. tcp_cleanup_rbuf(sk, skb->len);
  3722. }
  3723. if (!eaten) {
  3724. if (tcp_checksum_complete_user(sk, skb))
  3725. goto csum_error;
  3726. /* Predicted packet is in window by definition.
  3727. * seq == rcv_nxt and rcv_wup <= rcv_nxt.
  3728. * Hence, check seq<=rcv_wup reduces to:
  3729. */
  3730. if (tcp_header_len ==
  3731. (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
  3732. tp->rcv_nxt == tp->rcv_wup)
  3733. tcp_store_ts_recent(tp);
  3734. tcp_rcv_rtt_measure_ts(sk, skb);
  3735. if ((int)skb->truesize > sk->sk_forward_alloc)
  3736. goto step5;
  3737. NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS);
  3738. /* Bulk data transfer: receiver */
  3739. __skb_pull(skb,tcp_header_len);
  3740. __skb_queue_tail(&sk->sk_receive_queue, skb);
  3741. sk_stream_set_owner_r(skb, sk);
  3742. tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
  3743. }
  3744. tcp_event_data_recv(sk, skb);
  3745. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
  3746. /* Well, only one small jumplet in fast path... */
  3747. tcp_ack(sk, skb, FLAG_DATA);
  3748. tcp_data_snd_check(sk);
  3749. if (!inet_csk_ack_scheduled(sk))
  3750. goto no_ack;
  3751. }
  3752. __tcp_ack_snd_check(sk, 0);
  3753. no_ack:
  3754. #ifdef CONFIG_NET_DMA
  3755. if (copied_early)
  3756. __skb_queue_tail(&sk->sk_async_wait_queue, skb);
  3757. else
  3758. #endif
  3759. if (eaten)
  3760. __kfree_skb(skb);
  3761. else
  3762. sk->sk_data_ready(sk, 0);
  3763. return 0;
  3764. }
  3765. }
  3766. slow_path:
  3767. if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb))
  3768. goto csum_error;
  3769. /*
  3770. * RFC1323: H1. Apply PAWS check first.
  3771. */
  3772. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  3773. tcp_paws_discard(sk, skb)) {
  3774. if (!th->rst) {
  3775. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  3776. tcp_send_dupack(sk, skb);
  3777. goto discard;
  3778. }
  3779. /* Resets are accepted even if PAWS failed.
  3780. ts_recent update must be made after we are sure
  3781. that the packet is in window.
  3782. */
  3783. }
  3784. /*
  3785. * Standard slow path.
  3786. */
  3787. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  3788. /* RFC793, page 37: "In all states except SYN-SENT, all reset
  3789. * (RST) segments are validated by checking their SEQ-fields."
  3790. * And page 69: "If an incoming segment is not acceptable,
  3791. * an acknowledgment should be sent in reply (unless the RST bit
  3792. * is set, if so drop the segment and return)".
  3793. */
  3794. if (!th->rst)
  3795. tcp_send_dupack(sk, skb);
  3796. goto discard;
  3797. }
  3798. if (th->rst) {
  3799. tcp_reset(sk);
  3800. goto discard;
  3801. }
  3802. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  3803. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  3804. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3805. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  3806. tcp_reset(sk);
  3807. return 1;
  3808. }
  3809. step5:
  3810. if (th->ack)
  3811. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3812. tcp_rcv_rtt_measure_ts(sk, skb);
  3813. /* Process urgent data. */
  3814. tcp_urg(sk, skb, th);
  3815. /* step 7: process the segment text */
  3816. tcp_data_queue(sk, skb);
  3817. tcp_data_snd_check(sk);
  3818. tcp_ack_snd_check(sk);
  3819. return 0;
  3820. csum_error:
  3821. TCP_INC_STATS_BH(TCP_MIB_INERRS);
  3822. discard:
  3823. __kfree_skb(skb);
  3824. return 0;
  3825. }
  3826. static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
  3827. struct tcphdr *th, unsigned len)
  3828. {
  3829. struct tcp_sock *tp = tcp_sk(sk);
  3830. struct inet_connection_sock *icsk = inet_csk(sk);
  3831. int saved_clamp = tp->rx_opt.mss_clamp;
  3832. tcp_parse_options(skb, &tp->rx_opt, 0);
  3833. if (th->ack) {
  3834. /* rfc793:
  3835. * "If the state is SYN-SENT then
  3836. * first check the ACK bit
  3837. * If the ACK bit is set
  3838. * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
  3839. * a reset (unless the RST bit is set, if so drop
  3840. * the segment and return)"
  3841. *
  3842. * We do not send data with SYN, so that RFC-correct
  3843. * test reduces to:
  3844. */
  3845. if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt)
  3846. goto reset_and_undo;
  3847. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  3848. !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
  3849. tcp_time_stamp)) {
  3850. NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED);
  3851. goto reset_and_undo;
  3852. }
  3853. /* Now ACK is acceptable.
  3854. *
  3855. * "If the RST bit is set
  3856. * If the ACK was acceptable then signal the user "error:
  3857. * connection reset", drop the segment, enter CLOSED state,
  3858. * delete TCB, and return."
  3859. */
  3860. if (th->rst) {
  3861. tcp_reset(sk);
  3862. goto discard;
  3863. }
  3864. /* rfc793:
  3865. * "fifth, if neither of the SYN or RST bits is set then
  3866. * drop the segment and return."
  3867. *
  3868. * See note below!
  3869. * --ANK(990513)
  3870. */
  3871. if (!th->syn)
  3872. goto discard_and_undo;
  3873. /* rfc793:
  3874. * "If the SYN bit is on ...
  3875. * are acceptable then ...
  3876. * (our SYN has been ACKed), change the connection
  3877. * state to ESTABLISHED..."
  3878. */
  3879. TCP_ECN_rcv_synack(tp, th);
  3880. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  3881. tcp_ack(sk, skb, FLAG_SLOWPATH);
  3882. /* Ok.. it's good. Set up sequence numbers and
  3883. * move to established.
  3884. */
  3885. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  3886. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  3887. /* RFC1323: The window in SYN & SYN/ACK segments is
  3888. * never scaled.
  3889. */
  3890. tp->snd_wnd = ntohs(th->window);
  3891. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq);
  3892. if (!tp->rx_opt.wscale_ok) {
  3893. tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
  3894. tp->window_clamp = min(tp->window_clamp, 65535U);
  3895. }
  3896. if (tp->rx_opt.saw_tstamp) {
  3897. tp->rx_opt.tstamp_ok = 1;
  3898. tp->tcp_header_len =
  3899. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  3900. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  3901. tcp_store_ts_recent(tp);
  3902. } else {
  3903. tp->tcp_header_len = sizeof(struct tcphdr);
  3904. }
  3905. if (tp->rx_opt.sack_ok && sysctl_tcp_fack)
  3906. tp->rx_opt.sack_ok |= 2;
  3907. tcp_mtup_init(sk);
  3908. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  3909. tcp_initialize_rcv_mss(sk);
  3910. /* Remember, tcp_poll() does not lock socket!
  3911. * Change state from SYN-SENT only after copied_seq
  3912. * is initialized. */
  3913. tp->copied_seq = tp->rcv_nxt;
  3914. smp_mb();
  3915. tcp_set_state(sk, TCP_ESTABLISHED);
  3916. security_inet_conn_established(sk, skb);
  3917. /* Make sure socket is routed, for correct metrics. */
  3918. icsk->icsk_af_ops->rebuild_header(sk);
  3919. tcp_init_metrics(sk);
  3920. tcp_init_congestion_control(sk);
  3921. /* Prevent spurious tcp_cwnd_restart() on first data
  3922. * packet.
  3923. */
  3924. tp->lsndtime = tcp_time_stamp;
  3925. tcp_init_buffer_space(sk);
  3926. if (sock_flag(sk, SOCK_KEEPOPEN))
  3927. inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
  3928. if (!tp->rx_opt.snd_wscale)
  3929. __tcp_fast_path_on(tp, tp->snd_wnd);
  3930. else
  3931. tp->pred_flags = 0;
  3932. if (!sock_flag(sk, SOCK_DEAD)) {
  3933. sk->sk_state_change(sk);
  3934. sk_wake_async(sk, 0, POLL_OUT);
  3935. }
  3936. if (sk->sk_write_pending ||
  3937. icsk->icsk_accept_queue.rskq_defer_accept ||
  3938. icsk->icsk_ack.pingpong) {
  3939. /* Save one ACK. Data will be ready after
  3940. * several ticks, if write_pending is set.
  3941. *
  3942. * It may be deleted, but with this feature tcpdumps
  3943. * look so _wonderfully_ clever, that I was not able
  3944. * to stand against the temptation 8) --ANK
  3945. */
  3946. inet_csk_schedule_ack(sk);
  3947. icsk->icsk_ack.lrcvtime = tcp_time_stamp;
  3948. icsk->icsk_ack.ato = TCP_ATO_MIN;
  3949. tcp_incr_quickack(sk);
  3950. tcp_enter_quickack_mode(sk);
  3951. inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
  3952. TCP_DELACK_MAX, TCP_RTO_MAX);
  3953. discard:
  3954. __kfree_skb(skb);
  3955. return 0;
  3956. } else {
  3957. tcp_send_ack(sk);
  3958. }
  3959. return -1;
  3960. }
  3961. /* No ACK in the segment */
  3962. if (th->rst) {
  3963. /* rfc793:
  3964. * "If the RST bit is set
  3965. *
  3966. * Otherwise (no ACK) drop the segment and return."
  3967. */
  3968. goto discard_and_undo;
  3969. }
  3970. /* PAWS check. */
  3971. if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0))
  3972. goto discard_and_undo;
  3973. if (th->syn) {
  3974. /* We see SYN without ACK. It is attempt of
  3975. * simultaneous connect with crossed SYNs.
  3976. * Particularly, it can be connect to self.
  3977. */
  3978. tcp_set_state(sk, TCP_SYN_RECV);
  3979. if (tp->rx_opt.saw_tstamp) {
  3980. tp->rx_opt.tstamp_ok = 1;
  3981. tcp_store_ts_recent(tp);
  3982. tp->tcp_header_len =
  3983. sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
  3984. } else {
  3985. tp->tcp_header_len = sizeof(struct tcphdr);
  3986. }
  3987. tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
  3988. tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
  3989. /* RFC1323: The window in SYN & SYN/ACK segments is
  3990. * never scaled.
  3991. */
  3992. tp->snd_wnd = ntohs(th->window);
  3993. tp->snd_wl1 = TCP_SKB_CB(skb)->seq;
  3994. tp->max_window = tp->snd_wnd;
  3995. TCP_ECN_rcv_syn(tp, th);
  3996. tcp_mtup_init(sk);
  3997. tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
  3998. tcp_initialize_rcv_mss(sk);
  3999. tcp_send_synack(sk);
  4000. #if 0
  4001. /* Note, we could accept data and URG from this segment.
  4002. * There are no obstacles to make this.
  4003. *
  4004. * However, if we ignore data in ACKless segments sometimes,
  4005. * we have no reasons to accept it sometimes.
  4006. * Also, seems the code doing it in step6 of tcp_rcv_state_process
  4007. * is not flawless. So, discard packet for sanity.
  4008. * Uncomment this return to process the data.
  4009. */
  4010. return -1;
  4011. #else
  4012. goto discard;
  4013. #endif
  4014. }
  4015. /* "fifth, if neither of the SYN or RST bits is set then
  4016. * drop the segment and return."
  4017. */
  4018. discard_and_undo:
  4019. tcp_clear_options(&tp->rx_opt);
  4020. tp->rx_opt.mss_clamp = saved_clamp;
  4021. goto discard;
  4022. reset_and_undo:
  4023. tcp_clear_options(&tp->rx_opt);
  4024. tp->rx_opt.mss_clamp = saved_clamp;
  4025. return 1;
  4026. }
  4027. /*
  4028. * This function implements the receiving procedure of RFC 793 for
  4029. * all states except ESTABLISHED and TIME_WAIT.
  4030. * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
  4031. * address independent.
  4032. */
  4033. int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
  4034. struct tcphdr *th, unsigned len)
  4035. {
  4036. struct tcp_sock *tp = tcp_sk(sk);
  4037. struct inet_connection_sock *icsk = inet_csk(sk);
  4038. int queued = 0;
  4039. tp->rx_opt.saw_tstamp = 0;
  4040. switch (sk->sk_state) {
  4041. case TCP_CLOSE:
  4042. goto discard;
  4043. case TCP_LISTEN:
  4044. if (th->ack)
  4045. return 1;
  4046. if (th->rst)
  4047. goto discard;
  4048. if (th->syn) {
  4049. if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
  4050. return 1;
  4051. /* Now we have several options: In theory there is
  4052. * nothing else in the frame. KA9Q has an option to
  4053. * send data with the syn, BSD accepts data with the
  4054. * syn up to the [to be] advertised window and
  4055. * Solaris 2.1 gives you a protocol error. For now
  4056. * we just ignore it, that fits the spec precisely
  4057. * and avoids incompatibilities. It would be nice in
  4058. * future to drop through and process the data.
  4059. *
  4060. * Now that TTCP is starting to be used we ought to
  4061. * queue this data.
  4062. * But, this leaves one open to an easy denial of
  4063. * service attack, and SYN cookies can't defend
  4064. * against this problem. So, we drop the data
  4065. * in the interest of security over speed unless
  4066. * it's still in use.
  4067. */
  4068. kfree_skb(skb);
  4069. return 0;
  4070. }
  4071. goto discard;
  4072. case TCP_SYN_SENT:
  4073. queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
  4074. if (queued >= 0)
  4075. return queued;
  4076. /* Do step6 onward by hand. */
  4077. tcp_urg(sk, skb, th);
  4078. __kfree_skb(skb);
  4079. tcp_data_snd_check(sk);
  4080. return 0;
  4081. }
  4082. if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
  4083. tcp_paws_discard(sk, skb)) {
  4084. if (!th->rst) {
  4085. NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED);
  4086. tcp_send_dupack(sk, skb);
  4087. goto discard;
  4088. }
  4089. /* Reset is accepted even if it did not pass PAWS. */
  4090. }
  4091. /* step 1: check sequence number */
  4092. if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
  4093. if (!th->rst)
  4094. tcp_send_dupack(sk, skb);
  4095. goto discard;
  4096. }
  4097. /* step 2: check RST bit */
  4098. if (th->rst) {
  4099. tcp_reset(sk);
  4100. goto discard;
  4101. }
  4102. tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
  4103. /* step 3: check security and precedence [ignored] */
  4104. /* step 4:
  4105. *
  4106. * Check for a SYN in window.
  4107. */
  4108. if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
  4109. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN);
  4110. tcp_reset(sk);
  4111. return 1;
  4112. }
  4113. /* step 5: check the ACK field */
  4114. if (th->ack) {
  4115. int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH);
  4116. switch (sk->sk_state) {
  4117. case TCP_SYN_RECV:
  4118. if (acceptable) {
  4119. tp->copied_seq = tp->rcv_nxt;
  4120. smp_mb();
  4121. tcp_set_state(sk, TCP_ESTABLISHED);
  4122. sk->sk_state_change(sk);
  4123. /* Note, that this wakeup is only for marginal
  4124. * crossed SYN case. Passively open sockets
  4125. * are not waked up, because sk->sk_sleep ==
  4126. * NULL and sk->sk_socket == NULL.
  4127. */
  4128. if (sk->sk_socket) {
  4129. sk_wake_async(sk,0,POLL_OUT);
  4130. }
  4131. tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
  4132. tp->snd_wnd = ntohs(th->window) <<
  4133. tp->rx_opt.snd_wscale;
  4134. tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq,
  4135. TCP_SKB_CB(skb)->seq);
  4136. /* tcp_ack considers this ACK as duplicate
  4137. * and does not calculate rtt.
  4138. * Fix it at least with timestamps.
  4139. */
  4140. if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
  4141. !tp->srtt)
  4142. tcp_ack_saw_tstamp(sk, 0);
  4143. if (tp->rx_opt.tstamp_ok)
  4144. tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
  4145. /* Make sure socket is routed, for
  4146. * correct metrics.
  4147. */
  4148. icsk->icsk_af_ops->rebuild_header(sk);
  4149. tcp_init_metrics(sk);
  4150. tcp_init_congestion_control(sk);
  4151. /* Prevent spurious tcp_cwnd_restart() on
  4152. * first data packet.
  4153. */
  4154. tp->lsndtime = tcp_time_stamp;
  4155. tcp_mtup_init(sk);
  4156. tcp_initialize_rcv_mss(sk);
  4157. tcp_init_buffer_space(sk);
  4158. tcp_fast_path_on(tp);
  4159. } else {
  4160. return 1;
  4161. }
  4162. break;
  4163. case TCP_FIN_WAIT1:
  4164. if (tp->snd_una == tp->write_seq) {
  4165. tcp_set_state(sk, TCP_FIN_WAIT2);
  4166. sk->sk_shutdown |= SEND_SHUTDOWN;
  4167. dst_confirm(sk->sk_dst_cache);
  4168. if (!sock_flag(sk, SOCK_DEAD))
  4169. /* Wake up lingering close() */
  4170. sk->sk_state_change(sk);
  4171. else {
  4172. int tmo;
  4173. if (tp->linger2 < 0 ||
  4174. (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4175. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
  4176. tcp_done(sk);
  4177. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4178. return 1;
  4179. }
  4180. tmo = tcp_fin_time(sk);
  4181. if (tmo > TCP_TIMEWAIT_LEN) {
  4182. inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
  4183. } else if (th->fin || sock_owned_by_user(sk)) {
  4184. /* Bad case. We could lose such FIN otherwise.
  4185. * It is not a big problem, but it looks confusing
  4186. * and not so rare event. We still can lose it now,
  4187. * if it spins in bh_lock_sock(), but it is really
  4188. * marginal case.
  4189. */
  4190. inet_csk_reset_keepalive_timer(sk, tmo);
  4191. } else {
  4192. tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
  4193. goto discard;
  4194. }
  4195. }
  4196. }
  4197. break;
  4198. case TCP_CLOSING:
  4199. if (tp->snd_una == tp->write_seq) {
  4200. tcp_time_wait(sk, TCP_TIME_WAIT, 0);
  4201. goto discard;
  4202. }
  4203. break;
  4204. case TCP_LAST_ACK:
  4205. if (tp->snd_una == tp->write_seq) {
  4206. tcp_update_metrics(sk);
  4207. tcp_done(sk);
  4208. goto discard;
  4209. }
  4210. break;
  4211. }
  4212. } else
  4213. goto discard;
  4214. /* step 6: check the URG bit */
  4215. tcp_urg(sk, skb, th);
  4216. /* step 7: process the segment text */
  4217. switch (sk->sk_state) {
  4218. case TCP_CLOSE_WAIT:
  4219. case TCP_CLOSING:
  4220. case TCP_LAST_ACK:
  4221. if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
  4222. break;
  4223. case TCP_FIN_WAIT1:
  4224. case TCP_FIN_WAIT2:
  4225. /* RFC 793 says to queue data in these states,
  4226. * RFC 1122 says we MUST send a reset.
  4227. * BSD 4.4 also does reset.
  4228. */
  4229. if (sk->sk_shutdown & RCV_SHUTDOWN) {
  4230. if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
  4231. after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
  4232. NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA);
  4233. tcp_reset(sk);
  4234. return 1;
  4235. }
  4236. }
  4237. /* Fall through */
  4238. case TCP_ESTABLISHED:
  4239. tcp_data_queue(sk, skb);
  4240. queued = 1;
  4241. break;
  4242. }
  4243. /* tcp_data could move socket to TIME-WAIT */
  4244. if (sk->sk_state != TCP_CLOSE) {
  4245. tcp_data_snd_check(sk);
  4246. tcp_ack_snd_check(sk);
  4247. }
  4248. if (!queued) {
  4249. discard:
  4250. __kfree_skb(skb);
  4251. }
  4252. return 0;
  4253. }
  4254. EXPORT_SYMBOL(sysctl_tcp_ecn);
  4255. EXPORT_SYMBOL(sysctl_tcp_reordering);
  4256. EXPORT_SYMBOL(tcp_parse_options);
  4257. EXPORT_SYMBOL(tcp_rcv_established);
  4258. EXPORT_SYMBOL(tcp_rcv_state_process);
  4259. EXPORT_SYMBOL(tcp_initialize_rcv_mss);